Clinically aligned hierarchical machine learning for patient-level classification of valvular regurgitation severity and subtype from phonocardiograms (PCG)
H Yu, L Lui, X Xu, B Yan, K SoAbstract
Background
Echocardiography is the diagnostic standard for valvular disease but is not always accessible [1]. Low-cost phonocardiograms (PCG)-based screening and subtyping is challenging, and broad multi-lesion labels increase confusion and label noise [2].
Purpose
To test whether a hierarchical, clinically aligned machine-learning framework improves patient-level screening and subtyping of left-sided regurgitation from short PCG recordings.
Methods
We analyzed 415 PCG recordings from 228 patients collected in August–September 2025 (52.2% male; median age 68 years, IQR 60.0–76.0). Heart sounds were recorded with a 15-second Eko digital stethoscope protocol. Features included 88 openSMILE acoustic eGeMAPS descriptors, 31 handcrafted PCG features (band energies, periodicity, spectral slope), and 7 baseline covariates. Echocardiography reports provided five-level severity labels (None to Severe) across valve-related fields, and the analysis focused on aortic (AR) and mitral regurgitation (MR) as they share a backflow mechanism and produce murmur patterns captured by PCG. Methodologically meaningful regurgitation was defined as severity ≥ mild. We defined four patient-level classes: no significant regurgitation, mixed AR+MR (both ≥ mild), predominant AR (only AR ≥ mild), and predominant MR (only MR ≥ mild). The hierarchy used Stage A to screen for any regurgitation versus none, then Stage B to assign subtype (mixed, predominant AR, predominant MR) among Stage A positives, yielding an end-to-end four-class prediction. We also evaluated a focused Stage C task that excluded mixed cases and classified predominant AR vs MR. Models used a stacked ensemble strategy combining tree- and kernel-based learners (e.g. gradient-boosted trees and support vector machines), with nested tuning and validation-based threshold optimization; stage-wise probabilities were propagated through the hierarchy to produce patient-level predictions.
Results
On the test set, Stage A screening performed best (AUC 0.76, F1 0.88). Subtype classification was lower (Stage B: AUC 0.78, F1 0.52), and the end-to-end four-class hierarchy was similar (AUC 0.765, F1 0.450). Excluding mixed cases improved the predominance task (AR vs MR) (Stage C: AUC 0.86, F1 0.73), suggesting mixed AR+MR drives ambiguity. Stage A interpretability (waveform/spectrogram/band-power examples) identified consistent patterns linked to regurgitation risk.
Conclusion
A clinically aligned hierarchical PCG approach was most reliable for screening clinically meaningful left-sided regurgitation, while subtype assignment remained difficult. Better AR-versus-MR performance after excluding mixed cases supports a screening-first workflow to guide echocardiography referral. Next steps include multi-center validation, improved subtype labeling with standardized multi-position recordings and quantitative echo linkage, and longitudinal monitoring with repeated PCGs.Figure1Figure2