Methodological choices influence machine learning performance for predicting intracranial hypertension in traumatic brain injury

Intracranial hypertension (IH) is a major cause of secondary brain injury after traumatic brain injury (TBI) and is typically treated after onset. Machine learning (ML) approaches show promise for IH prediction, but methodological heterogeneity across studies limits comparability and clinical translation. We evaluated how key methodological choices influence ML performance for IH prediction. In this retrospective observational study, intracranial pressure (ICP), arterial blood pressure (ABP), and additional variables were collected from TBI patients in two Dutch hospitals and the CENTER-TBI high-resolution cohort. Six methodological categories were varied: IH definition (ICP ≥ 22 or ≥ 30 mmHg), input data configuration, labelling strategy (IH onset versus continuous labelling), temporal separation of control segments, observation window length (10–600 min), and model class (eXtreme Gradient Boosting, random forest, logistic regression, Gaussian naïve Bayes, and long short-term memory). Models were evaluated using patient-level tenfold cross-validation with area under the receiver operating characteristic curve (AUC) as primary outcome. We included 316 patients. Prediction performance varied substantially across methodological choices. Continuous labelling inflated performance, and optimal observation windows were 240–360 min. Across conditions, eXtreme Gradient Boosting achieved the highest performance (AUC 0.92). These findings highlight the need for standardized methodology and transparent reporting in IH prediction studies.