Anticipating type 1 diabetes in children using machine learning and routinely-collected data in primary care

Townson, J., Daniel, R., Jones, H., Francis, N. A., Paranjothy, S., Baldwin, B., Thayer, D. and Gregory, J. W. (2022) Anticipating type 1 diabetes in children using machine learning and routinely-collected data in primary care. Diabetic Medicine, 39 (S1), [e14809]. (doi:10.1111/dme.14809).

Abstract

Oral abstracts:Background and aim: Delayed or misdiagnosis are significant risk factors for children presenting in diabetic ketoacidosis (DKA) at diagnosis. Symptoms may be attributed to more common conditions, making diagnosis challenging. We aimed to develop and validate a predictive model to identify type 1 diabetes based on coded data within the primary care electronic medical record.
Methods: To develop the algorithm we used routinely collected primary care data (Secured Anonymised Information Linkage (SAIL) databank), linked with secondary care data (Brecon Register), in Welsh children <15 years, within 01/01/00 to 31/12/16. Codes relating to primary care contacts were grouped into 25 pre-determined categories. We used machine learning (SuperLearner) to train the predictive algorithm, and tested it using Clinical Practice Research Datalink (CPRD) (English primary care data) linked with Hospital Episode Statistics (type 1 diabetes diagnosis) data.
Results: The algorithm was trained on data from 1.2 million children, with 35 million general practitioner (GP) contacts (SAIL), 2051 (0.21%) diagnosed with type 1 diabetes (DKA 26%) (Brecon). The test data from CPRD included 1.5 million children and 43 million GP contacts. 1516 (0.10%) were determined newly diagnosed (DKA 26%). Using a definition of effort-benefit trade-off (benefit = patients diagnosed, effort = GP contacts) the estimated AUC was 0.906. Using a 10% effort threshold, 75% of children with type 1 diabetes were identified by the algorithm 90 days prior to diagnosis. 25% of children would have their diagnosis anticipated by at least a week

This record has no associated files available for download.

Contributors

Author: N. A. Francis

Download statistics