The aggravation of left ventricular diastolic dysfunction (LVDD) could lead to left ventricular remodeling, wall stiffness, and the reduced compliance, which make progression to heart failure with preserved ejection fraction (HFpEF). To achieve early diagnosis of LVDD, a non-invasive method is proposed, which utilizes the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) multi-scale sample entropy (MSE) characteristics and the logistic regression model. Firstly, the improved wavelet denoising method is used for heart sound signals preprocessing. Then, the non-stationary heart sound signals are decomposed into several intrinsic mode functions ( IMF) which reflect the characteristics of heart sound itself by the ICEEMDAN method. The mutual correlation coefficient criterion is used to select IMF. The MSE values of the selected IMFs are extracted to form the eigenvectors, which are used as the input into the classifier for identification. Finally, the logistic regression is applied for LVDD identification by the comparison of performances with other three models. Results show that the proposed method could effectively extract the features of heart sound with 89. 85% accuracy, 92. 17% sensitivity and 87. 63% specificity, which demonstrate the effectiveness of heart sound signals for LVDD diagnosis.