Development and comparison of ML and DL models for ECG anomaly classification in wearable devices

The increasing demand for low-cost, non-invasive systems capable of monitoring electrocardiogram (ECG) traces in non-hospital settings has consequently led to a rapid evolution of intelligent medical decision support systems for the automatic analysis of ECGs. These systems detect and classify cardiac anomalies early, reducing healthcare professionals' workload and improving diagnostic capabilities. This thesis focuses on developing and comparing two anomaly classification models in the ECG medical context using Machine Learning (ML) and Deep Learning (DL). Both models have been used on signals acquired from the CGM Hi 3 Leads ECG (Hi-ECG), a wearable ECG device commercialised by CGM and co-developed with STMicroelectronics. This device allows the acquisition, recording, and transmission of three ECG channels and other physiological signals sent via Bluetooth to an external device. A key component of this thesis is a study carried out in two residential care facilities (RCFs) in Piedmont, Italy, where the Hi-ECG has been used to acquire signals from a diverse group of volunteer subjects. This effort has resulted in the creation of an ex novo database called HiDB, characterised by ECG signals acquired exclusively with the Hi-ECG device and various clinical and non-clinical information, such as subject demographics and health profiles. The database comprises a variety of clinical cases, which will be labelled by expert clinicians in a future project. However, to provide an initial classification framework, the database has been provisionally labelled by non-experts into three general classes. This thesis work is divided into three phases. The first phase uses signals from HiDB and two online databases with manually labelled ECG waves to train a robust LSTM model for segmenting ECG signals into beats and identifying characteristic waves. The second phase verifies the clinical validity of both AI approaches using the MIT-BIH Arrhythmia Database, which contains manual annotations of the five Association for the Advancement of Medical Instrumentation (AAMI)-defined arrhythmia superclasses. Both approaches involve signal processing, with segmentation performed using the pre-trained LSTM model. For the ML approach, the segmented ECG beats were manually analysed to extract and select features from various domains, which were then used to train a Support Vector Machine (SVM) model. Regarding the DL approach, the segmented ECG beats were directly fed into an Autoencoder (AE) model, allowing for unsupervised feature extraction from each heartbeat. Then, the pre-trained Encoder is fine-tuned with a feedforward neural network for classification. This approach is well-suited for situations with limited labelled data, which is often challenging to acquire, particularly with wearable ECG devices. The final phase of the work involves using HiDB signals to train and test both ML and DL models, with a slightly different approach due to the variation in the number and categories of classes involved in the two databases. The methods proposed in this work have been chosen with a view to the future implementation of these AI algorithms within the Hi-ECG wearable device. A significant innovation of this study has been the creation of a comprehensive database ex novo, specifically for this device. The results have demonstrated these methods’ effectiveness in automatically detecting cardiac anomalies, highlighting the need for AI implementations to improve the timeliness and accuracy of diagnosis.