Abstract
Accurate irregularity identification in biological signals is essential for timely and reliable clinical monitoring. Traditional methods often rely on supervised learning and require large volumes of labeled dataset, which are rarely available in realistic clinical settings. In this study, we propose an self-supervised deep learning approach that leverages an autoencoder-based architecture and a novel probabilistic evaluation mechanism to detect and quantify abnormalities in biological dataset, such as cardiac rhythm. The approach captures patterns of normal behavior and flags deviations based on prediction deviation, while the irregularity importance is assessed using a hybrid evaluation system that considers both deviation magnitude and temporal irregularity rate. We evaluated the proposed method on a six-month dataset comprising over six million cardiac rhythm samples from 25 individuals, collected via wearable sensors. To enhance robustness, a combination of Isolation Forest, Kernel Density Estimation, and a sliding window technique was used during preprocessing to isolate high-confidence abnormalities. The approach’s effectiveness was validated through inter-subject cross-validation and further benchmarked using synthetic datasets with known irregularity rates. Results show high identification accuracy with minimal false positives, demonstrating strong generalization across individuals. The evaluation system provides interpretable insights by assigning context-aware importance scores, making the approach well-suited for real-time health monitoring applications. This approach supports early identification of critical biological changes, offering a promising pathway for intelligent and scalable clinical solutions.
Presenters
Maryam VatankhahStudent, Assistant Professor, City University of New York, New York, United States
Details
Presentation Type
Paper Presentation in a Themed Session
Theme
Interdisciplinary Health Sciences
KEYWORDS
Unsupervised Anomaly Detection, Signal Analysis, Deep Autoencoder Models, Wearable Devices
