This retrospective study across three US academic hospitals evaluated automated discharge CPC classification in adults comatose after cardiac arrest. Good outcomes were defined as CPC 1–3 and poor outcomes as CPC 4–5. An NLP logistic regression model trained on 249 patients was evaluated in 108 holdout patients. Using hospitalization-wide notes, accuracy was 81% and AUROC 0.90; using first-24-hour notes, these fell to 74% and 0.83. Precision for poor outcomes was higher, particularly among patients undergoing withdrawal of life-sustaining therapy, highlighting the importance of documentation context.
Automated discharge CPC classification showed promising holdout performance, but these results do not establish reliable early prognostication independent of clinician documentation or withdrawal-of-life-sustaining-therapy decisions.
Study limitation: Performance was evaluated in a 108-patient holdout set, and early documentation may reflect prognostic framing as well as patient biology.
AI-generated from the published abstract with automated source checks. For research discovery; consult the original paper before applying findings.