Machine learning classification of epicardial vessels using near-infrared spectroscopy in porcine and human hearts

Near-infrared spectroscopy (NIRS) offers potential for real-time, label-free identification of epicardial coronary vessels during ablation procedures. We present a machine learning classification algorithm for vessel detection using NIRS spectra acquired from porcine and human hearts ex vivo. Our methodology employs three optical index strategies: wavelength reflectance ratios, integrated reflectance, and derivative reflectance analysis. A systematic grid search over various wavelength combinations identified the optimal reflectance ratio of 1000nm/945nm, achieving the highest area under the receiver operating characteristic curve for vessel detection. Four classification models were evaluated: logistic regression (with/without LASSO regularization), support vector machines, and random forests. Models trained on porcine data achieved 80.91% sensitivity and 78.43% specificity when validated on human specimens, demonstrating robust cross-species generalizability. This technology could significantly improve epicardial ablation safety by enabling real-time vessel detection, helping clinicians avoid inadvertent coronary artery damage during catheter placement and energy delivery.