Authors - S M Mazharul Hoque Chowdhury, Ruth West, Stephanie Ludi Abstract - The prediction of liver disease through clinical data analysis faces difficulties because current machine learning methods fail to handle class imbalance and produce incorrect probability assessments. The existing supervised and ensemble methods use fixed decision thresholds together with heuristic weighting methods which results in biased predictions that compromise their ability to achieve balanced performance. The research introduces CAL-WE++ which serves as a Calibration- Weighted Ensemble system that uses an MCC-Optimized Threshold to forecast liver disease. The system employs five-fold stratified cross-validation without data leakage to produce out-of-fold probability results. The model weights are determined by evaluating both the model's ability to distinguish between outcomes (measured through ROC-AUC) and its accuracy in predicting probabilities (assessed through Expected Calibration Error ECE). The Matthews Correlation Coefficient (MCC) serves as the optimization method to determine the final classification threshold which helps to solve class imbalance problems. The Indian Liver Patient Dataset (583 records; 416 diseased, 167 non-diseased) experiments show that CAL-WE++ achieves a mean cross-validation MCC of 0.3474 and a test MCC of 0.4487 which exceeds the performance of baseline classifiers. The model achieves a ROC-AUC score of 0.8140 and a PR-AUC score of 0.9272 while maintaining a low ECE value of 0.0774 which demonstrates strong ability to distinguish between different outcomes and accurate probability assessments. The CAL-WE++ framework offers medical professionals a decision-making system that maintains balance between multiple criteria while delivering dependable outcomes for medical datasets with unequal class distributions.
Open Zoom