Development and validation of a multimodal deep learning framework for vascular cognitive impairment diagnosis

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Cerebrovascular disease (CVD) is a major cause of dementia globally.
• Accurate detection of vascular cognitive impairment (VCI) in CVD patients is challenging.

Purpose of the Study:

• To develop a multimodal deep learning framework for VCI detection in CVD patients.
• To create an accurate and interpretable clinical decision support tool.

Main Methods:

• Collected clinical and neuroimaging data from 307 CVD subjects.
• Developed a hybrid model combining vision transformer and extreme gradient boosting.
• Utilized a framework incorporating six clinical and two neuroimaging features.

Main Results:

• The hybrid model showed robust performance on internal and external datasets.
• Diagnostic accuracy was comparable to expert clinicians.
• The model identified key brain regions and clinical features contributing to VCI diagnosis.

Conclusions:

• An accurate and explainable AI tool for VCI detection in CVD patients has been developed.
• The tool enhances transparency and interpretability in VCI diagnosis.
• This approach aids in early identification and management of VCI in CVD.