Unsupervised Heart Abnormality Detection Based on Phonocardiogram Analysis with Beta Variational Auto-Encoders

Abstract

Heart Sound (also known as phonocardiogram (PCG)) analysis, is a popular way that detects cardiovascular diseases (CVDs). Most PCG analysis uses supervised way, which demands both normal and abnormal samples. This paper proposes a method of unsupervised PCG analysis that uses beta variational auto-encoder (β − VAE) to model the normal PCG signals. The best performed model reaches an AUC (Area Under Curve) value of 0.91 in ROC (Receiver Operating Characteristic) test for PCG signals collected from the same source. Unlike majority of β − VAEs that are used as generative models, the best-performed β − VAE has a β value smaller than 1. This fact demonstrates that the resampling process helps the improvements on anomaly PCG detection through reconstruction loss worth a heavier weight. Further investigations suggest that anomaly score based on reconstruction loss may be better than anomaly scores based on latent vectors of samples in PCG analysis based on VAE systems.

Unsupervised Phonocardiogram Analysis With Distribution Density Based Variational Auto-Encoders

“…As discussed by Higgins et al (12), in general β > 1 is necessary to achieve good disentanglement. However, as reported by Li et al (14), a smaller β value may help the performance of PCG analysis. As a result, this paper sets the β values to wider range: 0.01, 0.1, 1, 10, and 100 to test how the value of β effects the performance of proposed systems.…”

“…Then we test the performance of the proposed system when how the subsets are combined. The baseline system selected is a β-VAE based system (14), which follows the extract experiment design in this paper.…”

Using UMAP to Inspect Audio Data for Unsupervised Anomaly Detection under Domain-Shift Conditions

“…Second place was achieved by [15] (∼65%) with a simpler setup based on applying nonparametric inference methods (Local Outlier Factor (LOF) and KNN) to trained embeddings. We note that, while it was observed that Representation Learning (RepL)-based methods generally underperform reconstruction-based ones for UAD [16], this does not seem to be the tendency here: reconstruction objectives are barely present in the top ranks, possibly due to sensitivity to domain shifts, and the emphasis is on representations, e.g. the importance of spectrogram hyperparameters noted by [13] and the implications and effectiveness of different embeddings analyzed by [17] (3 rd place, ∼64.2%) which propose to use AdaCos [18].…”