Enhancing Hospital Mortality Prediction: A Comprehensive Approach Through Machine Learning and Model Interpretability

Accurate hospital mortality prediction is crucial for guiding clinical decisions, optimizing resource allocation, ensuring quality of care, and managing healthcare costs. However, challenges such as data preprocessing, dataset balancing, and model interpretability limit the effectiveness of existing approaches. This study addresses these gaps by developing a robust predictive framework. Using a comprehensive dataset of 59 numerical and categorical variables sourced from Kaggle, preprocessing techniques were applied to handle missing values, categorize variables, scale features, and balance classes. Principal Component Analysis was employed to reduce feature dimensionality and minimize multicollinearity, strengthening the models' predictive capabilities. Multiple machine learning algorithms, including AdaBoost, Naive Bayes, and Logistic Regression, were evaluated. To enhance model transparency, SHAP and LIME were integrated, offering valuable insights into feature influence and improving trust in the models' decisions. Additionally, a Classifier Evaluation Dashboard was developed to enable interactive visualization and better interpretability. The results revealed that while models like Random Forest, Extra Trees Classifier, and Decision Tree achieved high performance metrics, they were prone to overfitting. Logistic Regression provided balanced performance with lower overfitting risks with 84.06% accuracy, 84.82% precision, 83.23% recall, 84.02% F-1 score and 84.07% ROC AUC, making it more suitable for practical applications.