A Fuzzy Rule-Based Decision Framework with Conflict Index Analysis for Coronary Artery Disease Classification

Source abstract

Coronary artery disease (CAD) classification requires integration of clinical and diagnostic variables whose interpretations are often gradual rather than strictly binary. This study evaluates a fixed 45-rule Mamdani fuzzy classifier on the 13 variables of the processed Cleveland Heart Disease dataset (297 complete cases) as an auditable post-diagnostic evidence-integration framework rather than a replacement for optimized machine learning classifiers. Three repeats of five-fold nested cross-validation on the 237-record development cohort yielded mean accuracies of 83.7% for random forest, 82.3% for support vector machine (SVM), and 78.2% for multilayer perceptron, while the fixed fuzzy classifier achieved 79.7% selective accuracy with 99.6% coverage (bootstrap 95% CI for selective accuracy: 74.2–84.7%). On the descriptive 60-record internal holdout, fuzzy accuracy was 81.7% (95% CI: 70.1–89.4%), compared with 83.3% for SVM and random forest; exact paired McNemar tests did not show evidence of an accuracy difference (all p = 1.00). A conflict audit based on simultaneous Low and High rule supports was evaluated as an exploratory review flag. Applying the prespecified exploratory threshold C(x) ≥ 0.5 to the holdout retained 95.0% of cases and produced 82.5% selective accuracy, a small change from 81.7%. Additional analyses examine feature dependence, rule activation, membership-boundary perturbation, alternative t-norms and defuzzification procedures, and competing ambiguity measures. The results support computational auditability but not clinical validity: the author-specified rules, reliance on ca and thal, small historical dataset, and lack of external cardiologist validation remain important limitations.