A deep-learning model for predicting post-stroke cognitive impairment based on brain network damage.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Post-stroke cognitive impairment (PSCI) is a frequent complication after acute lacunar stroke (ALS).
• Early diagnosis of PSCI within 3 months of ALS is challenging due to limitations in current assessment and imaging techniques.
• This study addresses the need for effective and reliable methods for early PSCI detection.

Purpose of the Study:

• To develop a deep learning (DL) model for predicting PSCI within 3 months of ALS.
• To integrate neuroimaging data for a comprehensive analysis of brain network integrity.
• To improve the accuracy and reliability of early PSCI diagnosis.

Main Methods:

• 100 acute lacunar stroke (ALS) patients were analyzed (39 with PSCI, 61 without).
• Quantified 3D gray-matter damage (regional damage - RD) and white-matter tract disconnection (structural disconnection - SDC).
• Developed a ResNet18-based DL model integrating 3D RD, SDC, and diffusion-weighted imaging (DWI) to predict PSCI.

Main Results:

• The DL model achieved high predictive performance: accuracy (ACC) 0.820±0.024, AUC 0.795±0.068.
• The model outperformed existing methods in five-fold cross-validation.
• PSCI was associated with significant damage in salience, default mode, and somatic motor networks.

Conclusions:

• The developed DL model accurately predicts PSCI within 3 months of ALS.
• Identified specific brain networks affected in PSCI, offering potential targets for symptom-based treatments.
• Provides novel insights into the neurobiological underpinnings of PSCI.