Automatically detecting and identifying murmurs at an early stage can reveal the presence of Congenital heart defects and other cardiac valve defects. In this work, we devised an automatic screening system to categorize the multiple heart sound signals recorded from multiple auscultation sites on the body into the following three classes: (a) the presence, (b) absence, and (c) ambiguous cases of murmurs. The method is based on hand-crafted features and application of LightGBM classifier for the multi-class classification of heart sound signals. In total 366 features were extracted from each record, out of which 6 were demographic features, 135 tensor based features, and 45 are hand-crafted features from five sites of auscultation. Records with any missing sites are imputed with corresponding mean feature values to maintain the uniform feature length. Tucker tensor decomposition is applied on tensors formed out of spectrograms to obtain tensor related features. The 45 hand-crafted features comprise of 5 statistical features, and 40 frequency-domain features including mel-frequency cepstrum coefficients (MFCCs). In an unofficial phase challenge entry, our team ‘Medics’ achieved a challenge score of 2238. As an additional evaluation using the training data, cross-validation of 80–20 splits was carried out and average challenge score of 508 was obtained.