Piezoelectric actuators are widely used in medical robots due to the high positioning accuracy and repeatability. However, the inherent hysteresis characteristics seriously affect the positioning performance. Therefore, it is necessary to model the hysteresis characteristics to improve the positioning accuracy. This paper innovatively proposes a Modified K-means Gaussian Mixture Model - Gaussian Mixture Regression(Modified K-means GMM-GMR) model for hysteresis modeling. Firstly, hysteresis loops are separated into ascending curve and descending curve, and then classified by K-means respectively. Subsequently, each class of hysteresis characteristics is modeled using the trained sub-GMM-GMR, and the historical predicted output displacement is adopted to modify the current predicted output displacement. Compared with the Prandtl–Ishlinskii(PI) model and the Modified Bouc-Wen(MBW) model, the experimental results verify the superiority of the proposed method. In addition, compared with the traditional K-means GMM-GMR model, the experimental results show that the modeling error is reduced by about 10 \(\%\) on average, which verifies the effectiveness of the proposed method.