Synthetic Data-Augmented Machine Learning for 30-Day Readmission Prediction in Chronic Condition

Participants were adults (≥18 years) with a primary diagnosis of COPD, HF, or T2DM admitted between 2008 and 2019, who had complete 30-day follow-up and survived the index admission.

• The study aimed to create and assess an explainable machine learning framework augmented with synthetic data to predict unplanned 30-day hospital readmissions among adults with COPD, heart failure, or type 2 diabetes mellitus, using both structured electronic health record (EHR) data and information extracted from unstructured clinical notes.
• Design was a retrospective cohort analysis within ICU and general ward settings at a single tertiary academic center, drawn from the MIMIC-IV database.
• Synthetic data generation was employed to address outcome imbalance and to bolster model training, using resampling methods and deep learning-based techniques.

Model interpretability relied on established global and local explanation approaches.

• Predictors spanned six domains: demographics, comorbidities, clinical acuity, therapies, behavioral factors, and care continuity.
• The primary outcome was unplanned all-cause readmission within 30 days of discharge.
• Predictive performance was benchmarked across multiple machine learning algorithms with fivefold cross-validation.

• Ensemble-based machine learning models yielded the strongest predictive performance across all three diagnostic cohorts (COPD, HF, T2DM).
• Consistent with prior literature, top predictors included greater illness severity, higher comorbidity burden, medication non-adherence, gaps in preventive care, and limited social support.
• Models that incorporated synthetic data augmentation demonstrated improved discrimination relative to models trained solely on original data, suggesting a benefit from synthetic augmentation to address class imbalance and potentially enrich feature representation.

• The findings support the feasibility of deploying an explainable, synthetic-data–driven framework to predict 30-day readmissions using routinely captured EHR data, including information derived from unstructured notes.
• The study emphasizes integrating clinical, behavioral, and social determinants to enhance prediction models for readmission risk in chronic disease populations.

• The study is based on a single tertiary center and a specific database (MIMIC-IV), which may limit generalizability to other settings.