Screening of Cognitive Impairment Biomarkers Based on Machine Learning and SHAP Interpretability Analysis

This study aims to develop a cognitive impairment screening model based on the Clinical Dementia Rating (CDR) using the XGBoost and SHapley Additive exPlanations (SHAP) framework, with the goal of improving predictive performance while enhancing clinical interpretability. The Open Access Series of Imaging Studies (OASIS)-1 cross-sectional dataset was used, containing multimodal clinical and magnetic resonance imaging (MRI) data. After rigorous data cleaning, 235 subjects were included. Seven features with clear neurobiological significance were selected as input: age, education, socioeconomic status, Mini-Mental State Examination (MMSE), estimated total intracranial volume (eTIV), normalized whole brain volume (nWBV), and atlas scaling factor (ASF). The model was evaluated using a test set from OASIS-1 and further validated on an additional OASIS-2 cohort. The proposed model achieved high performance on the OASIS-1 test set (Accuracy=0.851, area under the receiver operating characteristic curve [AUC]=0.943) and demonstrated good generalization on OASIS-2 (Accuracy=0.867, AUC=0.938). SHAP analysis revealed that MMSE, nWBV, and age are the most influential features, with directionality consistent with established understanding of cognitive decline and neurodegeneration. An ablation study excluding MMSE showed a performance drop (AUC: 0.943 → 0.850), confirming that cognitive assessment is a critical component while structural features also provide substantial predictive information. The framework provides a transparent and lightweight screening tool with potential for clinical deployment, particularly in resource-limited settings where advanced biomarker assays are unavailable.