Multi-grained cascade forest model for automatic CAD characterization on ECG segments

Abstract This paper presents a high-performance algorithm for diagnosing coronary artery disease (CAD) through a blindfold strategy and subject-specific data with ECG signals.In the proposed method,multi-grained scanning with sliding window is constructed to ensure the validity of the features extracted in the ECG segments.Completely random forests and random forests are used for extracting diverse features automatically.This allows the proposed algorithm to achieve excellent detection accuracy with subject-specific data and different-scale training data. Moreover, the gain comparison of the cascade forests is used inherently to optimize model parameters automatically, thereby avoiding errors caused by the manual setting of parameters. The proposed algorithm achieves an accuracy with 100% when 10% of the training data are used. Even in the case where the training set ratio is 0.1%, the detection accuracy of the proposed model is 99.86%. Additionally, the classification performance of the proposed algorithm on subject-specific data reaches 99.98%. Due to its robustness to perturbations in the scale of the training data and efficiency regarding specific subject data, the proposed system is applicable to CAD diagnosis.