Interpretable Time-Series Anomaly Detection using Micro-β VAEs

Accurate diagnosis of heart diseases relies on determining any discrepancies in time-series data of electrocardiogram (ECG) signals. Though Variational Autoencoders (VAEs) deliver powerful probabilistic modeling tools for anomaly detection, their accuracy drops significantly when compared to deterministic models on clean, highly aligned datasets, which is caused by posterior collapse. This paper explores and theorizes the practical limits of posterior collapse and the effectiveness of Autoencoders against a Micro-β VAE architecture (β ∈ {0.001,0.1}) combined with a Kullback-Leibler (KL) annealing schedule over ECG signals. By integrating a feature-level attention mechanism, we improve clinical interpretability. Our approach improves on the currently available deterministic Autoencoders and achieves a mean F1-score of 0.9735 when it is evaluated over five independent initializations on the ECG5000 dataset. This demonstrates increased robustness to increasing synthetic Gaussian noise. Anomalous time steps are localized by attention weights despite the stochastic nature of the latent space.