Machine Learning Triage Models Post Total-body Gamma Irradiation

Background: In nuclear emergencies, the rapid estimation of radiation exposure is critical for informed medical treatment. Traditional methods like lymphocyte dynamics and chromosome aberration analysis are time-consuming and require skilled personnel, indicating a pressing need for efficient biodosimetry tools.

Method: This study employed targeted lipidomics to analyze lipid changes in rat plasma following total body irradiation (TBI) at doses of 0, 5, and 10 Gy over 14 days. Machine learning algorithms, including LASSO and decision trees, were utilized to develop triage models using seven lipid biomarkers previously identified.

Results: The LASSO model exhibited an F1 score of 0.93 in the training set, 0.82 in the test set, and reached an AUC of 0.99 in independent validation, indicating a highly effective classification ability between 0 Gy and >0 Gy. On days 1, 7, and 14 post-irradiation, significant temporal variations in lipid profiles were noted, with specific lipids demonstrating dose-dependent trends.

Conclusion: This research highlights the innovative use of lipid biomarkers combined with machine learning to enhance biodosimetry, suggesting that the LASSO algorithm can serve as a viable triage tool for rapid radiation dose assessment post-exposure. Limitations include the need for further validation in larger clinical cohorts to establish robustness.

Original citation address: https://www.besjournal.com/en/article/doi/10.3967/bes2025.157