Using Machine Learning to Predict Outcomes After Traumatic Brain Injury

Resumo

Introdução: O traumatismo cranioencefálico (TCE) apresenta alta morbimortalidade. Métodos prognósticos tradicionais têm limitações, e o aprendizado de máquina (ML) pode melhorar a acurácia preditiva. Objetivo: Avaliar modelos de ML na predição de mortalidade hospitalar em pacientes com TCE. Métodos: Estudo retrospectivo com 745 pacientes do HC-UFTM (2007–2017), dos quais 169 (22,7%) morreram. Foram testados seis modelos: Rede Neural, Random Forest, XGBoost, Gradient Boosting, Regressão Logística e Máquina de vetores de suporte (SVM). As métricas incluíram acurácia, sensibilidade, especificidade, valor preditivo positivo (VPP), valor preditivo negativo (VPN) e área sob a curva ROC (AUC), com validação cruzada estratificada (10 dobras). Resultados: O Gradient Boosting teve maior acurácia (0,820 ± 0,049), especificidade (0,846 ± 0,091) e VPP (0,629 ± 0,147). A Regressão Logística apresentou maior sensibilidade (0,805 ± 0,057) e VPN (0,931 ± 0,018). O SVM alcançou a maior AUC (0,863 ± 0,021), indicando melhor discriminação. A escala de coma de Glasgow na admissão foi o preditor mais relevante, seguida pelo tempo de internação, idade, creatinina e leucócitos. Conclusão: Os modelos de ML demonstraram bom desempenho na predição da mortalidade hospitalar no TCE. Gradient Boosting destacou-se em acurácia e especificidade; Regressão Logística, em sensibilidade; e SVM, na AUC.

Palavras-chave

Inteligência artificial; Lesões cerebrais traumáticas; Mortalidade hospitalar; Modelos logísticos; Machine learning; Random forest; Máquina de vetores de suporte

Abstract

Introduction: Traumatic Brain Injury (TBI) carries high morbidity and mortality. Traditional prognostic methods have limitations, and machine learning (ML) may improve predictive accuracy. Objective: To evaluate ML models for predicting in-hospital mortality in TBI patients. Methods: A retrospective study included 745 patients treated at HC-UFTM (2007–2017), with 169 deaths (22.7%). Six ML models were tested: Neural Network, Random Forest, XGBoost, Gradient Boosting, Logistic Regression, and Support Vector Machine (SVM). Performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the ROC curve (AUC), using 10-fold stratified cross-validation. Results: Gradient Boosting achieved the highest accuracy (0.820 ± 0.049), specificity (0.846 ± 0.091), and PPV (0.629 ± 0.147). Logistic Regression had the best sensitivity (0.805 ± 0.057) and NPV (0.931 ± 0.018). SVM achieved the highest AUC (0.863 ± 0.021), indicating superior discrimination. The Glasgow Coma Scale at admission was the strongest predictor, followed by hospital stay, age, creatinine, and leukocyte count. Conclusion: ML models demonstrated strong predictive performance for in-hospital mortality in TBI. Gradient Boosting excelled in accuracy and specificity, Logistic Regression in sensitivity, and SVM in AUC. Incorporating temporal data and biomarkers may further enhance predictions.

Keywords

Artificial intelligence; Brain injuries, traumatic; Hospital mortality; Logistic models; Machine learning; Random forest; Support vector machine

References

1. National Academies of Sciences, Engineering, and Medicine. Evaluation of the disability determination process for traumatic brain injury in veterans. Washington, D.C.: National Academies Press; 2019. http://doi.org/10.17226/25317.

2. Ribeiro Fernandes RN, Silva M. Epidemiology of traumatic brain injury in Brazil. Braz Neurosurg. 2013;32(3):136-42. http:// doi.org/10.1055/s-0038-1626005.

3. Almeida CER, Sousa JL Fo. Traumatic brain injury

epidemiology in Brazil. World Neurosurg. 2016;87:540-7. http:// doi.org/10.1016/j.wneu.2015.10.020. PMid:26485419.

4. Magalhães ALG, Barros JLVM, Cardoso MGF, et al. Traumatic brain injury in Brazil: an epidemiological study and systematic review of the literature. Arq Neuropsiquiatr. 2022;80(4):410-23. http:// doi.org/10.1590/0004-282x-anp-2021-0035. PMid:35476074.

5. Amorim RL, Oliveira LM, Malbouisson LM, et al. Prediction of Early TBI mortality using a machine learning approach in a LMIC population. Front Neurol. 2020;10:1366. http://doi.org/10.3389/fneur.2019.01366. PMid:32038454.

6. Uparela-Reyes MJ, Villegas-Trujillo LM, Cespedes J, Velásquez-Vera M, Rubiano AM. Usefulness of artificial intelligence in traumatic brain injury: a bibliometric analysis and mini-review. World Neurosurg. 2024;188:8392. http://doi.org/10.1016/j.wneu.2024.05.065. PMid:38759786.

7. Ferreira AMM, Santos JS, Marim AA, et al. Prognostic value of serum creatinine in traumatic brain injury. JBNC -. J Bras Neurocir. 2021;32(3):229-41. http://doi.org/10.22290/jbnc.v32i3.1976.

8. Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12(85):2825-30.

9. Maas AIR, Menon DK, Adelson PD, et al. Traumatic brain injury: integrated approaches to improve prevention, clinical care, and research. Lancet Neurol. 2017;16(12):987-1048. http://doi.org/10.1016/ S1474-4422(17)30371-X. PMid:29122524.

10. 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. http://doi.org/10.1371/journal.pmed.0050165. PMid:18684008.

11. Rosenfeld JV, Maas AI, Bragge P, Morganti-Kossmann MC, Manley GT, Gruen RL. Early management of severe traumatic brain injury. Lancet. 2012;380(9847):1088-98. http://doi.org/10.1016/S0140-6736(12)60864-2.

13. Wang R, Zhang J, Xu J, He M, Xu J. Incidence and Burden of Acute Kidney Injury among Traumatic Brain-Injury Patients. Risk Manag Healthc Policy. 2021;14:4571-80. http://doi.org/10.2147/RMHP. S335150. PMid:34795542.

14. Pesonen A, Ben-Hamouda N, Schneider A. Acute kidney injury after brain injury, does it exist? Minerva Anestesiol. 2021;87(7):823-7. http://doi.org/10.23736/S0375-9393.20.14991-5. PMid:33054019.

15. Mao B, Feng L, Lin D, et al. The predictive role of systemic inflammation response index in the prognosis of traumatic brain injury: a propensity score matching study. Front Neurol. 2022;13:995925. http://doi.org/10.3389/fneur.2022.995925. PMid:36408504.

16. Matsuo K, Aihara H, Hara Y, et al. Machine learning to predict three types of outcomes after traumatic brain injury using data at admission: a multi-center study for development and validation. J Neurotrauma. 2023;40(15-16):1694-706. http://doi.org/10.1089/ neu.2022.0515. PMid:37029810.

17. Song J, Shin SD, Jamaluddin SF, et al. Prediction of mortality among patients with isolated traumatic brain injury using machine learning models in asian countries: an international multi-center cohort study. J Neurotrauma. 2023;40(13-14):1376-87. http://doi.org/10.1089/ neu.2022.0280. PMid:36656672.

18. Abujaber A, Fadlalla A, Gammoh D, Abdelrahman H, Mollazehi M, El-Menyar A. Prediction of in-hospital mortality in patients with post traumatic brain injury using National Trauma Registry and Machine Learning Approach. Scand J Trauma Resusc Emerg Med. 2020;28(1):44. http://doi.org/10.1186/s13049-020-00738-5.

PMid:32460867.