• Milli, M.
• Frasconi, P.
• Bernardini, A.
• Marcucci, R.
• Berteotti, M.
• Giusti, B.
• Testa, S.
• Palareti, G.
• Antonucci, E.
• Bindini, L.

Abstract

<jats:title>Abstract</jats:title><jats:sec><jats:title>Background</jats:title><jats:p>Despite the availability of different risk scores, the accuracy of actual prediction tools for outcomes in patients with atrial fibrillation (AF) remains modest. Although Machine Learning (ML) has been used to predict outcomes in the AF population, evidences about outcome prediction in a population entirely on oral anticoagulation is lacking.</jats:p><jats:p>The aim of this study is to use ML to predict outcomes in anticoagulated patients with atrial fibrillation, processing data from the Italian AF START–2 Register.</jats:p></jats:sec><jats:sec><jats:title>Methods and aims</jats:title><jats:p>Different ML models were applied to predict all–cause death, cardiovascular (CV) death, major bleeding and stroke in anticoagulated patients with nonvalvular AF. Overall population, Vitamin K Antagonists (VKA) and Direct Oral Anticoagulants (DOACs) populations were considered for the analyses.</jats:p></jats:sec><jats:sec><jats:title>Results</jats:title><jats:p>11078 AF patients (male n=6029, 54.3%) were enrolled with a median follow–up period of 1.5 years [IQR 1.0–2.6]. Patients on VKA were 5135 (46.4%), while 5943 (53.6%) were on DOACs. During the follow–up, 785 patients died, of which 169 (21.6%) due to CV causes; 240 major bleeding events were recorded, and 50 strokes occurred. Using Multi–Gate Mixture of Experts (MmoE), a cross–validated AUC of 0.779 ± 0.016 and 0.745 ± 0.022 were obtained, respectively, for the prediction of all–cause death and CV death in the overall population. The best ML model outperformed CHA2DSVA2SC and HAS–BLED for all–cause death prediction (p&amp;lt;0.001 for both). When compared to HAS–BLED, Gradient Boosting improved major bleeding prediction in DOACs patients compared to HAS–BLED score (0.711 vs. 0.586, p&amp;lt;0.001). Body mass index, age, glomerular filtration rate, platelet count and hemoglobin levels resulted the most important variables for ML prediction.</jats:p></jats:sec><jats:sec><jats:title>Conclusions</jats:title><jats:p>In patients with AF, ML models showed good discriminative ability to predict all–cause death, regardless of the type of anticoagulation strategy, and major bleeding on DOAC therapy, outperforming CHA2DS2VASC and the HAS–BLED scores for risk prediction in these populations. Anemia, platelet count, and BMI emerged as new potential risk predictors in anticoagulated AF patients. The applications of ML prediction models in clinical practice could improve the risk assessment and subsequent management of anticoagulated patients with AF.</jats:p></jats:sec>

Topics

• impedance spectroscopy
• size-exclusion chromatography
• machine learning