A deep attention-based encoder for the prediction of type 2 diabetes longitudinal outcomes from routinely collected health care data
Enrico Manzini, Bogdan Vlacho, Josep Franch‐Nadal, Joan Escudero, Ana Génova, Elisenda Reixach, Erich Andrés, Israel Pizarro, Dı́dac Mauricio, Alexandre Perera-Lluna · Expert Systems with Applications · 2025
Recent evidence indicates that Type 2 Diabetes Mellitus (T2DM) is a complex and highly heterogeneous disease involving various pathophysiological and genetic pathways, which presents clinicians with challenges in disease management. While deep learning models have made significant progress in helping practitioners manage T2DM treatments, several important limitations persist. In this paper we propose DARE, a model based on the transformer encoder, designed for analyzing longitudinal heterogeneous diabetes data. The model can be easily fine-tuned for various clinical prediction tasks, enabling a computational approach to assist clinicians in the management of the disease. We trained DARE using data from over 200,000 diabetic subjects from the primary healthcare SIDIAP database, which includes diagnosis and drug codes, along with various clinical and analytical measurements. After an unsupervised pre-training phase, we fine-tuned the model for predicting three specific clinical outcomes: i) occurrence of comorbidity, ii) achievement of target glycemic control (defined as glycated hemoglobin < 7 % ) and iii) changes in glucose-lowering treatment. In cross-validation, the embedding vectors generated by DARE outperformed those from baseline models (comorbidities prediction task A U C = 0 . 88 , treatment prediction task A U C = 0 . 91 , HbA1c target prediction task A U C = 0 . 82 ). Our findings suggest that attention-based encoders improve results with respect to different deep learning and classical baseline models when used to predict different clinical relevant outcomes from T2DM longitudinal data. • Deep learning shows promise in managing T2DM and predicting its progression • We developed DARE, a transformer model for analyzing T2DM longitudinal records • DARE was trained on data from 200K diabetic patients spanning a 5-year period • DARE forecasts treatment changes, HbA1c targets, and diabetes comorbidities.