Unifesp Uses AI on SUS Hospital Records to Map Domestic Violence Against Women

Researchers at the Federal University of São Paulo (Unifesp) have used artificial intelligence on public health system records to map hospital admissions of women who were victims of violence, Band reported on October 24. According to the broadcaster, the work drew on about 190 million records and clinical notes from Brazil's Unified Health System (SUS) between 2008 and 2023 and points to an average of 15 hospital admissions a day of women who suffered domestic violence.

The study, published in the journal Ciência & Saúde Coletiva by the university's Department of Health Informatics with Hospital São Paulo, SPDM and Fiocruz Mato Grosso do Sul, analysed 90,798 admissions of women aged 20 to 59. Using unsupervised machine learning techniques such as topic modelling and association rules, it identified nine recurring admission profiles. The most frequent involves women aged 20 to 29, predominantly of mixed race, admitted as emergencies with surgical procedures, multiple injuries, chest injuries and attacks with sharp objects. Skull and face fractures were strongly linked with attacks using bodily force.

A Unifesp researcher told Band the technology matters because the information sits in clinical notes that cannot be analysed manually, with suspicions confirmed only at the end through the ICD-10 code for domestic violence. Many victims arrive with serious injuries but report falls or household accidents out of fear or coercion. Lead author André Massahiro Shimaoka said the methods reveal relationships that conventional analysis often misses.

What it means in practice

Health services are often the first place outside the home where a woman in danger can be reached, so patterns in hospital data can help staff recognise cases that would otherwise be logged as accidents and help managers target prevention. The work also shows the value of Brazil's large public datasets, a theme behind projects such as LNCC's open Portuguese-language model and INPE's AI deforestation alerts funded by the Amazon Fund.

Using sensitive health records also demands care. Work of this kind depends on careful handling of personal data and on rules from the data protection authority, which is building its own AI centre and preparing a second sandbox. The study describes patterns across admissions; it does not identify individual victims or predict who is at risk.

What to watch: whether the Ministry of Health or state health departments adopt the profiles in surveillance and staff training, and whether similar models are tested on more recent records.