SpectroCardioNet — Attention Network for Heart Valve Disease Detection

Abstract

The phonocardiogram (PCG) signal is used for the early detection of cardiovascular diseases (CVDs) as it captures the heart sound characteristics. In this article, a spectral attention-based deep learning network is proposed for the automatic detection of cardiac disease from the spectrograms of PCG signals, namely SpectroCardioNet. From a given PCG audio signal, in view of simultaneously utilizing both time- and frequency-domain information, spectrograms, delta-spectrograms, and double-delta-spectrograms are generated. The extracted triple-spectrogram representation is applied in the proposed network as a three-channel 2-D input, where it passes spectral and sequential feature paths. In the spectral feature path, a spectral attention block (SAB) is designed to emphasize some regions in the spectrograms based on a deep attention network and its output is then processed through the spectral pattern detectors (SpPDs). On the other hand, in order to extract the temporal behavior of the frequency components of the spectrograms, a 1-D convolution-based sequential feature extractor is also proposed. Extensive experimentation is carried out on two standard PCG datasets and very satisfactory performance is achieved in comparison to that obtained by some existing methods.

Highlights

A dual-path architecture: a spectral feature path (spectral attention + 2-D convolutional pattern detectors) and a sequential feature path (1-D convolutional pattern detectors) fused before classification — capturing both spatial structure and temporal evolution of the spectrogram.
A triple-spectrogram input — log spectrogram, delta, and double-delta — instead of the single spectrogram most prior work uses, giving the network both frequency content and how it changes over time.
A spectral attention block (SAB) that learns a soft spatial mask over the spectrogram, emphasizing diagnostically relevant time-frequency regions before pattern detection.
Only 2.4M parameters across 25 layers, trained from scratch with no transfer learning — two orders of magnitude smaller than VGG16 (138M), ResNet50 (25.6M), or DenseNet (121M) — with an estimated 22MB memory footprint small enough for a Raspberry Pi 4 or BeagleBone Black.
91.36% accuracy on the PhysioNet/CinC 2016 normal/abnormal benchmark and 97.77% on a 5-class valve disease dataset, beating prior published methods on both.

Method

Rather than treat a PCG spectrogram as an ordinary 3-channel color image and reuse a large pretrained vision backbone — the conventional approach — SpectroCardioNet treats it as a single-channel 2-D signal and builds a compact, purpose-built network around it, trained from scratch.

Full SpectroCardioNet architecture: a triple spectral feature generator feeds parallel sequential and spectral feature paths, which are concatenated and passed through a fully connected layer to a disease class output

Fig. 1 — Full architecture. The Triple Spectro-Temporal Transformer (TSTT) turns the 1-D PCG signal into a (129, 54, 3) triple-spectrogram (log, delta, double-delta). This feeds a sequential feature path (four 1-D SePD blocks) and a spectral feature path (SAB followed by two 2-D SpPD blocks and global average pooling); both paths are concatenated and passed through a fully-connected layer to predict the disease class.

Triple Spectro-Temporal Transformer (TSTT)

The raw PCG signal is converted to a spectrogram via short-time Fourier transform (256-sample frames, 32ms overlap, at a resampled 2kHz rate — PCG spectral content lives almost entirely below 500Hz). Delta and double-delta spectrograms are then derived to capture how the frequency content changes over time, giving a 3-channel, 2-D input analogous to how delta/delta-delta features are used in speech recognition.

Spectral Attention Block (SAB)

The spectrogram is expanded via transposed convolution, normalized, passed through a convolution that restores the original size, and squashed through a sigmoid into a soft [0,1] mask — upweighting diagnostically relevant time-frequency regions and downweighting the rest. This mask is then multiplied element-wise back onto the original input before further processing.

SAB architecture: input features pass through transpose convolution, batch norm, ReLU, convolution, batch norm, then a sigmoid branch produces an attention mask that is multiplied back onto the original input

Fig. 3 — The Spectral Attention Block (SAB): a learned sigmoid mask emphasizes the region of interest in the input spectrogram before it is passed to the spectral pattern detectors.

Spectral & Sequential Pattern Detectors

The SAB’s output passes through two Spectral Pattern Detector (SpPD) blocks — ordinary 2-D convolution, batch normalization, and ReLU — that extract spatial time-frequency correlations, finishing with global average pooling into a 1-D feature vector. In parallel, the sequential feature path runs the un-attended triple-spectrogram through four 1-D Sequential Pattern Detector (SePD) blocks (1-D convolution along the time axis), capturing how frequency-domain features evolve over time. The two resulting feature vectors are concatenated and passed through a fully-connected layer to produce the disease class prediction.

Results

Evaluated on two public PCG datasets: the PhysioNet/CinC 2016 Challenge dataset (3126 recordings, 2 classes — normal/abnormal, tested on the 301 recordings the dataset provider marks as most challenging) and a secondary 5-class dataset (1000 recordings: normal, aortic stenosis, mitral regurgitation, mitral stenosis, mitral valve prolapse).

PhysioNet/CinC 2016 (Normal vs. Abnormal)

Method	Accuracy	Sensitivity	Specificity	F1
SNMFNet	—	89.18%	77.66%	—
AlexNet+SVM	85.00%	—	—	—
DNN [27]	85.65%	86.73%	85.74%	—
DNN [28]	85.00%	—	—	—
TFDDL	85.16%	—	—	—
SpectroCardioNet (ours)	91.36%	93.33%	89.40%	91.50%

5-Class Valve Disease Dataset

Method	Classes	Classifier	Accuracy
WSST	4	RF	95.13%
MFCC	5	SVM	91.60%
DWT	5	SVM	92.30%
IFE	5	RF	94.80%
IFE	5	KNN	91.60%
SpectroCardioNet (ours)	5	DNN	97.77%

Per-class performance on the 5-class dataset (5-fold cross-validation) averages 97.77% accuracy, 96.57% precision, 96.66% recall, and 96.60% F1 across Normal, Aortic Stenosis, Mitral Regurgitation, Mitral Stenosis, and Mitral Valve Prolapse — with the confusion matrix below showing only a handful of cross-class errors (mostly between mitral stenosis and mitral valve prolapse) and only 5.31% false positives / 3.32% false negatives on the binary PhysioNet task, corresponding to an ROC AUC of 0.92.

Confusion matrices for the PhysioNet/CinC-2016 dataset (left, binary) and the 5-class GitHub dataset (right), showing high diagonal accuracy with few misclassifications

ROC curve for the PhysioNet/CinC-2016 dataset, bowing sharply toward the top-left with an area under the curve of 0.92

Fig. 6 (left) — Confusion matrices: PhysioNet/CinC-2016 (binary, left panel) and the 5-class GitHub dataset (right panel). Fig. 7 (right) — ROC curve for PhysioNet/CinC-2016 (AUC = 0.92).

Conclusion

SpectroCardioNet shows that a compact, purpose-built network — combining a triple-spectrogram input, a learned spectral attention mask, and parallel spectral/sequential feature paths — can outperform both classical hand-crafted-feature pipelines and larger transfer-learning-based networks on PCG-based heart valve disease detection, while using far fewer parameters and no pretraining. At an estimated 22MB memory footprint, the model is small enough to run on low-power single-board computers, making it a plausible building block for a low-cost clinical diagnostic tool or a remote-monitoring pipeline where a microcontroller streams PCG data to a server running the model.

Citation

@article{chowdhury2022spectrocardionet,
  title   = {{SpectroCardioNet}: An Attention-Based Deep Learning
             Network Using Triple-Spectrograms of {PCG} Signal for
             Heart Valve Disease Detection},
  author  = {Chowdhury, Sakib and Morshed, Monjur and Fattah, Shaikh Anowarul},
  journal = {IEEE Sensors Journal},
  volume  = {22},
  number  = {23},
  pages   = {22799--22807},
  year    = {2022},
  doi     = {10.1109/JSEN.2022.3196263}
}