Aims: The therapeutic window for preventing delayed encephalopathy after carbon monoxide poisoning (DEACMP) remains unclear. We aimed to define this temporal risk relationship and establish an intervention threshold using machine learning.
Methods: In this multicenter retrospective cohort study (n = 1755), a gradient boosting model for predicting DEACMP was developed (n = 1654) and externally validated (n = 101). Performance was assessed using the area under the receiver operating characteristic curve (AUC) and interpreted using Shapley Additive exPlanations (SHAP).
Results: The exposure-to-treatment interval was the most powerful predictor of DEACMP risk. Intervention within four hours emerged as the most critical variable influencing risk (SHAP analysis). The model demonstrated robust discrimination in the training (AUC = 0.944, 95% CI, 0.926-0.960), internal validation (AUC = 0.849, 95% CI, 0.785-0.905), and external validation (AUC = 0.872, 95% CI, 0.772-0.946) sets.
Conclusion: Treatment delay is the primary modifiable risk factor for DEACMP following CO poisoning. The identified critical four-hour therapeutic window provides the first quantitative, evidence-based benchmark to inform clinical guidelines and optimize emergency response strategies aimed at preventing delayed neurological sequelae.
Keywords: carbon monoxide poisoning; cohort study; delayed encephalopathy; machine learning; risk stratification; temporal factors.
© 2026 The Author(s). CNS Neuroscience & Therapeutics published by John Wiley & Sons Ltd.