Soft-voting classifier for outcome prediction in traumatic brain injury patients with clear consciousness: model development and error pattern analysis.

Accurate outcome prediction enables stratified treatment in traumatic brain injury (TBI), a field that has yet to achieve a major therapeutic breakthrough. We aimed to develop a model to predict unfavorable in-hospital outcomes in TBI patients presenting with clear consciousness— a particularly challenging and previously unexplored target. We developed a soft-voting ensemble machine learning (ML) model combining six algorithms, including transformer-based architectures, to predict unfavorable discharge outcomes, trained on 349 cases with 17 clinical parameters and tested on 90. To develop the model, we used data from patients aged 10 years or older with TBI who were admitted to Japanese hospitals with an initial Glasgow Coma Scale score of 15. Additionally, permutation importance and misclassification analyses were conducted. Patients with unfavorable outcomes at discharge accounted for 76 of 439 (17.3%). The ML model outperformed the statistical model in sensitivity (43.8% vs. 37.5%, p = 0.003), specificity (97.3% vs. 95.9%, p = 0.01), and accuracy (87.8% vs. 85.6%, p < 0.001). The area under the receiver operating characteristic curve showed no significant difference (0.843 vs. 0.796, p = 0.4). Permutation importance analysis revealed that the ML model assigned greater importance to D-dimer and systolic blood pressure compared to the logistic regression model. Misclassified cases had fewer acute subdural hematomas and more normal hemoglobin levels than did correctly predicted cases. Prospective validation and integration into emergency care workflows are necessary for clinical implementation of this newly developed prediction model for TBI patients with clear consciousness.