Synthetic ECG-guided temporal gating improves paediatric murmur grading, timing, and shape classification on the CirCor DigiScope dataset
Z Ahmad, N Perwaiz, S Khan, S Sultana, I HaqAbstract
Background/ Introduction
Non-invasive murmur detection from phonocardiograms (PCG) has advanced, but detection answers only whether a murmur is present, not how severe it is or what to do next. Cardiologists also need to know: what grade, what timing, and what shape? As target pathologies increase in number, their acoustic content overlaps within the 20–500 Hz PCG bandwidth. In-band noise occupies the same frequency range, so frequency-based denoising cannot separate artefacts from diagnostic content. Grading accuracy on the CirCor DigiScope benchmark reaches only about 52%. Characterisation needs temporal localisation of murmur energy within the cardiac cycle, but single-channel PCG gives no timing anchor for phase boundaries.
Purpose
We tested whether a synthetic electrocardiogram (ECG), reconstructed from PCG, can gate each cardiac cycle into phase-specific windows and support murmur grading, timing, and shape classification from a single acoustic sensor.
Methods
We trained a deep learning model (Pix2Pix conditional generative adversarial network) to synthesise ECG from PCG spectrograms using 2,267 paired PCG-ECG cycles from 23 healthy adult EPHNOGRAM subjects. From synthetic-ECG R-peaks, we divided each cardiac cycle into four phases by fixed ratios: S1 (0–15%), systole (15–50%), S2 (50–60%), and diastole (60–100%) and computed 464 temporal, spectral, and morphological features per phase. We compared two ECG-gated classifiers against a PCG-only 2D convolutional neural network with self-attention on 3,163 CirCor DigiScope recordings from 942 paediatric patients (age 0–21, 54% male): a multilayer perceptron (MLP) and a random forest (RF), both under stratified 10 and 5-fold cross-validation.
Results
ECG reconstruction reached 0.903 structural similarity (SSIM) and 95.4% R-peak detection F1-score. Compared with the PCG-only baseline, ECG-gated classification improved all characterisation metrics (paired t-tests across 10 folds): grading accuracy increased from 67.4% to 74.0% (p = 0.008), timing from 68.4% to 72.9% (p = 0.011), and shape from 58.8% to 70.6% (p < 0.001). Detection accuracy decreased from 93.4% to 90.3% (p = 0.042), expected given the 6.9-fold shorter gated window (580 ms versus 4,000 ms baseline). An ablation that replaced ECG-gated boundaries with equal-time segmentation on the same feature set showed that temporal alignment drives the gain: MLP characterisation accuracy dropped to 45.6–50.8% without phase-gated boundaries (all p < 10⁻⁴, paired t-test).
Conclusions
Temporal gating through synthetic ECG improves characterisation accuracy by 4.5–11.8 percentage points over the PCG-only baseline. This approach could let a single PCG sensor extract murmur grade, timing, and shape for bedside screening in settings without echocardiography. We recommend prospective validation against echocardiographic diagnosis.Results comparison tableCore methodology