Machine Learning for Multi-domain Data Integration and Phenotyping in Stroke Risk Assessment

By: Marta Saiz Vivó
Supervisor: Dr. Oscar Cámara & Dr. Gemma Piella

Date: June 19, 2026 - 10:30 h
Room: 55.309

Abstract

Ischemic stroke is a leading cause of global mortality and morbidity, resulting from critically reduced cerebral perfusion, most often due to thrombotic or embolic vascular occlusion. Multiple stroke etiologies exist, and clinical tools for assessing stroke recurrence risk across these etiologies remain limited. Atrial fibrillation (AF) is the main risk factor of cardioembolic stroke subtypes. In AF patients, thrombi most frequently form within the left atrium (LA), particularly the left atrial appendage (LAA). Current clinical risk scores, such as the CHA2DS2-VASc, estimate thromboembolic risk in AF based on demographic characteristics and comorbidities. However, growing evidence indicates that thrombus formation and stroke risk are governed by interactions across multiple patient domains, including clinical factors, LA morphology and hemodynamics. For AF patients with contraindications to oral anticoagulation, left atrial appendage occlusion (LAAO) has emerged as an alternative preventive strategy, although suboptimal device implantation may lead to device-related thrombus and increased stroke risk. The aim of this thesis is to gain deeper insights into ischemic stroke risk factors by analyzing multi-domain patient data and identifying data-driven phenotypes whose clinical relevance to ischemic events and thrombus formation is subsequently assessed. To this end, four objectives are followed: (i) the development of a computational pipeline for extracting explicit and implicit LA morphological descriptors with minimal manual intervention; (ii) the assessment of LAA morphology to identify clinically relevant features; (iii) the implementation of unsupervised learning techniques for multi-domain data integration to explore AF patient phenotypes; (iv) the extension of this framework to study stroke recurrence across heterogeneous stroke etiologies; and (v) the proposal of a gradient-based optimization framework for optimizing LAAO device configuration in patientspecific LA geometries to minimize post-procedural complications.

Multi-Scale Comparative Visual Analytics for Cardiovascular Hemodynamics

By: Jasna Nuhic
Supervisor: Dr. Oscar Cámara

Date: June 30, 2026 - 11:00 h
Room: 55.309

Abstract

Atrial fibrillation (AF) is a prevalent arrhythmia associated with increased thromboembolic stroke risk. In non-valvular AF, thrombi commonly originate in the left atrial appendage (LAA), motivating left atrial appendage occlusion (LAAO) for patients unsuitable for long-term anticoagulation. Preprocedural planning remains challenging because the LAA is anatomically variable and the hemodynamic consequences of alternative device configurations are difficult to anticipate from morphology alone. At the same time, broader cardiovascular research increasingly requires methods for comparing complex blood-flow patterns across patients, conditions, and analytical scenarios.

This thesis addresses that need through a comparative visual analytics framework for cardiovascular hemodynamics. It first situates the work through a state-of-the-art and bibliometric analysis of visual analytics in medicine, highlighting the growing methodological relevance of comparison, multimodal integration, and workflow-oriented analysis. Building on this foundation, the thesis develops comparison-ready data preparation and transformation pipelines for patient-specific LAAO simulations and introduces LAAOVis, a domain-specific visual analytics system for configuration-level comparison. LAAOVis combines synchronized views of 3D atrial anatomy, surface-mapped hemodynamic parameters, flow visualizations, statistical summaries, and case metadata with temporal standardization, vertex-consistent correspondence, in-application surface flattening, region-of-interest propagation, and integrated evidence capture and structured reporting.

The proposed framework is examined through controlled comparative case studies of device families, implantation positions, and modeling assumptions, supported by sensitivity analyses and comparison with available 4D flow MRI data, and its workflow is further assessed through expert evaluation. Beyond patient-specific intervention planning, the thesis demonstrates that the same comparative logic can be extended to exploratory cohort-level analysis through CoCluBFis, which combines pathline-derived descriptor abstraction, low-dimensional embedding, and coordinated detail-on-demand views. A cohort case study further shows how localized group structure and source- and component-specific descriptor differences can be related back to interpretable case-level flow behavior. Overall, the thesis shows how comparative visual analytics can strengthen the interpretability, traceability, and reproducibility of cardiovascular hemodynamic analysis across both patient-specific and cohort scales.

From ECG Digitization to Intracardiac Signal Modeling: Machine Learning Approaches for Electrophysiological Procedures

By: Álvaro José Bocanegra Pérez
Supervisor: Dr. Oscar Cámara & Dr. Gemma Piella

Date: July 3, 2026 - 10:30 h
Room: 55.309

Abstract

This thesis presents a comprehensive study on the application of artificial intelligence techniques to support the technification of electrophysiological procedures. The work addresses multiple stages of signal acquisition and interpretation, from surface electrocardiogram (ECG) digitization to intracardiac electrogram (EGM) analysis, with the aim of improving data usability, extracting clinically relevant information, and assisting decision-making during clinical workflows.

First, a robust ECG digitization tool is developed to enable the use of heterogeneous and legacy ECG data, demonstrating its applicability to real-world scenarios with significant template variability. The digitized signals are then used in a clinically motivated study of ST-segment elevation, showing consistency with established clinical patterns.

Second, machine learning models combining ECG features and patient data are evaluated for the classification of outflow tract ventricular arrhythmias. The results show that this approach achieves performance comparable to rule-based and existing machine learning methods, allowing reliable identification of the ventricle of origin, although not precise anatomical localization.

Finally, EGM-based analyses are explored to support the identification of clinically relevant ablation targets. While EGMs provide valuable information for locating the site of origin, the results highlight limitations in precision and the need for improved annotations and methodologies that account for multiple signals and their temporal evolution, better reflecting clinical practice.

Overall, the findings demonstrate that artificial intelligence methods can capture clinically meaningful patterns from electrophysiological signals, while also revealing the challenges that must be addressed to achieve robust and clinically applicable solutions.