Traditional probability neural network (PNN) has strong fault tolerance, simple learning process and fast training speed. To improve the performance of the traditional PNN in heart sound classification, we adopt least mean square (LMS) method to implement the optimization, thereby increasing the accuracy of heart sound classification and prediction. The LMS-PNN algorithm frames the heart sound signal using the window function, uses the double threshold method to determine the value of the data, employs the LMS algorithm to debug the corresponding parameters, and saves the denoised data in the format of mat file. It extracts the short-time autocorrelation coefficients and short- time power spectral densities of each heart sound, and uses PNN to extract 40 000 sample data for training. Each heart sound is graded and predicted. After inputting the training data from the mode layer of the PNN algorithm, experimental data verification shows that the prediction accuracy of LMS-PNN can reach more than 96%.