Abstract
Coronary artery disease remains one of the leading causes of death and disability worldwide. Timely diagnosis can reduce the incidence of complications and ease the burden on healthcare systems. However, traditional methods are often costly, invasive, and limited in accessibility. Recent studies confirm the potential of machine learning for clinical applications. This raises the question: is it possible to reliably predict the presence of disease using only clinical and demographic data, without imaging methods?The aim of this study was to evaluate the accuracy and practical value of such models. Using the UCI Heart Disease dataset (n = 920) under a unified protocol, the LightGBM algorithm was trained and achieved the following results: accuracy = 0.8696, precision = 0.8679, recall = 0.9020, F1-score = 0.8846. These findings complement previous research based on imaging approaches.The study compared multiple algorithms under identical preprocessing and validation conditions, assessed probability calibration, and applied SHAP for interpretability. The analysis revealed that the main predictors (e.g., ST-segment depression) aligned with established clinical knowledge. This confirms that the model can be used for initial screening and referral to additional diagnostics. Overall, calibrated and interpretable algorithms based on open clinical data can serve as a valuable tool for patient routing in resourcelimited settings.