Comparing Spectrogram Front-Ends for Abnormal Heart-Sound Detection with a Convolutional Neural Network

Technical Summary: Comparing Spectrogram Front-Ends for Abnormal Heart-Sound Detection

Problem Statement
Cardiovascular disease remains a leading cause of mortality, necessitating early, low-cost screening methods. While digital stethoscopes can capture phonocardiograms (audio signals of heart sounds), manually interpreting these signals is resource-intensive. This project addresses the binary classification task of determining whether a short heart-sound recording is "normal" or "abnormal." The core challenge identified is not the choice of the classifier, but the representation of the raw audio waveform. The authors posit that small variations in how a spectrogram is computed—specifically regarding analysis windows, frequency scaling, and dynamic-range compression—significantly influence the patterns a downstream Convolutional Neural Network (CNN) can learn.

Methodology
The study employs a controlled experimental design to isolate the impact of the input representation (front-end) on model performance.

• Dataset: The research utilizes the PhysioNet/CinC 2016 heart-sound dataset, comprising 3,240 recordings (79.5% normal, 20.5% abnormal). Data was split at the recording level to prevent leakage, resulting in a stratified test set of 1,984 clips (5-second duration, resampled to 2,000 Hz).
• Fixed Classifier: A compact 2D CNN was used for all experiments to avoid overfitting on the relatively small dataset. The architecture consists of three convolutional blocks (8, 16, and 32 filters), batch normalization, ReLU activations, and global average pooling.
• Variable Front-Ends: Three distinct spectrogram representations were tested on the exact same model, optimizer (Adam), and training schedule (20 epochs, weighted cross-entropy loss):
  1. Vanilla Log-Mel: A standard baseline using a 64-bin mel-scaled spectrogram with log-amplitude compression.
  2. PCEN (Per-Channel Energy Normalization): Replaces static log compression with an automatic gain control mechanism that normalizes each frequency bin against its recent background energy, aiming to highlight transient events like murmurs.
  3. Multi-Resolution Log-Mel: Stacks three log-mel spectrograms generated with different FFT window sizes (256, 512, and 1024 samples) as three input channels, providing the CNN with simultaneous time and frequency resolutions.
• Evaluation: Performance was measured using the official PhysioNet 2016 Modified Accuracy (MAcc), defined as the unweighted mean of sensitivity and specificity, alongside F1 score, raw accuracy, and confusion matrices. Grad-CAM was employed to visualize model attention.

Key Results

• Front-End Performance: All three front-ends achieved high sensitivity (~0.95), indicating strong detection of abnormal cases. However, PCEN and Multi-Resolution front-ends outperformed the Vanilla Log-Mel baseline on specificity and the primary MAcc metric.
  • Vanilla Log-Mel: MAcc = 0.910
  • PCEN: MAcc = 0.915
  • Multi-Resolution: MAcc = 0.916 (Best overall)
    The improvements were driven primarily by a reduction in false positives (flagging normal recordings as abnormal) rather than an increase in true positives.
• Architecture Robustness (Question 2): When the CNN architecture was reduced to a smaller two-block network, the performance gap widened. The Vanilla baseline suffered a significant drop in MAcc (to 0.826), whereas PCEN and Multi-Resolution remained stable (MAcc ~0.894). This suggests that adaptive front-ends provide structural inductive bias that compensates for reduced model capacity.
• Interpretability (Question 3): Grad-CAM analysis revealed that all models focused on the low-frequency bands corresponding to S1 and S2 heart sounds. However, the Vanilla model showed slightly more attention to high-frequency noise. PCEN focused on the regions immediately following S1/S2 spikes (likely murmurs), while the Multi-Resolution model exhibited a broader attention distribution across the signal, leveraging its multiple time-frequency views.

Significance and Claims
The paper claims that while a standard log-mel spectrogram is a strong baseline capable of detecting abnormal heart sounds, adaptive and multi-resolution front-ends offer a consistent, albeit modest, improvement in specificity and overall screening accuracy.

• Inductive Bias: The study demonstrates that in low-to-moderate data regimes, the quality of the input representation is as critical as model capacity. PCEN and Multi-Resolution front-ends inject domain-appropriate bias (suppressing background noise and capturing time-frequency trade-offs) that the CNN would otherwise struggle to learn from scratch.
• Clinical Relevance: The high sensitivity across all models aligns with the clinical goal of screening, where missing an abnormal case (false negative) is more costly than a false alarm. The models' focus on physiologically relevant S1/S2 bands, rather than dataset artifacts, validates the pipeline's reliability.
• Trade-offs: The authors note that the Multi-Resolution front-end achieved the highest metric but at the cost of triple the input channels and computational load. They conclude that while the gains are real, they are not "game-changing," suggesting that the choice of front-end depends on the specific constraints of the deployment environment.

Limitations
The authors acknowledge that the performance margins between front-ends are small relative to training variance (random seed effects). Additionally, the study was limited by computational constraints, preventing the testing of larger architectures or more extensive hyperparameter tuning. The dataset's imbalance and the binary nature of the labels (normal vs. abnormal) also limit the ability to diagnose specific cardiac conditions.