A Multi-Modal Machine Learning Framework Integrating Medical Records and Wearable Sensor Data for Early Diabetes Prediction

Diabetes mellitus is a rapidly emerging health challenge globally and delayed diagnosis of the disease tends to lead to severe complications, besides increasing the healthcare cost. Traditional screening methods are mainly based on periodic clinical tests and static medical records, limiting the ability of obtaining physiological changes in early stage states. Existent prediction models often use a single source of data which results in poorly accurate prediction in generalization across different populations. This study is to develop a multi-modal machine learning framework which consists of a combination of electronic medical records (EMRs) and continuous wearable sensor data that would enable prediction of diabetes as early as possible and with high accuracy. Some feature engineering techniques, temporal aggregation, and data normalization were applied prior to training the model. Ensemble and deep learning models, such as Random Forest, XGBoost and Long Short-Term Memory (LSTM) networks were tested based on stratified crossvalidation. The integrated multi-modal model achieved an accuracy of 94.2%, precision of 93.1%, recall of 95.4%, F1-score of 94.2% and an AUC of 0.97 which are 8-12% better than single modality-baselines. The results show that integration of medical records with wearable sensor data is able to improve prediction of diabetes early significantly in favor of proactive intervention and personalized healthcare delivery.