Real-time AI-based signal quality control enhances signal reliability and diagnostic stability in multi-site digital auscultation

Background Cardiac auscultation is crucial for cardiovascular screening but suffers from high dependence on technique, contact, motion artifact, and noise1. Signal instability often leads to repeat recordings, variable interpretation, and delayed referrals, reducing workflow efficiency. Our optical phonocardiogram platform2 allows multi-site acquisition, but standardized real-time quality assurance is vital for consistent signal capture, especially by non-physician healthcare professionals. Purpose To evaluate a real-time artificial intelligence-based quality assurance/quality control (QA/QC) system and assess its impact on variability in PCG acquisition by non-physician HCPs. We examined whether automated QC supports scalable, standardized data capture. Methods Expert-annotated and publicly available phonocardiogram datasets, harmonized to a Good/Poor quality definition, established ground-truth signal quality labels. Standardized preprocessing and predefined acoustic features (stability, spectral organization, entropy, envelope, periodicity) were extracted to train a gradient-boosted (XGBoost) classifier, evaluated using standard accuracy and agreement metrics, to separate good and poor recordings. The final QA/QC model was deployed for real-time assessment and validated on 3,655 dual-site recordings from seven anatomical sites (carotid arteries, suprasternal notch, mitral, tricuspid, pulmonary, and aortic regions), acquired by a novice. Signal quality classification was then analyzed for its association with diagnostic confidence and transthoracic echocardiography (TTE) findings. Results The XGBoost QA/QC classifier showed strong performance on validation data (accuracy 0.88, F1 0.84, ROC-AUC 0.95), consistently classifying good and poor recordings. In a prospective evaluation of 3,655 dual-site recordings across seven regions, 94% were good quality and 6% were poor quality. Poor quality recordings were associated with lower diagnostic confidence (0.74 vs 0.94) and exhibited greater spectral irregularity and reduced periodic stability. Subgroup analysis confirmed stable QA/QC performance across all seven anatomical regions and demographic groups. A higher proportion of poor quality recordings was observed in subjects with TTE-confirmed valvular heart disease (9.1% vs 3.1%), though the impact on downstream interpretation is yet to be determined. The model's consistent performance supports its reliability for real-time quality control during multi-site acquisition by non-expert operators. Conclusion The QA/QC model reliably assessed real-time cardiovascular acoustic recording quality across different anatomies, ages, sexes, and VHD statuses. Automated quality control facilitated a high proportion of usable recordings, supporting standardized, large-scale multi-site cardiovascular signal acquisition for research and clinical use. Independent validation is needed to confirm generalizability and robustness.