Adult healthcare is associated with more emergency healthcare for young people with life-limiting conditions

Background Children with life-limiting conditions receive specialist paediatric care in childhood, but the transition to adult care during adolescence. There are concerns about transition, including a lack of continuity in care and that it may lead to increases in emergency hospital visits. Methods A retrospective cohort was constructed from routinely collected primary and hospital care records for young people aged 12–23 years in England with (i) life-limiting conditions, (ii) diabetes or (iii) no long-term conditions. Transition point was estimated from the data and emergency inpatient admissions and Emergency Department visits per person-year compared for paediatric and adult care using random intercept Poisson regressions. Results Young people with life-limiting conditions had 29% (95% CI: 14–46%) more emergency inpatient admissions and 24% (95% CI: 12–38%) more Emergency Department visits in adult care than in paediatric care. There were no significant differences associated with the transition for young people in the diabetes or no long-term conditions groups. Conclusions The transition from paediatric to adult healthcare is associated with an increase in emergency hospital visits for young people with life-limiting conditions, but not for young people with diabetes or no long-term conditions. There may be scope to improve the transition for young people with life-limiting conditions. Impact There is evidence for increases in emergency hospital visits when young people with life-limiting conditions transition to adult healthcare. These changes are not observed for comparator groups - young people with diabetes and young people with no known long-term conditions, suggesting they are not due to other transitions happening at similar ages. Greater sensitivity to changes at transition is achieved through estimation of the transition point from the data, reducing misclassification bias.


Supplementary material S2: Classification of records as adult or paediatric
Healthcare use is known to vary with age. Age has the potential to be a confounder to the variable of interest (transition status) as those in adult care will generally be older than those in paediatric care. Sufficient years of data are needed to separate associations of the outcomes with age and with transition status. Inclusion of too many years decreases sample size (as it restricts to the models to individuals present for longer) and may underestimate short term associations with transition (effects of transition may be mostly short term, lasting a few years). Longer sampling timeframes may also introduce bias: for example, young people may change primary health care provider and leave the dataset around age 18 years if they move for education or employment and may differ from those who remain.
At least three years of data are needed, with at least two either in adult or paediatric healthcare to ensure age and transition status are not entirely collinear (over only two years of data, associations with age and transition would be indistinguishable). An a priori decision was made to use the last two years of paediatric data and first two years of adult data per cohort member and it was required that all cohort members were present in the data for at least these years. For the regressions, sensitivity analyses were conducted with all possible combinations of two, three and four years of paediatric data and one, two, three or four years of adult data.

Effects of sampling timeframe
Regressions were run with 2-4 years of data while in paediatric care and 1-4 years of data while in adult care.

Effects on incidence rate ratios dependent on transition status and age in year
Incidence rate ratios for the outcomes of emergency inpatient admissions and Emergency Department visits dependent on being in adult (compared to paediatric) healthcare and per year of age for males and females are shown in Figure S2, for the range of sampling timeframes tested.
The major differences are between regressions using one year of data when the young person was in adult healthcare or more than one year of data. Other variations in pre-and post-transition sampling years produce incidence rate ratios that have overlapping confidence intervals for the outcomes dependent on being in adult compared to paediatric care across all categories of condition and for the outcomes dependent on year of age except for the Emergency Department visits in the no long-term conditions group.
A key point is that when only one year of adult data is used, compared to those when more years of adult data are used, incidence rate ratios for the outcomes differ in opposing directions for being in adult (compared to paediatric) healthcare and age. This suggests a different model fit when only one year of adult data is used, with change sin outcomes being associated with changes in age rather than changes in healthcare. It is to be expected that too short a time period of data would make it more difficult to distinguish between associations with age and associations with transition status. This, and the apparent stability of incidence rate ratios once two or more years of adult data are used, suggests that it is the regressions using two or more years of data that most closely match the real associations between the outcomes, age and transition status.

Observations on model fit to all available data
Observed mean emergency inpatient admissions ( Figure S3) and Emergency Department visits per person year ( Figure S4) are plotted for the whole cohort using all data available from age 12 to 23 years, along with expected values from the regression models for the same population and age range. It should be noted that the models were developed on a subset of these data: 2-4 years of data while in paediatric care and 1-4 years of data while in adult care, as indicated in the figures. Models and observations therefore sometimes diverge for ages for which there were few individuals included in the sample used to generate the models -for example, transition for young people with no long-term conditions was set to 16 years, so no young people were included when over 19 years of age (the largest sampling timeframe was four years of adult data, ending for this group at age 19 years).
Figures S3 and S4 demonstrate poor fit when only one year of adult data is used, particularly for emergency inpatient admissions (Figure 2). Differences for other sampling timeframes are much less marked, although there is observable worse fit for some of the lines using two years of paediatric data compared to three or four years of paediatric data (in particular, for emergency inpatient admissions for females with diabetes ( Figure 2) and for Accident and Emergency visits for males with diabetes or no long-term conditions (Figure 3).

Appropriate choices of sampling time frame
Given the above, and the objectives set out in the main text regarding sampling timeframe selection, at least two years of paediatric and adult data appear to be required and the exact choice of sampling timeframe beyond that does not greatly change the conclusions drawn . Figure S2: Incidence rate ratios for emergency inpatient admissions and Emergency Department visits associated with being in adult healthcare (versus paediatric healthcare) and per year of age for males and females, depending on the years of data pre-and post-transition used in the regressions. Shaded regions indicate 95% confidence intervals. Figure S3: Fits of predicted numbers of emergency inpatient admissions per person per year by age from the models (with indicated years of data used) against all available data for cohort members. Figure S4: Fits of predicted numbers of Emergency Department visits per person per year by age from the models (with indicated years of data used) against all available data for cohort members.
Minimum age at exit from dataset required to have transition assigned The results were insensitive to variations in the required minimum age at exit from the dataset required for transition to be set to 16 years in the absence of sufficient data to estimate transition (Table S3 and S4). In practice, few individuals were present for long enough to have data for the last two years of paediatric care and first two years of adult care with a transition age of 16 years (i.e. present from at least age 14-17 years) for these decisions to have much impact. Exclusion of ethnic group and deprivation group Inclusion or exclusion of ethnic group and deprivation group had little effect on incidence rate ratios for other variables in the models (Table S5 and S6). Ethnic group and deprivation group were both at level 1 (individual) and so were partly accounted for by the use of a random intercept in the models.  Inclusion of year of birth Inclusion or year of birth for possible cohort effects (e.g. a young person born in 1992 experiencing different care and transition or being at a different stage of condition at transition age to a young person born in 2001) had little effect on incidence rate ratios for other variables in the models and on the combined associations of the outcomes with transition for each condition group (Tables S7 and S8).