Predictive MRI signatures of alcohol-related risk diverge from group-level structural differences in firefighters

Group-level structural differences may not identify features most relevant to individual prediction. In this cross-sectional study of 686 firefighters, high alcohol-related risk was defined as an Alcohol Use Disorders Identification Test score ≥ 16. We compared structural effect sizes associated with alcohol-related risk with SHAP-derived importance at feature and region levels in a model combining MRI radiomics with demographic, cognitive, and laboratory features.

The primary stratified 70/30 split was fixed at seed 42 before held-out performance was examined. In this prespecified split, multimodal CatBoost achieved an AUC of 0.743 (95% CI 0.666–0.819), higher than the clinical-only and radiomics-only representative models. Across 10 training-stage runs on the same held-out test set, its mean AUC was 0.681 (SD 0.054; range 0.555–0.743), and it ranked highest in 6 of 10 runs; 0.743 was the maximum.

Cortical features showed larger group differences, whereas SHAP-derived importance highlighted distributed subcortical features in the primary model; thalamic, ventral diencephalic, and ventricular features were among stable repeated-run contributors. Structural effect sizes and SHAP importance showed weak feature-level correspondence (r = 0.22, p = 0.044) and no detectable region-level correspondence (r = 0.03, p = 0.844). Within the selected radiomics set, these findings support evaluating group-level association and model-dependent predictive contribution as complementary properties of candidate imaging biomarkers.

This work was supported by the Korea Institute for Advancement of Technology (KIAT) grant funded by theKorean government (MOTIE) (RS-2025-02214034, HRD Program for Industrial Innovation) and the NationalResearch Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (RS-2024-00457381). Dong Ok Kim and MyeongGyun Jang: These authors contributed equally to this work. Department of Biomedical Informatics, Korea University College of Medicine, 161 Jeongneung-ro, Seongbuk-gu, Seoul, 02708, Republic of Korea Dong Ok Kim, MyeongGyun Jang & Hwamin Lee Ewha Brain Institute, Ewha Womans University, 52 Ewhayeodae-gil, Seodaemun-gu, Seoul, 03760, Republic of Korea Department of Brain and Cognitive Sciences, Ewha Womans University, Seoul, Republic of Korea Correspondence to Sujung Yoon or Hwamin Lee.

The authors declare no competing interests. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Below is the link to the electronic supplementary material.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material.