Prediction of post-stroke cognitive impairment using brain FDG PET: deep learning-based approach

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Cerebrovascular Diseases

Background:

• Post-stroke cognitive impairment affects up to one-third of stroke survivors, significantly impacting quality of life.
• An objective, quantitative biomarker is needed for early prediction of dementia following stroke.
• Current methods for assessing cognitive decline lack objective quantitative measures.

Purpose of the Study:

• To develop a deep learning (DL)-based signature using positron emission tomography (PET) for objective evaluation of cognitive decline in stroke patients.
• To establish a DL model capable of predicting dementia after stroke.
• To validate the transferability of a DL model trained on Alzheimer's disease data to a stroke cohort.

Main Methods:

• A DL model was trained to differentiate Alzheimer's disease (AD) from normal controls (NC) using fluorodeoxyglucose (FDG) PET data.
• The trained DL model was directly applied to a prospectively enrolled cohort of stroke patients.
• Model performance was assessed using area under the curve of receiver operating characteristic curves (AUC-ROC), and correlations with clinical variables were analyzed.

Main Results:

• The DL model achieved an AUC-ROC of 0.94 for AD vs. NC classification.
• The transferred model discriminated stroke patients with dementia with an AUC-ROC of 0.75.
• The DL-based cognitive decline score was an independent risk factor for post-stroke dementia (hazard ratio, 10.90).

Conclusions:

• A DL-based cognitive signature using FDG-PET was successfully transferred to an independent stroke cohort.
• This DL-based cognitive evaluation shows potential as an objective biomarker for cognitive dysfunction in cerebrovascular diseases.
• The findings support the use of DL and FDG-PET for early identification of dementia risk in stroke survivors.