DESIGN AND EXPLAINABILITY OF MACHINE LEARNING ALGORITHMS FOR THE CLASSIFICATION OF CARDIAC ABNORMALITIES FROM ELECTROCARDIOGRAM SIGNALS

The research activity contained in the present thesis work is devoted to the development of novel Machine Learning (ML) and Deep Learning (DL) algorithms for the classification of Cardiac Abnormalities (CA) from Electrocardiogram (ECG) signals, along with the explanation of classification outputs with explainable approaches. Automated computer programs for ECG classification have been developed since 1950s to improve the correct interpretation of the ECG, nowadays facilitating health care decision-making by reducing costs and human errors. The first ECG interpretation computer programs were essentially developed by emph{translating into the machine} the domain knowledge provided by expert physicians. However, in the last years leading research groups proposed to employ standard ML algorithms (which involve feature extraction, followed by classification), and more recently emph{end-to-end} DL algorithms to build automated ECG classification computer programs for the detection of CA. Recently, several research works proposed DL algorithms which even exceeded the performance of board-certified cardiologists in detecting a wide range of CA from ECGs. As a matter of fact, DL algorithms seem to represent promising tools for automated ECG classification on the analyzed datasets. However, the latest research related to ML and DL carries two main drawbacks that were tackled throughout the doctoral experience. First, to let the standard ML algorithms to perform at their best, the proper preprocessing, feature engineering, and classification algorithm (along with its parameters and hyperparameters) must be selected. Even when end-to-end DL approaches are adopted, and the feature extraction step is automatically learned from data, the optimal model architecture is crucial to get the best performance. To address this issue, we exploited the domain knowledge of electrocardiography to design an ensemble ML classification algorithm to classify within a wide range of 27 CA. Differently from other works in the context of ECG classification, which often borrowed ML and DL architectures from other domains, we designed each model in the ensemble according to the domain knowledge to specifically classify a subset of the considered CA that alter the same set of ECG physiological features known by physicians. Furthermore, in a subsequent work, toward the same aim we experimented three different Automated ML frameworks to automatically find the optimal ML pipeline in the case of standard and end-to-end DL algorithms. Second, while several research articles reported remarkable results for the value of ML and DL in classifying ECGs, only a handful offer insights into the model’s learning representation of the ECG for the respective task. Without explaining what these models are sensing on the ECG to perform their classifications in an explainable way, the developers of such algorithms run a strong risk of discouraging the physicians to adopt these tools, since they need to understand how ML and DL work before entrusting it to facilitate their clinical practice. Methods to open the emph{black-boxes} of ML and DL have been applied to the ECG in a few works, but they often provided only explanations restricted to a single ECG at time and with limited, or even absent, framing into the knowledge domain of electrocardiography. To tackle such issues, we developed techniques to unveil which portions of the ECG were the most relevant to the classification output of a ML algorithm, by computing average explanations over all the training samples, and translating them for the physicians' understanding. In a preliminary work, we relied on the Local Interpretable Model-agnostic Explanations (LIME) explainability algorithm to highlight which ECG leads were the most relevant in the classification of ST-Elevation Myocardial Infarction with a Random Forest classifier. Then, in a subsequent work, we extended the approach and we designed two model-specific explainability algorithms for Convolutional Neural Networks to explain which ECG waves, a concept understood by physicians, were the most relevant in the classification process of a wide set of 27 CA for a state-of-the-art CNN