Predictive Signatures of Violence Against Women in Brazilian Health Records: A Comparative Study Using Classifier Committees and Logistic Regression

• Rafaela M. Lucca Universidade de São Paulo (USP) https://orcid.org/0009-0001-8161-6311
• Roseli A. F. Romero Universidade de São Paulo (USP) https://orcid.org/0000-0001-9366-2780

DOI: https://doi.org/10.5753/sbbd_estendido.2026.249686

Resumo

Violence against women is a serious public health problem in Brazil, yet predictive studies on its notified records remain scarce, and most existing work treats violence as a single phenomenon. This paper investigates whether the three main types of violence registered in the Brazilian Notifiable Diseases Information System (SINAN), namely physical, psychological, and sexual, have distinct predictive signatures. We work with 2.22 million notifications from women between 2018 and 2024 and compare a regularized logistic regression against two classifier committees, Random Forest and LightGBM, evaluated under a temporal train and test split. Our analysis prioritizes interpretability and calibration over raw performance. Results show that each type of violence has its own distinguishable signature, the non-linear gain over the linear model is marginal, and class balancing substantially degrades probability calibration in imbalanced outcomes. The model fails to detect roughly 44.9 percent of psychological violence cases, exposing how automated tools tend to mirror the under-reporting already present in the source data.

Palavras-chave: Conformal prediction, Local coverage, Public health, Violence against women, Logistic regression, SINAN

Referências

Bernardo, A. M. C. S. and Rodrigues, A. L. L. d. S. M. (2024). Violência sexual contra criança e adolescente na amazônia: um panorama no arquipélago do marajó. Fonte Segura, Fórum Brasileiro de Segurança Pública.

Breiman, L. (2001). Random forests. Machine Learning, 45(1):5–32.

Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1–3.

Coelho, M. d. M. F., Oriá, M. O. B., Tamboril, A. C. R., Moreira, W. C., Oliveira, B. A. d., Martins, M. C., et al. (2025). Temporal evolution and projections of violence against women in Brazil: Scenarios from 2013 to 2033. Journal of Interpersonal Violence, page 08862605251355969.

González-Prieto, Á., Brú, A., Nuño, J. C., and González-Álvarez, J. L. (2021). Machine learning for risk assessment in gender-based crime. arXiv preprint arXiv:2106.11847.

Harrell, F. E. (2001). Regression modeling strategies: with applications to linear models, logistic regression, and survival analysis, volume 608. Springer.

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., and Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30.

Kozhevnikova, S., Yukhnenko, D., Scola, G., and Fazel, S. (2025). Machine learning for violence prediction: a systematic review and critical appraisal. arXiv preprint arXiv:2511.23118.

Lima, G. C. C., Passos, C. M. d., Pinheiro, A. L. S., Ribeiro, Í. J. S., and Maia, E. G. (2025). Temporal trend and epidemiological profile of notifications of violence against women in Brazil: 2014–2023. Epidemiologia e Serviços de Saúde, 34:e20240475.

Souto, H., Mello, R., and Furtado, A. (2019). An acoustic scene classification approach involving domestic violence using machine learning. In Encontro Nacional de Inteligência Artificial e Computacional (ENIAC), pages 705–716. SBC.

Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1):267–288.

Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society Series B: Statistical Methodology, 67(2):301–320.