Deep Learning for Early Diagnosis of Chronic Conditions Using Electronic Health Records

Abstract

Early diagnosis of chronic diseases remains a major challenge in healthcare, especially given the complexity and volume of longitudinal Electronic Health Records (EHR). This study proposes a deep learning framework based on Long Short-Term Memory (LSTM) networks enhanced with attention mechanisms to identify early onset patterns of chronic conditions such as Type 2 Diabetes Mellitus (T2DM), Hypertension, Chronic Kidney Disease (CKD), and Congestive Heart Failure (CHF). Trained on a dataset of 72,593 patient records, the model achieved a high overall F1-Score of 90.8% and AUROC of 96.2%, significantly outperforming traditional models like logistic regression, random forest, and XGBoost. Condition-wise analysis showed strongest performance in T2DM (F1-Score: 92.0%), attributed to the model’s ability to track lab and medication sequences. The framework demonstrated robustness across demographics, with F1-Scores exceeding 88% across age, gender, and ethnic groups, confirming its fairness and general applicability. Ablation studies validated the essential roles of temporal learning and attention components, while visualization of attention weights provided meaningful interpretability aligned with clinical reasoning. Generalization experiments on MIMIC-III and eICU datasets yielded F1-Scores of 88.8% and 86.5%, respectively, underscoring the model’s resilience to domain shifts. These results support the deployment of the proposed deep learning framework as a reliable, equitable, and interpretable tool for early chronic disease diagnosis. Future extensions will target integration