Abstract

Hospital readmission prediction for Chronic Obstructive Pulmonary Disease (COPD) patients remains a critical challenge in clinical decision support. Our prior work (Papers 1 and 2) established baseline models and proposed the Feature-Weighted Cost-Sensitive Random Forest (FW-CSRF) algorithm, demonstrating improved recall and AUC over conventional classifiers. However, limitations persist in temporal data utilization, model explainability depth, and adaptability to heterogeneous patient trajectories. This paper presents an Enhanced Explainable Predictive Framework (EEPF-COPD) that integrates three key advances: (i) SHAP (SHapley Additive exPlanations)-guided dynamic feature optimization, (ii) Temporal Sequence Modeling using a Bidirectional LSTM (BiLSTM) for capturing longitudinal patient patterns, and (iii) Bayesian Hyperparameter Optimization for model tuning. Experimental evaluation on MIMIC-IV data demonstrates that EEPF-COPD achieves AUC-ROC of 0.91, recall of 0.86, and F1-score of 0.83, outperforming FW-CSRF and all prior baseline models. Qualitative clinical interpretability analysis via SHAP summary and force plots provides actionable insights for clinicians. The proposed framework represents a mature, clinically deployable enhancement suitable for real-world hospital decision support systems.

Authors

S.K. Althaf Rahaman, K. Vedavathi
GITAM Deemed to be University, India

Keywords

COPD, Readmission Prediction, Explainable AI, SHAP, BiLSTM, Temporal Modeling, Bayesian Optimization, Predictive Analytics