Use of multi-modal artificial intelligence model for prediction of atrial fibrillation in hospitalized patients

Atrial fibrillation (AF) tends to be a common cardiac arrhythmia in hospitalized patients with acute illness. Although some general predictors are known, a systematic and integrated prediction schema that incorporates clinical predictors along with machine learning (ML) is lacking. In our study, we aimed to develop a multimodal artificial intelligence-based approach as a potential predictive tool for AF in hospitalized patients.

In our study, we sought to develop 3 different ML models for AF prediction in hospitalized patients.

We used the Symile-MIMIC dataset to develop 3 ML models: (1) clinical only, (2) electrocardiogram (ECG)-based, and (3) a multimodal fusion model.

We used a total of 13,817 ECGs from the Symile-MIMIC-IV dataset. These ECGs were then split into a training set (8923 ECGs, 65%), a validation set (675 ECGs, 5%), and an independent testing set that included a total of 4219 ECGs (30%). The clinical-only model had 250 predictor variables, and based on Shapley Additive exPlanations analysis to assess explainability of this model, age was the strongest risk factor for the prediction of AF. Although our ML models based on clinical- and ECG-only achieved a reasonable degree of predictive ability, the multimodal fusion model demonstrated the best predictive value (area under the receiver operating characteristic curve: 0.81) for the diagnosis of AF.

In our study, we developed 3 ML models based on clinical factors, 12-lead ECGs, and a multimodal fusion model combining both clinical and ECGs. Based on assessment with various performance metrics and different thresholds, the multimodal fusion model demonstrated the best predictive ability for AF.