On the Value of MRI in Early Cognitive Risk Prediction: Survival Analysis Comparing Statistical and Machine Learning Models

Billichova, M, Bruno, D ORCID: 0000-0003-1943-9905, Sharifian, F, Czanner, S and Czanner, G On the Value of MRI in Early Cognitive Risk Prediction: Survival Analysis Comparing Statistical and Machine Learning Models. BMC Medical Informatics and Decision Making. ISSN 1472-6947 (Accepted)

Text BMC-Manuscript.pdf - Accepted Version Access Restricted Available under License Creative Commons Attribution. Download (433kB)

Abstract

Background: Early identification of individuals at risk of developing amnestic Mild Cognitive Impairment (aMCI) is essential for timely intervention in Alzheimer’s disease (AD). For such identification, MRI data may be included to capture brain changes and identify clinical symptoms. Sex differences have been reported in cognitive decline and Alzheimer’s disease progression. Therefore, we additionally examined whether MRI associations with aMCI differed between men and women.

Methods: This study evaluated the predictive value of MRI variables for conversion to aMCI using data from the National Alzheimer’s Coordinating Center, including 742 women and 451 men. Time-to-event outcomes were analysed using Cox proportional hazards (CoxPH) regression models. The effect of MRI variables, right and left hippocampus volume (RHIPPO, LHIPPO), total intracranial volume (NACCICV), and left entorhinal mean cortical thickness (LENTM), was investigated. The models were adjusted for demographic and clinical variables, while interaction with sex was also considered. Model performance was assessed using the C-index, Brier score and time-dependent AUC. To compare with machine learning, alternative survival modelling approaches, including Random Survival Forests and Gradient Boosting Survival models, were also applied.

Results: Higher RHIPPO and LHIPPO volumes and greater LENTM were associated with a lower risk of conversion to aMCI. However, incorporating MRIderived features only slightly improved predictive performance, increasing the CoxPH model C-index from 85.26% to 85.94% and the 4-year time-dependent AUC from 86.9% to 88.0%. Machine learning–based survival models achieved comparable discrimination and slightly lower Brier scores, indicating lower prediction error for 4-year survival probabilities.

Conclusion: Our study indicates that although some MRI-derived structural variables were significantly associated with conversion to aMCI, they provided only a modest additional improvement in predictive performance. Sex-stratified analyses showed that associations were directionally consistent across sexes, and formal interaction analyses revealed no statistically significant sex differences. Further research is needed to clarify potential sex-specific effects.

Item Type:	Article
Uncontrolled Keywords:	0806 Information Systems; 1103 Clinical Sciences; Medical Informatics; 4203 Health services and systems
Subjects:	B Philosophy. Psychology. Religion > BF Psychology Q Science > QA Mathematics > QA76 Computer software T Technology > T Technology (General)
Divisions:	Computer Science and Mathematics Psychology (from Sep 2019)
Publisher:	BMC
Date of acceptance:	1 September 2026
Date Deposited:	08 Sep 2026 12:33
Last Modified:	08 Sep 2026 12:33
URI:	https://researchonline.ljmu.ac.uk/id/eprint/29340

View Item