Machine learning for the classification of atrial fibrillation utilizing seismo- and gyrocardiogram
Mehrang, Saeed (2023-12-12)
Machine learning for the classification of atrial fibrillation utilizing seismo- and gyrocardiogram
(12.12.2023)
Turun yliopisto
Julkaisun pysyvä osoite on:
https://urn.fi/URN:ISBN:978-951-29-9543-1
https://urn.fi/URN:ISBN:978-951-29-9543-1
Tiivistelmä
A significant number of deaths worldwide are attributed to cardiovascular diseases (CVDs), accounting for approximately one-third of the total mortality in 2019, with an estimated 18 million deaths. The prevalence of CVDs has risen due to the increasing elderly population and improved life expectancy. Consequently, there is an escalating demand for higher-quality healthcare services. Technological advancements, particularly the use of wearable devices for remote patient monitoring, have significantly improved the diagnosis, treatment, and monitoring of CVDs.
Atrial fibrillation (AFib), an arrhythmia associated with severe complications and potential fatality, necessitates prolonged monitoring of heart activity for accurate diagnosis and severity assessment. Remote heart monitoring, facilitated by ECG Holter monitors, has become a popular approach in many cardiology clinics. However, in the absence of an ECG Holter monitor, other remote and widely available technologies can prove valuable. The seismo- and gyrocardiogram signals (SCG and GCG) provide information about the mechanical function of the heart, enabling AFib monitoring within or outside clinical settings. SCG and GCG signals can be conveniently recorded using smartphones, which are affordable and ubiquitous in most countries.
This doctoral thesis investigates the utilization of signal processing, feature engineering, and supervised machine learning techniques to classify AFib using short SCG and GCG measurements captured by smartphones. Multiple machine learning pipelines are examined, each designed to address specific objectives. The first objective (O1) involves evaluating the performance of supervised machine learning classifiers in detecting AFib using measurements conducted by physicians in a clinical setting. The second objective (O2) is similar to O1, but this time utilizing measurements taken by patients themselves. The third objective (03) explores the performance of machine learning classifiers in detecting acute decompensated heart failure (ADHF) using the same measurements as O1, which were primarily collected for AFib detection. Lastly, the fourth objective (O4) delves into the application of deep neural networks for automated feature learning and classification of AFib.
These investigations have shown that AFib detection is achievable by capturing a joint SCG and GCG recording and applying machine learning methods, yielding satisfactory performance outcomes. The primary focus of the examined approaches encompassed (1) feature engineering coupled with supervised classification, and (2) iv automated end-to-end feature learning and classification using deep convolutionalrecurrent neural networks.
The key finding from these studies is that SCG and GCG signals reliably capture the heart’s beating pattern, irrespective of the operator. This allows for the detection of irregular rhythm patterns, making this technology suitable for monitoring AFib episodes outside of hospital settings as a remote monitoring solution for individuals suspected to have AFib. This thesis demonstrates the potential of smartphone-based AFib detection using built-in inertial sensors. Notably, a short recording duration of 10 to 60 seconds yields clinically relevant results. However, it is important to recognize that the results for ADHF did not match the state-of-the-art achievements due to the limited availability of ADHF data combined with arrhythmias as well as the lack of a cardiopulmonary exercise test in the measurement setting.
Finally, it is important to recognize that SCG and GCG are not intended to replace clinical ECG measurements or long-term ambulatory Holter ECG recordings. Instead, within the scope of our current understanding, they should be regarded as complementary and supplementary technologies for cardiovascular monitoring.
Atrial fibrillation (AFib), an arrhythmia associated with severe complications and potential fatality, necessitates prolonged monitoring of heart activity for accurate diagnosis and severity assessment. Remote heart monitoring, facilitated by ECG Holter monitors, has become a popular approach in many cardiology clinics. However, in the absence of an ECG Holter monitor, other remote and widely available technologies can prove valuable. The seismo- and gyrocardiogram signals (SCG and GCG) provide information about the mechanical function of the heart, enabling AFib monitoring within or outside clinical settings. SCG and GCG signals can be conveniently recorded using smartphones, which are affordable and ubiquitous in most countries.
This doctoral thesis investigates the utilization of signal processing, feature engineering, and supervised machine learning techniques to classify AFib using short SCG and GCG measurements captured by smartphones. Multiple machine learning pipelines are examined, each designed to address specific objectives. The first objective (O1) involves evaluating the performance of supervised machine learning classifiers in detecting AFib using measurements conducted by physicians in a clinical setting. The second objective (O2) is similar to O1, but this time utilizing measurements taken by patients themselves. The third objective (03) explores the performance of machine learning classifiers in detecting acute decompensated heart failure (ADHF) using the same measurements as O1, which were primarily collected for AFib detection. Lastly, the fourth objective (O4) delves into the application of deep neural networks for automated feature learning and classification of AFib.
These investigations have shown that AFib detection is achievable by capturing a joint SCG and GCG recording and applying machine learning methods, yielding satisfactory performance outcomes. The primary focus of the examined approaches encompassed (1) feature engineering coupled with supervised classification, and (2) iv automated end-to-end feature learning and classification using deep convolutionalrecurrent neural networks.
The key finding from these studies is that SCG and GCG signals reliably capture the heart’s beating pattern, irrespective of the operator. This allows for the detection of irregular rhythm patterns, making this technology suitable for monitoring AFib episodes outside of hospital settings as a remote monitoring solution for individuals suspected to have AFib. This thesis demonstrates the potential of smartphone-based AFib detection using built-in inertial sensors. Notably, a short recording duration of 10 to 60 seconds yields clinically relevant results. However, it is important to recognize that the results for ADHF did not match the state-of-the-art achievements due to the limited availability of ADHF data combined with arrhythmias as well as the lack of a cardiopulmonary exercise test in the measurement setting.
Finally, it is important to recognize that SCG and GCG are not intended to replace clinical ECG measurements or long-term ambulatory Holter ECG recordings. Instead, within the scope of our current understanding, they should be regarded as complementary and supplementary technologies for cardiovascular monitoring.
Kokoelmat
- Väitöskirjat [2894]