EHR Based Patient’s Severity Prediction using Machine Learning
Students & Supervisors
Student Authors
Supervisors
Abstract
Electronic Health Records (EHRs) provide valuable longitudinal clinical data that can be used to predict patient severity and facilitate timely clinical interventions. This study presents a machine learning framework designed to predict pa- tient severity levels (mild, moderate, and severe) using a balanced EHR dataset that includes vital signs, laboratory values, demo- graphics, and ICU admission data. To address class imbalance, the synthetic minority over-sampling technique (SMOTE) was applied to ensure balanced representation across severity classes. The models were trained and evaluated by stratified sampling, early stopping, and fairness-aware validation. The LightGBM and XGBoost models were implemented, achieving high accuracy, with LightGBM reaching 99% accuracy and XGBoost achieving 99.9% accuracy. Both models showed impressive AUROC scores, highlighting their strong predictive performance. The models were trained efficiently, with LightGBM achieving a faster training time of 100s and XGBoost for taking 120s. Our findings demonstrate the effectiveness of tree-based ensemble methods in capturing nonlinear feature interactions, while maintaining interpretability in severity classification. This study emphasizes the potential of EHR-based machine learning models to enhance clinical decision-making, improve accuracy, and enable proactive patient care.
Keywords
Publication Details
- DOI: DOI is assigned once the paper is indexed in IEEE Explore
- Type of Publication:
- Conference Name: IEEE Conference on Biomedical Engineering, Computer and Information Technology for Health 2025
- Date of Conference: 29/11/2025 - 29/11/2025
- Venue: Eastern University, Dhaka, Bangladesh
- Organizer: IEEE Bangladesh Section