Hospital Readmission Prediction

Tech Stack

The Problem

• Hospitals faced rising 30-day readmission rates triggering CMS financial penalties and reduced reimbursements.
• Clinicians had no validated, quantitative risk scoring tool to identify high-risk patients at discharge.
• Severe class imbalance in patient data caused standard models to under-detect the high-risk minority class.
• Manual risk assessment was subjective, slow, and highly inconsistent across clinical care teams.
• No automated pipeline integrated clinical notes, lab values, and admission history into a unified risk score.

Our Solution

• Built a comprehensive ML pipeline with SMOTE oversampling to correct critical class imbalance in readmission data.
• Performed extensive feature engineering on clinical variables, lab results, and length-of-stay metrics.
• Selected CatBoost as the primary model for its exceptional performance with clinical categorical variables.
• Applied Optuna hyperparameter optimisation maximising recall to ensure high-risk patients are never missed.
• Delivered probability-scored risk outputs to assist clinical discharge planning and post-care interventions.

Impact

Identified high-risk patients with strong predictive accuracy, enabling targeted post-discharge care programs.