Admission Glasgow Coma Scale Captures Most of the Measurable Prognostic Signal in Intracranial Suppurative Infections: An Explainable Machine-Learning Cohort Study

Abstract

Background.
Prognostic studies rank factors by the p value, which measures the compatibility of an association with the null hypothesis rather than the contribution of a variable to prediction. The actual added value of machine-learning models is rarely assessed against the best single clinical predictor.

Objective.
To quantify, using several complementary definitions of importance, the concentration of the predictive signal among the variables available at admission, and to assess the added value of multivariable models relative to a prespecified clinical comparator.

Methods.
Retrospective cohort of 150 admission episodes corresponding to 148 patients managed for intracranial suppurative infection at Bouaké University Hospital between 1 January 2016 and 31 January 2026; 142 episodes had an evaluable Glasgow Outcome Scale at discharge. Primary outcome: unfavourable functional outcome at discharge (GOS ≤ 3). Twenty-seven admission variables were entered into four prespecified algorithms — penalised logistic regression, random forest, XGBoost, LightGBM — compared with a model containing the Glasgow Coma Scale (GCS) alone, prespecified as the clinical comparator. The episode was the unit of prediction and the patient the unit of partitioning and resampling. Out-of-fold probabilities from 20 repetitions of grouped stratified 5-fold cross-validation were averaged; uncertainty was estimated by patient-level paired percentile bootstrap (1000 replications), the principal comparative inference. Discrimination, probabilistic accuracy, calibration and net benefit were assessed jointly, and importance was quantified by SHAP values, permutation importance and mutual information.

Results.
Fifty-seven of the 142 episodes (40.1%) had an unfavourable outcome, including 16 deaths. GCS alone achieved an AUC of 0.896 (95% bootstrap CI 0.834–0.945) and a Brier score of 0.110, versus 0.849 to 0.869 and 0.134 to 0.146 for the four multivariable models. No model showed superior performance: paired Brier score differences excluded zero for all four comparisons, and AUC differences excluded zero versus penalised logistic regression (+0.047; +0.007 to +0.093) and LightGBM (+0.036; +0.003 to +0.073). GCS alone offered the highest net benefit across the range of thresholds explored. All three importance approaches ranked GCS first, but the estimated magnitude of concentration varied markedly with the metric. GCS ranked first in all 1000 internal resamples; beyond that, ranks were unstable. After selection recalculated within each training fold, the AUC fell from 0.903 with one variable to 0.854 with twenty-seven. The advantage of GCS persisted after exclusion of episodes with GCS 3–8 (0.851 versus 0.797–0.821) and within the range of clinical uncertainty, GCS 9–14 (0.769 versus 0.671–0.736). A prespecified parsimonious clinical model combining GCS, collection size and midline shift reached 0.911, without the difference from GCS alone being distinguishable from zero.

Conclusion.
In this single-centre cohort, admission GCS captured most of the measurable predictive signal contained in the baseline variables, and additional algorithmic complexity did not improve validated performance. These observations support systematically comparing any complex model with a prespecified simple clinical predictor. External evaluation is required before any clinical implementation or methodological generalisation.