Machine Learning Models

The top-performing model used all six views, feature concatenation, temporal difference features, and an MLP classifier. The best models achieved AUC values ranging from 0.91 to 0.95. Gradient boosting was the best model for COPD and T2DM by F1 score, while XGBoost was the best model for HF. Models trained on TVAE-augmented data outperformed baseline models, and ensemble methods were stronger than logistic regression. The deep TSM embeddings performed better than the 38-dimensional biomarker features in the best reported comparison. The modelling experiments compared expert-selected variables, data-driven variables and combined variable sets. The LGB Machine was chosen because prior work showed strong performance and interpretability. A frozen TSM network with a ResNet-18 backbone generated 512-dimensional spatiotemporal embeddings for downstream classification. The LGB algorithm handled missing values natively, so the study did not impute missing data. The study trained logistic regression, random forest, gradient boosting, and XGBoost models on original and augmented datasets. Model selection prioritized recall and F1 because missed high-risk patients were considered more clinic…