Temporal Feature-Based AI Models Excel Over Traditional Machine Learning for Predicting Diabetes

Deep learning applied to longitudinal electronic medical record (EMR) data for early diagnosis and prevention of type 2 diabetes mellitus. In order to evaluate the prediction power of deep learning models vs sophisticated machine learning models for temporal risk factors, this study compares both. For predicting diabetes, several deep learning models and characteristics were examined. Over 19,000 patients' longitudinal EMR data from a private hospital were used to apply the framework. Research results show that deep learning models routinely beat conventional machine learning models, obtaining over 91% prediction accuracy without overfitting. Hemoglobin A1c, body mass index, and fasting blood sugar are important indicators of the start of diabetes. With more training data density and patient visit history, the models perform better. In order to improve early diabetes prediction using longitudinal EMR data, this work verifies the effectiveness of LSTM deep learning in incorporating temporal risk aspects.