Predicting stroke outcome: A case for multimodal deep learning methods with tabular and CT Perfusion data.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Acute ischemic stroke significantly impacts global morbidity and disability.
• Accurate patient prognosis is vital for effective stroke rehabilitation planning.
• Deep learning offers potential for improved predictions by integrating diverse data sources.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning approach for predicting functional outcomes in acute ischemic stroke patients.
• To assess the efficacy of combining tabular data and CT perfusion imaging for prognosis.

Main Methods:

• Experiments were conducted using deep learning architectures (TabNet, ResNet-10) on tabular and imaging data, respectively.
• A multimodal deep learning architecture (DAFT) was implemented to integrate both data types.
• The study utilized data from 98 acute ischemic stroke patients with CT perfusion scans.

Main Results:

• TabNet achieved an AUC of 0.71 on tabular data; ResNet-10 achieved an AUC of 0.70 on imaging data.
• The multimodal DAFT architecture yielded superior results with an AUC of 0.75 and an F1 score of 0.80.
• The model demonstrated high performance with fewer parameters and a smaller dataset compared to existing studies.

Conclusions:

• The study confirms the feasibility of predicting functional outcomes for ischemic stroke patients.
• Multimodal deep learning architectures are effective for stroke outcome prediction.