Atrial fibrillation (AF) is the most common clinically significant arrhythmia, often severely disrupting cardiac hemodynamics and drastically increasing the risk of thromboembolic events. Around 90% of such intracardiac thrombus formation in AF patients takes place in the left atrial appendage (LAA). Such thrombus have been related to blood stasis, which at the moment, can be only assessed through noisy imaging data from transesophageal echocardiography (TEE) at one single point in space and time, vastly oversimplifying the characterization of the complex 4D nature of blood flow patterns. Alternatively, attempts have been made to relate LAA morphology to the risk of thrombi formation, some studies suggesting reduced risk of thrombosis on chicken-wing morphologies. Nonetheless, such classification of the LAA morphology has been found to be highly inconsistent and subjective, excluding as well, several fundamental morphological parameters such as the ostium size or the pulmonary vein (PV) orientation among others. More recently, computational fluid dynamics (CFD) have been employed on the left atrium (LA), seeking to assess the risk of thrombogenesis more quantitatively. CFD has proven to be an invaluable tool in establishing a mechanistic relation between patient-specific organ morphology and its characteristic hemodynamics. In fact, it has long been implemented in other human tissues, such as the coronary arteries, cerebral aneurysms and the aorta with unparalleled success, enabling early diagnosis and risk assessment of various cardiovascular diseases. Nevertheless, traditional CFD methods are renowned for their large memory requirements and long computing times, which severely hinders its suitability for time-sensitive clinical applications. Hence, this thesis seeks to harness the immense potential of deep learning (DL) by developing a deep neural network (DNN), with the objective of generating a fast and accurate surrogate of CFD, capable of instantaneously evaluating the risk of thrombus formation in the LAA. Already having revolutionized fields such as data processing, it has only recently begun to employ DNNs in high-dimensional, complex dynamical systems such as fluid dynamics. In fact to our knowledge, this study represents the first successful implementation of a DL surrogate of CFD analysis in a structure as complex as the LAA, which had only been previously attempted in the aorta. For that purpose, two DL architectures have been successfully designed and trained, which receive the specific LAA geometry as an input, and accurately predict its corresponding endothelial cell activation potential (ECAP) map, parameter linked to the risk of thrombosis. The first approach, is based on a simple fully-connected feedforward network, while the latter, also embeds unsupervised learning. An statistical shape model (SSM) of the LAA was created to generate the training dataset, encompassing 210 virtual shapes, on which CFD simulations were performed to attain the ground truth ECAP mappings. Once trained, the final D...

Detection and delineation are key steps for retrieving and structuring information of the electrocardiogram (ECG), being thus crucial for numerous tasks in clinical practice. Digital signal processing (DSP) algorithms are often considered state-of-the-art for this purpose but require laborious rule readaptation for adapting to unseen morphologies. This work explores the adaptation of the the U-Net, a deep learning (DL) network employed for image segmentation, to electrocardiographic data. The model was trained using PhysioNet’s QT database, a small dataset of 105 2-lead ambulatory recordings, while being independently tested for many architectural variations, comprising changes in the model’s capacity (depth, width) and inference strategy (single- and multi-lead) in a fivefold cross-validation manner. This work features several regularization techniques to alleviate data scarcity, such as semi-supervised pre-training with low-quality data labels, performing ECG-based data augmentation and applying in-built model regularizers. The best performing configuration reached precisions of 90.12%, 99.14% and 98.25% and recalls of 98.73%, 99.94% and 99.88% for the P, QRS and T waves, respectively, on par with DSP-based approaches. Despite being a data-hungry technique trained on a small dataset, a U-Net based approach demonstrates to be a viable alternative for this task.

Automatic detection and delineation of the electrocardiogram (ECG) is usually the first step for many feature extraction tasks. Although deep learning (DL) approaches have been proposed in the literature, those employ nonoptimal and outdated architectures. Thus, rule-based algorithms remain as state-of-the-art. Nevertheless, those may not generalize on other datasets and require difficult offline tuning for adapting to new scenarios. This work frames this task as a segmentation problem for using an adaptation of the U-Net architecture, a fully convolutional network. The detection performance shows a precision of 89.27%, 98.18% and 93.60% for the detection of the P, QRS and T waves, respectively, and a recall of 89.07%, 99.47% and 95.21%. This work shows promising results, outperforming existing DL approaches while being more generalizable than rule-based methods.

In order to determine the site of origin (SOO) in outflow tract ventricular arrhythmias (OTVAs) before an ablation procedure, several algorithms based on manual identification of electrocardiogram (ECG) features, have been developed. However, the reported accuracy decreases when tested with different datasets. Machine learning algorithms can automatize the process and improve generalization, but their performance is hampered by the lack of large enough OTVA databases. We propose the use of detailed electrophysiological simulations of OTVAs to train a machine learning classification model to predict the ventricular origin of the SOO of ectopic beats. We generated a synthetic database of 12-lead ECGs (2,496 signals) by running multiple simulations from the most typical OTVA SOO in 16 patient-specific geometries. Two types of input data were considered in the classification, raw and feature ECG signals. From the simulated raw 12-lead ECG, we analyzed the contribution of each lead in the predictions, keeping the best ones for the training process. For feature-based analysis, we used entropy-based methods to rank the obtained features. A cross-validation process was included to evaluate the machine learning model. Following, two clinical OTVA databases from different hospitals, including ECGs from 365 patients, were used as test-sets to assess the generalization of the proposed approach. The results show that V2 was the best lead for classification. Prediction of the SOO in OTVA, using both raw signals or features for classification, presented high accuracy values (>0.96). Generalization of the network trained on simulated data was good for both patient datasets (accuracy of 0.86 and 0.84, respectively) and presented better values than using exclusively real ECGs for classification (accuracy of 0.84 and 0.76 for each dataset). The use of simulated ECG data for training machine learning-based classification algorithms is critical to obtain good SOO predictions in OTVA compared to real data alone. The fast implementation and generalization of the proposed methodology may contribute towards its application to a clinical routine.

Obtaining per-beat information is a key task in the analysis of cardiac electrocardiograms (ECG), as many downstream diagnosis tasks are dependent on ECG-based measurements. Those measurements, however, are costly to produce and time-consuming to process in bulk, especially in recordings that change throughout long periods of time. Currently, ECG delineation is performed either using digital signal processing (DSP), which are able to produce high-quality delineations but are difficult to generalize, and machine learning (ML), which commonly produces increased performance at the cost of needing large databases of annotated data. However, the existing annotated databases for ECG delineation are small, being insufficient in size and in the array of pathological conditions they represent. This article delves into the latter with two main contributions. First, a pseudo-synthetic data generation scheme was developed, based in probabilistically composing unseen ECG traces given "pools" of fundamental segments cropped from the original databases and a set of rules for their arrangement into coherent synthetic traces. The generation of conditions is controlled by imposing expert knowledge on the generated trace, which increases the input variability for training the model. Second, two novel segmentation-based loss functions have been developed, which attempt at enforcing the prediction of an exact number of independent structures and at producing closer segmentation boundaries by focusing on a reduced number of samples. The best performing model obtained an F 1 -score of 99.38% and a delineation error of 2.19 ± 17.73 ms and 4.45 ± 18.32 ms for all wave's fiducials (onsets and offsets, respectively), as averaged across the P, QRS and T waves for three distinct freely available databases. The excellent results were obtained despite the heterogeneous characteristics of the tested databases, in terms of lead configurations (Holter, 12-lead), sampling frequencies (250, 500 and 2, 000 Hz) and represented pathophysiologies (e.g., different types of arrhythmias, sinus rhythm with structural heart disease), hinting at its generalization capabilities, while outperforming current state-of-the-art delineation approaches.