Machine Learning Models to Differentiate Hemorrhagic from Ischemic Stroke: insights from the International Stroke Trial

Modelos de Aprendizado de Máquina para Diferenciar AVC Hemorrágico de Isquêmico: insights do International Stroke Trial

• Samuel Pedro Pereira Silveira    
• Gustavo del Rio Lima    
• Gustavo Branquinho Alberto    
• Luiza Carolina Moreira Marcolino    
• Larissa Batista Xavier    
• Carlos Umberto Pereira    
• Murillo Martins Correia
• Roberto Alexandre Dezena    

• https://doi.org/10.22290/jbnc.2026.02031047
• J Bras Neurocirur 37 (2):169-183, 2026

Resumo

Introdução: O AVC hemorrágico, embora menos frequente que o isquêmico, causa grande mortalidade e incapacidade. Diferenciar precocemente os subtipos é fundamental. Objetivo: Avaliar modelos de aprendizado de máquina que combinem variáveis clínicas e um marcador radiológico inicial (infarto visível na TC) para classificar AVC hemorrágico vs. isquêmico no banco de dados do International Stroke Trial (IST). Métodos: Foram incluídos 12.229 pacientes (97% isquêmicos; 3% hemorrágicos). Utilizou-se validação cruzada aninhada, múltiplos algoritmos (regressão logística, SVM, floresta aleatória, boosting, MLP, XGBoost) e técnicas de balanceamento (SMOTE, ADASYN, Tomek Links, ENN). Resultados foram calibrados e interpretados por SHAP. Resultados: O melhor desempenho foi do Gradient Boosting com ENN (AUC-ROC 0,746; sensibilidade 0,796; especificidade 0,580; AUC-PR 0,091). A precisão permaneceu baixa (≈0,055), refletindo o desequilíbrio de classes. O valor preditivo negativo foi tipicamente elevado na maioria dos modelos (frequentemente ≥0,98). O preditor mais relevante foi “infarto visível na TC”, seguido por pressão arterial sistólica e déficits neurológicos. Conclusão: Modelos multimodais alcançam discriminação moderada, mas precisão limitada. Podem apoiar triagem e decisão clínica em cenários com recursos restritos, sem substituir a neuroimagem. Validações contemporâneas e prospectivas são necessárias. 

Palavras-chave

Acidente vascular cerebral; Hemorragias intracranianas; AVC isquêmico; Inteligência artificial; Aprendizado de máquina

Abstract

Introduction: Hemorrhagic stroke, although less frequent than ischemic stroke, accounts for a disproportionate share of stroke-related mortality and disability. Early differentiation is critical for management. Objective: To evaluate machine learning models integrating routine clinical variables and early radiological markers (visible infarct on CT) for classifying hemorrhagic vs. ischemic stroke using the International Stroke Trial (IST) dataset. Methods: After exclusions, 12,229 patients were analyzed (11,866 ischemic; 363 hemorrhagic). Nested stratified cross-validation assessed logistic regression, support vector machines, random forest, gradient boosting, multilayer perceptron, and XGBoost are the methodologies used. To address severe class imbalance (~3% hemorrhagic), multiple resampling techniques (SMOTE, ADASYN, Tomek Links, ENN, and hybrids) were compared. Probabilities were calibrated and interpretability evaluated with SHAP values. Results: Gradient Boosting with Edited Nearest Neighbors achieved the best performance (AUC-ROC 0.746; sensitivity 0.796; specificity 0.580; AUC-PR 0.091). Precision was low (≈0.055), reflecting the rarity of hemorrhage. Negative predictive value was typically high for most models (often ≥0.98). SHAP identified “visible infarct on CT” as the dominant predictor, followed by systolic blood pressure and selected neurological deficits. Conclusion: Machine learning models achieve moderate discrimination but limited precision under extreme imbalance. They may complement, but not replace, neuroimaging, particularly in resource-limited settings. External validation in contemporary cohorts remains essential.

Keywords

Stroke; Intracranial hemorrhages; Ischemic stroke; Artificial intelligence; Machine Learning

References

1. Feigin VL, Abate MD, Abate YH, et al. Global, regional, and national burden of stroke and its risk factors, 1990–2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet Neurol. 2024;23(10):9731003. https://doi.org/10.1016/S1474-4422(24)00369-7. PMid:39304265.

2. Parry-Jones AR, Krishnamurthi R, Ziai WC, et al. World Stroke Organization (WSO): global intracerebral hemorrhage factsheet 2025. Int J Stroke. 2025;20(2):145-50. https://doi.org/10.1177/17474930241307876. PMid:39629687.

3. International Stroke Trial Collaborative Group. The International Stroke Trial (IST): a randomised trial of aspirin, subcutaneous heparin, both, or neither among 19 435 patients with acute ischaemic stroke. Lancet. 1997;349(9065):1569-81. https://doi.org/10.1016/ S0140-6736(97)04011-7. PMid:9174558.

4. Issaiy M, Zarei D, Kolahi S, Liebeskind DS. Machine learning and deep learning algorithms in stroke medicine: a systematic review of hemorrhagic transformation prediction models. J Neurol. 2025;272(1):37. https://doi.org/10.1007/s00415-024-12810-6. PMid:39666168.

5. Asadi F, Rahimi M, Daeechini AH, Paghe A. The most efficient machine learning algorithms in stroke prediction: a systematic review. Health Sci Rep. 2024;7(10):e70062. https://doi.org/10.1002/ hsr2.70062. PMid:39355095.

6. Goh B, Bhaskar SMM. Evaluating machine learning models for stroke prognosis and prediction in atrial fibrillation patients: a comprehensive meta-analysis. Diagnostics. 2024;14(21):2391. https:// doi.org/10.3390/diagnostics14212391. PMid:39518359.

7. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. 2002;16:321-57. https://doi.org/10.1613/jair.953.

8. He H, Bai Y, Garcia EA, Li S. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In: Proceedings of the IEEE International Joint Conference on Neural Networks (IJCNN); 2008 Jun 1-8; Hong Kong, China. Piscataway: IEEE; 2008. p. 1322-8.

9. Tomek I. Two modifications of CNN. IEEE Trans Syst Man Cybern. 1976;6(11):769-72.

10. Wilson DL. Asymptotic properties of nearest neighbor rules using edited data. IEEE Trans Syst Man Cybern. 1972;2(3):408-21. https:// doi.org/10.1109/TSMC.1972.4309137.

11. Batista GE, Prati RC, Monard MC. A study of the behavior of several methods for balancing machine learning training data. SIGKDD Explor. 2004;6(1):20-9. https://doi.org/10.1145/1007730.1007735.

12. Lemaître G, Nogueira F, Aridas CK. Imbalanced-learn: A Python toolbox to tackle the curse of imbalanced datasets in machine learning.

J Mach Learn Res. 2017;18(1):559-63.

13. Berkson J. Application of the logistic function to bio-assay. J Am Stat Assoc. 1944;39(227):357-65.

14. Breiman L. Random forests. Mach Learn. 2001;45(1):5-32. https:// doi.org/10.1023/A:1010933404324.

15. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20(3):273-97. https://doi.org/10.1023/A:1022627411411.

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

17. Friedman JH. Greedy function approximation: A gradient boosting machine. Ann Stat. 2001;29(5):1189-232. https://doi.org/10.1214/ aos/1013203451.

18. Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating errors. Nature. 1986;323(6088):533-6. https:// doi.org/10.1038/323533a0.

---

1Faculty of Medicine, Universidade Federal do Triângulo Mineiro, Uberaba, MG, Brazil.

2Center for Mathematics, Computing and Cognition – CMCC, Universidade Federal do ABC, Santo André, SP, Brazil.

3Hospital das Clínicas, Universidade Federal do Triângulo Mineiro, Uberaba, MG, Brazil.

4Neurosurgery Division, Universidade Federal do Sergipe – UFS, Aracaju, SE, Brazil.

5Neurosurgery Division, Universidade Federal do Triângulo Mineiro, Uberaba, MG, Brazil.

6Discipline of Neurosurgery, Hospital das Clínicas, Universidade Federal do Triângulo Mineiro, Uberaba, MG, Brazil.

 

Received Sep 2, 2025

Corrected Feb 23, 2026

Accepted Feb 24, 2026