Abstract: The applicability of traditional human reliability analysis methods is constrained by the complex scenarios and intense human-robot interactions inherent in motion rehabilitation robot systems. Consequently, a tailored human reliability analysis model is developed to achieve precise prediction of human error probabilities. A modified cognitive reliability and error analysis method (CREAM) is proposed by improving the evaluation of specific common performance conditions (CPCs) using quantitative rehabilitation data. The decision making trial and evaluation laboratory (DEMATEL) method, integrated with spherical fuzzy numbers, is employed to convert expert linguistic evaluations into quantitative data. This approach analyzes the interactions among CPC factors while quantifying their impact on human reliability. A Bayesian network model is constructed to infer operator cognitive control modes and human error probabilities during robot usage. Experimental results indicate that the proposed model successfully predicts human error probabilities, accurately reflecting sensitivity to changes in CPC states. The improved CREAM framework achieves an adaptive extension for complex human-machine interaction scenarios. The constructed model provides a quantifiable and explainable tool for risk assessment and control decisions, verifying its applicability and effectiveness in rehabilitation training contexts.