Adding brain MRI-derived features to a clinical model improved prediction of atrial fibrillation detected after stroke (ROC-AUC 0.75 vs 0.67; p<0.01), but did not improve an age and HRV model.
Cohort (n=1,689)
Does the addition of brain MRI-derived features improve the prediction of atrial fibrillation detected after ischemic stroke in patients with ischemic stroke?
Brain MRI-derived features do not add significant predictive value for post-stroke atrial fibrillation beyond simple clinical variables like age and heart rate variability.
Absolute Event Rate: 0.75% vs 0.67%
p-value: p=<0.01
Abstract Background and aims Improved risk stratification for post-stroke atrial fibrillation (AF) is urgently needed. We previously developed a machine-learning (ML) model predicting AF using age and heart rate variability (HRV). Whether brain magnetic resonance imaging (MRI) provides additional predictive value remains controversial. MRI-derived features may differentially inform AF detected after stroke (AFDAS) versus previously known AF (KAF). Methods We extracted interpretable MRI-derived features including lesion volume, infarct distribution across vascular territories, insular involvement, and central autonomic network (CAN) involvement. Tree-based ML models were developed using MRI-derived features, clinical variables, and HRV metrics, with separate analyses for AFDAS and KAF. Incremental predictive value of MRI features was quantified, key predictors were identified using Shapley additive explanations (SHAP), and external validation was performed in an independent cohort. Results The primary cohort included 1,227 patients (69 AFDAS 5.6%, 287 KAF 23.4%); the external cohort included 462 patients (17 AFDAS 3.7%). Models based solely on MRI-derived features showed modest discrimination. SHAP analysis identified insular involvement, lesion volume, and CAN involvement as key predictors for AFDAS. Adding MRI-derived features significantly improved a clinical-feature model for AFDAS (ROC-AUC 0.67 vs. 0.75; p0.01), but not for KAF (ROC-AUC 0.74 vs. 0.77). MRI-derived features did not improve performance of the previously published age+HRV model in internal and external validation (ROC-AUC 0.83 vs. 0.81). Conclusions MRI-derived features provide complementary, biologically meaningful information for AFDAS beyond clinical variables, supporting the concept of a neurogenic AF subgroup. However, for pragmatic post-stroke AF risk stratification in clinical practice, age and HRV alone may be sufficient. Conflict of interest Maximilian Schöls: nothing to disclose
Schöls et al. (Fri,) conducted a cohort in Ischemic stroke (n=1,689). Brain MRI-derived features vs. Clinical variables or age and heart rate variability (HRV) alone was evaluated on Prediction of atrial fibrillation detected after stroke (AFDAS) (p=<0.01). Adding brain MRI-derived features to a clinical model improved prediction of atrial fibrillation detected after stroke (ROC-AUC 0.75 vs 0.67; p<0.01), but did not improve an age and HRV model.