Predicting cancer aetiology in patients with stroke using brain imaging alone.

Background: Ischaemic stroke due to cancer is becoming more common with high mortality rate. There is an unmet need to predict cancer aetiology in patients with stroke using brain imaging alone, to facilitate early diagnosis and treatment. Aims: To describe unique brain imaging patterns of cancer-related stroke and thereby develop a predictive model for diagnosis of stroke due to cancer. Our hypothesis was that such imaging patterns would be more closely associated with metastatic cancer. Methods: Retrospective cohort study at two neighbouring sites in Australia from 2014 to 2022. The cohort group included patients with acute ischaemic stroke due to active cancer. The control group consisted of conventional stroke aetiologies and no active cancer diagnosis. Brain imaging using MRI diffusion-weighted imaging sequences classified lesions by distribution, vascular territories, lesion number and whether multifocal, crossed territories or scattered. Statistical analysis used adjusted logistic regression models. Machine learning-based predictive modelling evaluated the predictive capacity on a diagnosis of metastatic cancer. Results: There were 138 patients available for analysis. A three-territory sign was the only significant imaging predictor for cancer aetiology in the multivariable analysis (OR 20.79; 95% CI 2.44 to 177.47, p=0.006). In the machine learning modelling, both logistic regression and support vector machines models predicted the presence of metastatic cancer well with a balanced accuracy of 75% and 79%, respectively, and area under the curve (receiver operating characteristic area) scores of 0.85 and 0.87, respectively. Conclusions: Brain imaging alone might potentially predict cancer aetiology with good accuracy in patients with stroke, especially in metastatic cancer.