Cardiac auscultation is a non-invasive technique to detect congenital and acquired heart diseases from pediatric populations. The goal of the George B. Moody PhysioNet Challenge 2022 is to develop algorithms to detect the presence or absence of heart murmurs for each pediatric subject from multiple phonocardiogram (PCG) recordings from different auscultation locations. In this work, we develop a mixture of deep experts model to classify each PCG recording into one of the three classes: present, absent, and unknown. Our approach was motivated by the following two hypotheses: 1) hand-crafted PCG features can utilize prior clinical knowledge in identifying noisy recordings that are labeled as unknown by clinicians; 2) both the raw time series and their spectrogram representations provide complementary information in identifying the presence and absence of heart murmurs. Therefore, our approach combined three local deep experts trained on different data representations: 1) a residual neural network trained on raw time series of the recording; 2) a Transformer network trained on spectrogram representation of the recording; 3) a multiple layer perceptron network trained on a subset of predictive hand-crafted PCG features, which were selected by ranking their mutual information with the class label. Different local experts were dynamically selected by the gating function for different PCG recordings. The subject was predicted as present if any of the multiple PCG recordings from different auscultation locations was predicted as present. Using the official scoring metric, our first entry in the unofficial phase achieves a score of 885.228 (team name = prna) on the held-out test set.