Contribuciones de las técnicas machine learning a la cardiología. Predicción de reestenosis tras implante de stent coronario

Abstract

Background: There is currently no other hot topic comparable to the capacity of the current technology to develop similar capabilities as human beings, even in medicine.This ability to simulate the processes of human intelligence with computer systems is known as artificial intelligence. One of the fields of artificial intelligence with greater application today in medicine is that of prediction, recommendation or diagnosis, where machine learning techniques are applied. Furthermore, there is growing interest in precision medicine techniques as machine learning, which can deliver individually adapted medical care.Percutaneous coronary interventions (PCI) with stent implantation has become routine practice for the revascularization of coronary vessels with significant obstructive atherosclerotic disease. Furthermore, PCI is the gold-standard of treatment in patients with acute myocardial infarction; reducing the rates of death and recurrent ischemia as compared to medical treatment. The long-term success of this procedure may be limited by stent restenosis, a pathological process leading to recurrent arterial narrowing at the site of PCI. Identifying which patients will do restenosis is an important clinical challenge; since it can manifest itself as a new acute myocardial infarction or force a new revascularization of the affected vessel, and that in cases of recurrent restenosis represents an important therapeutic dilemma.Objectives: Starting with a review of artificial intelligence techniques applied to medicine and in greater depth, of machine learning techniques applied to cardiology, the main objective of this doctoral thesis has been to develop a machine learning model to predict restenosis in patients with acute myocardial infarction undergoing PCI with stent implantation. In addition, secondary objectives have been to compare the developed machine learning model with classical predictive clinical scores, and to develop software able to transfer this machine learning contribution to daily clinical practice. In order to develop an easily-applicable model, we performed our predictions without any additional variables than those obtained in daily practice. Material: The dataset, obtained from the GRACIA-3 trial, consisted of 263 patients with demographic, clinical and angiographic characteristics; 23 of them presented restenosis at 12-months after stent implantation. All software development has been made in Python and cloud computing has been used, specifically AWS (Amazon Web Services).Results: Our best performing model was developed with an extremely randomized trees classifier; which significantly outperformed (0.77; area under the ROC curve) three clinical scores; PRESTO-1 (0.58), PRESTO-2 (0.58), and TLR (0.62). Precision-Prefacio recall curves gave a more accurate picture of the extremely randomized trees model´s performance showing an efficient algorithm (0.96) for non-restenosis, returning high precision as well as high recall. For a threshold considered as optimal, out of 1,000 patients undergoing s...