Dynamic Data Augmentation for Early Prediction of Non-Communicable Diseases with Sequence Structure

Non-communicable diseases (NCDs), such as cardiovascular diseases, cancers, chronic respiratory diseases, and diabetes, are responsible for 71% of global deaths.Early prediction of these diseases is critical for effective management and cost reduction in healthcare.However, early prediction with EHR data is difficult because the records are sparse, sequential, and heavily imbalanced.Traditional models do not handle these issues well, especially when the data represents single events, which reduces predictive performance.This study presents a data augmentation method that integrates the sliding-window approach with Borderline-SMOTE to address these challenges, using MIMIC-IV as a proof-of-concept evaluation.The method converts sparse records into sequential inputs and balances the class distribution, aiming to improve early prediction accuracy.By improving the quality of sequential data inputs, the proposed technique facilitates timely interventions and better patient outcomes.The research assesses the performance of different machine learning models trained on augmented datasets including LSTM, GRU, Random Forest, XGBoost and KNN, with XGBoost achieving the best performance, reaching 94.6% accuracy, precision, recall, and F1-score.These results highlight the effectiveness of the proposed approach in the early prediction of NCDs.