Semantic Anomaly Detection in Medical Time Series

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Signal Processing

Background:

• Accurate detection of anomalies in physiological signals like ECG is crucial for medical diagnosis.
• Existing outlier recognition methods often require labeled data or struggle with complex time series patterns.

Purpose of the Study:

• To develop and assess a new unsupervised deep learning approach for identifying anomalous intervals in time series data.
• To evaluate the efficacy of this method on real-world electrocardiogram (ECG) signals.

Main Methods:

• Utilized denoising autoencoders with recurrent neural networks (RNNs) and gated recurrent units (GRUs) for signal encoding.
• Employed cluster analysis in the code space, specifically cluster-based similarity partitioning combined with DBSCAN, for anomaly detection.
• Applied the developed system to expert-labeled real-world ECG data.

Main Results:

• The unsupervised deep learning approach achieved an adjusted Rand index of 0.11 on ECG data.
• The system demonstrated a precision and recall of approximately 0.72 for anomaly detection.
• Outperformed several established state-of-the-art outlier recognition methods.

Conclusions:

• The proposed general unsupervised deep learning method is effective for differentiating normal and anomalous time series intervals.
• This approach shows promise for various medical time series analyses and can serve as a foundation for specialized detectors.
• The method offers a robust alternative for unsupervised or semi-supervised anomaly detection in complex data.