Prediction of Violence Against Women in Rio Grande do Norte state, Brazil

The Maria da Penha Law is a landmark piece of legislation in the criminalization of violence against women in Brazil, establishing clear definitions of what constitutes physical, psychological, sexual, economic, and moral violence (INSTITUTO MARIA DA PENHA, 2026). In addition, it has also contributed to defining gender-based violence as a responsibility of the State. Subsequently, laws were enacted to mandate that the Unified Health System (SUS)—which provides free access to healthcare for the Brazilian population—offer mandatory care to victims of violence against women. Furthermore, it was mandated that cases handled by health services must be reported to the police (Rio Grande do Norte, 2020), and submitted through the Notifiable Diseases Information System (SINAN). Today, there is a great deal of research linking the main factors identified among the profiles of women who have experienced violence in Brazil and around the world. However, it is a fact that the social context of each region is also a determining factor in the profile of women in that state. For instance, Shimaoka et al. (2026) identified Rio Grande do Norte as one of the states with the highest proportional rate of hospital admissions due to violence against women; their research utilized public data from the Hospital Admissions System (SIH), analyzing clinical and demographic profiles. This study aims to predict the type of violence against women based on the descriptors found in the SINAN database (with 160 attributes) and the Observatório da Mulher Potiguar (OMP) database (with 19 attributes). First, in the context of machine learning, we evaluated baseline supervised algorithms such as k-NN, Decision Tree, Naive Bayes, and MultiLayerPerceptron, as well as the following ensemble methods: Bagging, Boosting, Stacking, Random Forest, and XGBoost—the latter were trained using the best dataset identified in the analysis of the supervised algorithms. The analytical pipeline incorporated decomposition techniques (One-vs-One and One-vs-Rest), methods for addressing class imbalance (Synthetic Minority Oversampling Technique), and feature selection (Feature Selection via Random Forest). These techniques were used to refine the processing of data from the databases. Furthermore, to evaluate the impact of attribute space complexity, feature selection and reduction techniques were performed, resulting in three reduced datasets for each ecosystem (SINAN and OMP). The central hypothesis (H1) of this study is that the demographic, epidemiological, and socioeconomic attributes available in the SINAN database are sufficient for predictive modeling, reinforcing the patterns widely documented in the literature. Furthermore, the second hypothesis (H2) posits that the attributes specific to the OMP reveal a distinct profile that complements the health data, while also questioning the sufficiency of these attributes for multilabel prediction. Empirical results showed that the Decision Tree model achieved the best individual performance (86%) and the Stacking ensemble with 20 classifiers achieved the best overall result (96%). As a result, H1 was confirmed, highlighting that the demographic and socioeconomic attributes in the SINAN database were important for classification. However, H2 was not supported, leading to the conclusion that the OMP database does not contain sufficient attributes to predict the type of violence suffered by women. Furthermore, the results also show that the application of attribute reduction and feature selection contributed significantly to the Naive Bayes model in the OMP dataset. The procedure of transferring the MLP configuration from Reduced Dataset 1 to Reduced Dataset 2 and Original also contributed significantly to the algorithm’s performance. Finally, the case study using SMOTE showed that the MLP was the algorithm that benefited most from the SINAN database. For future work, there is potential to develop new hypotheses based on the available data that has not been sufficiently utilized in the Original SINAN database.