Limited Predictability of Traumatic Intracranial Hemorrhage from Routine Pre-CT Clinical Variables in Older Adults...

Background/Objectives: Traumatic intracranial hemorrhage (tICH) in older adults following low-energy falls (LEF) represents a common yet diagnostically challenging condition in the emergency department (ED), where predicting injury prior to computed tomography (CT) remains difficult. Machine learning (ML) has been proposed to support CT decision-making, but its feasibility using routinely available pre-CT clinical variables in this specific population remains unclear. This study presents a systematic exploratory benchmarking of ML pipeline configurations for pre-CT tICH prediction in a well-defined retrospective cohort of older emergency patients following LEF. Methods: We performed a secondary analysis from a retrospective observational bicentric study from two university hospital EDs, including 2250 patients aged ≥65 years presenting after an LEF and undergoing cranial CT. Clinical data were extracted manually from electronic health records (EHRs). Eighteen pre-CT clinical features retrieved from electronic health records were selected based on routine availability and ≤10% missingness. Overall, 1224 valid ML pipeline configurations, combining nine classification algorithms, six imputation strategies, four class-balancing approaches, and optional hyperparameter tuning, were evaluated using 10-fold stratified cross-validation on a training set. The 20 highest-ranked configurations by cross-validation AUC were then assessed on a previously inspected exploratory hold-out test set (n = 563); training-derived rule-out operating points were evaluable for 17 of these 20, as three tuned SVM configurations lacked stored out-of-fold predictions. Results: tICH prevalence was 7.0% (n = 158). Across the 20 highest-ranked configurations, hold-out AUC ranged from 0.517 to 0.585, with Matthews correlation coefficient near zero and balanced accuracy of approximately 50% throughout, indicating differences in operating point rather than in discriminative ability. Some of these top-ranked pipelines reached higher cross-validation AUC (up to 0.679) but detected no cases at the default 0.5 threshold-an effect of the decision threshold under class imbalance rather than of the models' rank-order discrimination, which was itself limited (hold-out AUC of 0.517-0.585). Conclusions: Despite comprehensive exploratory benchmarking across 1224 ML pipelines, routinely available pre-CT clinical features did not provide sufficient discriminatory signal to develop a clinically useful tICH prediction model in this cohort of CT-imaged older adults following LEF. These findings indicate that none of the evaluated configurations produced clinically adequate performance; this near-chance result persisted across all pipelines and most plausibly reflects a combination of limited feature signal, low outcome prevalence, and a sample size below the level required for reliable model development at this event rate. Future studies should target substantially larger prospective multicenter cohorts and evaluate additional feature domains, including structured clinical examination findings, point-of-care biomarkers, and imaging features.