Evaluation of unsupervised machine learning as a support in pre-hospital electrocardiography

Cardiovascular diseases (CVDs) remain the leading cause of disease burden globally, contributing to premature mortality, disability, and rising health care costs. In this spectrum of cardiac conditions, myocardial ischemia and acute coronary syndrome (ACS) are particularly notable, which are joined by almost the same causes, with the atherosclerosis being the most prevalent. The latter has as main consequence the intraluminal thrombi formation, which occludes the blood vessel resulting in acute myocardial ischemia. When occurring in the coronary circulation, it manifests as ACS. Prompt decisions and accurate diagnosis of these conditions, already in the pre-hospital phase, are crucial for preserving cardiac function as much as possible. In the emergency department (ED), electrocardiography (ECG) is the most widely used initial diagnostic test to support clinical diagnosis and aid in risk stratification, as recommended by the current clinical guidelines for screening patients presented with chest pain and anginal equivalents. The standard 12-lead ECG is the criterion standard for electrocardiographic detection of acute myocardial ischemia/injury and is reported to be the single most important method to rapidly identify ACS in the ED. Nevertheless, due to the unpredictable and dynamic nature of acute ischemic changes, a single ECG snapshot may be inadequate. Thus, it is recommended a comparison of the acute ECG, which is under suspicion, to a previously recorded non-acute ECG of the same patient, to account for interindividual variability. With the improvement in pattern recognition abilities of artificial intelligence and the urgency of large-scale data management, the unsupervised machine learning (ML) can be employed for quickly discovering data distributions and relevant trends, learning the underlying structures, facilitating diagnosis-making procedures, without any predefined labels. The analysis focuses on ten clustering methods, which are applied to data from SUBSTRACT study, which are characterized by class imbalance. More specifically, the dual purpose of this thesis consists of: (1) demonstrating the effectiveness of the clustering techniques in revealing the underlying structure of data by correctly distinguishing pathological cases from healthy individuals; (2) evaluating the role of both ECG features set, 18 direct measurements and 28 serial features, on the performance of those algorithms and their ability to support accurate diagnosis of myocardial ischemia and ACS. First, the comparative analysis of clustering methods, assessed through evaluation metrics, reveals that the implemented techniques perform well with both databases, especially with myocardial ischemia database, achieving high accuracy scores exceeding 70,33%. Notably, the CLARA methodology results in the most effective clustering method across both conditions and feature sets; with accuracy and F1 score consistently above 74,65% and 69,75%, respectively. Secondly, in ACS database the serial features are found to be less relevant for the diseases detection, differently from the myocardial ischemia, where some algorithms benefit from their inclusion. These findings underscore the importance of database characteristics, within parameters choice, in influencing the algorithm performance. In summary, integrating advanced ML methods into routine diagnostic processes for ECG recordings in pre-hospital care has the potential to support healthcare professionals in quickly analyzing medical reports and making diagnoses, thereby reducing the risk of health complications and ultimately improving patient outcomes.