EXPLAINABLE PROGNOSIS OF ISCHEMIC HEART DISEASE: CALIBRATED ML WITH KNN BASELINES AND EXTERNAL VALIDATION

Abstract

Ischemic heart disease (IHD) remains a leading cause of global morbidity and mortality. Traditional risk scores often lack personalization and transparency, limiting their clinical utility. This study develops and evaluates explainable, calibrated machine learning (ML) models for IHD prognosis, benchmarked against k-nearest neighbors (kNN) baselines and validated across independent datasets. A quantitative, survey-based approach was combined with predictive modeling. Data from 400 purposively sampled physicians were analyzed using SPSS to assess perceptions of ML explainability, calibration, trust and adoption intention. Reliability, descriptive statistics, correlations and multiple regression were performed. ML models, including kNN, logistic regression, ensemble methods and an explainable calibrated ML framework, were trained and evaluated using internal and external validation. Performance was measured using AUC, Brier scores and calibration curves. Survey analysis showed strong physician support for explainability (M = 4.21, SD = 0.58) and high behavioral intention to adopt ML tools (M = 4.18, SD = 0.59). Regression confirmed explainability (β = .33, p < .001), usefulness (β = .30, p < .001) and trust (β = .22, p < .001) as the strongest predictors of adoption. Calibration literacy also contributed positively (β = .11, p = .008). In model comparisons, kNN achieved an AUC of 0.70 (external validation) while the explainable calibrated ML model outperformed all others, with AUC = 0.87 and the lowest Brier score (0.148), demonstrating both accuracy and reliability across datasets. Explainability, calibration and external validation are critical enablers of physician trust and adoption of ML prognosis tools for IHD. When combined with organizational support and training, these models hold significant potential to enhance prognostic accuracy and improve patient care in cardiology.