Abstract
Designed prevention and clinical support ensembles for early cardiac screening on the Dangerous Heartbeat Dataset (CHSC2011). Heart sounds were resampled at 4 kHz, segmented into 1-second windows, described with MFCC, chroma, spectral and temporal descriptors, and reduced from 338 to 41 features via Spearman-based filters. The prevention ensemble keeps false normals under control (ROC-AUC 0.96, TPR 43.4% at 1% FPR) while the five-class support ensemble delivers macro F1 81.6 with per-class risk analysis and SHAP explanations.
About
The study processes the Dangerous Heartbeat Dataset (CHSC2011) by resampling heterogeneous recordings at 4 kHz, slicing 1-second windows and extracting 338 temporal, spectral and cepstral descriptors (MFCC, chroma STFT, RMS, ZCR, CQT, spectral centroid/bandwidth/roll-off). A two-step Spearman filter removes weakly correlated and redundant attributes, shrinking the footprint to 41 features. Two complementary pipelines are trained: a prevention ensemble (Random Forest + MLP Ultra + MLP Rollercoaster) that evaluates normal-vs-rest performance across strict FPR thresholds, and a five-class diagnostic ensemble (Random Forest + MLP Ultra) that balances macro F1, MCC and per-class risk. SHAP values and permutation importance explain decisions and highlight waveform regions that drive predictions.
Key results
- 0.82
- F1-Score
- 0.96
- ROC-AUC
- 43.4%
- TPR @1% FPR
- 74.3%
- TPR @5% FPR
- 86.6%
- TPR @10% FPR
- 95.8%
- TPR @20% FPR
- 41 of 338
- Features Retained
- 81.53
- Support MCC
Key findings
- The Dangerous Heartbeat Dataset (CHSC2011) was resampled to a uniform 4 kHz, windowed at 1 s and enriched with MFCC, chroma, spectral and temporal descriptors before Spearman filtering removed 87.9% of redundant variables.
- Complementary prevention (normal vs rest) and clinician-support (5-class) ensembles were trained from Random Forest and diverse MLP backbones to balance low false normals with detailed disease classification.
- Explainability with permutation importance and SHAP highlighted waveform regions influencing predictions, surfacing where murmurs and extra systoles overlap normal beats and guiding future data collection.
Challenges
- Dataset size and lack of a dedicated validation split increase variance risk in the reported metrics.
- One-second windows may miss full cardiac cycles, limiting performance on murmurs and extra systoles.
- Explainability relies on averaged features; medical validation is still required before clinical deployment.
Methodology
- 01Heart sound clips from CHSC2011 were resampled at 4 kHz, segmented into 1-second windows, described with 338 spectral/temporal features, pruned to 41 via two-stage Spearman filtering, and used to train prevention and five-class support ensembles whose behaviour was analysed with risk metrics, permutation importance and SHAP.