Baseline cognition accounts for most predictive information on cognitive recovery during rehabilitation after carbon monoxide poisoning

Background Cognitive change during rehabilitation after carbon monoxide (CO) poisoning may differ from prognosis at acute presentation. We examined whether information recorded in a rehabilitation pathway added predictive information beyond baseline cognition. Methods This single-center retrospective cohort included 200 patients who completed 30 repetitive transcranial magnetic stimulation (rTMS) sessions and serial cognitive assessments between January 1, 2018, and January 31, 2026. Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) scores were assessed before rTMS and after 20 and 30 sessions. A MoCA increase of at least 5 points was the study-defined classification outcome; this threshold has not been validated for CO poisoning. Four exploratory algorithms were evaluated against a baseline-cognition benchmark. Performance was estimated from stratified five-fold out-of-fold predictions, with paired bootstrap comparisons of area under the curve (AUC) and Brier score. Results MMSE increased from 23.73 ± 2.85 to 27.83 ± 2.15 and MoCA from 21.58 ± 3.57 to 27.02 ± 2.78 after 30 sessions (both p < 0.001); 137 patients (68.5%) met the classification criterion. The baseline-cognition-only random forest had an AUC of 0.836 (95% CI 0.758–0.905) and a Brier score of 0.122, compared with 0.860 (95% CI 0.798–0.908) and 0.119 for the expanded model. Neither the paired AUC difference (0.024, 95% CI − 0.013 to 0.064; p = 0.219) nor Brier-score improvement (0.003, 95% CI − 0.008 to 0.014; p = 0.581) was statistically detectable. Conclusion Baseline cognition dominated prediction of a score-change outcome that is mathematically linked to baseline and constrained by ceiling effects. The additional routine measures and coarse imaging categories available here showed no detectable incremental value. These preliminary findings describe selected rehabilitation completers and require external validation before clinical application.