A framework to address key issues of neonatal service configuration in England: the NeoNet multimethods study

Adjustment of length of stay

The NNRD data are summarised for each day of care. Total care days add up to more than the actual LOS as any part day is counted as a full day in these raw data. We therefore adjusted the days in each level of care in proportion to the total LOS calculated from admission and discharge times (Figure 1). The level used in this study was the BAPM 2011 level of care. 2

FIGURE 1.

Calculation of LOSs in different care levels.

Geographic data coverage and predicting neonatal demand

The birth model was based on HES data; therefore, our models do not include the small proportion (≈2%) of births that take place at home. 43 Only 2% of births in England take place in freestanding midwife-led units,24 and, although these will be in the HES data set, we do not seek to model freestanding midwife-led units as they are currently such a small part of the neonatal care scene.

Permission was given to use data from all NICUs and 85% and 90% of LNUs and SCUs, respectively. Overall, 90% of units gave permission to use their neonatal care data. Because data are missing from LNUs and SCUs, there is the possibility of bias in the statistics used that underlie the model. In order to minimise any bias, statistical summaries were based on only those infants with a complete data record and who live in an area where all closest levels of care locations are present within the data set. Infants whose closest Level 2 or 3 care locations are not in the data set are not used in the analysis (this does not apply to Level 1 NICU care as all NICUs are present in the data set).

As neonatal demand was not available for all hospitals and LSOAs, a regression model was used to predict demand per LSOA. This model was based on LSOAs that had all closest unit types present in the neonatal data set. These represent 75.4% of all LSOAs in England. A demand regression model for neonatal admissions based on births is also less likely to overfit to very specific geographic neonatal demand patterns that may be present simply because of the lower numbers of admissions per geographic area.

A regression model was built based on births and IMD scores. IMD score was chosen as an independent variable because a link between social deprivation and incidence of preterm births has been demonstrated. 44

The regression model predicted total neonatal admissions for each LSOA. The admissions of VLBW infants and all intensive care, high-dependency care and special care admissions were based on a fixed proportion of total predicted neonatal admissions, based on the number of VLBW infants and use-of-care levels (see Chapter 5, Use-of-care levels and length of stay by gestational age).

Table 1 shows the coefficients obtained from the regression analysis. Neonatal demand is positively correlated with births and IMD score.

Figure 2 shows the results of the regression analysis when predicting demand at the upper-layer super output area (there are 325 upper-layer super output areas in England). The prediction of demand for VLBW infant admissions, intensive care admissions and high-dependency care admissions all produced correlations with the R² of > 90%, showing good regional consistency in predicting these levels of demand. Prediction of special care admissions had an R² of 78%. The slightly lower predictive accuracy of special care admissions may be attributable to more regional variation in the clinical judgement on whether or not an infant needs admitting to neonatal care (with much less variation in the perceived need for neonatal care for higher levels of care).

FIGURE 2.

Correlation between actual and predicted neonatal admissions at upper-layer super output area. (a) VLBW admissions; (b) infants with ≥ 24 hours of intensive care; (c) infants with ≥ 24 hours of high-dependency care; and (d) infants with ≥ 24 hours of special care.

Assumptions

We assume that all units have an unlimited capacity. However, the maximum number of admissions to any single unit is kept as low as possible by criterion 6 stated in Decision criteria.

We assume that neonatal care can only take place in the existing neonatal units. Considering that the existing network is already quite dense and the tendency is towards a reduction of the number of units, then this assumption is reasonable.

Decision criteria

The aim of this study is to optimise the location of neonatal units in order to improve the outcome of the infants, the experience of the parents and to align the neonatal network with the current demand. These objectives translate into the following criteria:

1. Minimise the average distance from a mother’s place of residence to an available unit.

2. Minimise the maximum distance for any mother to an available unit.

3. Maximise the proportion of mothers living within 30 minutes of the nearest available unit (other travel time limits, such as 45 and 60 minutes, are also evaluated, but the 30-minute range is used for optimisation as this range is more discriminating between options).

4. Maximise the proportion of births taking place in units with more than a given number, A_min, of admissions per year.

5. Maximise the minimum number of admissions (for any single unit) below a given number of admissions per year.

6. Minimise the maximum number of admissions (for any single unit) over a given number of admissions per year.

7. Maximise the proportion of mothers within 30 minutes of the nearest available unit and going to a unit with more than a given number, A_min, of admissions per year.

The first three coverage criteria, very common in facility location problems, aim to align facility locations with the population distribution. They rely on the assumption that women will go the nearest care facility. 45 The other four criteria are based on the number of admissions to each unit, once every woman has been assigned to an available location. In particular, criterion 4 is specific to this study and favours the creation of big units in order to improve clinical outcomes. 22 Criteria 3 and 4 can conflict in areas with a sparse population. Finally, criterion 6 limits the maximum size of care units. Possible criteria values define the objective space.

Dealing with multiple criteria: Pareto dominance

Our study tackles a multiobjective optimisation (MOO) problem. When solving an optimisation problem based on one objective, the optimal solution is given by the configuration with the best (highest or lowest) objective value. In the case of MOO, comparing several solutions requires reference to the notion of dominance: a vector a of the objective space dominates another vector b if all criteria of a are better or equal to the criteria of b and a ≠ b. 46 Then, there is no single best solution but a set of non-dominated solutions, called the Pareto front.

Complexity of the decision space

If the problem is to find the best combination of h units in a list of H possible locations, then one common method is to compute all objectives for all possible combinations (brute force). The number of h-combinations (combinations of size h) is the binomial coefficient:

if h ≤ H, or equals zero if h > H. With H = 161 possible locations, there are, for instance, 12,880 combinations of two locations and the maximum number of configurations is 1.82 × 10⁴⁷, reached with half of locations (80) available. With such a large number of combinations, the brute force method cannot be used and the optimisation problem must be solved by heuristic methods. In practice, the size of the decision space means that it is impossible to give an exact solution to this optimisation problem; only an approximate solution can be given.

Greedy algorithm

The greedy algorithm is a deterministic technique that leads to one local optimum. The greedy algorithm cannot identify a Pareto front, but seeks to find the best solutions based on a composite score (the ‘fitness value’) of all objectives. The solution identified is highly dependent on how the different objectives are weighted. The algorithm starts with the location with the highest fitness value. At each iteration, the algorithm selects the location that brings the most improvement and adds it to the combination. Although this technique is intuitive and fast, it provides only one solution and it strongly relies on weights and fitness.

Steepest-ascent hill climbing

The hill-climbing method is an efficient method to find local optima of the fitness function. From a given individual, candidates are generated by mutating only one gene of the original. Then, the fitness value for each candidate is computed. The individual is then replaced by the candidate that brings the highest improvement (steepest ascent). If no improvement is possible, the local optimum is recorded and the exploration is restarted from another random point.

Simulated annealing

Simulated annealing is a stochastic local search method. 47 Here, annealing refers here to the cooling of materials with a controlled temperature. The probability of accepting combinations with a worse fitness decreases with the temperature, so that the search converges to an optimum.

Genetic algorithms

Genetic algorithms manage a population of individuals encoded as vectors through a given number of generations. At each generation, ‘good’ parents are selected from the population depending on their fitness. Parents are then combined, using a crossover operator, to create children that are finally mutated. Genetic algorithms differ in the parent selection process, in the crossover and mutation processes, and in the way that the population is archived.

Convergence metrics

The number of generations was determined with a steady-state detection-based termination criterion (a chi-squared test based on the generational distance of objective values, or on the Hamming distance of integer vectors) adapted from Wong et al. 59

Data

Data on the number of births in existing maternity units were provided by the HES database (see Chapter 4, Birth data). Birth numbers per year were averaged using data from 2013 to 2015.

Travel times for patients between the LSOAs of England (country divisions with similar population sizes as defined in 2011) and the 161 existing maternity units were estimated (see Chapter 4, Travel time data). Maternity units were selected from those hospitals that had any level of neonatal care unit.

Estimation

The NSGA-II method56 and the criterion functions defined in Decision criteria were implemented using MATLAB programming software (MathWorks, Natick, MA, USA). In particular, since the Royal College of Obstetricians and Gynaecologists has recommended that units have ≥ 6000 births per year,22 criteria 4 and 1 were applied with a number of births per year defined by A_min = 6000.

Criterion functions were normalised between 0 and 1 using values from Table 5, with 1 being the ideal value. Fitness was computed as the weighted average of normalised objective values, with equal weights for all objectives.

The first generation was generated randomly using an integer uniform distribution in 〚1; h〛, so that there was no location repetition inside combination chromosomes and no duplicates in the population.

The population size was set to P = 200 and remained constant through generations. This value was chosen as a compromise between the exploration of the decision space and the computing time.

The population was propagated through G = 200 generations; this value would provide a steady state as described in Convergence metrics.

The MOO process was run independently for all h ∈ 〚1; H〛. Note that the same optimisation algorithm can be applied with a binary location encoding (u₁, . . . , u_H) ∈ {0; 1}^H. In that case, the number of open locations is flexible and only one optimisation run provides a Pareto front with various numbers of locations. However, the range of location numbers in the final solutions is not easily controlled with such encoding and depends on score quality. The integer encoding allows a more thorough exploration of the decision space.

Parents were selected from the population using the tournament selection process. 56 To do so, pairs of individuals are randomly drawn from the population and the one with the highest fitness is selected as a parent with a probability of p_t = 0.75. The tournament is an elitist selection process as it favours high fitness values but it allows lower fitness values with a probability of 1 – p_t.

Parents are then combined using a single-point crossover; the crossover point location is randomly chosen following a uniform distribution.

The alleles of every child were mutated with a probability of p_m = 0.001. After crossover and mutations, additional mutations took place to ensure that children did not contain location repetitions.

The outputs of the optimisation process are the final Pareto front layers up to P individuals and their corresponding objective values. The minimum number of admissions for each combination is also recorded in order to evaluate the impact of the redistribution of locations.

Results

Access to care

Figure 6a shows the relationship between the average and maximum travel time for mothers, assuming that they go to the closest unit, as a function of the number of available obstetric units. Each dot represents the criterion value of one configuration or set of units discovered in the Pareto front of solutions.

FIGURE 6.

Influence of the number of maternity units on the access to care. (a) Travel time (minutes) (potential configuration: minimum and maximum travel times); (b) proportion of women in target time (potential configuration: proportion of women within 30, 45 or 60 minutes of a maternity unit); (c) number of admissions (potential configuration: minimum and maximum admissions to any single unit); (d) births in units with ≥ 6000 births per year (%) (potential configuration: proportion of births in a unit with ≥ 6000 admissions per year); and (e) proportion of births in target (potential configuration: the proportion of births within 30 minutes, proportion of births in a unit with ≥ 6000 admissions per year, and proportion of births in a unit with 6000 births per year and within 30 minutes). Quantified by the distance to units and the proportion of women in target time, on the number of admissions and on the proportion of births occuring in large units (≥ 6000 births per year).

With the current number of 161 obstetric units, the average travel time is 15 minutes and the maximum travel time is 82 minutes. As expected, the travel time increases when the number of units decreases; for instance, reducing the number of units by half to 80 increases the average travel time to 21 minutes (+6 minutes) and the maximum travel time to 99 minutes (+17 minutes). Note that removing one unit only affects the patient access in some LSOAs and the average travel time is computed for all LSOAs. As a consequence, the effect on the average travel time is limited.

In Figure 6b, the proportion of women living within 30/45/60 minutes of the closest obstetric unit increases with the number of units. With 161 units, 93% of women live within 30 minutes and 100% live within 45 minutes of the closest unit. With only half of the units (80), the proportion of women (1) within 30 minutes of the closest unit is reduced to 84% (–9%), (2) within 45 minutes is reduced to 90% (–10%) and (3) within 60 minutes is reduced to 95% (–5%). Similar to the average travel times, the effect of changing the number of units on the proportion of women within a target travel time is rather smooth because it only affects a part of the population.

Regional population projections

The 10-year projection for changes in women of child-bearing age (considered to be those aged 15–39 years) ranged from a reduction of about 1% in the North West to an increase of about 4% in London (Table 6).

These projections were made before the UK voted to leave the European Union (expected to formally take place in March 2019). There may, therefore, be significant uncertainty about these projections as the age group of interest coincides with the mobile working-age population. Owing to current uncertainty, we have not sought to build projections into our modelling; rather, when considering the output of the modelling, it should be remembered that admission numbers may change to be between 1% lower and 4% higher over the course of 10 years.

Data

Model

To study the location of NICUs in England, we adapted the model criteria presented in Decision criteria to focus on improving the clinical outcome for VLBW infants. To do so, we aimed to include ≥ 100 VLBW infants per year in all modelled NICUs. As a result, the set of optimised criteria is a combination of travel time criteria and VLBW number criteria, as detailed in Table 5.

Estimation

In the first stage, the location analysis was restricted to the H = 45 existing NICUs. In the second stage, the location optimisation was extended to all H = 161 existing neonatal care locations.

For both stages, the MOO based on the NSGA-II method56 and the location model (see Model) was applied to the data described in Data, similar to the computation of optimal maternity locations (see Maternity unit location analysis).

The population size was set to P = 100 and remained constant through generations. The population was propagated through G = 200 generations; this value would provide a steady state, as described in Convergence metrics.

The MOO process was run independently 10 times for all h ∈ 〚1; H〛 with different first generations. Such iterations improve the reliability of analysing the content of the Pareto front configurations.

Results

The analysis of the Pareto front configurations provided by the optimisation process enlightens the relationship between the number of NICUs in England, the access to neonatal intensive care and the size of units.

Access to care

In Figure 7a, the best achievable average and maximum travel times from mother to unit both decrease as the number of NICUs increases. This relationship is tenuous, as the best achievable maximum travel time remains constant in the case of existing NICUs and it decreases by only 5 minutes between 45 and 60 units in the case of all possible locations. This can be explained: the optimisation criteria include the average travel time; as a result, Pareto front configurations are optimal for a majority of mothers. Such comparison highlights that the maximum travel time from any mother to the closest unit could be improved significantly by changing the location of units, reducing the maximum travel time from 142 minutes to 86 minutes with the current number of 45 NICUs.

FIGURE 7.

Influence of the number of NICUs on the access to care. (a) Travel time (minutes) (configuration: average and maximum travel times); (b) VLBW admissions (configuration: minimum and maximum number of VLBW admissions per year); (c) proportion of VLBW infants in target (configuration: proportion of admissions within 30 minutes of mother’s home, proportion of VLBW admissions in a unit with ≥ 100 VLBW admissions per year, and proportion of VLBW infants in a unit with ≥ 100 VLBW admissions per year and within 30 minutes of home); and (d) proportion of VLBW infants in target [same configuration as (c) but only showing the best performing configurations]. Quantified by the travel time to units and the proportion of women in target time, on the number of admissions and on the proportion of births occuring in large units. Dotted lines show configurations in which NICUs can be chosen from any neonatal care location; solid lines show configurations in which choice of location is limited to current NICU locations.

Example of an alternative neonatal intensive care unit configuration

Going beyond the performance metrics, it is interesting to study what the discovered Pareto front configurations mean for the patients locally. To do so, an example of an alternative NICU configuration was selected using the following method:

1. Select the Pareto front configuration in the highest quartile of the proportion of patients both attending units with ≥ 100 VLBW infants per year and living within 30 minutes of the closest unit. This allows the selection of the best-performing configurations.

2. Select the Pareto front configurations with all patients attending units with ≥ 100 VLBW infants per year.

3. Select the Pareto front configurations with the highest number of units.

From the 88,800 discovered configurations, the selection led to two configurations with 48 units, of which the one that was closest to the current configuration was chosen.

In Table 7, the criterion values of this alternative configuration are compared with the model of the current state regarding the care of VLBW infants. The admission numbers are also provided for the infants receiving > 24 hours of intensive care (in an ICU), regardless of their weight.

Overall, the alternative configuration, shown in Figure 8, offers a more homogeneous set of units in terms of admissions. Indeed, although the number of VLBW infant admissions, if VLBW infants attend their closest NICU, varies from 29 to 450 per year in the current model, it varies from 101 to 241 per year in the alternative configuration. The latter has the advantage of increasing the proportion of patients attending units with ≥ 100 VLBW infants per year from 90% to 100% while keeping a similar number of units. Furthermore, the proportion of mothers within 30 minutes of the closest unit is raised from 65% to 73%.

FIGURE 8.

Example of optimal configuration in England with 48 NICUs. (a) Background map © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright); (b) number of admissions per year by unit location; and (c) key performance indicators.

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

For instance, the area including Somerset, Bristol, Wiltshire and Gloucestershire, in the alternative configuration, is covered by one NICU in Bristol (with 143 VLBW infants), one NICU in Bath (with a predicted admission of 101 VLBW infants per year) and one NICU in Gloucester (with 139 VLBW infants), whereas it is currently covered by two NICUs in Bristol (with a predicted admission of 149 and 178 VLBW infants per year) if VLBW infants attend their closest NICU. This local alternative configuration offers faster access to the population outside Bristol, whereas the impact on the Bristol population is limited.

In the current model (assuming that VLBW infants attend their closest NICU), the Greater London area (as shown in Figure 9) is covered by Chelsea and Westminster Hospital (with a predicted admission of 32 VLBW infants per year), St Thomas’ Hospital (with a predicted admission of 29 VLBW infants per year), Homerton (with a predicted admission of 450 VLBW infants per year), King’s College (with a predicted admission of 136 VLBW infants per year), Queen Charlotte’s (with a predicted admission of 256 VLBW infants per year), Royal London Hospital (with a predicted admission of 210 VLBW infants per year), St George’s (with a predicted admission of 162 VLBW infants per year), St Peter’s (with a predicted admission of 379 VLBW infants per year) and University College London (with a predicted admission of 77 VLBW infants per year). In the alternative configuration, intensive care would be provided by Barnet Hospital (with a predicted admission of 127 VLBW infants per year), Chelsea and Westminster (with a predicted admission of 155 VLBW infants per year), North Middlesex University Hospital in Edmonton (with a predicted admission of 241 VLBW infants per year), Newham General Hospital (with a predicted admission of 140 VLBW infants per year), Northwick Park (with a predicted admission of 159 VLBW infants per year), Queen Elizabeth Hospital in Woolwich (with a predicted admission of 117 VLBW infants per year), Royal London (with a predicted admission of 153 VLBW infants per year), St Helier Hospital (with a predicted admission of 191 VLBW infants per year), Kingston upon Thames (with a predicted admission of 119 VLBW infants per year) as well as Wexham Park Hospital in Slough (with a predicted admission of 124 VLBW infants per year) and Darrent Valley Hospital in Dartford (with a predicted admission of 131 VLBW infants per year) on the periphery of Greater London. The number of VLBW infants per year would vary from 117 to 241 in Greater London as opposed to 29 to 379 with the current mapping and VLBW infants attending their closest NICU. As a result, all patients in Greater London could attend a unit with ≥ 100 VLBW infants per year. Moreover, as can be seen in Figures 9–11 and Table 7, the LSOAs in which mothers can reach a NICU within 30 minutes would be extended from central London to the peripheral areas (see Appendix 1).

FIGURE 9.

Example current configurations for neonatal units with a focus on the Greater London area, with 45 NICUs in England. Admission numbers assume that the closest appropriate unit is used. Chelsea and Westminster (32 VLBW infants), Guy’s and St Thomas’ (29 VLBW infants), Homerton (450 VLBW infants), King’s College (136 VLBW infants), Queen Charlotte’s (256 VLBW infants), Royal London Hospital (210 VLBW infants), St George’s (162 VLBW infants), St Peter’s (379 VLBW infants) and University College London (77 VLBW infants). © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright).

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

FIGURE 10.

Example alternative configuration for neonatal units in the Greater London area, with 48 NICUs in England. Admission numbers assume that the closest appropriate unit is used. Barnet (127 VLBW infants), Chelsea and Westminster (155 VLBW infants), Edmonton (241 VLBW infants), Newham (140 VLBW infants), Northwick Park (159 VLBW infants), Queen Elizabeth – Woolwich (117 VLBW infants), Royal London (153 VLBW infants), St Helier (191 VLBW infants), Kingston upon Thames (119 VLBW infants), Wexham Park – Slough (124 VLBW infants) and Dartford (131 VLBW infants). © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright).

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

FIGURE 11.

Example centralised configuration for neonatal units in the Greater London area, with 30 NICUs in England. Admission numbers assume that the closest appropriate unit is used. Queen Charlotte’s (231 VLBW infants), Croydon (238 VLBW infants), Dartford (343 VLBW infants), St Peter’s (356 VLBW infants), Edmonton (440 VLBW infants), University College London (176 VLBW infants) and Watford (214 VLBW infants). © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright).

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

A caveat of the observations around London is that we have used estimated road travel times to choose between hospitals. It is possible that the order of choice may be different if public transport is used (as is common in London). However, the observation remains that the hospitals in Greater London are currently concentrated in the centre of London despite many mothers living in the more outlying regions.

Example of a centralised neonatal intensive care unit configuration

An example of a centralised NICU configuration with 30 units was selected using the following method:

1. Select the Pareto front configurations with 30 units.

2. Select the Pareto front configurations in the best half of the proportion of patients both attending units with ≥ 100 VLBW infants per year and living within 30 minutes of the closest unit.

3. Select the Pareto front configurations with all patients attending units with ≥ 100 VLBW infants per year.

4. Select the Pareto front configurations with the highest number of existing NICUs.

If reducing the number of NICUs to 30 optimal locations, the impact on the access of care would be limited as the average travel time would increase to 29 minutes (+1 minute), the maximal travel time would increase to 153 minutes (+11 minutes) and the proportion of mothers within 30 minutes of the closest unit would decrease to 64% (–1%). The main impact would be focused on the minimum number of admissions, which would increase from 29 to 104 VLBW infants per year. Furthermore, changing the location of NICUs by upgrading existing LNUs would equalise the distribution of units in the country, which would then reduce the maximum number of VLBW infant admissions from 450 to 440 per year.

Building a resilient network

Resilience charts

Figure 12 shows the resilience or optimality probability of the locations of four Operational Delivery Networks (ODNs), as well as their existing level of care. The equivalent charts for other ODNs are available in Appendix 1. If the resilience probability is high, then the location is likely to contribute to building a NICU network resilient to changes. For instance, the unit in Dartford has a very high resilience probability, which may be attributable to the hospital being easily accessible by road. The unit in Plymouth also has a high resilience probability because there is a high demand in the local area. It is interesting to compare the probability graph with the existing level of care. Some units with a high resilience probability, such as Dartford, Haywards Heath and Gloucester, do not currently provide intensive care. In areas such as the South London Neonatal ODN, the resilience probability is rather low because several units are present in the restricted geographic area. As a consequence, selecting one unit or another would not significantly change the access performances.

FIGURE 12.

Optimality probability and existing level of neonatal care of locations in South East England. Neonatal ODN (first dark green group), South London Neonatal ODN (first medium green group), South West Neonatal ODN (second dark green group) and Southern West Midlands Maternity and Newborn Network (second medium green group). For a full explanation of this chart, see Appendix 1.

Data

The demand for high-dependency care was modelled as the number of neonatal admissions needing > 24 hours of high-dependency care. It was estimated in all LSOAs as the number of neonatal admissions predicted by the regression analysis (see Chapter 4, Geographic data coverage and predicting neonatal demand) and multiplied by a factor of 0.185.

The lists of existing neonatal units, geographic data and travel times were the same as those used in the intensive care location analysis (see Neonatal intensive care location analysis).

Example of an alternative local neonatal unit configuration

Following the example of an alternative NICU configuration in Neonatal intensive care location analysis, this section focuses on selecting 78 LNU locations to build an optimal high-dependency network. The number 78 corresponds with the current number of LNUs in England.

Example of a centralised local neonatal unit configuration

Following the example of a centralised NICU configuration (see Neonatal intensive care location analysis, Example of a centralised neonatal intensive care unit configuration), this section focuses on selecting 30 LNU locations to build an optimal centralised LNU network.

Data

The demand for special care was modelled as the number of neonatal admissions requiring > 24 hours of special care. It was estimated in all LSOAs as the number of neonatal admissions predicted by the regression analysis (see Chapter 4, Geographic data coverage and predicting neonatal demand) and multiplied by a factor of 0.825.

The lists of existing neonatal units, geographic data and travel times were the same as those used in the intensive care location analysis (see Neonatal intensive care location analysis, Data).

Example of a centralised special care unit configuration

Maternity units

For both maternity care and neonatal care, choosing an optimal configuration of units is a compromise between fair access to local care for every patient and access to experienced medical staff in units with a threshold number of patients.

The Royal College of Obstetricians and Gynaecologists recommends that all maternity units should ideally receive ≥ 6000 babies per year in order to guarantee the 24/7 (24 hours a day, 7 days a week) presence of a consultant on site. 61 The maternity location analysis (see Maternity unit location analysis) has shown that in order to achieve 100% of births occurring in such units, the number of birth centres would need to be reduced from 161 to approximately 72. Although such change is unrealistic, reducing the number of birth centres from 161 to 140 optimal locations could increase the proportion of births in 24/7 units from 30% to 50%.

Reducing the size of the network by a small number of units does not significantly change the average travel time if the new national configuration is optimally chosen, but it increases the maximal travel time. Average travel times have the potential to hide large effects that only affect a few people in more remote areas.

Access to local care is indicated by the proportion of patients within 30 minutes of the closest unit. Achieving this criterion conflicts with achieving all births in 24/7 consultant-led units. By reducing the number of birth centres from 161 to approximately 65, the proportion of patients meeting both the travel time target and the unit size target would be increased from 24% to 82%: the maximum achievable based on our results. Such a low number of birth centres is not necessarily ideal, but it must be expected that, as the number either increases or reduces, the proportion of patients meeting both targets will reduce either because distances become too large or because units become too few.

Neonatal units

An alternative neonatal configuration was analysed in Neonatal intensive care location analysis, Example of a centralised neonatal intensive care unit configuration. By changing the level of care provided by some units, but not the total number of units, it should be possible to admit all VLBW infants to units with ≥ 100 VLBW infants per year. This configuration would guarantee access to experienced staff to all patients. The neonatal units would also be more homogeneous in size, with fewer units receiving a very low or a very high number of VLBW infants (or other infants requiring intensive or high-dependency care). Note that the nurse capacity needed at the national scale for this configuration would remain similar to the current capacity, as the number of units for each level of care would remain almost identical; reconfiguration of this type would therefore be about improving care and access for patients rather than reducing costs to service providers. These results show that the current location of NICUs is not optimal for either access (closeness) or optimising the use of large NICUs admitting ≥ 100 VLBW infants per year; it is theoretically possible to improve both access to and coverage of large NICUs simultaneously if there was the opportunity to upgrade some units and downgrade others.

The analysis of the Greater London area has shown that current NICUs are concentrated in the city centre, probably for historical reasons. With the alternative configuration, the optimal locations for NICUs would be more evenly spread in neonatal units in the conurbation, closer to residential areas and main road axes. The analysis is similar in the Bristol area, where the local population would benefit from NICUs in Bath (rather than two NICUs in Bristol, the current configuration). It is likely that this observation is transferable to other conurbations, such as Birmingham and Manchester.

London is likely to be somewhat of a special case because of the greater provision and use of public transport than in other regions. We have estimated travel times based on typical road travel times. A more focused study of London is likely to benefit from use of public transport times in addition to road distances. Nevertheless, the finding that hospitals in London are not focused on the location where parents live is likely to still be significant; access to care is likely to improve by some movement of locations of care away from central London towards the population centres further away from the city centre.

General comments

Precision of model

The model had a warm-up period of 1 year, followed by 10 years of run-time. Variation between individual years in the base-case model (current configuration without resource constraints) were as follows:

• average distance from home – 0.6% coefficient of variation (CV)

• average number of infants in surgical phase intensive care – 6.4% CV

• average number of infants in Level 1 care – 3.0% CV

• average number of infants in Level 2 care – 1.1% CV

• average number of infants in Level 3 care – 0.6% CV

• average number of infants in Level 4 care – 1.1% CV

• average total number of infants in any neonatal care – 1.3% CV

• average nurse workload – 1.1% CV.

All 95% confidence limits of the mean estimated values above were less than ± 5% from the mean value. All results are presented as the mean across the 10 years.

Comparison between model and actual data

The average number of infants present and the average workload were compared between the model and the actual average admissions in the NDAU data. For this validation, the model was run with the assumption that all infants attend their closest appropriate unit.

The difference between predicted and actual admissions depended on the closeness of a unit to its nearest neighbouring unit (Figure 15). For units that were ≥ 15 minutes away from their nearest neighbouring unit, the accuracy of prediction of the number of infants present was typically ± 20–30%, or ± 2–3 infants, and prediction of nurse workload was typically ± 15–20% or ± 1 nurse equivalent workload.

FIGURE 15.

Violin plots showing accuracy of predicting neonatal unit occupancy and workload by proximity of a unit to its nearest neighbouring neonatal unit. Units are binned by proximity to the nearest neonatal unit: 0–15 minutes (18 units), 15–30 units (77 units), 30–45 minutes (35 units), 45–60 minutes (10 units), 60–75 minutes (2 units), and 75–90 minutes (2 units). Bars show the range of error in terms of absolute difference between model and actual occupancies and workload, with the middle cross-bar showing median error. The shaded regions shows distribution of error. The charts show error in occupancy [(a) and (c)] and nurse workload [(b) and (d)] expressed either as differences in numerical occupancy and workload [(a) and (b)] or percentage error from the actual occupancy and workload [(c) and (d)].

Median travel times from home to place of care compared well between the model and actual data. The median time from home for Level 1 care was 24 minutes in the model (compared with 21 minutes actual), Level 2 care was 14 minutes (compared with 16 minutes actual) and Level 3 care was 12 minutes (compared with 13 minutes actual).

Providing further credence to the principle of this geographic modelling, that infants will usually attend their closest neonatal unit (or one close to it), is that 95.3% of infants who only required local Level 3 care (special care) were cared for in either their closest unit or a unit no more than 15 minutes further than their closest neonatal unit.

Literature review

During the last 20 years, many models have been developed to estimate infant and neonatal mortality. The majority of these models explored the impact of infant characteristics on mortality and estimated mortality for either the VLBW infants, weighing 800–1500 g, or very preterm infants, with a gestational age of 22–32 weeks. Most studies used a logistic regression approach with covariates that have been found to be strong predictors for neonatal mortality (i.e. of gestational age, small for gestational age, sex, birthweight and birthweight z-score). 70–73 A few studies considered sociodemographic and socioeconomic variables including race and ethnicity,18,20,74–76 education, insurance status, percentage of inhabitants living below the poverty line18,20,77 and the lowest decile of the IMD score. 19

A smaller number of studies have investigated the impact of service configuration in terms of (1) working patterns and staffing and (2) organisational level. One of the first studies that looked at the impact of workload and staffing was by the Tucker and UK Neonatal Staffing Study Group,78 which explored the impact of the availability of consultants (high availability defined as ≥ 2 consultants) and nurses relative to the BAPM nurse-to-cot ratio guidelines (high availability is defined as a nurse-to-infant ratio of ≥ 0.84). The UK Neonatal Staffing Study Group78 used three workload measures, including occupancy, which measured the maximum number of infants present in a unit over their study period. The authors reported that for every 10% increase in percentage of maximum occupancy at admission, the odds of mortality increased by 1.09 (95% CI 1.01 to 1.18). Rautava et al. 79 explored the impact of working patterns on mortality and found that the risk of mortality for very preterm infants increased when the infant was born in non-office hours, and that mortality rates could be consequently improved by an increase of resources. Finally, Watson et al. 14 estimated the effect of the 1 : 1 nurse-to-patient ratio for intensive care neonates and found that a 1 : 1 nurse-to-patient ratio reduces infant mortality. A further set of studies explored the degree to which mortality varied depending on the care levels by which services were organised (e.g. NICUs, LNUs and SCUs in the UK). For example, Cifuentes et al. 20 estimated the mortality of low-birthweight infants for different levels of care in the California area and found that the level of care in NICUs can influence the probability of survival. The 2010 systematic review by Lasswell et al. 80 explored the association between the designation level of hospital and VLBW infant mortality based on studies that evaluated the regionalisation of perinatal services for very preterm or VLBW infants. Most studies that looked at the impact of organisation in terms of staffing or hospital level used logistic regression, with the exception of Watson et al. ,19 who used an instrumental variable approach. Care is needed in interpreting estimation of the logistic model when there is a small number of events, as is the case in neonatal mortality, and so these estimates may be affected by small-sample bias.

Other studies have aimed to look at the impact of organisational factors, such as the number or volume of infants treated in neonatal units, based on the idea that staff can become more experienced and skilful as they treat a larger number of different and complex infants. A high-volume unit has been defined as one that treats at least a fixed number of VLBW or very preterm infants per year76,78 or as one that is in the top quartile of all neonatal units. 19 Rogowski et al. 76 explored the impact of hospital-level determinants of mortality among VLBW infants and showed that the number of VLBW infants admitted to a neonatal unit reduced infant mortality. More recently, studies have aimed to look at the causal relationship between infants born in high-volume units and infant mortality using instumental-variable (IV) approach methods. 14,19,77 The effect of designation and volume of neonatal care for preterm birth reported in Watson et al. 19 had a significant effect, especially on neonatal mortality and in-hospital morbidities (bronchopulmonary dysplasia, necrotising enterocolitis and retinopathy of prematurity). For infants with < 33 weeks of gestational age, the risk of death was reduced by 2.6 percentage points if admitted to a high-volume unit, and this effect was higher if infants had a gestational age of < 27 weeks. Finally, some studies explored the impact of hospital volume within different levels of care. For instance, Phibbs et al. 18 examined the impact of hospital volume among different levels of neonatal care units and found that volume and level had a significant impact on risk of infant mortality, showing that the delivery of VLBW infants in NICUs with a high volume can reduce neonatal mortality.

Table 14 summarises the neonatal mortality models that explored the impact of volume on mortality.

Thus, the few studies that have attempted to estimate causal effects suggest that there is a positive effect of birth in higher-volume hospitals, whereas there is no evidence that birth in a hospital with a NICU has any effect on mortality. However, previous studies have only compared NICUs against all other hospital designation categories combined without distinguishing between SCU and LNU hospitals. Furthermore, there is no study in England that has evaluated the causal impact on LOS and reimbursement costs. Our aim was to estimate the impact of volume and designation level on mortality and costs between NICUs, SNUs and LNUs, and separately between high-volume units and other hospitals of birth.

Data

Data relative to neonatal care were collected in 2014/15 from units in England as part of the BadgerNet data set. The distance in time and miles was evaluated using LSOAs of the mother and the postcode of the hospital. Mortality was defined as mortality during the in-hospital period from the admission to the discharge.

Mortality was registered between 2014 and 2015 for a total of 2010 infants, out of a total number of 165,450 admissions to neonatal units. Out of all registered deaths, 52% were for infants born with a gestational age of < 28 weeks. By adding infants born between 29 and 32 weeks of gestational age, 65% of deaths are covered; including infants born between 33 and 36 weeks, up to 83% of deaths are accounted for. Figure 18 illustrates the rate of death and survival per gestational age of infants admitted to neonatal units in our study.

FIGURE 18.

Mortality and survival rates of infants by gestational age.

Method

High-risk infants tend to be treated in high-volume units, so in order to estimate the effect of volume on mortality, it is important to measure this effect in a representative sample of infants, including both high- and low-risk infants. More recent studies have aimed to estimate the causal effect of volume using an IV approach to control for confounding. The proximity to care can be used as an instrument that determines the chances of receiving care (e.g. birth in a hospital with a NICU) and should be independent of infant mortality. Proximity to care can be used to control confounding by estimating the difference in mortality between those who live close to NICUs and those who live far from NICUs, and both groups should be made up of a comparable mixture of high- and low-risk infants. The strength of an instrument can be tested by looking at the relationship between the instrument (i.e. travel time) and attendance at high-volume NICUs. In our case, a strong instrument would imply that those who live nearest to high-volume units are more likely to attend at those hospitals (i.e. the t-statistic on travel time, or the F-statistic if there is more than one instrument, is > 10).

Watson et al. 19 used an IV approach based on travel distance alongside eight other instruments [i.e. surgical facilities (1 = yes, 0 = no), high volume (1 = yes, 0 = no), Level 3 (1 = yes, 0 = no), Level 2 (1 = yes, 0 = no), distance × surgical facility, distance × high volume, distance × Level 3 and distance × Level 2]. Level 3 represents NICUs and Level 2 represents LNUs.

The main analyses will use only one instrument (travel time or travel distance) to facilitate interpretation and to eliminate the need to identify and remove weak instruments when multiple instruments are used. The secondary analysis will use multiple instruments. We control for the following covariates in the model: age and age squared at birth, sex, deprivation of residence (quintiles of multiple deprivation), mode of delivery (emergency caesarean without labour, emergency caesarean with labour, vaginal non-spontaneous, elective section, unknown or vaginal spontaneous) and fetus number. A similar approach is used to estimate the impact of high volume on total LOS (sometimes referred as the super stay) and the associated reimbursement costs by level of care (BAPM) actually received, for infants referred to high-volume units, where length of hospital stay is defined as the number of days from admission to hospital discharge or death, whichever took place first. The LOS results are shown in the evaluation section of this report (see Chapter 9).

Two sets of neonatal mortality models are estimated: semiparametric and parametric models. The semiparametric model is a structural mean model (SMM)82 and serves to estimate the treatment effect on the treated, thus allowing for the possibility of different treatment effects between the treated (NICU-born) and untreated (non-NICU-born) infants. The parametric model instead adopts a bivariate (probit) distribution for mortality outcomes and the exposure status (birth in a hospital with a NICU vs. birth in a hospital without a NICU) and allows us to estimate the average treatment effect (i.e. the effect in the whole infant population, at the cost of imposing the assumption of homogeneity of treatment effect and normal distribution of unobservable confounding factors). We present both sets of results, but in the main discussion we focus on the semiparametric results, given that these results are based on less restrictive assumptions and so are more robust estimates of the effect. We have instrumented both the semiparametric and parametric models using an IV approach and, for comparison, run a linear probability model (LPM) that has no instrument [i.e. ordinary least squares (OLS)].

Two different units were considered: high-volume and tertiary units (i.e. NICUs). High-volume units are characterised by a minimum number of 100 admissions per year of infants with a birthweight of < 1500 g. Tertiary units are represented by NICUs that are the highest level of neonatal care, providing a service dedicated to babies needing respiratory support (ventilation) weighing < 1000 g, born at < 28 weeks’ gestation or needing significant continuous positive airway pressure support.

Like Watson et al. ,19 we estimated infant mortality for infants born at a gestational age of < 32 weeks in a high-volume unit or hospital with a NICU. In addition, we conducted analysis to explore the effect of neonatal transfers to a NICU hospital from a lower-level hospital in infants born before 32 weeks’ gestational age. We conducted a sensitivity analysis on the results for those born between 26⁺⁰ and 31⁺⁶ weeks of gestational age, to account for the possibility of bias due to the effect of imbalance in the distribution of extremely premature babies across treatment (hospital of birth) groups.

Secondary analyses of the relative effects of birth in a hospital with a NICU versus a SCU and a LNU were conducted using three available instruments of travel time to these three types of hospitals. These IV models were formulated as seemingly unrelated equations and estimated by simulated maximum likelihood. 83 In these analyses, neonatal mortality was analysed using multivariate probit distributions. Because we did not have complete information on the closest LNU and SCU units to some infants in the data set, we conducted these analyses excluding infants with incomplete data.

Results

We first checked if travel time was correlated with the exposure variable (i.e. birth in a hospital with a high-volume neonatal unit, and birth in a hospital with a NICU), thus supporting its use as an IV. Second, we looked at whether or not travel time is correlated with any other sample characteristics, such as birthweight and sex, to check if any possible effect of travel time on mortality operating through the exposure variable is confounded by other variables.

The descriptive statistics in Table 15 summarise sample characteristics by tertiles of travel time for high-volume units and NICUs and shows that in most cases there are no systematic differences of sample characteristics across the travel time tertiles. In addition to the exposure variables (delivery at hospital with a NICU and delivery at hospital with a high-volume unit), systematic differences arise only for deprivation of residence and unknown delivery mode; this suggests the need to control for possible confounding by these variables in our analyses.

Table 16 shows the mortality model using LPM, semiparametric IVs and parametric IV bivariate probit model (marginal effect) for the high-volume units and hospitals with NICU. The controlled covariates used in these models are gestational age and age squared at birth, birthweight, sex, deprivation of residence, mode of delivery and fetus number.

The instrument strength reported for the IV model estimates, both in the linear SMM and bivariate probit models, shows that the instrument is strong (e.g. the t-score test statistic on travel time is 32 for high-volume units; rule of thumb is that a robust instrument has an F-test statistic of > 10). In addition, travel time affects the treatment variable (the probability of delivery at a hospital with a high-volume unit in one analysis and delivery at a hospital with a NICU in another) in the expected direction and is negative, implying that the longer travel times are associated with a lower likelihood of birth in hospitals with high-volume units and of birth in hospitals with NICUs.

The estimated causal effects of the IV models indicate that delivery in high-volume units reduces neonatal mortality by 1.2 percentage points in accordance with the bivariate probit model and by 5.0 percentage points with the linear SMM. In contrast, the LPM reports high-volume units having an increased risk of death, but this result is likely to be affected by confounding issues (i.e. high-volume units treat high-risk infants and so are likely to have higher mortality), as indicated by the Hausman test statistic. Sensitivity analyses excluding infants born at < 26 weeks’ gestational age were also conducted and they confirmed these findings (see Appendix 2).

Birth in a hospital with a NICU does not appear to result in any difference in terms of mortality relative to other hospitals, as summarised in Table 13. These results are based only on the use of one instrument: travel time to closest NICU hospital. We ran additional analysis to explore the effectiveness of NICUs for those infants born in a hospital with a NICU or transferred within the first 48-hour period. Of all neonatal transfers with a recorded transfer time (n = 1519), 65% took place within 48 hours of birth. Infants who were born in a hospital with a NICU or who were transferred to a hospital with a NICU had no detectable effect on mortality, using the IV approach based on the single instrument of travel time to closest NICU (or distance to closest NICU). The same results were obtained when the sample was limited to those infants born between 26⁺⁰ and 31⁺⁶ weeks’ gestational age (see results in Appendix 2).

Secondary analysis of the relative effects of birth in a hospital with a NICU compared with a hospital with a SCU or a LNU were conducted using three available instruments of travel time to these three types of hospitals. In these additional analyses, we find that the NICU does appear to reduce mortality, compared with the other levels of care, by 2 percentage points, and so suggests that NICUs in themselves have some beneficial impact on mortality compared with other levels of care (see Appendix 2).

Discussion

We estimate infant mortality for infants born at a gestational age of < 32 weeks as a function of exposure to high-volume unit or hospital with a NICU at birth. We find that exposure to a high-volume unit at birth reduces mortality relative to other neonatal units. A very preterm infant born in a high-volume unit (≥ 100 babies weighing < 1500 g per year) has a 5-percentage-point lower risk of death in this unit than in other neonatal units when travel time is used as the instrument and mortality is estimated using a semiparametric method (linear SMM). This estimate drops to 1.2 percentage points when the same analysis is run using a parametric method (bivariate probit). There is a debate to be had about which result should be given greater credence. The semiparametric approach is based on less-restrictive assumptions and, therefore, is potentially more robust to violations of assumptions underpinning the analysis, so we have chosen to emphasis this result here; to allow comparisons with existing literature in this area, it is necessary to further discuss the parametric result. The parametric result is also the one that we use in the evaluation section of this report to calculate the incremental cost-effectiveness of high-volume units, to avoid potential problems with predicted values outside the 0–1 probability range.

Watson et al. 19 found that a preterm infant born in a high-volume unit (defined as those in the top quartile of all neonatal units) has a 2.6-percentage-point lower risk of death than in other neonatal units when travel distance is used as the instrument and mortality is estimated using a parametric method. This percentage point risk reduction is not immediately apparent from the paper, but can be calculated from the reported OR of 0.68 for in-hospital mortality reported in the paper (which approximates the risk ratio in cases like this when the deaths are rare), and the in-hospital mortality for high-volume units reported as 5.5 percentage points in the descriptive statistics [giving an estimated percentage point reduction of approximately = (5.5/0.68) – 5.5 = 2.6 percentage points]. The 2014 estimate of Watson et al. 19 is higher than the 1.2 percentage point reduction found by our parametric approach. An obvious explanation for the differences is the definitions used for high-volume units, but similar results were found when we defined high volume as those in the top quartile of all neonatal units. Another reason for the differences is the instruments used, given that Watson et al. 19 used travel distance and we explored the use of both travel distance and travel time. We report here only the results based on travel time as an instrument because there is strong support for travel time accurately representing access to health-care services. Sensitivity analysis excluding infants born at < 26 weeks of gestational age halves the mortality effect of birth in high-volume units compared with other units (2 vs. 5 percentage points in all the infants aged < 32 weeks’ gestational age), but the estimates are imprecise.

A baby being born at < 32 weeks’ gestational age in a hospital with a NICU does not appear to result in any difference in terms of the risk of death compared with other units. Similar results are also found by Watson et al. 19 We ran additional analysis to explore the effectiveness of NICUs for those infants born in a hospital with a NICU or transferred within the first 48-hour period. Birth at a hospital with a NICU or being transferred to a NICU within a 48-hour period had no detectable effect on mortality when using the IV approach based on the single instrument of travel time to closest NICU (or distance to closest NICU).

In our interviews, policy-makers raised queries about the robustness of the finding that NICUs did not affect the risk of death compared with other hospitals. To check the robustness of the result, we present additional analyses comparing the effectiveness of NICUs with other levels of care (i.e. NICU vs. LNU and NICU vs. SCU). All other results for the NICU were based on only one instrument: travel time to closest NICU. The advantage of considering other levels of care is that it opened up the approach to using three instruments: travel time to the closet NICU, LNU and SCU. A slight disadvantage of the approach is that the data set was less complete, because we did not have complete information on the closest LNU and SCU for some infants in the data set; therefore, we were restricted to performing these additional analyses on a smaller data set that excluded infants with incomplete data. In these additional analyses, we find that NICU does appear to reduce mortality by 2 percentage points, compared with the other levels of care.

A limitation of the data and analysis is that it is currently not possible to estimate the impact of mortality for those infants who are transferred into NICUs. The numbers of transfers are not insignificant; for example, for all infants with a gestational age of < 33 weeks at birth and who have ≥ 7 days of BAPM Level 1 (intensive) care, 61.2% are born in a hospital with a NICU, 18.9% are born in a hospital without a NICU and are transferred to a NICU and 19.8% are born in a hospital without a NICU and are not transferred to a NICU. We are therefore unable to separate out the benefits of antenatal care taking place prior to birth. It is clear that birth in a high-volume unit leads to improvements in mortality, and closing down low-volume units might lead to more infants being born in high-volume units, but the impact of changes in transfers is unclear.

Literature review

There are a number of papers that look at the impact of gestational age on hospital neonatal costs and families. Rogoswki et al. 87 explored the impact of gestational age on neonatal and perinatal cost in the American Vermont Oxford hospital network for NICUs. Costs were classified into accommodation costs and ancillary costs; ancillary costs were divided into five subcategories: respiratory therapy, laboratory, radiology, pharmacy and other ancillary. The authors show how costs vary within gestational age, birthweight, location of birth and discharge status. The category of infants who have the highest costs are those born between 24 and 26 weeks, with a birthweight of < 1000 g and born outside the hospital. The study also shows an inverse relationship between costs and gestational age; this result is also confirmed for neonatal and childhood costs for extreme preterm88 and preterm births. 89 Further work by Petrou et al. 88 estimated costs for extremely preterm birth for families using evidence from a population study. Results show that extremely preterm births are associated with higher public sector costs and there is an inverse relationship between costs and gestational weeks. Several sociodemographic covariates were included in the model, but only long-term unemployment is associated with an increase in costs.

For service reorganisations, it is important to consider the impact on costs as services change and the effects of economies of scale and scope. One approach used to address this question is to look at the elements that make up the costs and analyse how they vary by case mix. The UK Neonatal Staffing Study Steering Group developed a cost function to evaluate the nature and the degree of economies of scale in the provision of care for NICUs. 90 The economic analysis shows that volume and case mix interact to determine the degree of economies of scale, even if the determinants of costs and efficiency in neonatal costs have a high complexity. The treatment of the sickest infants centralised at a regional level can take advantage of economies of scale. Another study by O’Neill et al. 91 investigated the relationship between activity (total days of care provided and total days of intensive care provided) and costs (clinical staffing, support services and overheads) using a multivariate regression model. They found an inverse relationship between average cost per day and scale of services provided, confirming the benefits of centralisation of intensive care in larger units. The authors91 also show that the adoption of a different form of estimation (i.e. the log–log or double-log function) provided the best fit to the data.

Another approach is to simply look at the costs of high-volume units compared with low-volume units. Watson et al. 92 costed NICU services using the tariffs paid to hospitals to cost high-volume NICUs compared with low-volume NICUs, and compared their effectiveness in terms of reductions in the risk of mortality to estimate the cost-effectiveness of moving £100 to high-volume NICUs. The study estimated an incremental cost per life saved of £420,000 per life-year saved. 92

Reference costs and how they inform Healthcare Resource Groups

Historically, HRGs for infant intensive care tended to be too low and HRGs for infant special care tended to be too high. Intensive care for infants should use similar costs to intensive care for adults and so should be much higher. For example, the ratio of HRGs in 2014/15 were intensive care = 2.8 × special care, high dependency = 2 × special care, special care = transitional care. This anomaly has arisen because of the way units have submitted reference costs; there is a tendency to average the nursing over all infants rather than to apportion nurses’ costs to the care needs of infants based on BAPM guidelines. 2 Differences between units’ costs have also arisen as a result of the way that units apportion:

• costs between neonates and paediatrics

• costs between the different neonatal unbundled HRGs

• diagnostic costs

• layout and organisation costs, etc.

Reference cost submissions inform HRGs, but there is a lag whereby HRGs change more slowly than the reference cost submissions. Payments continue to be based on the 2001 HRGs despite the new BAPM classification in 2011. An update to BAPM 2011 was agreed in 2015, and this went through the appropriate systems to flow into the NHS data collection systems from December 2016. There are now two sets of data being collected: HRG 2001 for price and payment, and HRG 2016 for reference costs. Figure 19 shows the information flows relative to the two BAPM classifications.

FIGURE 19.

Example of information flows for 2014/15 data.

Length of stay and costs

Parent costs

Systematic review

The current guidance for attribute selection in DCEs emphasises the need to achieve a balance between the competing objectives of the participants and the decision-maker, the relevance of the research question(s) and if attributes are related to one another. 106–108 However, the challenges faced in selecting appropriate attributes for service provision, outcome or both have received less attention. The systematic review examines the extent to which researchers investigating antenatal and neonatal care through DCEs consider and justify in their design whether the attributes are for service provision, outcome or both (see Report Supplementary Material 1).

Methods

Systematic electronic searches of prespecified terms were performed in EconLit, EMBASE HMIC (Healthcare Management Information Consortium), MEDLINE, NHS EED (Economic Evaluation Database), PsycINFO and Web of Science databases, from 2000 to June 2016. DCE studies investigating topics on premature infants, neonates, newborns, or mothers/fathers/parents were included. Studies were excluded if the participants were children or adolescents aged < 16 years.

Results

After removing duplicates, 9701 unique results were identified but only 299 DCEs for antenatal and neonatal care were initially considered, of which 13 met the inclusion criteria. The selection of studies is presented in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram109 (see Appendix 3, Figure 31).

Most of the studies were conducted in industrialised countries (UK, n = 4;110–113 Australia, n = 3;114–116 the Netherlands, n = 2;117,118 and Canada, n = 1119), and three were conducted in developing countries. 120–122

The objective of the elicitation exercise varied in all studies from types of obstetric services,117 screening tests for Down syndrome,113 induction of labour114 and delivery,118 perinatal experiences119 and supplementary diet,121 among others. Most of the studies involved women participants, and health-care providers were included in only two studies. 115,122 Almost 80% (10/13) of the studies focused on antenatal care, including one that considered both antenatal and postnatal care. The number of individuals participating in the DCEs varied from 56 to 1464 (mean 284, median 130). The mode of administration was also diverse, with web-based questionnaires;110,115,118 postal questionnaires,111,117 and pen-and-paper questionnaires at clinics and teaching hospitals,113,114,119,121 with the rest reporting using a questionnaire but without specifying the mode of administration. 112,116,120,121 The reduction in studies selected (from 299 to 27) due to eligibility is because the vast majority related to paediatric rather than neonatal care and/or did not relate directly to DCE. Overall, the studies showed a heterogeneous picture of selection of attributes for service provision, outcomes or a mix of both, but there is a lack of consideration or justification for these choices.

Summary of approaches used to develop attributes within these discrete choice experiments

Attributes were selected by reviewing existing studies,111 as part of the characteristics of the intervention being evaluated in a randomised controlled trial,113,118 qualitative interviews with patients and/or stakeholders,112,113,117,118,122–124 literature review and qualitative interviews,110,115,116,120,125–127 and international guidelines. 121 However, in some studies it was unclear how the authors reduced the number of attributes obtained in qualitative research to a manageable set of attributes. 116

Focus group

At the beginning of the study, we started with a set of attributes suggested by the pilot work, which included one process outcome (access to neonatal services) and three health outcomes: maternal mental health, infant death and risk of childhood health problems. We presented these attributes to the first focus group before conducting the qualitative interviews, giving examples of how they might be shown to parents:

• maternal mental health – reduce anxiety and depression

• risk of infant death for those born at 24 weeks of gestational age – reduce from 5 in 1000 to 2 in 1000 infants

• risk of childhood health problems – eyesight problems increase from 1–2 in 100 to 2–5 in 100 infants

• access to neonatal services – increases travel time for some families from 60 to 120 minutes.

Although some of the outcomes, like access and travel times, are more straightforward, the issues of risk and uncertainties over morbidity are clearly complex, and so the research set out to explore the feasibility of DCEs, or other approaches, to collect data to inform the weights that decision-makers can apply in this context. Examples of framing considered for uncertainty in focus groups were as follows:

• Low confidence – our confidence in this effect estimate is limited.

• Medium confidence – we are moderately confident in this effect estimate.

• High confidence – we are very confident that the true effect lies close to our estimate.

Or:

• Between 1 and 10 infants, most probably 5 infants, out of 100 will experience eyesight problems.

Most parents felt that the presentation of these options was difficult and not easy to understand and compare, and two parents preferred to use the confidence levels rather than the values reported for risks. We also presented parents with a table to illustrate risk with a pictogram with 100 faces, showing 2% chance as two faces of different colours. Generally, it was felt to be a clear presentation of risk. We also showed parents the protocols we were aiming to use in the qualitative study, which aimed to discuss the parents’ experiences focusing on a set of open questions. Little additional feedback was given on these protocols in the focus group.

Patient preference interview

The qualitative study built on earlier public and patient involvement work,33 which aimed to identify outcomes that were important to families of children requiring neonatal care. The qualitative research explored these outcomes further and aimed to refine the language that is used to describe these outcomes, to explore in what ways these outcomes are perceived to vary by parents and to identify if any attribute might prove problematic in the DCE.

Methods

Results

Ten families were interviewed using the flexible topic guide (see Report Supplementary Material 2), which explored the following characteristics:

• Hospital environment.

• Travel time and means of transport.

• Impact on the family (emotional and financial).

• What is important for the family?

• Language used, sensitivity, medical language, how parents explain the events, services and treatment.

• Understanding of risks – infant survival and long-term disabilities.

• Mitigating aspects – family support, SNUG and BLISS support and hospital staff support.

• Background – gestation weeks, LOS and other children in family.

The families that were interviewed all had very different stories. The gestational time ranged from 24 to 34 weeks and the LOS ranged from 2 to 17 weeks. For half of the families, this was their first child. Some of the births were in a NICU and some were in a LNU; some of the babies born in a LNU were transferred to a NICU.

A summary of the synthesis under each theme as presented to the PPI group is presented in the following sections. We identified the themes in the coding process (see Report Supplementary Material 2).

Discussion of qualitative study

The results of the qualitative interviews show that interviewees talked more about the infant as a whole, rather than separating the risks of death and childhood health problems. This notion of a combined attribute is similar to the idea of an aggregate measure of length of quality of life used by NICE,133 which may be more meaningful if extrapolated forward to consider the longer-term impacts on infants, and not just short-term prognoses. In addition, the families made a connection between the baby’s and the mother’s health, and the importance of the mother’s health was considered to be equal to that of the child by the focus group.

The qualitative interviews also raised other process outcomes including communication with the families and family support.

Furthermore, the qualitative study raised questions about the ability and willingness of parents to trade off health attributes for the process attributes. Although no trade-off questions were asked in the patient interviews, it became clear that mothers were unlikely to want to sacrifice these ‘core’ aspects of their baby’s health for improvements in process outcomes. Parents stated in interviews that once they knew that their baby was receiving the best care, they could then perhaps consider other factors. This raised questions about including a DCE with combined health and non-health outcomes, because parents would always choose the configuration that favoured the best health outcomes for the infant (lexicographic preferences). However, in configurations that maintain a high level of care but affect process outcomes, trade-offs and DCEs become more feasible.

Methods

Our study estimates the cost-effectiveness of an intervention that leads to infants being born in high-volume units rather than in other neonatal units. We also calculate the effectiveness of NICUs versus LNUs and versus SCUs. In contrast, Watson et al. 92 estimated the cost-effectiveness of investing an additional £100 of resources in NICUs per day. Their approach also differs from the one taken here in the way that they estimated costs. In their approach, the costs of neonatal services were based on the HRG for each hospital, which is part of the national reference cost submission and is used to inform future HRG tariffs. They investigate the extent to which hospitals with higher HRG unit (per diem) costs have lower mortality rates. In contrast, our approach is more in line with the traditional cost-effectiveness approach in that we focus on the impact of volume on outcome and LOS and then cost the LOS using the corresponding fixed tariff HRG per diem reimbursement for the number of days spent in each level of care (intensive care, high-dependency care, special care and transitional care).

The methods used to estimate LOS are described in detail in Chapter 8, Methods. In Chapter 9, Results, we bring together the LOS and mortality estimates from these two sections to calculate the incremental cost-effectiveness ratio (ICER).

Results

Table 20 summarises the IV results for LOS and costs for (1) high-volume units compared with other units and (2) unit designation (i.e. NICUs vs. LNU and NICUs vs. SCU). Table 20 shows that birth in a high-volume unit increased the total LOS by almost 9 days, representing an additional mean reimbursement cost to commissioning groups of £5715 relative to births in non-high-volume units. Dividing this additional cost by the reduction in neonatal mortality in high-volume units (£5715/0.012) results in a cost per neonatal life saved of £460,887. If we suppose that an infant survives for 81 years, in line with the average life expectancy at birth for the English population,134 and they remain in full health for the entire period, then we can convert the cost per life saved into cost per life-year gained. We discount costs and effects using an approach recommended in the HM Treasury Green Book for longer-term interventions,135 and the discount rate is 2.5% for first 30 years of life, 3% from the 31st to the 75th year of life and 2.5% for 76th to the 81st year of life. This analysis results in an ICER per life-year gained of £15,620. If data become available on the life expectancies or quality-adjusted life-years of infants admitted to neonatal units, then this figure can be further updated.

The results of the analysis for hospital designation show that birth in a NICU exposes infants to a lower risk of neonatal death than birth in a LNU (a 1.9-percentage-point risk reduction), whereas there is a statistically insignificant effect of NICUs relative to SCUs, which may be attributable to the small number of SCU cases (10%), the different type of care provided and the imprecision of the IV method. Furthermore, birth in a NICU results in an additional 1–3 total days in hospital than birth in a LNU or SCU, and higher reimbursement costs for NICUs compared with SCUs, but the effects are imprecisely estimated (with overlapping CIs).

Overall, a NICU is more costly and no more effective than a SCU, whereas a NICU is more effective and cost saving when compared with a LNU with an incremental cost per neonatal life saved of –£43,096 compared with a LNU. If we suppose that an infant survives for 81 years, and again discount costs and effects, the cost per life-year saved is –£2279 for a NICU compared with a LNU; however, more investigation is needed regarding these results to check the robustness of the findings.

Discussion

Births in high-volume units compared with births in other neonatal units have an incremental cost per life saved of £460,887. Comparing this initiative with other similar initiatives that save lives suggests that the intervention is likely to be cost-effective. 92,136,137 When we compare NICUs with LNUs, we find NICUs to be cost saving (although not statistically significant) and to reduce mortality, and so NICUs are likely to be cost-effective.

We also estimated the ICER per life-year gained using a set of simplifying assumptions and a declining long-term discount rate. Births in high-volume units compared with births in other neonatal units have an ICER per life-year gained of £15,620. Currently, NICE uses a threshold range of £20,000 to £30,000 per quality-adjusted life-year gained,133 but there have been a number of projects to further investigate this threshold. 138–140 The ranges from this study of cost per life-year gained seem to fall within these current thresholds, but future work should aim to explore the quality-adjusted life-years for neonates who survive to provide more accurate estimates.

In conclusion, the estimation of LOS analysis seems to be a useful and appropriate metric against which to explore the impact of changes in service delivery, and, given the data available, was used here to estimate the commissioner’s reimbursement costs. However, we are aware that these commissioning costs are likely to be lower than the actual service delivery costs, and that there are recent initiatives within the NHS to collect data that more accurately reflect the BAPM nursing requirement for each level. There are now two sets of data being collected: HRG 2001 for price and payment, and HRG 2016 for reference costs (which are more in line with BAPM nursing requirements). Future estimates of the impact of LOS on actual costs will be greatly improved by the ability to estimate actual costs of service delivery.

This analysis also pointed to areas for future research. For example, recent literature questions the use of a threshold to assess intervention efficiency when there are economies of scale or scope,141 an issue that is most relevant to interventions looking at the cost-effectiveness of service delivery. In addition, service delivery interventions tend to cover a wider range of costs and effects, which need to be fully reflected in the opportunity costs. 142 As we move forward in the evaluative frameworks for service delivery, it seems important that such wide considerations are taken into consideration.

Methods

Discussions with policy-makers and PPI groups and the review of the literature led to a list of key elements that were likely to be affected by service reconfiguration. In this report, we focus on three main categories that we have data on and can cost: nursing, travel times and transfers. In the current evaluation, we were unable to incorporate childcare costs (because the numbers of parents with other dependent children were unclear), costs of doctors (because, as revealed in the site visits, there are several differences arising from staff composition: the ANNPs and junior consultants have similar tasks but different salaries and categories) and overnight stays (without further consensus when these should be provided). Food, parking and unpaid leave policies also affected the costs for families, but these are not included because they are unlikely to greatly change with service configuration.

Mortality can be estimated by regression analysis, with information on birth weight (continuous, which means we can work with any meaningful categories); sex; mode of delivery (emergency caesarean not labour induced, emergency caesarean labour induced, vaginal spontaneous or unknown); quintiles of multiple deprivation; multiple foetuses (≥ 2 vs. 1); and level of hospital of birth (NICU or other) or, alternatively, high-volume units of birth (≥ 100 infants born weighing < 1500 g per year). We do not include the mortality estimates in these provisional estimations, given concerns about how best to assess the impact of transfers of infants into the analysis.

Predictions of NHS costs are based on the level of care received and a microcosting based on WTE nurses. Unit costs for WTE nurses are based on Curtis. 38 Although each band of nurses has a range of salaries attached to it, the cost of nurses was calculated based on the average unit cost of bands 5, 6 and 7 for neonatal nurses with and without specialisation for 2014/15, which amounted to £48.50 per hour. To cost infant transfers, we used the national reference costs associated with the HRG code XA06Z, which covered Transportation in Neonatal Critical Care and amounted to £1100.97 in 2015.

The costs to families that we include in the current model are travel time and vehicle operating costs. The unit costs applied to travel time are based on the Department for Transport’s non-business costs of travel. 39 Travel for non-business is based on a study estimating the willingness to pay of travellers for shorter travel times. 40 This study found that travellers gave the same value to time in all modes of transport. The Department for Transport costs business travel at a higher rate, which varies between modes of transport based on lost workplace productivity and the wages of employees who typically use that mode of transport. It is recognised that travel may also take place in work time for spouses of women who have undergone caesarean sections (and are advised not to drive in the first 6 weeks), who may take time off work to drive the mothers to the units. The cost of travel time per hour in non-working time is £5.56 and the cost of travel time per hour in working time is £31.98; here we used an average cost of £14.99 per hour. The unit cost of vehicle operating is based on motoring costs set out by the AA (Automobile Association),143 and includes an allowance for car parking.

Results

Table 21 summarises the costs of three different service reconfiguration scenarios under no capacity constraints and in which units run at ≈80–85% of maximum capacity (based on the results of the modelling shown in Table 12). The current configuration represents the actual configuration in England, whereas the centralised configuration considers a large reduction of units at each level of care. An alternative configuration was designed to minimise travel distances while having all NICUs receiving ≥ 100 VLBW infants per year. Table 21 reports the presence of no capacity constraint and the constraint of running units at 80–85% of maximum capacity. On-duty nurse recourses for each unit are set at the next whole number up from a theoretical 85% of capacity utilisation (e.g. if running at 85% of capacity is calculated as requiring 5.4 nurses present at any time, then unit resources are set to six nurses at any one time). The number of on-duty nurses is given here assuming that capacity is capped in accordance with the BAPM recommendations. If there is allowed working beyond the BAPM workload recommendations,2 then nurse numbers will scale down proportionally.

It is clear from the figures in Table 21 that nursing costs are the largest cost component, approximately 18 times higher than travel costs and 33 times higher than transfer costs. Nursing costs also reduce during centralisation, because of economies of scale, and so are likely to be a key driver for any observed changes in overall costs. For example, in the unconstrained system, greater centralisation (from 48 to 45 to 30 NICUs) reduces nurse costs and slightly increases transfers costs, resulting in cost reductions from a NHS perspective. Incorporation of a wider set of costs, for example travel time for families, results in families paying more in travel costs, but overall the more centralised configurations still reduce costs from a societal perspective. A similar result is found for the constrained system, although the reductions in costs are lower.

Discussion

The three scenarios (current, centralised and alternative configurations) show the potential impact on neonatal costs from service reorganisation. Nursing costs are the largest cost component and are reduced by centralisation; overall, this leads to reduced costs from a NHS and societal perspective. To assess the cost-effectiveness of these scenarios, more information is needed on the impact that these service reconfigurations have on mortality. High-volume units reduce mortality, but we can only show that these benefits are accrued for those born within those units. To assess service reconfigurations more fully, future research needs to explore the impact of high-volume units on infants transferred into those units after birth.

Our analysis has a number of limitations. The current analysis identifies potential savings from service reorganisations assuming that resources will be made available. We are, however, aware that some geographic areas may already be experiencing shortages of experienced nurses and consultants, and this analysis does not take into account staff availability. We have also assumed that the changes are instantaneous, and we have not modelled the transitional costs of moving from one type of configuration to another. To minimise disruption to the service, it is likely that reconfigurations will take the form of multiple sequential steps that take place over several years. The approach we have adopted for the economic evaluation is similar to that used by NICE, which does not typically take capacity or transitional costs into account in the economic evaluations underpinning clinical guidelines. However, for implementation, it is clear that these issues have to be factored in on a case-by-case basis.

Commissioners and policy-makers

One clear community of interest for our study is policy-makers and commissioners in health care who are keen to understand the issues and evidence supporting different options for neonatal care and to use this information to inform policy decisions as well as justify such decisions to others. Responses from our interviews with policy-makers in this study highlighted the importance of communication as key to promoting change, justifying policy initiatives and for exploring options of new models of care.

A clear need here is for relatively complex data to be conveyed in accessible formats that can be shared in a decision-making context. Such data often bear on the relationships between different (sometimes opposing) parameters of interest; for instance, the relationship between cost and outcome or geographic centralisation and localisation for alternative service delivery options. In such circumstances, the use of graphs or maps to represent a range of policy scenarios and demonstrate the modelled impact of these alternatives against key metrics can be essential.

In this context, graphical standards can be very important in supporting clear communication and understanding between groups; for example, the extensive adoption of forest plots in health-care meta-analysis and systematic reviews146 and cost-effectiveness acceptability curves in health technology assessments147 have greatly facilitated understanding and discourse, especially for relatively complex informational needs in policy contexts. In this project, we would point to our development of the Villeneuve chart (see Appendix 1) as a novel and potentially valuable graphical method of conveying important information in relation to location preferences and resilience of service centres. These charts could be applicable in a variety of contexts. The use of ‘violin plots’ (see Figure 15) demonstrates another use of information graphics to convey relatively complex aspects of the data analysis.

When maps and other graphical media are used to represent important findings, it is important to be aware of potential perceptual biases that can result. This is especially the case when interest groups may wish to promote or advocate different policy alternatives. One example is the undue prominence that can be given to rural areas in shaded heat maps. Rural areas are typically shown as overly large visual areas relative to their population. Cartograms, which scale geographic regions by variables, such as population (rather than physical area), can be used to overcome this perceptual bias,148,149 although such representations produce an unfamiliar and highly distorted image relative to a standard map of the UK. Although we did not explore the use of cartograms, this is clearly an area of interest that could be explored in this context.

Researchers

Researchers and academics commonly need to be able to extract clear and precise technical information from a study. For this reason, relevant data should always be made available in numerical form (e.g. in formatted tables). In addition, however, graphical and visualisation tools can be important to reveal trends, patterns and relationships in outputs, especially when these are not immediately apparent from the numerical data. Descriptive analyses of large data sets can particularly benefit from such methods,144 which often provide a useful synopsis of the key findings.

Parents and the general public

In discussion with parents in the PPI workshops (see Chapter 11, The workshops), it seemed clear that map-based presentations were most helpful when describing the various scenarios for geographic service configuration. In particular, parents could easily relate to maps that described services in their local regional area because they had direct and experiential knowledge of these services. In addition, parents readily understood narratives and examples of pathways of care because stories and case study examples can be readily understood. Line graphs and charts, for example those that describe trade-offs in centralisation versus localisation of services, often needed substantial explanation to be fully appreciated by parents. In such cases, it seems clear that careful consideration needs to be given to the mode and design of communication (e.g. the design of information presentations as well as forms for interviews and choice experiments).

For communication with the general public, it is notable that some of the publicity literature150 makes extensive use of information graphics, such as pictograms, to convey basic statistics about neonatal care in England. This approach seems particularly appropriate to help ensure that the information is both accessible and attractive to general readers and was adopted when communicating with parents in this project for some elicitation processes (see Chapter 8, Focus group).

Importantly, the use of visualisation methods in communication with parents and the public can be valuable in eliciting information as well as in presenting information. In our study, for example, pictorial representations were utilised to identify key concerns in a preference elicitation process. Further work could be useful in determining whether or not such graphical techniques could be deployed in the presentation of options and scenarios for preference elicitation within the proposed framework.

June 2016 workshops

These initial workshops were used to introduce the parents to operational research and its potential application to the planning of neonatal services. This was done by demonstrating how a simple operational model might be generated to analyse the flow of customers through a restaurant. We then applied this line of thinking to analysing the distribution of neonatal services and illustrated this with findings from our initial study of neonatal services in South West England. 33 It became apparent that the parents in the group were willing and able to understand how this type of research might be used to plan services and how issues relevant to parents might be built in to such a model so that it could reflect a number of different parameters set by various stakeholders including the parents.

During these workshops, the parents exchanged stories/experiences of using neonatal services. This had two benefits. First, it helped the parents to quickly gel as a group. Second, it was helpful for the researchers involved to appreciate the lived experience of the parents and how this might be taken into account in their work. The issues that were raised fell into the following themes:

• Daily living. Maintaining routines such as work and siblings attending school. Getting maternity and paternity leave, finding accommodation and food if living away from home for extended periods.

• Barriers to maintaining daily living. Distance travelled to access services and LOS. The strains placed on families increased as both of these factors increased. Coping with uncertainty was difficult and there was the impact of all of these factors on physical and mental health.

• Support.

  • Practical. Provision of accommodation [location (e.g. on the ward, in the hospital or in the town)]. Financial support to help with the cost of travel, parking and food.

  • Psychological. Stress on mother and on marital relationships and impact on siblings. A need was identified for ongoing psychological support in some cases, which could be provided through peer support.

It became apparent that when considering the potential of centralising neonatal care to improve outcomes for neonates, thought needs to be given to measures to mitigate the potential negative impacts on parents and families. This is likely to have a positive impact on the longer-term health and well-being of the neonate.

These workshops also highlighted the different experiences of living in semirural locations compared with urban locations. Cities potentially offer good public transport links. For people in rural areas, using public transport can greatly increase travel times. Finding accommodation for overnight stays within a hospital may also be more difficult in some urban areas and older hospitals.

Our health economics team also presented a planned DCE to the group. The parents raised concerns about some of the terminology used and the assumptions made: for example, that morbidity had been explained as ‘your baby having eye problems’. The group suggested that the health economics team needed to review its approach to ascertaining the ‘cost’ of using neonatal care (see Chapter 8, Results). Eventually, it was decided to employ a qualitative researcher to carry out interviews with a sample of parents to develop this further.

November 2016 workshops

We held a second pair of workshops in November 2016 in both Exeter and London. Parents were presented with the initial findings from the study. We explored with the parents how best to present differing outputs from the operational research. This included the use of graphs, maps and Pareto fronts. The parents found maps and Pareto fronts very visual and relatively easy ways to understand the data. Graphs were also seen as useful but these needed careful explanation. The parents found Pareto fronts a particularly useful way of visualising the balance that might be struck between the differing demands of providing accessible services and centralising neonatal units to improve clinical outcomes.

The group also discussed the need to explain the methodology that produced the results. The parents did not feel that this needed to be done in great detail. However, they felt that the process behind the production of these results had to be made transparent to deal with concerns that units may be being closed merely to save money. Given the high numbers of potential scenarios generated by operational research, the parents found the evolutionary analogy helpful in explaining how these scenarios might be narrowed down, but care needed to be taken with language and terminology (e.g. referring to variations of a scenario as ‘child’ scenarios).

The parents felt that once a limited number of scenarios had been developed by operational research methods, they should be involved in any final decision-making about neonatal reorganisation. They felt that their local knowledge and experiential expertise would be crucial in making a fully informed final decision. Our work in this workshop demonstrated that it would be feasible and practical to do this.

Parents also felt that, when possible, the transfer of babies between neonatal units should be minimised. They reported that there are sometimes significant differences in the way care is delivered between units, which can create additional stress for families (e.g. because parents have to learn how a new unit works).

In this workshop, a proposed health economics interview schedule was also reviewed. The parents expressed concern about some of the terminology used and the order of questions. After the workshop, Sue Prosser facilitated revisions to the interview schedule, working closely with the qualitative researcher to reorder and rework the interview schedule with parental input. This included a practice/mock interview in a role play with Sue Prosser. The parents also reviewed the flyer that was used to attract volunteer interviewees. Copies of the advertising flyer and interview schedules used for the qualitative interviews with parents are provided in Report Supplementary Material 2.

April 2017 workshop

This was held in Exeter with parents in London able to join via Skype and was a final workshop to present the finding of our work to the parents and to discuss possible dissemination activities. The parents confirmed that they would potentially be interested in being involved in further dissemination work. We discussed what we had learnt and how parents might be involved in the decision-making about potential reorganisations of neonatal care. The parents were very clear that they felt that they could and should be involved in this decision-making process. They felt that the evidence for the clinical benefits of centralisation, including the weaknesses in this evidence, needed to be presented to parents in an unbiased way. They also felt that operational modelling potentially offered opportunities to make the decision-making process more open and transparent but that any final decisions should incorporate their local knowledge alongside the knowledge derived from operational research.

Publications

A range of publications and presentations are planned on the basis of this work. These develop the themes and present them in more technical detail or explore some of the key methodological issues. These publications and presentations comprise:

• a detailed analysis of the geographic analysis applied to maternity units in the context of current guidance from the Royal College of Obstetricians and Gynaecologists

• more detailed and technical approaches to the mathematical methods used for the location analysis in this study

• an expansion of the technical aspects of the health economic modelling and a comparison of alternative approaches to cost and mortality modelling

• an expanded examination of the use of visualisation tools to communicate research output based in a thorough user-centred requirements analysis

• a paper that investigates the differences between PPI workshops and qualitative research approaches, using neonatal care and this study as a case study example.

Tools

The methodological framework presented in this report, we believe, provides the basis for a tool set that could be used to support an evidence base for policy-makers in neonatal health and care. Although considerable further development would be required to fully implement this framework, we believe that it has the potential to be readily adapted for general use.

Definition of the resilience probability

In order to analyse the possible scenarios, it is interesting to highlight which locations are more likely to appear in the optimal configurations (i.e. those configurations that make up the Pareto front). To do so, we compute the resilience probability of every location in accordance with the following method:

1. Select the Pareto front configurations with 35 to 55 units (current number 45 ± 10). This allows us to select configurations close to the current state.

2. Select the Pareto front configurations in the highest quartile of the proportion of patients both attending units with ≥ 100 VLBW infants per year and living within 30 minutes from the closest unit. This allows us to select the best-performing configurations.

3. Compute the probability p(u_i|h) of each unit u_i to appear in the selected ‘h-configurations’ (configurations with h units):

Such probability verifies:

1. Compute the probability p(u_i) of each unit u_i to appear in the selected configurations:

Such probability verifies:

The locations that appear the most often in the optimal configurations will have a higher resilience probability. Such an indicator is useful in order to select the locations that are the most resilient to network changes.

Reading a Villeneuve resilience chart

These charts display the resilience or optimality probability of the locations of four ODNs, as well as their existing levels of care.

The Villeneuve resilience chart provides a means to assess the relative probability that a neonatal unit location appears in an optimal configuration (as output by the genetic algorithm used for location analysis) and is therefore likely to be a good location to place a NICU. In Figure 20, unit locations in four separate networks are shown (represented by the upper level of bars). For each location, the probability of each existing unit appearing in an optimal configuration (coloured bar) as well as the unit’s existing care level [NICU, LNU or SCU (lower level of bars)] are shown.

FIGURE 20.

Elements within a Villeneuve chart.

In the example in Figure 20, it can be seen that some lower level units (e.g. Dartford and Haywards Heath) are very likely to be in favourable locations for the establishment of a NICU.

FIGURE 21.

Annotated elements of Villeneuve charts.

Resilience charts for other networks

FIGURE 22.

Villeneuve charts for specific neonatal network areas.

Definition of the resilience scenario score

The optimisation process provides thousands of Pareto front scenarios. To select a representative scenario u = {u₁, . . . , u_h}, we can compute the resilience score R(u) defined as follows for all scenarios:

Access to high-dependency care

FIGURE 26.

Configuration with current 45 NICUs and current 78 LNUs with high-dependency demand (Tables 22a and b). HD, high dependency. (a) Background map © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright); and (b) key performance indicators.

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

FIGURE 27.

Example of optimal configuration with 48 NICUs and 78 LNUs with high-dependency demand (Tables 23a and b). HD, high dependency. (a) Background map © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright); and (b) key performance indicators.

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

FIGURE 28.

Example of optimal configuration with 30 NICUs and 30 LNUs with high-dependency demand (Tables 24a and b). HD, high dependency. (a) Background map © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright); and (b) key performance indicators.

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Access to special care

FIGURE 29.

Current configuration with 45 NICUs, 78 LNUs and 38 SCUs with special care demand (Tables 25a–c). SC, special care. (a) Background map © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright); and (b) key performance indicators.

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

FIGURE 30.

Example of optimal configuration with 30 NICUs, 30 LNUs and 30 SCUs with special care demand (Tables 26a–c). SC, special care. (a) Background map © OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright); and (b) key performance indicators.

© OpenStreetMap contributors;60 the data are available under the Open Database Licence and the cartography is licenced as CC BY-SA (www.openstreetmap.org/copyright)© Queen’s Printer and Controller of HMSO 2018. This work was produced by Villeneuve et al. under the terms of a commissioning contract issued by the Secretary of State for Health and Social Care. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.