This paper introduces a method for the automatic classification of the phonocardiogram. The method utilizes a multi layer perceptron for detecting the presence of the systolic murmur and a spectral analysis for separating innocent murmur from pathological murmur. The effectiveness of the method was confirmed by applying the method to the data collected from nationwide phonocardiogram screening of the elementary school children conducted in Japan. Data include 44 pathological systolic murmur, 61 innocent murmur and 36 normal data. Average correct detection rate of 90.8% was achieved for systolic murmur detection. Spectral analysis then successfully separated musical murmur from pathological systolic murmur. The method could be an effective assistance for medical doctors make their final diagnoses.