Acute-Phase T2-FLAIR Radiomics in Acute Carbon Monoxide Poisoning: Deep White Matter Vulnerability and Lucid-Interval Prediction of Delayed Encephalopathy

BACKGROUND AND PURPOSE: Delayed encephalopathy after acute carbon monoxide poisoning (DEACMP) remains difficult to predict during the lucid interval. We tested whether routine acute-phase T2-FLAIR radiomics could detect subvisual deep-white-matter vulnerability and improve lucid-interval DEACMP risk stratification. MATERIALS AND METHODS: We retrospectively analyzed 786 adults with acute carbon monoxide poisoning, including 201 acute-phase T2-FLAIR MRI scans split chronologically into imaging-training and held-out imaging-test sets. Six-compartment radiomics and clinical-radiomic stacking used training-restricted preprocessing and model development. Repeated MRI assessed temporal behavior; blinded visual MRI benchmarking and MRI-training downsampling were used as support analyses. RESULTS: Deep white matter showed the dominant DEACMP radiomic signal: 8/9 atlas-guided candidate bundles were cross-sectionally significant after false-discovery-rate correction, and 7/9 showed repeated-MRI group-level elevation without DEACMP-by-time divergence. Pallidal texture supported a secondary DEACMP-dependent urinary-incontinence phenotype. In the held-out imaging-test set, stacking achieved AUC 0.876, Brier score 0.170, sensitivity 0.800, and specificity 0.839. CONCLUSIONS: Acute-phase T2-FLAIR radiomics delineated a deep-white-matter-dominant imaging phenotype associated with DEACMP risk and a separable DEACMP-dependent pallidal texture phenotype linked to urinary incontinence. Clinical-radiomic stacking improved lucid-interval DEACMP risk stratification, supporting T2-FLAIR radiomics as a research-stage subvisual risk-stratification approach.