Multimodal deep learning for near term prediction of new onset atrial fibrillation in hospitalized patients

Background

New-onset atrial fibrillation (AF) is common in patients who are hospitalized with acute illness and is associated with overall poor outcomes. Current clinical and demographic predictors of new-onset AF have limitations. Electrocardiogram (ECG) waveforms, combined with these factors as part of multimodal approach could help improve prediction of new-onset AF.

Purpose

The purpose of our study was to develop a multimodal, deep-learning system that combines ECG waveforms with clinical data for prediction of new-onset AF in hospitalized patients.

Methods

We developed and compared two separate 1-dimensional models with a fusion model for prediction of new-onset AF.

(a): one-dimensional convolutional neural network (1D-CNN) that learns patterns from the raw ECG waveforms.

(b): a Multi-Layer Perceptron (MLP), unsupervised machine learning model which utilizes data from structured variables such as: age, heart rate, routinely obtained laboratory data and medications. The two modules each produce latent feature vectors, which are then placed end-to-end and fed into a final Multi- MLP head that outputs the probability of New-onset AF during hospitalization. We then compared this fused MLP model with the two simpler unimodal models based solely on ECG waveforms and clinical data respectively.

Results

On the held-out test set data, the feature-level fused model achieved an Area Under the Curve (AUC) of 0.81. This significantly outperformed both the ECG-alone model (0.77) and the clinical-alone model (0.74). By adjusting the threshold, we can fine-tune the tool at a threshold of 0.5, it flags 83% of patients who later develop AF (207 of 248) with a precision of 10%, a setting that could be useful for wider screening. By adjusting the threshold of 0.98 for a high-precision alert configuration, it has the ability to predict 46% of future occurrence of AF cases (113 of 248) while reducing false alarms by 81% (from 1,797 to 328), increasing precision to 26% and achieving an overall accuracy of 89%, a setting useful when resources for follow-up are limited.

Conclusions

Combining ECG signals with clinical data significantly enhances the model’s ability to detect new-onset AF in hospitalized patients as compared to a uni-dimensional CNN model. The dynamic nature of ML model also allows for adjustment of its decision threshold which can aid in broader screening and improving adverse outcomes related to occurrence of AF.ROC Curve for Fusion model