TITLE:
Clinical Phenotypes of Intracranial Suppurative Infections Identified by Unsupervised Machine Learning
AUTHORS:
Koffi Yves Soress Dongo, Faozo Stéphane Landry Teti, Kouadio Jean-Baptiste Keke, Yao Bernard Fionko, Aderehime Haidara
KEYWORDS:
Intracranial Suppuration, Brain Abscess, Subdural Empyema, Unsupervised Machine Learning, Clinical Phenotypes, Côte d’Ivoire
JOURNAL NAME:
Neuroscience and Medicine,
Vol.17 No.4,
October
14,
2026
ABSTRACT: Background: Intracranial suppurative infections (ISIs) are conventionally classified by anatomical compartment—brain abscess, subdural empyema, extradural empyema—a taxonomy that captures neither the heterogeneity of routes of infection nor the diversity of care pathways. No previous study has examined whether unsupervised learning can redefine clinically meaningful phenotypes of ISI beyond conventional anatomical categories, and none has done so in an African cohort. Methods: We analysed 150 consecutive observations (148 patients) of ISI managed at Bouaké University Hospital, Côte d’Ivoire, between 2016 and 2026. In the primary analysis, nineteen pre-therapeutic variables spanning host factors, symptom-to-admission delay, neurological severity, grouped portal of entry, lesion type and lesion burden were entered into a factor analysis of mixed data (FAMD); the origin of the suppuration, almost collinear with the portal of entry, was removed, and the specification including both was kept as a sensitivity analysis. Phenotypes were derived by K-means clustering on the retained factor coordinates, with Ward clustering on the same Euclidean coordinates and partitioning around medoids (PAM) on Gower distances as concordance checks. The reproducibility of the whole pipeline was assessed by repeated nested cross-validation, imputation, FAMD and K-means being re-estimated in every training set. In-hospital mortality and functional status were analysed as separate outcomes, and the incremental prognostic value of phenotype beyond GCS and age was assessed with optimism correction by bootstrap (1000 replicates). Results: In the primary non-redundant specification, a four-class solution was retained (silhouette 0.296). Its concordance with the dual-coded partition was moderate (adjusted Rand index [ARI] 0.325), and the frontal-interhemispheric sinogenic subdural empyema phenotype was the most robust (28 of 33 observations recovered). PAM on Gower distances identified four groups of 48, 36, 34 and 32 observations (silhouette 0.312; ARI with FAMD-K-means 0.671). Out-of-sample agreement of the complete pipeline was 87.6% ± 4.8% (mean ARI 0.79 ± 0.09). The dual-coded specification, used for detailed characterisation, separated community-acquired brain abscess of ENT-dental or cryptogenic origin (n = 54), delayed post-traumatic suppuration (n = 33; median delay 29 days), sinogenic subdural empyema (n = 33; craniotomy 91%, external ventricular drainage 24%) and post-operative suppuration (n = 30; Gram-negative-enriched microbiology). In-hospital mortality ranged from 5.6% to 18.2% across phenotypes (p = 0.29). Among 126 evaluable survivors, 41 (32.5%) had a GOS ≤ 3, ranging from 19.2% to 40.0% across phenotypes (p = 0.38). Adding phenotype to GCS and age raised the optimism-corrected AUC by 0.009 only, for mortality (0.754 to 0.763) as for functional status (0.794 to 0.803), without gain in Brier score or calibration. Conclusions: Unsupervised learning delineated clinically recognisable ISI phenotypes, organised mainly by the mechanism of infection and the care pathway and corroborated by a microbiological distribution that was never used to build them. Once the dual coding of the mechanism was removed, only the sinogenic subdural empyema phenotype was consistently reproduced; the other groupings remained conditional on the weight given to the mechanism. Phenotype added little prognostic information to admission GCS and age. Phenotype-aware triage, rather than phenotype-based prognostication, is the defensible clinical application of these results, and external validation in multicentre African cohorts is required before implementation.