Synthetic data-augmented machine learning for 30-day readmission prediction in patients with chronic conditions: a retrospective real-world study.

Figures

• 

  Figure 1. Conceptual framework for integrated clinical prediction and economic modelling using structured and unstructured EHR data. Notes: This figure illustrates the end-to-end pipeline for predictive modelling using data from the MIMIC-IV database. The pipeline begins with structured and unstructured clinical data extraction, followed by data augmentation using techniques such as SMOTE, CTGAN or TVAE to address class imbalance. Key clinical features, including demographics, medications, vital signs and clinical notes, are incorporated. Feature engineering and selection are applied prior to training machine learning models such as XGBoost. NLP techniques are used to extract informative signals from unstructured text. AUC, area under the receiver operating characteristic curve; CTGAN, conditional tabular generative adversarial network; EHR, electronic health record; F1, F1 score; GBM, gradient boosting machine; LR, logistic regression; MIMIC-IV, Medical Information Mart for Intensive Care IV; NLP, natural language processing; RF, random forest; SHAP, Shapley additive explanations; SMOTE, synthetic minority oversampling technique; TVAE, tabular variational autoencoder; XGBoost, extreme…

• 

  Figure 2. Receiver operating characteristic curves illustrating discrimination performance of the best-performing classifiers trained on TVAE-augmented data in chronic obstructive pulmonary disease, heart failure and type 2 diabetes mellitus cohorts. Notes: This figure illustrates the discrimination capability (AUC) of the best-performing models for each condition based on TVAE-augmented data. Gradient boosting was selected as the optimal model for COPD, XGBoost for HF and both gradient boosting and XGBoost performed comparably well for T2DM. While other augmentation strategies (eg, SMOTE, CTGAN) and classifiers were tested, only the top-performing models from the TVAE-enhanced datasets are visualised here due to their superior predictive performance. AUC values were computed using stratified fivefold cross-validation. AUC, area under the receiver operating characteristic curve; COPD, chronic obstructive pulmonary disease; CTGAN, conditional tabular generative adversarial network; HF, heart failure; SMOTE, synthetic minority oversampling technique; T2DM, type 2 diabetes mellitus; TVAE, tabular variational autoencoder; XGBoost, extreme gradient boosting.

• 

  Figure 3. Model interpretability using SHAP and LIME values for the best-performing classifiers trained on TVAE-augmented data in chronic obstructive pulmonary disease, heart failure and type 2 diabetes mellitus cohorts. Notes: The figure displays interpretability outputs from SHAP and LIME for the best-performing machine learning models in each cohort—Gradient Boosting for COPD and T2DM, and XGBoost for HF. The plots highlight key features contributing to model predictions, such as age, ARNI, angiotensin receptor-neprilysin inhibitor; BNP, brain naturetic peptide; SOFA score, CCI, ICU admission and medication non-adherence. CCI, Charlson Comorbidity Index; CHF, congestive heart failure; COPD, chronic obstructive pulmonary disease; HF, heart failure; ICU, intensive care unit; LIME, Local Interpretable Model-Agnostic Explanations; PCOS, polycystic ovary syndrome; SHAP, SHapley Additive exPlanations; SOFA, Sequential Organ Failure Assessment; T2DM, type 2 diabetes mellitus; TVAE, Tabular Variational Autoencoder; XGBoost, extreme gradient boosting.