Personalized and Zero-Shot Electrocardiogram Arrhythmia Monitoring System
Duman, Mehmet Mert (2024)
2024
Tietojenkäsittelyopin maisteriohjelma - Master's Programme in Computer Science
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2024-12-17
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024120510828
https://urn.fi/URN:NBN:fi:tuni-2024120510828
Tiivistelmä
Electrocardiograms (ECG) are a non-intrusive and highly accurate tool for monitoring cardiac abnormalities. They are widely used in clinical settings and can be measured on wearable devices such as smartwatches and Holter monitors. Early detection of cardiac arrhythmias through continuous ECG monitoring has the potential to reduce mortality rates and improve lives. This thesis presents a comprehensive and efficient pipeline for personalized and zero-shot ECG arrhythmia detection suitable for low-end wearable devices. Earlier methods for arrhythmia detection require labeled healthy and anomalous heartbeats for training; however, such data is often unavailable, particularly for healthy individuals with no history of cardiac disorders.
This thesis addresses a real-world scenario where a healthy individual requires continuous monitoring for early detection and prevention of potential cardiac issues. First, we propose a lightweight representation error-based arrhythmia detection method using the nullspaces of the learned sparse representation dictionaries. Our lightweight method reduces computational complexity while maintaining detection accuracy. Then, we improve representation error-based detection performance via compact dictionaries learned through single hidden-layer sparse autoencoders. Then, we propose a sparse representation-based domain adaptation technique to learn morphology transformation matrices for the target individual, which are used to adapt the morphology of existing ECG data from other patients to the domain of the target. This enables a zero-shot classifier to be trained for a healthy individual without any anomalous beats. Then, we design an ensemble classifier by combining the lightweight representation error-based classifier with the zero-shot classifier and achieve an accuracy of 98.2% and an F1-Score of 92.8%. Finally, we propose an energy-efficient monitoring system for mobile devices capable of automatically classifying up to 40% of the test samples with minimal computation, making continuous monitoring feasible for wearable sensors.
This thesis addresses a real-world scenario where a healthy individual requires continuous monitoring for early detection and prevention of potential cardiac issues. First, we propose a lightweight representation error-based arrhythmia detection method using the nullspaces of the learned sparse representation dictionaries. Our lightweight method reduces computational complexity while maintaining detection accuracy. Then, we improve representation error-based detection performance via compact dictionaries learned through single hidden-layer sparse autoencoders. Then, we propose a sparse representation-based domain adaptation technique to learn morphology transformation matrices for the target individual, which are used to adapt the morphology of existing ECG data from other patients to the domain of the target. This enables a zero-shot classifier to be trained for a healthy individual without any anomalous beats. Then, we design an ensemble classifier by combining the lightweight representation error-based classifier with the zero-shot classifier and achieve an accuracy of 98.2% and an F1-Score of 92.8%. Finally, we propose an energy-efficient monitoring system for mobile devices capable of automatically classifying up to 40% of the test samples with minimal computation, making continuous monitoring feasible for wearable sensors.