Evaluation of vascular cognitive impairment and identification of imaging markers using machine learning: a multimodal MRI study

Haoying He¹Dongwei Lu²Sisi Peng²Jiu Jiang³Fan Fan¹Sun Dong¹Tianqi Sun¹Xu Zhi Peng²Ping Zhang¹Xiaoxiang Peng⁴Ming Lei⁵Junjian Zhang^1*

• ¹Department of Neurology, Zhongnan Hospital, Wuhan University, Wuhan, China
• ²Department of Neuropsychology, Zhongnan Hospital of Wuhan University, Wuhan, Hebei Province, China
• ³Electronic Information School, Faculty of Information Sciences, Wuhan University, Wuhan, Hubei Province, China
• ⁴Hubei Provincial Third People's Hospital (Zhongshan Hospital), Wuhan, Hubei Province, China
• ⁵General Hospital of The Yangtze River Shipping, Wuhan Brain Hospital, Wuhan, Hubei Province, China

Background Vascular cognitive impairment (VCI) is prevalent but underdiagnosed due to its heterogeneous nature and the lack of reliable diagnostic tools. Machine learning (ML) enhances disease evaluation by enabling accurate prediction and early detection from complex data. This study aimed to develop ML models to detect VCI using clinical data and multimodal MRI, and to explore the associations between imaging markers and cognitive function.The study enrolled 313 participants from Wuhan and surrounding areas, including 157 patients with VCI (age 62.38±6.62 years, education 10.83±3.00 years) and 156 cognitively normal individuals with vascular risk factors (age 59.93±6.74 years, education 13.97±3.19 years). An independent dataset of 82 participants was used for external validation. Clinical data, neuropsychological assessments, and MRIs (T1, T2-FLAIR, and DTI) were collected. After imaging processing and preliminary model selection, optimal models using various data modalities were constructed.Model reduction was undertaken to simplify models without sacrificing performance.SHapley Additive exPlanations and moDel Agnostic Language for Exploration and eXplanation were used for model interpretation.The comprehensive final model integrating clinical and multimodal MRI measures achieved the best performance with eight input variables (AUC of 0.956, 95%CI 0.919–0.988 for internal and 0.919, 95%CI 0.866–0.966 for external validation). During external validation, DTI demonstrated more stable performance than T1 and T2-FLAIR imaging, highlighting its potential importance over conventional imaging markers. Key imaging markers, especially along the lateral cholinergic pathway, were highlighted for their importance in diagnosing VCI and understanding its manifestation. Conclusion Our study developed and validated accurate ML models for VCI detection, emphasizing the importance of DTI. The identified imaging markers, particularly those derived from DTI, underscoring the potential in enhancing diagnostic accuracy and understanding cognitive impairments related to vascular changes.