Improving Sepsis Mortality Prediction with Machine Learning Using Full Region Synthetic Sampling Approach

Background and Aims: Sepsis is a life-threatening condition and remains a leading cause of mortality in Intensive Care Units (ICUs). Accurate mortality prediction is crucial for optimizing ICU resource allocation. However, severe class imbalance in ICU datasets hampers the generalization performance of machine learning models.

Methods: This study proposes a Full Region Synthetic Sampling Approach (FRSSA), a novel data augmentation method that dynamically balances the minority class distribution based on regional density. Additionally, we introduce Adaptive Synthetic Sampling Tuning (ASST), an optimization-based strategy that adjusts augmentation weights to enhance model performance. To evaluate model fairness and clinical utility, we propose the Balanced Performance Score (BPS), which integrates accuracy, precision, and recall for personalized ICU risk assessment. Also, we compare two augmentation strategies: 1) Pre-Splitting: Augmentation occurs before dataset splitting, and 2) Post-Splitting: Augmentation is applied only to the training set to ensure fair evaluation. We utilize publicly available ICU datasets from both the MIMIC-IV and eICU-CRD databases and evaluate the performance of Random Forest, XGBoost, and LightGBM models, with hyperparameters optimized using RandomizedSearchCV on the training set.

Results: The Pre-Splitting strategy achieved 89.64% accuracy and an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.968, but posed a risk of test-set contamination. In contrast, the Post-Splitting strategy yielded 78.31% accuracy and an AUROC of 0.7241, ensuring better real-world generalization. The ASST mechanism optimally balanced 65% interpolation (FRSSA) and 35% expansion (FRSSA), reducing false positive rates and enhancing model fairness.

Conclusion: FRSSA preserves regional data distribution by generating synthetic samples near the imbalanced regions. ASST dynamically adjusts augmentation ratios to maximize classification performance and generalization. The BPS score-fusion metric offers a flexible evaluation framework, accommodating varying clinical priorities in ICU settings. Our findings demonstrate that Post-Splitting augmentation with FRSSA and ASST produces a fairer and more reliable ICU mortality prediction model.