Artificial intelligence for prediction of atrial fibrillation in the stroke unit: a retrospective derivation validation cohort study

Abstract

Background: Paroxysmal atrial fibrillation (AF) is a major cause of stroke but is often undetected in routine clinical practice. Effective stratification is needed to identify patients with stroke who might benefit the most from intensified AF screening. Several artificial intelligence models have been proposed to predict AF based on ECG in sinus rhythm, but broad implementation has been limited. The most valuable input features and most effective model design for AF prediction are also unclear.

Methods: We developed and tested AF prediction models utilising continuous electrocardiogram monitoring (CEM) recordings from the first 72 h after admission and multiple clinical input features from patients with stroke hospitalised at Charité, Berlin, Germany, between September 2020 and August 2023. We compared different models and input data to identify the best-performing model for prediction of AF. The relative contributions of different input data sources were assessed for explainability. A final model was externally validated using the first hour of monitoring data from the intervention group of the prospective multicentre MonDAFIS study.

Findings: The derivation dataset included 2068 patients with acute ischaemic stroke, of whom 469 (22.7%) had AF, first detected before or during the index hospital stay (366 vs. 103). In predicting newly detected AF, a Bayesian fusion model emerged as best, achieving a ROC-AUC of 0.89 (95% CI: 0.80, 0.96). Model introspection indicated that HRV was the main driver of the model's predictions. A final, simplified tree-based ensemble model using age and HRV parameters of the first hour of CEM data achieved similar performance (ROC-AUC 0.88, 95% CI: 0.79, 0.95). The final model consistently outperformed the AS5F score in a real-world scenario external validation on the MonDAFIS dataset (1519 patients, thereof 36 (2.37%) with AF; ROC-AUC 0.79 vs. ROC-AUC 0.69, p = 4.69e-03).

Interpretation: HRV appears to be the most informative variable for predicting AF. A computationally inexpensive model requiring only 1 h of single-lead CEM data and patients' age supports prediction of AF after acute ischaemic stroke for up to seven days. Such a model may enable risk-based stratification for cardiac monitoring, prioritising efforts where most needed to enhance AF screening efficiency and, ultimately, secondary stroke prevention.

Funding: This study was supported by the German Federal Ministry of Education and Research and the German Research Foundation.

Keywords: Artificial intelligence; Atrial fibrillation; Heart rate variability; Machine learning; Prediction; Stroke.

Fig. 1

Study design. a) AI-based risk stratification during stroke unit work-up. Data routinely collected during stroke unit workup serve as input variables; a ML model automatically stratifies patients according to their individual risk of underlying AF for PCM. b) Multimodal data used during comparative analyses. c) Optimal model development and external validation. Different ML models using different types and combinations of input data and different types of model architecture were systematically explored using 5-fold nested cross-validation. A final model was selected for internal and external validation based on a trade-off between predictive performance and ease-of-use. Sensitivity, positive predictive value, and negative predictive value at a predefined specificity threshold of 90% were calculated to evaluate clinical applicability. AI: Artificial intelligence; TIA: Transient ischaemic attack; AF: Atrial fibrillation; PCM: Prolonged cardiac monitoring; ICA: Internal carotid artery; GFR: Glomerular filtration rate; HRV: Heart rate variability; CEM: Continuous electrocardiogram monitoring; ML: Machine learning.

Fig. 2

Flow chart of patient selection. HDL: Health Data Lake; AF: Atrial fibrillation.

Fig. 3

Results of comparative testing of different models. a) Schematic depiction of the different classes of features and models. b, c, d) ROC curves of DNN, ensemble and fusion models for patients without AF vs. newly detected AF (1702 patients, thereof 1599 without and 103 with newly detected AF). e, f, g) ROC curves for patients without AF vs. newly detected AF and pre-known AF (2068 patients, thereof 1599 without AF, 103 with newly detected AF and 366 with pre-known AF). Inset Figure 3b and e: Inclusion of more ECG segments increased performance up to 20 segments. All curves are plotted with 95% confidence intervals across folds. ROC-AUCs and respective 95% confidence intervals are provided in the legends. ROC: Receiver operating characteristic; DNN: Deep neural network; CEM: Continuous electrocardiogram monitoring; HRV: Heart rate variability; AF: Atrial fibrillation; AUC: Area under the curve.

Fig. 4

Explainability analysis. a) Distribution of predictions generated by the DNN model based on distinct CEM segments for 28 example patients. Each column corresponds to a unique patient and shows the predictions for 10 ECG segments (duration 5.12 s each) for this patient. Segments labelled as sinus rhythm by human experts are depicted in blue, while those labelled as potential atrial arrhythmia are shown in pink. The dashed horizontal line shows the 90% specificity threshold. b) SHAP analysis on three raw CEM example segments. Icons in the upper right corner mark the corresponding segments in panel A. Top: segment from a patient with AF diagnosis which was classified as atrial arrhythmia by the human expert and as indicative for underlying AF by the model. Middle: segment from a patient with AF diagnosis which the human expert categorised it as sinus rhythm while the DNN classified it as indicative for underlying atrial fibrillation. Bottom: segment from a patient without AF diagnosis that was categorised as sinus rhythm by the human expert while the model did classify it as not indicative for underlying AF. c) SHAP analysis of the relative predictive performance of individual features applied to the ensemble model (HRV + clinical). DNN: Deep neural network; CEM: Continuous electrocardiogram monitoring; AF: Atrial fibrillation; SHAP: SHhapley Additive exPlanations; HRV: Heart rate variability; TP: Total power; HF: High frequency, LF: Low frequency, RMSSD: Root mean square of successive differences; DFA: Detrended fluctuation analysis; pNN50: Proportion of the number of pairs of successive normal heartbeats (NN) that differ by more than 50 ms divided by the total number of NN; MSE: Multiscale entropy; SDNN: Standard deviation of NN intervals; S2N: Number of supraventricular extrasystoles divided by the number of normal heartbeats. For more detailed information see Supplementary Table S1.

Fig. 5

External validation. a) Validation on the MonDAFIS-dataset, using only the first hour of CEM data. b) Validation on the MonDAFIS-dataset, incorporating different lengths of CEM data for HRV calculation. p-value for final model vs. AS5F was 4.69e-03 [DeLong's test]. CEM: Continuous electrocardiogram monitoring; AUC: Area under the curve; Thr: Threshold; Sens: Sensitivity; Spec: Specificity; PPV: Positive predictive value; NPV: Negative predictive value.