Predicting time to first emergency hospital admission in adults with type 2 diabetes in England: development and internal validation of multivariable survival models.
This project developed and internally validated statistical and machine-learning survival models using linked national healthcare data to predict time to first all-cause emergency hospital admission in adults with type 2 diabetes in England, with the aim of supporting earlier identification of people who may benefit from proactive care.
This project was completed by Annu Joy, Higher Information Analyst, in the Clinical Audit Team, as part of the Data Science MRes at the University of Leeds.
The project developed and internally validated models to predict time to first all-cause emergency hospital admission during follow-up in adults with type 2 diabetes in England. The rationale was to support earlier identification of people who might benefit from proactive review and targeted care. The literature review found that much previous work focused on readmission, specific diabetes complications or selected populations, rather than national prediction of first emergency admission using routinely available primary-care information. This informed the decision to compare statistical and machine-learning survival models, while deliberately excluding previous admissions as predictors to assess the value of current patient characteristics.
The study used 2022 National Diabetes Audit data linked to Hospital Episode Statistics and Office for National Statistics mortality records, covering 3,014,590 adults followed from January 2023 to March 2025. The work involved constructing and quality-checking the cohort, deriving clinical and care-process predictors, assessing missing data, and developing Cox proportional hazards, CatBoost accelerated failure time and Royston–Parmar flexible parametric models. Models were trained on 60% of the cohort and evaluated in a separate 40% test dataset, with smaller training subsets used for computationally intensive tuning. Evaluation considered discrimination, calibration and prediction error, followed by decision-curve analysis of the preferred model. An additional exploratory clustering analysis investigated whether the data identified reproducible patient profiles.
During 27 months of follow-up, 886,158 participants (29.4%) experienced an emergency admission. All three models showed moderate discrimination. CatBoost achieved the highest two-year AUC (0.707), but its risk predictions were less well calibrated. The Royston–Parmar model provided the best overall balance, achieving the highest C-index (0.716), a near-ideal calibration slope (1.006) and the lowest integrated Brier score (0.1193). Decision-curve analysis suggested potential benefit at moderate risk thresholds. Fewer completed care processes, poorer glycaemic control and chronic kidney disease were among the reported markers of increased risk; these were predictive associations, not evidence of causation. Clustering identified two reproducible profiles distinguished mainly by care-process completion and data availability, but cluster membership explained less than 1% of variation in predicted risk.
Overall, the project provides promising evidence that routinely collected primary-care data could support risk-informed diabetes care, while showing that a flexible statistical model can offer a more favourable overall performance than a more complex machine-learning approach. Important limitations included incomplete clinical recording, computational and linkage restrictions, internal validation only, and treating death as censoring rather than a competing event. External validation and prospective evaluation are needed before clinical implementation, including assessment of whether model-supported reviews improve outcomes without creating excessive workload or inequitable targeting.