Interpretable Independent Recurrent Networks for Forecasting Stroke in Atrial Fibrillation

Background

Atrial fibrillation (AF) is a major risk factor for transient ischemic attack (TIA)/ischemic stroke (IS).

Objectives

Given the dynamic nature of IS risk, this study aimed to predict IS risk in AF patients using a high-dimensional time-series model.

Methods

We conducted a cohort study at the National Taiwan University Hospital from 2014 to 2019, including 7,710 AF patients, with external validation in 6,822 patients from the National Taiwan University Hospital Yunlin Branch. The Forecasting Strokes via Interpretable Independent Networks (ForeSIIN) model, based on gated recurrent units, was proposed. Kaplan-Meier analysis with log-rank test evaluated risk group differences.

Results

The annual TIA/IS incidence rate ranged from 181.96 (95% CI: 164.42-200.93) to 15.81 (95% CI: 12.38-20.18) per 1,000 person-years, with an overall incidence of 42.40 (95% CI: 39.60-45.39). The ForeSIIN model achieved the best prediction with an area under the receiver-operating characteristics curve of 0.764 (95% CI: 0.722-0.810), compared with the CHA₂DS₂-VASc score (AUC: 0.650; 95% CI: 0.596-0.699) and other nonsequential models: extreme gradient boosting AUC: 0.722 (95% CI: 0.676-0.769), support vector machine AUC 0.691 (95% CI: 0.637-0.741), random forest AUC: 0.689 (95% CI: 0.637-0.742). External validation showed area under the receiver-operating characteristics curve of 0.646 (95% CI: 0.618-0.671) and area under the precision-recall curve of 0.222 (95% CI: 0.184-0.259). Feature impact analysis identified the top 5 factors: history of TIA/IS, estimated glomerular filtration rate, C-reactive protein, hematocrit, and plasma fasting glucose. Kaplan-Meier analysis showed significant risk differences between ForeSIIN groups (log-rank P < 0.001).

Conclusions

The innovative ForeSIIN model demonstrated accurate stroke prediction in AF patients and enhanced the interpretation of dynamic risk factors over time.