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

Abstract

Background: Post-stroke cognitive impairment (PSCI) is a common and severe complication following acute lacunar stroke (ALS). Due to the limitations of current assessment tools and imaging methods, the early diagnosis of PSCI within 3 months of ALS remaining challenging. This study aimed to develop an effective, reliable, and accurate deep-learning method to predict PSCI within 3 months of ALS.

Methods: In total, 100 ALS patients were enrolled in the study, of whom 39 were diagnosed with PSCI and 61 were non-PSCI. First, we quantified three-dimensional (3D) gray-matter damage and white-matter tract disconnection, providing both regional damage (RD) and structural disconnection (SDC) higher-dimensional insights into brain network disruption. Second, we developed a novel deep-learning model based on ResNet18, integrating 3D RD, SDC, and diffusion-weighted imaging (DWI) to provide a comprehensive analysis of brain network integrity and predict PSCI. Finally, we compared the performance of our method with other methods, and visualized brain network damage associated with PSCI.

Results: Our model showed strong predictive performance and had a mean accuracy (ACC) of 0.820±0.024, an area under the curve (AUC) of 0.795±0.068, a sensitivity (SEN) of 0.746±0.121, a specificity (SPE) of 0.869±0.044, and a F1-score (F1) of 0.760±0.050 in the five-fold cross-validation, outperforming existing models. In the PSCI patients, brain network damage significantly affected the salience, default mode, and somatic motor networks.

Conclusions: This study not only established a model that can accurately predict PSCI, it also identified potential targets for symptom-based treatments, offering new insights into PSCI.