PCEN and multi-resolution spectrograms boost heart sound detection to 0.916 accuracy

A new study from arXiv (cs.CY 2607.16220) compares three spectrogram front-ends for binary classification of abnormal heart sounds using a convolutional neural network. The researchers—Abhinav Pala and Dhanush Pala—kept the exact same CNN architecture, optimizer, and random seed while varying only how raw audio was converted into spectrograms. They tested a standard logmel spectrogram, PCEN (per-channel energy normalization that normalizes each frequency bin over time), and a multi-resolution version stacking multiple window sizes.

On the PhysioNet 2016 dataset, all three methods showed high sensitivity (around 0.95) for detecting abnormal cases. However, PCEN and multi-resolution spectrograms achieved official PhysioNet accuracy scores of 0.915 and 0.916, slightly but consistently outperforming the baseline logmel at 0.910. Grad-CAM heatmaps revealed the model consistently focused on low-frequency regions corresponding to S1 and S2 heart sounds, suggesting the CNN learned physiologically meaningful features. This work demonstrates that preprocessing choices matter even with fixed architectures, and PCEN or multi-resolution approaches could improve automated cardiac screening tools at minimal computational cost.