Diabetes Risk After Treatment for Childhood and Young Adult Cancer Free

Improvements in the survival of individuals with childhood and young adult cancer (CYAC) are giving rise to an ever-growing population of adult survivors, with 35,000 CYAC survivors currently resident in the U.K. (1) and approximately 496,000 in the U.S. (2). The impact of a cancer diagnosis on the long-term health of these individuals remains substantial. By age 40 years, almost 75% of pediatric cancer survivors will have developed one or more chronic health conditions (3,4). The premature acquisition of modifiable cardiovascular risk factors such as diabetes has been shown to potentiate the risk for severe and fatal cardiac events in CYAC survivors and survivors of adult-onset cancer (5,6).

Although several studies have demonstrated that CYAC survivors are at increased risk of developing diabetes compared with the general population (7–11), published prevalence estimates vary. With the risks and development of diabetes in CYAC survivors varying according to patient and treatment-related factors (8,11,12), heterogeneity in reported risk estimates could be attributed to differing case-mix across these cohorts.

Large-scale retrospective cohorts have been influential in advancing understanding of the risk of diabetes in adult survivors of childhood and adolescent cancers. The Childhood Cancer Survivor Study (CCSS), and the French-U.K. and the Dutch cancer survivor cohorts have similarly reported an increased risk of diabetes after radiation exposure to the abdomen (including para-aortic lymph nodes, spleen, and, more specifically, the pancreatic tail) in a dose-dependent manner (12,14). Despite the wealth of consistent and corroborating evidence ascertained by multiple retrospective cohorts, a characteristic of many of these studies is their reliance on postal questionnaire surveys to obtain outcome data. Although studies often aim to recruit eligible survivors across multiple institutions to broaden generalizability, the possibility remains that certain high-risk or ethnic minority groups may be traditionally under-represented, leading to nonparticipation bias. Other methodologic limitations of cohort studies include the use of self-reported outcome measures, which may produce invalid inferences leading to over- or underestimation of diabetes risk (4,8,10,12).

Population-based studies with robust, validated criteria for defining diabetes have the potential to overcome these methodological shortcomings, provide different insights, and potentially more robust evidence. To date, only two studies have investigated the risk of objectively measured diabetes in a population-based cohort of childhood cancer survivors; these were conducted across Scandinavia and Ontario (7,9). Findings from these studies validated previous results from the CCSS (8), showing that cancer survivors have a 55–60% increased risk of developing diabetes compared with a matched general population or sibling control participants (7–9). Among both population-based cohorts, survivors treated for leukemia (hazard ratio [HR] 2.29; 95% CI 1.68–3.14), lymphoma (HR 1.61; 95% CI 1.12–2.31) (9), and Wilms’ tumors (7) had the highest cause-specific hazard of diabetes compared with control participants (8,11,12).

However, the current consensus is that the treatments received for these cancers contribute most strongly to the increased risk for diabetes, rather than disease-specific factors (8). Because treatment data were not available in previous population-based studies (7,9), it remains unclear which treatment modality confers the greatest risk outside of a cohort study framework. Determining the risk of diabetes after CYAC and how this risk varies according to cancer treatments is of critical importance because the prevention or timely management of diabetes could play a central role in mitigating cardiovascular-related morbidity and mortality burdens in this population.

In this study, we explored treatment-related factors for the development of diabetes in a hospital-based cohort of CYAC survivors. The primary treatment-related exposure variables explored were informed by the existing evidence base, which largely comprises findings from large-scale retrospective studies (5,8,12). We aimed to validate these findings in a well-characterized, multiethnic cohort with an objective outcome measure and detailed treatment information.

A retrospective study using population-based cancer registration data linked to national and regional-level administrative health care databases in the U.K. was performed. Individuals diagnosed with cancer (aged 0–29 years at the time of cancer diagnosis) between 1 January 1974 and 31 December 2016 were identified from the Yorkshire Specialist Register of Cancer in Children and Young People (YSRCCYP) (15). The YSRCCYP is a population-based database of all children (aged 0–14 years) diagnosed with any malignant cancer or benign central nervous system tumor while resident in the Yorkshire and the Humber regions since 1974, and all individuals aged 15–29 years diagnosed from 1990 onward (15). The study population, comprised 1) individuals diagnosed with cancer (i.e., those aged 0–14 years) between 1974 and 2016 and 2) individuals aged 15–29 years at the time of their cancer diagnosis who were diagnosed between 1990 and 2016. The Yorkshire region covers a diverse, multiethnic population of 5.4 million people, 2 million of whom are younger than 30 years (16). Children and young people diagnosed with cancer in the Yorkshire region are primarily treated at one of two regional principal treatment centers in Leeds or Sheffield (17).

Follow-up data were derived from the Leeds Teaching Hospital’s Patient Pathway Management System (18) and through linkage with hospital admissions data. YSRCCYP records were submitted for linkage with Admitted Patient Care (APC) Hospital Episode Statistics (HES) (19), covering inpatient admissions occurring between 1 April 1997 and 31 March 2023. APC HES data were coded using ICD-10 and include a primary diagnosis field and up to 19 secondary diagnostic codes comprising details of all reported comorbidities.

Through linkage with Leeds Teaching Hospitals Trust (LTHT), additional sources of outcome data were obtained, such as results of clinical investigations from PPM+ and diagnostic information from the Diabetes Management Service (DMS). A total of 4,238 individuals were included in the final analysis (Supplementary Fig. 1). Information on vital status and date of death were acquired via linkage with the National Disease Registration Service (15,20). Anthropometric data and information on lifestyle factors were not available.

Data on treatment were obtained from the YSRCCYP. The primary treatment-related exposures of interest were 1) total-body irradiation (TBI), 2) cranial radiation, 3) abdominal radiation, 4) exogeneous corticosteroids, and 5) hematopoietic stem cell transplantation (HSCT). Sensitivity analysis was used to explore the effect of allogeneic versus autologous stem cell transplantation on the risk of diabetes and death. A detailed breakdown of transplant-specific characteristics are presented in Supplementary Table 1. Patients who had received radiotherapy to the cranium and abdomen were defined based on ICD-O3 topography codes (Supplementary Table 2). Cranial and abdominal irradiation did not include TBI.

The primary outcome was incident diagnosis of diabetes. Inpatient hospital clinical coding was used as the primary source of identification (19). ICD-10 codes were reviewed to identify all relevant diagnostic codes for diabetes and develop a coding definition (Supplementary Table 3).

Data enhancement was undertaken where more data sources were available from PPM+ or the DMS. Where available (for n = 858 individuals; 20.2%) (Supplementary Table 4) HbA_1c values were used to identify individuals with diabetes who lacked relevant clinical codes in HES data. All HbA_1c results were assessed algorithmically to identify instances of results ≥48 mmol/mol (6.5%) (21). The earliest date indicator of diabetes diagnostic data across available sources was used as the date of first diabetes detection. It was not possible to distinguish between type 1 diabetes and type 2 diabetes. Figure 1 represents the diabetes data definitions used in the analysis.

Individuals were followed up from date of primary cancer diagnosis until any of the following events occurred: detection of diabetes, all-cause death, or end of follow-up (31 March 2023). Individuals with a diagnosis of diabetes prior to the date of cancer diagnosis were excluded (Supplementary Fig. 1). To avoid misclassifying patients with transient hyperglycemia during cancer treatment [e.g., induced by glycemic-altering cancer agents such as immune checkpoint inhibitors and high-dose glucocorticoids (22)], any indications of incident diabetes detected by clinical coding or HbA_1c within 3 years of the date of primary cancer diagnosis were not included as a diagnosis of diabetes. This mirrors approaches used in previous population-based studies (9).

The total causal effect of each treatment-related exposure on the risk of incident diabetes was defined using adjusted cause-specific cumulative incidence functions (CIFs) in the presence of competing risks (all-cause death) (23,24). This enabled us to compare the difference in risk of developing diabetes by a specific time point under a specific treatment exposure of interest.

The minimal sufficient adjustment set for modeling the total effect of each treatment-related exposure on the risk of diabetes and of death was informed using directed acyclic graphs within DAGitty software, version 3.1 (25) (Supplementary Fig. 2). Flexible parametric models were adjusted for International Classification of Childhood Cancer (Third Edition) (ICCC-3) (26) diagnostic group (leukemia [I], lymphoma [II], central nervous system [CNS] tumors [III], and other solid tumors [IV–XII]), cancer stage or grade, age, and calendar year of primary cancer diagnosis. Age at cancer diagnosis was modeled as a continuous variable, and a spline (2 df) captured nonlinear association between age and survival. Scaled Schoenfeld residuals found no evidence of nonproportional hazards for primary treatment exposures in the diabetes model. Within the competing-risks death model, TBI, corticosteroid drugs, HSCT, and allogeneic HSCT were included as time-varying components. The baseline log cumulative hazard was modeled using restricted cubic splines (n = 4 internal knots) as determined by Akaike information criterion. Model fit was assessed using the stpm3km command.

Standardized CIFs for diabetes and death were estimated for each treatment exposure using the standsurv postestimation function (27). The cause-specific CIFs were adjusted for age and year of cancer diagnosis, cancer type, and stage or grade. The differences in standardized survival estimates reflect the average causal effect of a treatment on diabetes or death risk, assuming the covariates adequately control for confounding (28). Models were additionally stratified by age at cancer diagnosis (0–14 years and 15–29 years) to account for differential data availability and years of follow-up.

White blood cell count was used as a proxy for stage when modeling leukemia survival; lymphoma stage followed the Ann Arbor system (29); CNS tumors were categorized by World Health Organization grade (low I–II/high III–VI) (30); and other solid tumors were staged using the TNM classification (31). Missing staging data were imputed with 20 data sets via chained equations under the missing-at-random assumption. The imputation model included all analysis variables, including outcome variables (i.e., Nelson–Aalen survival estimator, censoring indicator), and binary treatment indicators (i.e., surgery, radiotherapy, chemotherapy, TBI). All analyses were conducted using Stata/MP version 18 (StataCorp, College Station, TX).

An additional analysis was conducted to estimate the standardized CIFs for diabetes among individuals who had received a key treatment exposure of interest (ie, exposures 1–5), compared with individuals who underwent surgery (excluding biopsy) only (Supplementary Table 5). Individuals who underwent neurosurgical tumor intervention were excluded from this surgery-only control cohort, due to potential links to hypothalamic–pituitary axis dysfunction and endocrine alterations (32).

The data sets analyzed in this study are not publicly available, due to privacy and ethical restrictions. Data are available subject to review, with the appropriate ethical and information governance approvals (15). Stata code can be found in the following repository: https://github.com/medkcro/Treatment-effects.

During a median follow-up of 14.3 (interquartile range [IQR] 8.2–22.6) years, 163 of 4,238 CYAC survivors (3.8%) were identified as having diabetes. Among these cases, 71 (43.6%) were identified from clinical coding alone, 52 (31.9%) by HbA_1c only, and 40 (24.5%) through hybrid definitions. Specifically, 27 (16.6%) were identified by HbA_1c and clinical coding in HES, 3 (1.9%) from diagnostic information in the DMS and HbA_1c; and 10 (6.1%) across all three data sources. Median time to diabetes detection was 11.9 years. Median age at detection was 30.0 years for individuals diagnosed with cancer when they were 0–29 years old, 23.1 years for individuals diagnosed with cancer between the ages of 0 and 14 years (Supplementary Table 6A), and 34.6 years for those diagnosed between the ages of 15 and 29 years (Supplementary Table 6B).

Table 1 presents the clinical and demographic characteristics of the cohort by diabetes status. Comparing cancer survivors with and without diabetes, a greater percentage of individuals in the former group had a primary diagnosis of leukemia (36.2% [n = 59] vs. 22.0% [n = 896]), were of South Asian ethnicity (16.6% [n = 27] vs. 8.8% [n = 357]) and lived in the most socioeconomically deprived neighborhoods at the time of their cancer diagnosis (32.5% [n = 53] vs. 27.9% [n = 1,136]). Median age at primary cancer diagnosis was 17.7 years for CYAC survivors in the diabetes group and 16.1 years for individuals in whom diabetes was not detected (Table 1).

A total of 928 individuals (21.9%) only underwent surgery as part of their CYAC treatment (Table 1). A diagnosis of diabetes was detected in 2.0% of individuals (n = 19) in the surgery-only cohort.

Figure 2 shows the standardized CIFs (95% CI) of diabetes and death, respectively, by prespecified treatment group (treatment exposures 1–5). Supplementary Figure 3 shows the corresponding difference in standardized CIFs between treatment arms (untreated and treated) (exposures 1–5) by time since primary cancer diagnosis (years).

Cumulative incidence of diabetes increased progressively by time since diagnosis and differed by treatment arm (TBI versus no TBI) (Fig. 2, Supplementary Fig. 3). The 10-, 20-, 30-, and 40-year standardized CIFs were equal to 4.4% (2.8–6.9), 11.3% (7.5–16.9), 16.5% (11.0–24.8), and 21.0% (13.8–31.9) with TBI and 1.4% (1.1–1.8), 4.0% (3.3–4.9), 6.2% (4.9–7.9), and 8.4% (6.1–11.5) with no TBI (Fig. 2).

The standardized CIF of diabetes in CYAC survivors who underwent cranial irradiation was equal to 1.6% (0.09–2.6) at 10 years after diagnosis, 4.3% (2.6–7.0) at 20 years, 6.6% (4.1–10.9) at 30 years, and 8.8% (5.2–14.7) at 40 years. There were no notable differences in the cumulative risk of diabetes comparing these individuals to those who did not receive cranial radiotherapy (Supplementary Fig. 3).

The standardized CIFs for patients who underwent abdominal irradiation who were diagnosed with cancer between the ages of 0 and 29 years (10 years, 2.1% [0.9–4.7]; 20 years, 5.7% [2.6–12.2]; 30 years, 8.8% [4.1–18.7]; 40 years, 11.7% [5.5–25.0]) were marginally greater than those of patients who did not undergo abdominal radiotherapy (10 years, 1.6% [1.3–2.1]; 20 years, 4.4% [3.7–5.3]; 30 years, 6.9% [5.6–8.7]; 40 years, 9.3% [6.9–12.6]) (Fig. 2).

When stratified by age at cancer diagnosis, abdominal radiotherapy was associated with an increased risk of diabetes in individuals diagnosed with cancer when they were 0–14 years old (Supplementary Fig. 4A). The standardized CIFs for childhood cancer survivors who underwent abdominal irradiation was 4.0% (1.4–11.3), 11.1% (4.3–28.7), 19.2% (8.0–45.9), and 27.7% (12.4–62.1) at 10, 20, 30, and 40 years after diagnosis, respectively. An association between abdominal radiotherapy and diabetes risk was not apparent in individuals diagnosed with cancer at ages 15–29 years (Supplementary Fig. 4B).

Diabetes risk steadily increased over time in individuals treated with corticosteroids, resulting in a pattern of standardized CIFs of 3.1% (2.0–4.7), 8.2% (5.6–12.1), 12.5% (8.5–18.5), and 16.3% (10.1–24.9) at 10, 20, 30, and 40 years after diagnosis, respectively (Fig. 2). At 10 years after diagnosis, the difference in standardized CIFs (corticosteroids versus no corticosteroids) was 1.6 (0.4–2.8). By 40 years after diagnosis, the difference in risk between treatment arms increased to 7.7 (1.7–13.7) (Supplementary Fig. 3).

The standardized CIF of diabetes in the HSCT group increased from 4.1% (2.8–6.0) at 10 years after diagnosis to 19.6% (13.4–28.6) at 40 years after diagnosis. Comparable estimates in the non-HSCT group were 1.4% (1.1–1.7) and 8.2% (6.0–11.3), respectively (Fig. 2).

Subgroup analysis revealed differing diabetes risks for allogeneic and autologous transplant survivors. Standardized CIF for diabetes in CYAC survivors who received allogeneic HSCT equaled 5.9% (3.7–9.2) at 10 years, and this increased to 25.7% (17.4–38.0) at 40 years, with a risk difference compared with survivors who received nonallogeneic HSCT of 4.5 (1.9–7.1) to 17.5 (8.2–26.8) over the same period (Supplementary Fig. 5A and B). The standardized CIF of diabetes 40 years postdiagnosis was 7.9% (3.3–19.1) with autologous HSCT, compared with 9.3% (6.9–12.5) with no autologous HSCT.

Supplementary Figure 6 shows the standardized CIFs of diabetes and death under key treatment exposures of interest (exposures 1–5) and under surgical intervention only. The standardized CIFs of diabetes for those treated with surgery only were 1.0% (0.5–1.6), 2.1% (1.3–3.6), 3.1% (1.8–5.3), and 3.9% (2.2–7.1) at 10, 20, 30, and 40 years after cancer diagnosis.

To our knowledge, this is the first study to assess the long-term diabetes risk in a representative U.K. hospital-based cohort of CYAC survivors. Our findings illustrate the significant impact of CYAC treatments up to 40 years after cancer-diagnosis, particularly with TBI and allogeneic HSCT. Associations with abdominal radiotherapy were apparent for survivors diagnosed when they were aged 0–14 years but were not evident in those diagnosed aged 15–29 years. Novel findings include varying diabetes risks between patients who received allogeneic and autologous HSCTs and identification of temporal diabetes trends for the first time according to cancer treatment received.

Our findings mirror those of previous nonpopulation-based studies showing an increased risk of diabetes among CYAC survivors treated with TBI (8,33,34) and exogeneous corticosteroid (8,34). CCSS reported sevenfold higher risk (odds ratio 7.2; 95% CI 3.4–15.0) of self-reported diabetes in TBI-treated survivors compared with sibling control participants. TBI-related differences in risk were apparent in our population-based cohort from 10 years after diagnosis, with the cumulative incidence nearly three times higher by 40 years in TBI-treated survivors. These findings emphasize the impact of treatment exposures during childhood and adolescence on the long-term metabolic health of this population.

CCSS (8,12), French-U.K., and Dutch cohorts (14) reported stark contrasts in risk for childhood cancer survivors who underwent abdominal irradiation compared with those who did not receive abdominal radiotherapy. The risk associated with abdominal radiation is thought to be potentiated in those receiving higher doses of radiation to the pancreatic tail (12) and treated at younger ages. This study provides support for the concept of an interaction effect with age, because associations between abdominal radiotherapy and diabetes risk were only evident when pediatric patients were analyzed separately. The French-U.K. cohort reported a minimal latency period of 20 years for diabetes occurrence (14). The median follow-up of 12.6 years for survivors diagnosed with cancer when they were aged 15–29 years in this study may not be sufficient to accurately capture diabetes occurrence in this group. Longer-term follow-up of the cohort may result in higher effect estimates that align more closely with the majority of the evidence base (12,14,35).

Previous population-based studies evaluated the risk of diabetes in childhood cancer survivors compared with population controls (7,9). Therefore, in this study, we focused on quantifying within-treatment group differences in diabetes risk. Direct comparisons of overall diabetes risk in the CYAC population relative to the general population were outside the scope of this study and not feasible due to a lack of viable external comparator cohorts. To mitigate this, a surgery-only cohort was used to facilitate internal comparisons, assessing the extent of risk in each treatment group compared with individuals who received surgery only. Use of this cancer comparator group is likely to account for some of the psychosocial factors attendant upon living with and beyond a cancer diagnosis, which may disproportionately affect diabetes risk.

The prevalence of diabetes among CYAC survivors in this study differed from that in equivalent studies, likely due to varying outcome detection methods. Diabetes was detected in 3.8% of CYAC survivors in Yorkshire, compared with 1.5% in population-based studies from Ontario, Canada, and Scandinavian countries (7,9), and 2.5% via self-report in the CCSS cohort (10). A recent study detected diabetes via clinical assessment (fasting glucose 100–125 mg/dL or HbA_1c 5.7–6.4%) in 6.5% of childhood cancer survivors (36). Unlike many studies relying on self-reported outcomes, this study objectively used a hybrid definition of diabetes, combining HbA_1c and clinical coding data to increase the accuracy of outcome data capture (7,37).

Within the study cohort, 20.2% of participants had at least one HbA_1c measurement. We recognize the potential for detection bias, particularly regarding the frequency of HbA_1c screening among certain high-risk survivor subgroups, such as those exposed to TBI or cranial radiotherapy, or those with a higher risk of cancer recurrence who may be monitored more closely. The differential likelihood of screening in these groups may affect observed associations, with larger effect estimates reflecting an increased surveillance in some cases, rather than a true increase in risk.

HbA_1c measurements were unavailable from Sheffield Children’s Hospital and Sheffield Teaching Hospitals, thus excluding a third of CYAC survivors treated in Yorkshire (17). The hybrid diabetes definition limited the cohort to individuals diagnosed, treated, or followed up at LTHT only. Cancer registration data were unavailable for individuals aged 15–29 years and diagnosed before 1990 (15). Linked HES APC data were not available until 1997, and HbA_1c values from 2002. Application of the National Data Opt-out, linkage errors, or the absence of an APC episode in the data collection period may have also led to missing HES data for a small number of individuals. These data gaps could lead to a delay in or failure to capture a diagnosis of diabetes in individuals who developed the condition before the availability of HES or PPM. Differential availability of data for CYAC survivors diagnosed during certain periods, age groups, or geographies, and the longer duration of follow-up for individuals diagnosed in earlier periods may have obscured observed treatment effects and led to an under or overestimation of diabetes prevalence in certain groups. Some individuals flagged as having diabetes based on HbA_1c alone may have been misclassified due to conditions like iron deficiency anemia.

Nonetheless, this study provides novel and valuable insights into a multiethnic, representative regional population of CYAC survivors in the U.K. (38). Although the data collected are limited to a subset of the Yorkshire region, intelligence generated is of benefit to national and international health services, in lieu of an equivalent U.K.-wide study.

Our study identified key treatment-related exposures that increased the risk of diabetes in patients with CYAC. Varying treatments have been associated with differing etiological mechanisms for diabetes. Abdominal irradiation is linked to direct radiation damage to the pancreatic tail, affecting insulin secretion (8,12,14), and radiation exposure to the hypothalamic–pituitary axis may lead to hyperinsulinemia and insulin resistance after TBI (33). We did not have anthropometric data to identify obesity in the study cohort. Previous studies have shown differential associations between CYAC treatments and diabetes risk, with evidence to suggest that diabetes risk may be independent of overall adiposity in some groups of survivors (8). Therefore, the lack of this variable recorded in the data set may not have greatly influenced observed findings. However, measures of overall and centralized adiposity are undeniably important covariates that should be included in future studies, where available.

This study demonstrates the value of regional specialist registers that collect detailed demographic, clinical, and follow-up data that are not available nationally (15). Survivors of South Asian heritage and those resident in the most socioeconomically deprived areas at the time of their primary cancer diagnosis had a higher prevalence of diabetes compared with their White and more affluent counterparts. This observation aligns with broader epidemiologic data indicating a predisposition to diabetes in ethnic minorities and more socioeconomically deprived populations (39) and after a childhood cancer diagnosis (40). Future work should assess the need for culturally tailored follow-up care and diabetes prevention strategies in patients with CYAC, considering the interaction among treatment and genetic, lifestyle, and demographic factors that may influence long-term risk.

Patients with CYAC who underwent HSCT, particularly allogeneic HSCT, had a higher risk of diabetes compared with those who had autologous HSCT or no HSCT. Of allogeneic transplant survivors in this cohort, 60.5% were also exposed to TBI, compared with just 4.5% of autologous transplant survivors. This suggests that the observed elevated risk of diabetes in the allogeneic-HSCT group may be attributable to preparative conditioning regimes comprising TBI (34). The role of high-dose chemotherapy could not be studied. Although our study confirmed the association among TBI, allogeneic HSCT, and diabetes risk, small within-group numbers meant we could not investigate the role that these conditioning therapies played in modifying the overall HSCT treatment effect.

This study addresses an important deficit in knowledge about diabetes’ late effects, supporting the evidence that diabetes is a late effect of cancer diagnosed in the pediatric and young adult age range, and that the risk and timing of diabetes onset varies according to cancer treatment received. Findings highlight the need for lifelong vigilance for the development of treatment-related complications in patients with long-term CYAC and provide evidence to implement risk-stratified screening.

Acknowledgments. This work uses data provided by patients and collected by the U.K. National Health Service as part of their care and support.

Funding. This research was supported by Child Health Outcomes Research at Leeds (choralresearch.org.uk/). The Yorkshire Register is supported by the Leeds Candlelighters’ Trust (grant RG.EPID.100811). K.J.C. is funded by the Emma and Leslie Reid Research Scholarship (University of Leeds).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. K.J.C., R.G.F., A.W.G., and R.D.M. were involved in the conception and design of the study. R.D.M., R.A.A., and N.F.H. were involved in clinical interpretation of the results. K.J.C. performed the analysis and drafted the manuscript. All authors critically revised, reviewed, and approved the final version of the manuscript. K.J.C. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Handling Editors. The journal editors responsible for overseeing the review of the manuscript were Cheryl A.M. Anderson and Amalia Gastaldelli.