The training and evaluation of cardiopulmonary resuscitation (CPR) are crucial for enhancing rescuers’ CPR proficiency, thereby potentially improving survival rates for cardiac arrest patients. To address limitations in conventional CPR training and quality monitoring, this study presented a type of intelligent algorithms for evaluating CPR actions based on wearable devices. After obtaining and preprocessing the raw data of EMG and IMU, we constructed a series of time-series features, and calculated frequence-domain features by fast Fourier transformation, which will be optimized through principal component analysis for dimensionality reduction. Then, we systematically evaluated comprehensive classification performance of multiple models based on these features and specific-category performance, in terms of five metrics: Accuracy, Precision, Recall, F1 and AUC. Further, we analyzed the contributions of EMG and IMU features and the effect of dimension reduction, respectively. The experimental results demonstrated that the models with the EMG&IMU features possessed high-quality performance for recognition of CPR actions; both EMG and IMU features have positive contributions, while their fusion can further enhance the recognition ability of CPR actions. Moreover, the dimension reduction can significantly improve computational efficiency while retaining generalizability. This study provides novel theoretical and technical support for intelligent CPR training and evaluation, demonstrating its potentials of applications in emergency rescue scenarios.