Enhancing Patient Survival Prediction via Cutting-Edge Data Mining Techniques and Future Prospects

Abstract

This research explores utilizing data mining of electronic health records to accurately predict hospital patient mortality. A dataset containing over 100,000 episodes of hospitalizations with extensive clinical variables was used to develop machine-learning models for survival classification. The significant class imbalance between survivor and non-survivor outcomes was handled in preprocessing via down sampling to prevent prediction bias. Feature engineering selected 15 key predictors from the hundreds available, including factors such as age and blood. pressures and disability scores. The extreme gradient boosting XGBoost classifier achieved the highest test accuracy of 84.75 percent. However, limitations around model interpretability through explainable AI techniques and rigorous temporal validation across recent periods persist. Enhancing reproducibility, transparency, and precision remains imperative before any clinical integration. The technical feasibility of distilling useful mortality risk insights from high-dimensional, heterogeneous patient data is demonstrated but significant challenges hamper real-world viability currently. This research highlights the overarching complexity but also the importance of data mining for unlocking reliable, trustworthy predictive insights to save lives in healthcare.