Abstract:
Cardioembolic stroke is a dangerous subtype of ischemic stroke. The patients with this subtype need special treatments to prevent recurrent events. The prevention is vital since only one more event could result in fatal damage. Hence, the classification into the categories of cardioembolic and non-cardioembolic subtypes is essential. We developed a multimodal machine learning model that can integrate the basic clinical information and non-contrast CT images to predict the risk of cardioembolic stroke. Our method reached the areas under the receiver operating characteristic curve (ROC-AUC) of 0.840 by using a dataset of only 227 samples of stroke patients. Besides the capability to classify the stroke subtypes, the method can provide the interpretability of the model decision in the forms of the heatmap for large infarct localization and the feature impacts for interpretation. Our approach can be widely applied since we need only the basic clinical information and non-contrast CT which are commonly available in general hospitals.
