AI Heart Scanner Hits 98.95% Accuracy Detecting Five Deadly Valve Diseases
A Gabor dictionary paired with elastic net regularization feeds a 1D+2D CNN-LSTM classifier, hitting 98.95% on five heart valve conditions.
PRO- New PCG classifier hits 98.95% on five heart valve conditions using Gabor dictionary and elastic net features.
- Replaces greedy matching pursuit with elastic net regularization solved by ADMM for atom selection.
- 1D coefficient vectors are reshaped into 2D time-frequency matrices to feed a 1D+2D CNN-LSTM.
- Best setup uses high-time low-frequency Gabor atoms (beta=2) and mostly-ridge regularization (alpha=0.1).
- Beats raw-signal CNN-LSTM by 30.92% relative error reduction and VMD pipeline by 22.22%.
- Works on whole recordings, avoiding costly per-cardiac-cycle segmentation required by prior methods. Paper
Automated heart auscultation has been slowly creeping toward clinical viability, and a new study pushes multi-class phonocardiogram (PCG) classification to 98.95% accuracy on a five-way valvular disease task. The recipe blends a classical signal processing idea, a sparse Gabor dictionary tuned with elastic net regularization, and a deep CNN-LSTM classifier. The result is a compact, interpretable feature pipeline that outperforms recent raw-signal and variational-mode-decomposition baselines on the same dataset.
Why heart sounds resist clean modeling
An estimated 17.9 million people die each year from cardiovascular diseases, and the PCG signal, essentially a stethoscope recording, carries acoustic clues about valve function that ECG cannot. The catch is that heart sounds are nonstationary. Once the analysis window size is chosen in an STFT, the resolution is constant over the whole time-frequency plane, forcing a hard trade-off: short windows catch fast transients like the S1 and S2 heartbeats but blur nearby frequencies, while long windows resolve frequencies but smear time.
Wavelets help, and so do overcomplete dictionaries where signals are decomposed as sparse combinations of localized time-frequency atoms. Prior work leaned on matching pursuit, a greedy iterative approximation. This paper swaps that step out for something more statistically grounded.
Gabor atoms meet elastic net
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
