Interpretable Deep Learning for Type 2 Diabetes Risk Prediction in Women Following Gestational Diabetes

Women with gestational diabetes mellitus (GDM) face a 7-10 times elevated risk of developing Type 2 Diabetes Mellitus (T2DM), yet current predictive models demonstrate limited accuracy (AUC-ROC: 0.70-0.85) and insufficient interpretability for clinical adoption. This study addresses the critical need for accurate, transparent risk prediction tools by developing an interpretable deep learning framework integrating bidirectional long short-term memory (BiLSTM) networks with attention mechanisms and SHapley Additive exPlanations (SHAP). Using a synthetic dataset of 6,000 simulated post-GDM women with 28 clinical risk factors, the BiLSTM-Attention model was evaluated through stratified 10-fold cross-validation against five baseline models. The proposed model achieved exceptional performance with 98.45% accuracy, 98.80% precision, 98.30% recall, 98.55% F1-score, 96.85% MCC, and 0.9968 AUC-ROC, significantly outperforming all baselines (p < 0.05). SHAP analysis identified recurrent GDM history, elevated HbA1c, and impaired glucose tolerance as primary predictors, while highlighting modifiable factors including physical inactivity, dietary habits, and obesity as actionable intervention targets. This proof-of-concept demonstrates the methodological feasibility of combining high-performance deep learning with explainable AI for T2DM risk stratification. However, synthetic data represents a significant limitation; comprehensive real-world clinical validation across diverse populations is essential before clinical implementation. The publicly available computational framework enables future validation studies to advance this approach toward clinical utility.