Robust Anomaly Detection of Adventitious Auscultation Signals using Bayesian Belief Tracking

Stethoscope screening serves as a primary method for diagnosing pulmonary infections, with medical professionals actively listening for signs of pathologies in breathing sounds like wheezing and crackling, which carry different clinical interpretations. Environmental conditions during auscultation recordings often share similarities with these abnormal lung sounds, and can mask or confound their presence making their detection highly sensitive to surrounding factors. To automate this process, a robust anomaly detection scheme with resilience to ambient backgrounds and high precision is essential. In this study, we propose an unsupervised framework for anomaly detection where statistics of a deep neural network embeddings are tracked using a Bayesian belief model in order to flag variations that are deemed anomalous, hence facilitating detection of adventitious auscultation events. The proposed scheme leverages two key principles: (1) learning of statistics of normal auscultation patterns using variational constraints, and (2) tracking changes in the statistics using Bayesian beliefs that interpret anomalies as deviations from normal statistics. This approach is shown to be very effective in detecting adventitious auscultations under various noise levels hence ensuring its resilience to environmental conditions.