Development and validation of machine learning models in cardiology

This thesis focuses on various techniques for training and validating machine learning models in the field of cardiology, specifically for aortic valve replacement, coronary artery disease, atrial fibrillation, and Phospholamban gene mutation. While prediction models are commonly used in medical decision-making, it is important to ensure accurate evaluation to avoid overestimating their effectiveness. This thesis discusses different methods for training and evaluating outcome prediction models, diagnostic and decision-support models, and the challenges faced in the process. A line of studies on this thesis focuses on TAVI mortality prediction. Initially, a model using data from a single center was developed. In addition, finetuning, distributed learning and model combination techniques were applied to use data from two medical centers without data being shared. Finally, it was also performed a temporal validation using data from a national registry, assessing the performance and stability of models trained repeatedly over time compared to models trained once using the oldest available data. Regarding coronary artery disease, we proposed the prediction of insufficient contrast enhancement on Coronary CT Angiography including test-bolus variables, which are not commonly used to adjust contrast delivery protocols. For atrial fibrillation, a model to predict recurrence following thoracoscopic surgery and a subgroup analysis was performed. Finally, a diagnostic support model for identifying patients with Phospholamban gene mutation based on ECGs was proposed using a pre-trained model to improve the model’s accuracy.