Machine learning models for stratifying traumatic brain injury severity based on interleukin-6 temporal dynamics
Extreme Medicine · 10.06.2026 · 28 (2) · 267-276
DOI: 10.47183/mes.2026-438 ↗Kurz gefasst
Fachartikel in der Zeitschrift Extreme Medicine, online erschienen am 10.06.2026. Verfasst von E. I. Balakin und 5 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben.
Abstract (Originalsprache)
Introduction
Traumatic brain injury (TBI) remains a serious healthcare problem due to its heterogeneity and unpredictable outcomes. In TBI, secondary inflammatory cascades are frequently associated with the level of interleukin-6 (IL-6). However, its diagnostic and prognostic value at various time points after injury requires further investigation in athletes.
Objective
To evaluate the diagnostic accuracy and prognostic significance of IL-6 in stratifying TBI severity at various time points after injury with the use of machine learning models. Materials and methods. A prospective cohort study was conducted among 89 male athletes, with a mean age of 28.7 ± 5.3 years. These athletes were actively engaged in contact sports (mixed martial arts and kudo) and had documented evidence of concussion or mild to moderate brain contusion. Sequential blood plasma collection was performed at 3, 6, 12, and 24 h post-injury. IL-6 concentration was determined using validated enzyme-linked immunosorbent assay (ELISA) protocols. In order to classify injury severity, models were trained using gradient boosting (XGBoost), logistic regression, and random forest algorithms. The accuracy of model responses was evaluated using ROC analysis. To identify the most significant time points, the SHAP (SHapley Additive exPlanations) feature importance method was applied. For patient prediction and stratification, a model based on a long short-term memory (LSTM) recurrent neural network was implemented. Statistical validation used the Kruskal–Wallis H-test (H = 31.77; p < 0.001) and Spearman’s correlation analysis (rs = 0.81; p < 0.001).
Results
The greatest discriminatory ability of IL-6 concentrations for moderate brain contusion was noted at 6 and 12 h. The XGBoost model achieved an area under the curve of 0.92 [95% CI: 0.88; 0.96], with a sensitivity of 87% and a specificity of 84%. The SHAP analysis revealed that IL-6 values at 6 and 12 h had the greatest impact on the model’s predictions. Logistic regression and random forest yielded areas under the curve of 0.84 and 0.88, respectively. The identified diagnostic window between 6 and 12 h post-injury coincides with the peak of neuroinflammatory activity.
Conclusions
The level of IL-6 measured within 6–12 h after TBI represents a specific biomarker for early stratification of injury severity. The integration of explainable machine learning approaches provides robust and clinically relevant decision support in neurotraumatology.
Zitieren
APA
Balakin, E. I., Yurku, K. A., Malsagova, K. A., Butkova, T. V., Kaysheva, A. L., & Pustovoit, V. I. (2026). Machine learning models for stratifying traumatic brain injury severity based on interleukin-6 temporal dynamics. Extreme Medicine, 28(2), 267-276. https://doi.org/10.47183/mes.2026-438
Vancouver
Balakin EI, Yurku KA, Malsagova KA, Butkova TV, Kaysheva AL, Pustovoit VI. Machine learning models for stratifying traumatic brain injury severity based on interleukin-6 temporal dynamics. Extreme Medicine. 2026;28(2):267-276. doi:10.47183/mes.2026-438
BibTeX
@article{Balakin2026,
author = {Balakin, E. I. and Yurku, K. A. and Malsagova, K. A. and Butkova, T. V. and Kaysheva, A. L. and Pustovoit, V. I.},
title = {Machine learning models for stratifying traumatic brain injury severity based on interleukin-6 temporal dynamics},
journal = {Extreme Medicine},
year = {2026},
volume = {28},
number = {2},
pages = {267--276},
doi = {10.47183/mes.2026-438}
}
EUVIMED