Predicting cancer aetiology in patients with stroke using brain imaging alone

Area of Science:

• Neurology
• Oncology
• Radiology

Background:

• Ischaemic stroke due to cancer is increasingly prevalent, associated with high mortality.
• Accurate prediction of cancer aetiology in stroke patients using only brain imaging is needed for timely intervention.
• Metastatic cancer is a significant concern in cancer-related strokes.

Purpose of the Study:

• To identify unique brain imaging patterns associated with cancer-related stroke.
• To develop a predictive model for diagnosing stroke caused by cancer.
• To investigate the association between specific imaging patterns and metastatic cancer.

Main Methods:

• Retrospective cohort study (2014-2022) comparing patients with cancer-related stroke to controls.
• Analysis of brain MRI diffusion-weighted imaging lesions: distribution, vascular territories, number, and pattern (multifocal, crossed territories, scattered).
• Statistical analysis included multivariable logistic regression and machine learning models (logistic regression, support vector machines) for predictive capacity assessment.

Main Results:

• The 'three-territory sign' was a significant predictor of cancer aetiology (OR 20.79, p=0.006).
• Machine learning models demonstrated good predictive performance for metastatic cancer.
• Balanced accuracy reached 75% for logistic regression and 79% for support vector machines, with AUC scores of 0.85 and 0.87, respectively.

Conclusions:

• Brain imaging alone shows potential for accurately predicting cancer aetiology in stroke patients.
• This approach may be particularly effective for identifying metastatic cancer as the cause of stroke.
• Early diagnosis and treatment of cancer-related stroke can be facilitated by advanced imaging analysis.