ORIGINAL
Introdução: Modelos prognósticos de traumatismo cranioencefálico (TCE), como o IMPACT, derivam de coortes moderadas-a-graves e não se aplicam às lesões leves, que são predominantes. Objetivo: Desenvolver um modelo de aprendizado de máquina para prever o desfecho funcional em seis meses no espectro de gravidade do TCE. Métodos: Dos 2.545 adultos do TRACK-TBI, 1.749 tinham desfecho válido em seis meses. Um XGBoost usou demografia, exame clínico, 13 achados tomográficos e cinco biomarcadores sanguíneos para prever desfecho desfavorável (GOSE-TBI 1–4), com validação cruzada e comparação ao IMPACT no subgrupo moderado a grave. Resultados: O desfecho desfavorável ocorreu em 242 pacientes (13,8%). Calibração boa (Brier 0,051; slope 0,95). AUROC intra-estrato: 0,85 (leve) e 0,89 (moderado a grave); a agrupada (0,944), inflada pela gravidade. A discriminação aumentou a cada modalidade (AUPRC 0,751→0,787→0,814) e superou o IMPACT neste benchmark intra-coorte (ΔAUROC +0,108 Core; +0,070 Extended), permanecendo mais acurada após recalibração na coorte (Brier 0,134 vs 0,189); decomposição pré-especificada atribuiu a vantagem aos preditores multimodais. Conclusão: Um único modelo calibrado e interpretável prediz o desfecho em seis meses no espectro de gravidade do TCE, discrimina melhor que o IMPACT e fornece estimativas de risco calibradas para a maioria leve, para a qual o IMPACT não gera nenhuma. A recuperação incompleta, desfecho prevalente no TCE leve, não foi modelada; validação externa é necessária antes do uso clínico.
Introduction: Prognostic models for traumatic brain injury (TBI), such as IMPACT, were derived in moderate-to-severe cohorts and do not apply to predominant mild injuries. Objective: To develop a machine-learning model predicting six-month functional outcome across the TBI severity spectrum. Methods: Of 2,545 TRACKTBI adults, 1,749 had a valid six-month outcome. An XGBoost model used demographics, clinical examination, 13 CT findings, and five blood biomarkers to predict unfavorable outcome (GOSE-TBI 1–4), evaluated by cross-validation and benchmarked against IMPACT in the moderate-to-severe subset. Results: Unfavorable outcomes occurred in 242 patients (13.8%). Calibration was good (Brier 0.051; slope 0.95). Within-severity AUROC was 0.85 (mild) and 0.89 (moderate-to-severe); the pooled 0.944 was severity-inflated. Discrimination rose with each added modality (AUPRC 0.751→0.787→0.814) and exceeded IMPACT in this in-cohort benchmark (ΔAUROC +0.108 Core; +0.070 Extended), remaining more accurate after in-cohort recalibration (Brier 0.134 vs 0.189); a pre-specified decomposition attributed this advantage to the multimodal predictor set. Conclusion: A single calibrated, interpretable model predicts six-month outcome across the TBI severity spectrum, discriminates better than IMPACT, and provides calibrated risk estimates for the mild-TBI majority, for which IMPACT produces none. The prevalent mild-TBI endpoint (incomplete recovery) was not modeled, and external validation is required before clinical use.
Lancet Neurol. 2017;16(12):987-1048. https://doi.org/10.1016/S14744422(17)30371-X. PMid:29122524.
2. Steyerberg EW, Mushkudiani N, Perel P, et al. Predicting outcome after traumatic brain injury: development and international validation of prognostic scores based on admission characteristics. PLoS Med.
2008;5(8):e165, discussion e165. https://doi.org/10.1371/journal. pmed.0050165. PMid:18684008.
3. Perel P, Arango M, Clayton T, et al. Predicting outcome after traumatic brain injury: practical prognostic models based on large cohort of international patients. BMJ. 2008;336(7641):425-9. https://doi.org/10.1136/bmj.39461.643438.25. PMid:18270239.
4. Nelson LD, Temkin NR, Dikmen S, et al. Recovery after mild traumatic brain injury in patients presenting to US level I trauma centers: a Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) study. JAMA Neurol. 2019;76(9):1049-59. https://doi.org/10.1001/jamaneurol.2019.1313. PMid:31157856.
5. Yue JK, Lee YM, Sun X, et al. Performance of the IMPACT and CRASH prognostic models for traumatic brain injury in a contemporary multicenter cohort: a TRACK-TBI study. J Neurosurg. 2024;141(2):417-29. https://doi.org/10.3171/2023.11.JNS231425. PMid:38489823.
6. Nelson LD, Magnus BE, Yue JK, et al. Beyond mild, moderate, and severe traumatic brain injury: modelling severity from clinical, neuroimaging, and blood-based indicators. EBioMedicine. 2025;121:106001. https://doi. org/10.1016/j.ebiom.2025.10600 . PMid:41192220.
7. Bazarian JJ, Biberthaler P, Welch RD, et al. Serum GFAP and UCH-L1 for prediction of absence of intracranial injuries on head CT (ALERT-TBI): a multicentre observational study. Lancet Neurol. 2018;17(9):782-9. https://doi.org/10.1016/S1474-4422(18)30231-X. PMid:30054151.
8. Gravesteijn BY, Nieboer D, Ercole A, et al. Machine learning algorithms performed no better than regression models for prognostication in traumatic brain injury. J Clin Epidemiol. 2020;122:95-107. https://doi.
org/10.1016/j.jclinepi.2020.03.005. PMid:32201256.
9. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. https://doi.org/10.1136/bmj-2023-078378. PMid:38626948.
10. Nelson LD, Magnus BE, Yue JK, et al. TRACK-TBI Acute Imaging, Blood Biomarker, and Clinical Indicators of TBI Severity in Humans [dataset]. Open Data Commons for Traumatic Brain Injury; 2025. ODC-TBI:1168. https://doi.org/10.34945/F5088C.
11. Wilson JT, Pettigrew LE, Teasdale GM. Structured interviews for the Glasgow Outcome Scale and the Extended Glasgow Outcome Scale: guidelines for their use. J Neurotrauma. 1998;15(8):573-85. https://doi. org/10.1089/neu.1998.15.573. PMid:9726257.
12. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (KDD ’16);
2016; New York. New York: ACM; 2016. p. 785-94. https://doi.
org/10.1145/2939672.2939785.
13. Van Calster B, McLernon DJ, van Smeden M, et al. Calibration: the Achilles heel of predictive analytics. BMC Med. 2019;17(1):230. https://doi.org/10.1186/s12916-019-1466-7. PMid:31842878.
14. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565-74. https://doi.org/10.1177/0272989X06295361. PMid:17099194.
15. Riley RD, Ensor J, Snell KIE, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. https://doi.org/10.1136/bmj.m441. PMid:32188600.
16. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Guyon I, von Luxburg U, Bengio S, et al., editors. Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems; 2017 Dec 4-9; Long Beach, CA. Red Hook (NY): Curran Associates, Inc.; 2017. p. 4765-74.
1Faculty of Medicine, Universidade Federal do Triângulo Mineiro – UFTM, Uberaba, MG, Brazil.
2Center for Mathematics, Computing and Cognition – CMCC, Universidade Federal do ABC – UFABC, Santo André, SP, Brazil.
3Neurosurgery Division, Clinics Hospital, Universidade Federal do Triângulo Mineiro – UFTM, Uberaba-MG, Brazil.
4Department of Neurosurgery, Northwell Health, New Hyde Park, NY, USA.
5Neurosurgery Division, Universidade Federal do Sergipe – UFS, Aracaju, SE, Brazil.
Received Jul 11, 2026
Accepted Jul 21, 2026
