Artificial Intelligence Models to Predict Clinically Relevant Cardiovascular Outcomes

Overview

Atrial fibrillation (AF) is a frequent and clinically relevant problem among the events that may occur during the hospitalization period in patients with cardiovascular disease. AF, indeed, is a determinant or aggravating condition of serious adverse events, such as myocardial infarction, heart failure, and thromboembolic stroke. The occurrence of AF in hospitalized patients, such as those admitted for coronary intervention, results in prolonged length of hospitalization, increased likelihood of discharge on anticoagulants, and increased 30-day risk of bleeding. It is noteworthy that while the incidence of AF in the general population is about 1-2 cases per 1000 people per year, this is much higher in patients hospitalized for acute myocardial infarction (AMI) (about 10% over the hospitalization period) or in patients undergoing coronary artery bypass grafting (CABG) (about 25% over the hospitalization period). Thus, identifying patients at high risk of AF during the hospitalization period could allow experimental testing of the efficacy and safety of preventive interventions (e.g., tailored anesthetic or surgical approaches, drug-prevention, etc.). It can be hypothesized that the clinical and nonclinical variables useful in estimating the risk of AF will change depending on the type of patients and that the identification and integration of these variables will require more complex predictive analysis systems than the regression models classically used to develop risk scores. On the other hand, the risk of recurrence of coronary events throughout the first years after CABG remains high (about 20% at 5 years) despite effective revascularization and early secondary prevention.Although some scores have been developed for estimating the risk of coronary event recurrence in secondary prevention using multivariate regression models, these algorithms consider a limited number of predictors, do not take into account possible interactions between different factors, and their actual predictive ability is not reported in the literature. With advances in Artificial Intelligence (AI) technology together with the rapid development of digital clinical datasets, machine learning has the potential to analyze substantial amounts of data and recognize patterns to predict AF onset and recurrence of coronary events within a defined time horizon (e.g., in-hospital event) in selected populations in a way that improves the predictive ability of conventional methods.

PerCard is a retrospective and prospective observational study. The study aims to develop and validate models for prediction of intrahospital AF and recurrence of coronary events in a long-term follow-up using Artificial Intelligence. The development and internal validation of predictive models of AF involve two retrospective cohorts: * Cohort A: 1258 patients underwent CABG at Centro Cardiologico Monzino (CCM) between 2002 and 2016 * Cohort B: 2445 patients admitted for AMI STEMI or NSTEMI to CCM between 2010 and 2018 The development and internal validation of predictive models of coronary event recurrence in long-term follow-up involve a third retrospective cohort: -Cohort C: 1248 patients underwent CABG at CCM between 2002 and 2014 . External validation of the predictive models of in-hospital AF involves a cohort of patients admitted with AMI STEMI or NSTEMI, who will be prospectively enrolled at Coronary Intensive Care Unit of Centro Cardiologico Monzino. In the different prediction models, clinical and instrumental variables specific to patients with AMI (e.g., infarcted area), variables that are common to patients with any form of coronary revascularization (e.g., how many and which coronary vessels have been revascularized), or variables that are common to patients and individuals without established coronary artery disease (e.g., age, sex, history of hypertension, particular gene polymorphisms related to AF, signals from the ECG, etc.) will be included, where available. In addition, the contribution of 16 gene polymorphisms associated with predisposition to intrahospital onset of AF has been previously evaluated in cohort A and will be evaluated and compared in the prospective cohort at the Immunology and Functional Genomics Research Unit of Centro Cardiologico Monzino.