Searching for underlying atrial fibrillation using artificial intelligence-assisted MRI images from ischemic stroke patients

Abstract

Abstract Background Nearly 30% of ischemic strokes are attributed to atrial fibrillation (AF) related cardiac thromboembolism. This subset of strokes are associated with higher recurrence rate, morbidity and mortality. As a result, current recommendation suggests a cardiac monitoring of variable intensity and duration to search for unknown AF after a stroke so that optimal secondary prevention could be administered. Unfortunately, it is reported that 24h Holter ECG monitoring for underlying AF is completed routinely in only a minority of stroke centers, making a huge gap between AF detection rate in clinical practice and estimated actual data (2–3% vs 23–30%). So, it is obvious that paroxysmal AF remains undetected in a significant proportion of patients after a stroke and an early and effective identification approach of cardiac disturbances is needed. Purpose We aim to develop an innovative end-to-end artificial intelligence model to rapidly search for atrial fibrillation in post ischemic stroke patients using MRI imaging data. Methods 489 patients diagnosed with ischemic stroke, 174 of which had a prior known history of AF or new AF detected by intensive ECG monitoring poststroke, from January 2018 to October 2021 were enrolled in this study. The MRI images obtained after their admission in the stroke unit were meticulously evaluated and acute ischemic stroke lesions on DWI sequences were segmented and confirmed by two senior neuroradiologists independently. These processed images were randomly split into training (n=315) and testing set (n=174) for the implementation and validation of our AI screening model. And the algorithm was based on the combination of radiomic features and semantic features extracted from convolutional neural network (CNN). The model performance was evaluated by accuracy, recall, precision, F-Measure and AUC. Also, heatmaps which indicate the attention mechanism of the model were generated for interpreting underlying patterns. Results The AUC of our model (COM) reached 0.8 (Figure 1A). Also, the algorithm yielded values for the accuracy, recall, precision and F-Measure of 70%, 92.5%, 63.8% and 75.5%, respectively, which showed satisfactory classification results. As to the interpretability of the AI model, we found that more “attention” was paid to the main lesions (Figure 1B) and radiomic features which indicated the minimum gray level intensity and the sphericity of the lesions (Figure 1C, D) were crucial to the classifier. Conclusions Our work revealed a potential correlation between brain ischemic lesion pattern on DWI images and underlying etiology of AF. Moreover, the AI model we developed may serve as a rapid screening tool for AF in clinical practice of stroke units. Funding Acknowledgement Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Natural Science Foundation of China