EEG-based machine-learning prognostication in comatose patients with indeterminate outcome after cardiac arrest

Background: Neurological prognostication after cardiac arrest remains challenging despite multimodal approaches recommended by the European Resuscitation Council and the European Society of Intensive Care Medicine (ERC/ESICM). Prognostic certainty is achieved in only a subset of comatose patients, leaving a substantial proportion with an indeterminate prognosis. In this population, the incremental prognostic value of electroencephalography beyond established predictors remains insufficiently characterized. Methods: We conducted a prospective 5-year (2021–2025) single-center observational study of adult comatose patients after cardiac arrest admitted to the Intensive Care Unit (ICU) of the University Hospitals of Geneva. Patients presenting with fewer than two ERC/ESICM poor outcome factors were included. Continuous electroencephalogram (EEG) recordings obtained within 72 h after cardiac arrest were analyzed. Quantitative EEG features reflecting spectral power, functional connectivity, signal complexity, and background continuity were extracted and combined with automated estimates of rhythmic and periodic patterns derived from a previously validated neural network. These features were integrated into a random forest classifier to predict neurological outcome at 3–6 months, dichotomized as good (Cerebral Performance Category 1–2) or poor (3–5). Model performance was evaluated using repeated stratified cross-validation, and feature contributions were assessed using Shapley value analysis. Results: Of a total of 313 patients, 77 met the inclusion criteria (mean age 61, 21 female), of whom 35 (45%) achieved a good neurological outcome. The model demonstrated good discriminative performance, with a mean Receiver Operating Characteristic – Area Under the Curve (ROC-AUC) of 0.80 and a Precision–Recall Area Under the Curve (PR-AUC of 0.83). Delta-band electroencephalography features accounted for most of the predictive contribution. Higher delta functional connectivity, higher delta relative power, and higher probability of periodic discharges were associated with poor outcome, whereas preserved relative alpha power and higher probability of rhythmic delta activity were associated with good outcome. Conclusions: In comatose patients after cardiac arrest with ERC/ESICM guideline-defined indeterminate prognosis, an interpretable EEG-based prognostic model provides clinically relevant information. This approach may complement current multimodal prognostication strategies by refining early risk stratification in a population characterized by substantial prognostic uncertainty. Graphical abstract: Supplementary Information: The online version contains supplementary material available at 10.1186/s13054-026-05982-2.