Explainable machine learning on weighted connectivity networks across frequencies for outcome prediction in comatose patients
Abstract
Background
Accurate early prediction of neurological outcomes in comatose patients after cardiac arrest is critical for guiding therapeutic decisions and improving individualized care. Current electroencephalographic (EEG) approaches typically rely on threshold-based or binarized measures of functional connectivity, which may overlook key subtleties of brain dynamics and limit clinical interpretability. We aimed to develop an explainable machine learning framework using weighted EEG connectivity and topological network features to improve early outcome prediction after cardiac arrest.
Methods and Findings
We analyzed EEG recordings from comatose patients within the first 24 hours after cardiac arrest, recruited from multiple intensive care units across Switzerland. Weighted functional connectivity networks were computed using the debiased weighted phase-lag index (dwPLI) across 1–40 Hz. We extracted multi-frequency topological metrics describing global integration and local segregation, and used these as features in an explainable machine learning classifier to predict long-term neurological outcome. Model interpretation was performed using Shapley additive explanations (SHAP). The classifier achieved up to 95% accuracy in distinguishing patients with favourable versus unfavourable outcomes, matching or surpassing existing approaches. SHAP analyses identified delta-band features, particularly path length and clustering coefficient, as the most informative predictors, highlighting differences in large-scale integration and local segregation between outcome groups. Main limitations include the moderate cohort size and the dependency on expert-guided artefact rejection during EEG preprocessing.
Conclusions
Our findings demonstrate that weighted network topology and multifrequency EEG analysis provide valuable, interpretable biomarkers of coma outcome. The proposed explainable AI framework offers a transparent, quantitative, and clinically meaningful approach for early neurological prognostication after cardiac arrest and may inform the design of future decision-support tools in critical care neurology.
Author summary
After a cardiac arrest, many patients remain unconscious in intensive care. Families and clinicians face difficult decisions, yet early and reliable signs of recovery are scarce. In this study, we analyze routine bedside EEG recordings by viewing the brain as a network. We do not study individual connections. Instead, we transform the data into network topological features, simple summary measures of the network’s overall shape and organization. Using these features across a range of slower and faster brain rhythms, we train a transparent prediction model to estimate each patient’s chance of regaining consciousness. For each patient, the model indicates which network features and which rhythm ranges most influenced its estimate, in clear figures that clinicians can examine. We find that these network summaries add useful information to current bedside assessments and may help structure earlier, more consistent conversations about prognosis. We present a step-by-step analysis designed with clinical workflows in mind. Our goal is to support, not replace, medical judgment, and to help build explainable tools for disorders of consciousness.
Related articles
Related articles are currently not available for this article.