Strategic Framework for Disease Classification: Leveraging Benefits and Costs in Multiclass Analysis

Health datasets are often imbalanced, with a skew towards the healthy class, making traditional performance metrics like accuracy inadequate for evaluating minority class (disease-affected) predictions. Moreover, standard classifiers typically assume equal costs and benefits for decision outcomes, which is unsuitable for healthcare where misclassification implications vary. This paper addresses these challenges by integrating cost–benefit considerations into both the training and evaluation of classifiers through a generalized, data-driven framework for deriving benefit and cost parameters based on clinical indicators such as survival probabilities, treatment outcomes, and healthcare expenditure. We propose modifications to Naive Bayes and Logistic Regression algorithms for multiclass classification and compare them against hierarchical and ensemble cost-sensitive algorithms, including Hierarchical Cost-Sensitive Kernel Logistic Regression (H-CSKLR), SAMME.C2, BAdaCost, and NP-MC. Experiments across fetal health, vertebral column, and hepatitis C datasets show that the proposed multiclass Benefit-Based Logistic Regression achieved the highest benefit values—17.96 for fetal health, 37.19 for vertebral column, and 4.80 for HCV—outperforming H-CSKLR (15.36, -12.54, -3.44) and traditional accuracy-based Logistic Regression in most cases (19.01, 35.10, 4.60). A Latin Hypercube Sampling–based sensitivity analysis further confirmed the robustness of the proposed framework across diverse benefit–cost configurations. These results highlight the transparency, stability, and clinical relevance of the approach.