Abstract—Prevailing studies employing the UCI Diabetes 130-US Hospitals dataset report classification performance exceeding Area Under the Curve (AUC) > 0.99; however, such estimates are systematically inflated by data leakage the inadvertent incorporation of information unavailable at prospective prediction time. This study introduces a rigorous, leakage-free evaluation framework for predicting diabetes medication initiation among hospitalized patients. Analyzing 101,766 inpatient encounters (1999–2008) from 71,290 unique patients, we enforced strict calendar-based temporal separation and excluded all 24 medication-specific features to eliminate both target and temporal leakage sources. Six models were evaluated spanning logistic regression, random forests, Extreme Gradient Boosting (XGBoost), One-Dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM), and a CNN-LSTM hybrid under identical leakage-free conditions. The 1D-CNN achieved the highest discrimination (AUC = 0.824), surpassing LSTM and all baselines, while demonstrating superior probability calibration (Brier score = 0.149; ECE = 0.028) and 3.2× faster inference. Primary predictors included HbA1c > 7%, elevated admission glucose > 180 mg/dL, advanced age (≥70 years), and frequent prior hospitalizations. These results confirm that lightweight convolutional architectures can outperform recurrent models on sparse, short-sequence Electronic Health Record (EHR) data a dataset-specific finding that should not be generalized without qualification. Dataset characteristics, scope constraints, and generalizability limitations are addressed fully in the methods and discussion sections.
 
Keywords—diabetes prediction, electronic health records, data leakage, temporal validation, model calibration, clinical decision support, convolutional neural networks


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).