Artificial intelligence predictive model for risk of paroxysmal atrial fibrillation

LI Pan-pan; HAN Yu-chen; LI Feng; CHEN Yu; GUO Jun

doi:10.3969/j.issn.1672-5301.2024.03.002

Abstract：Objective To develop an integrated model utilizing 24-hour electrocardiogram data to accurately predict the risk of atrial fibrillation in high-risk populations and provide real-time prediction for the onset of atrial fibrillation in patients with paroxysmal atrial fibrillation.Methods Consecutively,a total of 310 patients diagnosed with paroxysmal atrial fibrillation by electrocardiographic report at the First Affiliated Hospital of Jinan University were retrospectively collected from January 1,2018,to December 31,2021,and a total of 124 patients were enrolled in this study as the atrial fibrillation group after screening.Additionally,a non-atrial fibrillation group consisting of 496 patients with normal ECG reports was randomly selected at a ratio of 1∶4.Subsequently,both groups were randomly divided into three sets for ECG model training:a training set(n=434),a validation set(n=62),and a test set(n=124)in a ratio of 7∶1∶2.The establishment process involved developing an ECG neural network model and heart rate neural network model separately.Finally,logistic regression was employed to combine these models into an integrated model.Results After validation and testing,the AI algorithm achieved an AUC of 0.94(95%CI 0.75-0.94),with sensitivity,specificity,accuracy,precision,and F1score of 56.0%,98.0%,90.0%,93.0%,and 0.70 respectively.Compared with the clinical risk model and the existing AF prediction model,HARMS2-AF score,the artificial intelligence algorithm had a superior performance(P<0.01).Conclusions Artificial intelligence-integrated algorithms seem to be an effective method for predicting AF risk and forecasting AF episodes in real time.This could have important clinical implications for screening for AF and developing personalized anticoagulation plans.