Decoding heart signals through the ear: machine learning methods for arrhythmia detection and monitoring
Author(s)
Occhipinti, Edoardo
Type
Thesis
Abstract
Timely detection and continuous monitoring of heart conditions, such as arrhythmias, is essential for minimizing the risk of severe complications like stroke and heart failure. 12-Lead electrocardiogram (ECG) is the gold standard technique to assess the electrical activity of the heart. Despite significant advancements in wearable health technologies, current alternative solutions such as wrist-worn, chest-straps or hand-held devices show limitations in diagnostic accuracy, comfort and long-term monitoring ability.
Recent studies reveal that the cardiac dipole propagates from the chest to the head, such that an ECG signal can be recorded by measuring the potential difference between the two ear canals. In-ear ECG offers an alternative continuous cardiovascular monitoring solution merging user-convenience and functionality.
This thesis addresses the challenges of in-ear ECG associated with low signal-to-noise ratio and interference from noncardiac physiological signals, mainly EEG, which have limited its clinical applications so far.
The PhD work reported in this thesis can be summarised in three main contributions. First, a signal denoising model is introduced, leveraging a convolutional autoencoder to significantly improve the SNR of in-ear ECG signals, while preserving subject-specific morphological waveforms and enabling downstream tasks such as R-peak detection.
Second, heart rate variability (HRV) features extracted from denoised in-ear ECG signals are used to explore the relationship between autonomic nervous system and physiological states, showcasing its application in the classification of breathing rates.
Third, the thesis develops a robust framework for atrial fibrillation (AFIB) and atrial flutter (AFL) detection through a clinical study. The proposed machine learning framework achieves performance comparable to gold-standard ECG systems. The feasibility of transfer learning, along with AI model interpretability, highlight the effectiveness of in-ear ECG in a clinical setting for diagnosis of AFIB/AFL patients.
Potential future research avenues are explored to offer insights into the future applications of in-ear ECG.
Recent studies reveal that the cardiac dipole propagates from the chest to the head, such that an ECG signal can be recorded by measuring the potential difference between the two ear canals. In-ear ECG offers an alternative continuous cardiovascular monitoring solution merging user-convenience and functionality.
This thesis addresses the challenges of in-ear ECG associated with low signal-to-noise ratio and interference from noncardiac physiological signals, mainly EEG, which have limited its clinical applications so far.
The PhD work reported in this thesis can be summarised in three main contributions. First, a signal denoising model is introduced, leveraging a convolutional autoencoder to significantly improve the SNR of in-ear ECG signals, while preserving subject-specific morphological waveforms and enabling downstream tasks such as R-peak detection.
Second, heart rate variability (HRV) features extracted from denoised in-ear ECG signals are used to explore the relationship between autonomic nervous system and physiological states, showcasing its application in the classification of breathing rates.
Third, the thesis develops a robust framework for atrial fibrillation (AFIB) and atrial flutter (AFL) detection through a clinical study. The proposed machine learning framework achieves performance comparable to gold-standard ECG systems. The feasibility of transfer learning, along with AI model interpretability, highlight the effectiveness of in-ear ECG in a clinical setting for diagnosis of AFIB/AFL patients.
Potential future research avenues are explored to offer insights into the future applications of in-ear ECG.
Version
Open Access
Date Issued
2025-12-20
Date Awarded
01/01/2026
License URL
Advisor
Mandic, Danilo P.
Peters, Nicholas S.
Sponsor
UK Research and Innovation
Grant Number
EP/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)