Scalable CT-based prognostic modeling of dementia conversion in mild cognitive impairment

초록

Mild cognitive impairment (MCI) is a heterogeneous condition, with up to 40% of patients progressing to dementia within three years of clinical diagnosis. While MRI, PET, and CSF biomarkers have shown prognostic value, their limited accessibility restricts routine use. We aimed to develop and validate a CT-based machine learning model to predict dementia conversion in patients with MCI. We analyzed 791 MCI patients with baseline CT and longitudinal neuropsychological assessments. Candidate predictors included demographic factors, APOEgenotype, Clinical Dementia Rating-Sum of Boxes (CDR-SB), and CT-derived W-scores from regional CSF and ventricular volumes. CT-derived W-scores represent age-, sex-, and modality-adjusted z-scores of regional CSF and ventricular volumes. Converters were older (73.4 vs. 70.7 years, p < 0.001), more often female (60.4% vs. 51.4%, p = 0.015), and more likely APOE epsilon 4 carriers (52.5% vs. 40.0%, p < 0.001). Logistic regression with SMOTEENN sampling achieved the best performance (Area Under the Curve [AUC]= 0.840; accuracy = 0.743; sensitivity = 0.847; specificity = 0.674), with robust generalizability in the independent test set (AUC= 0.823). Feature selection and SHAP analysis identified six key predictors, including CDR-SB, APOE epsilon 4, age, and three CT-derived volumetric markers (W-score in the left inferior lateral ventricle, left parietal CSF, and left occipital CSF). Our findings demonstrate that CT-derived volumetric analysis, combined with clinical and genetic features, enables accurate and interpretable prediction of dementia conversion in MCI.

키워드