Strictly Personalized Zero-Shot ECG Arrhythmia Detection Using Sparse Geometry-Regularized Autoencoders
Haque, Asma (2026)
2026
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-06-30
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606298037
https://urn.fi/URN:NBN:fi:tuni-202606298037
Tiivistelmä
Cardiovascular diseases remain the leading cause of mortality worldwide, and cardiac arrhythmias represent a major clinical concern due to their potential association with stroke, heart failure, and sudden cardiac death. While deep learning has achieved remarkable success in ECG arrhythmia classification, most existing approaches rely on large labelled datasets containing both normal and abnormal heartbeats. This assumption is often unrealistic in personalized monitoring scenarios, where a new user may provide only a short period of healthy ECG activity and no examples of future arrhythmias.
This thesis investigates a strictly personalized zero-shot ECG anomaly detection framework in which only the first five minutes of healthy ECG data from each patient are used for model development. No abnormal beats are used during training, validation, threshold calibration, or model selection. The problem is formulated as reconstruction-based anomaly detection, where abnormalities are identified through deviations from a learned representation of patient-specific healthy cardiac activity.
Three reconstruction-based models are evaluated under the same experimental protocol: a standard Autoencoder (AE), a Variational Autoencoder (VAE), and a proposed Sparse Geometry Autoencoder. The proposed model incorporates latent sparsity and geometry-preserving regularization to encourage compact and structured representations of healthy ECG morphology. A comprehensive ablation study investigates the effects of latent dimensionality, masking strategies, weight tying, sparsity regularization, geometry regularization, and threshold calibration.
Experiments are conducted on the MIT-BIH Arrhythmia Database using patient-specific training and evaluation. Performance is assessed using accumulated confusion matrices and a macro-style F1-score to account for class imbalance. The proposed Sparse Geometry Autoencoder achieves the strongest overall performance, obtaining a macro-style F1-score of 0.8399, an accuracy of 0.9265, a precision of 0.7280, a recall of 0.7162, and a specificity of 0.9588. The results demonstrate that meaningful arrhythmia detection can be achieved using only a short healthy baseline and without requiring abnormal examples during model development.
The findings indicate that compact latent representations combined with sparsity and geometry-preserving regularization improve reconstruction-based anomaly detection under strict personalized zero-shot constraints. More broadly, the study demonstrates the potential of patient-specific representation learning as a foundation for future personalized cardiac monitoring systems.
This thesis investigates a strictly personalized zero-shot ECG anomaly detection framework in which only the first five minutes of healthy ECG data from each patient are used for model development. No abnormal beats are used during training, validation, threshold calibration, or model selection. The problem is formulated as reconstruction-based anomaly detection, where abnormalities are identified through deviations from a learned representation of patient-specific healthy cardiac activity.
Three reconstruction-based models are evaluated under the same experimental protocol: a standard Autoencoder (AE), a Variational Autoencoder (VAE), and a proposed Sparse Geometry Autoencoder. The proposed model incorporates latent sparsity and geometry-preserving regularization to encourage compact and structured representations of healthy ECG morphology. A comprehensive ablation study investigates the effects of latent dimensionality, masking strategies, weight tying, sparsity regularization, geometry regularization, and threshold calibration.
Experiments are conducted on the MIT-BIH Arrhythmia Database using patient-specific training and evaluation. Performance is assessed using accumulated confusion matrices and a macro-style F1-score to account for class imbalance. The proposed Sparse Geometry Autoencoder achieves the strongest overall performance, obtaining a macro-style F1-score of 0.8399, an accuracy of 0.9265, a precision of 0.7280, a recall of 0.7162, and a specificity of 0.9588. The results demonstrate that meaningful arrhythmia detection can be achieved using only a short healthy baseline and without requiring abnormal examples during model development.
The findings indicate that compact latent representations combined with sparsity and geometry-preserving regularization improve reconstruction-based anomaly detection under strict personalized zero-shot constraints. More broadly, the study demonstrates the potential of patient-specific representation learning as a foundation for future personalized cardiac monitoring systems.