Pioneering study in Recife (city in state of Pernambuco, Brazil) was realized by Vital Strategies in partnership with FrameNet Brasil, computational linguistics lab of the Federal University of Juiz de Fora (UFJF) and analyzed records of 13 thousand women victims of violence in the city.
Gender-based violence is not an issue solely of public security, but instead an intersectoral challenge. Data from the health sector, such as prevalence, hospitalization and mortality due to violence can provide valuable indicators for early identification of violence, allowing interventions before cases reach grave and fatal outcomes, such as hospitalization and femicide. The health sector has a privileged role in the identification of victims, since it is the entryway for many cases of violence into the public services. Vital Strategies has been working for years on data analyses that aim to map the trajectories of women victims of violence in the health data systems in Brazil. The goal is to identify patterns in these women’s visits to healthcare units that can predict the aggravation of violence before it is notified.
Several national and subnational studies have been conducted analyzing data from the same women in different databases of the Ministry of Health – the Notifiable Diseases and Conditions Information System (Sinan)*, Mortality Information System (SIM), Hospital Information System (SIH/SUS), Electronic medical records (e-SUS) – and of Public Security, such as police reports.
These analyses show that many cases of gender-based violence in Brazil are not notified or are notified too late. While only around a fifth (19.5%) of notifications in Sinan are made by basic health services, the majority (80%) is made by hospitals or emergency rooms. That is, violence is only notified when it becomes graver, even if the woman who is its target has had other visits of less gravity in primary care due to the same situation, which is most common.
Late notification is a lost opportunity for action in avoiding hospitalization or death for these women. A study from Vital Strategies, conducted in partnership with UFMG (Federal University of Minas Gerais), revealed that women who suffered assaults and did not have these incidents notified presented a risk of death that was eight times greater than those who did.
Pilot project in Recife
The project in Recife, conducted by Vital Strategies in partnership with FrameNet Brasil, the computational linguistics lab of the Federal University of Juiz de Fora (UFJF), analyzed the records of 13 thousand women victims of violence. These 13 thousand women were identified in Sinan, confirming they were victims of violence. Then, data from e-SUS (medical records filled in primary care units) were analyzed to map what is the volume of visits of these women to healthcare units and what are the reasons for these visits. This information was then compared to the set of women that visited primary care units in the municipality in the same period, but who did not have a notification of violence in Sinan.
For the victims of femicide, identified by matching the violence and mortality databases, the analysis showed that, in more than 60% of cases, homicide occurs 30 days after a notification of violence has been made in the healthcare system (Sinan). This comparison showed that 92 days before a woman has a notification of violence made in her name in Sinan, she starts seeking primary care services more often.
“The data show there is a considerable change in the pattern of use of primary care services for women who are about to suffer a graver form of violence. This shows a recurring pattern in the cycles of violence, in which it escalates and increases its gravity gradually. If this pattern change is identified early, measures can be implemented to interrupt this escalation, avoiding graver outcomes such as hospitalization and even death,” explains Sofia Reinach, Deputy Director of Vital Strategies’ Violence Prevention Program.
Semantic analysis and artificial intelligence: new data to map the trajectory of women victims of violence in the health systems
As a second stage in the project that analyzed the data of those 13 thousand women victims of violence, Vital Strategies and FrameNet Brasil utilized innovative technology to incorporate another great volume of information useful to identifying patterns in these women’s visits to healthcare units, which can help predict grave violent incidents. Through semantic analysis and artificial intelligence methodologies, it was possible to “read” the texts written by health professionals in the open fields of patients’ electronic medical records and in Sinan.
Besides information brought by the parametrized data available in the e-SUS databases, the mapping of these women’s trajectories was detailed through the inclusion of patterns identified by artificial intelligence.
Usage of these new data showed changes in the reports of women victims of violence in this 92-day window prior to the notification of violence. “That which health professionals write down in the medical records of women victims of violence from 92 days prior to the aggravation of the assaults also changes. Some motivations for seeking care grow, while others diminish. Beyond that, the semantic analysis also shows that these women are not only reporting violent situations, as in detailing how they occur.
In this timeframe, the reports of physical assaults and abuse grow, with mentions of the aggressor and the body part that is the focus of the assaults. When data matching and semantic analysis are combined, it is possible to see not only a change in the frequency of visits, but also in the motivations that led to that visit,” explains Tiago Torrent, coordinator of the FrameNet Brasil Laboratory of Computational Linguistics.
“With these data, we have a clear indicator that it is necessary to create strategies to anticipate the notification of violence in Sinan and ‘gain’ these 90 days to strengthen assistance for women in situations of violence and prevent femicides. In that sense, the Recife Health Secretariat is already, alongside partners, at the implementation stage of a data visualization dashboard, which will support the identification of cases of underreporting. The goal is to implement measures of sensitization, awareness-raising and qualification of frontline professionals so they may conduct the detection/identification of violence, as well as issue notifications, reduce underreporting and act by referring these identified victims to protection services,” explains Luciana Albuquerque, Recife’s Secretary of Health.
Opportunity to improve and test AI model for public benefit responsibly and ethically
Data anonymization: to ensure that no patient or victim’s potentially identifying information was accessed, a protocol for anonymization of e-SUS and Sinan’s open text fields was created and put into practice. Proper names, addresses and phone numbers were automatically substituted for anonymization tags, such as [NAME], [NUMBER], etc.
Ethical commitment: automated language models (LLMs) are widely used to analyze and interpret text in various areas. However, when handling sensitive and specific subjects, such as health and gender-based violence, it is essential to incorporate the human element to ensure that the results are precise and ethical. This was done throughout the entirety of the Recife project.
This human intervention allows a more precise and judicious classification of data. The AI is trained on specific areas of the subject of health, which ensures that results are more trustworthy and aligned with the needs of the context. Beyond that, the AI does not make decisions; it only delivers information which is then reviewed by humans, ensuring greater precision and transparency.
About Vital Strategies
Vital Strategies is a global health organization that believes all people should be protected by strong and equitable health policies and systems. We partner with governments, communities and organizations around the world to reimagine public health so that health is supported in all the places we live, work and play. The result is millions of people living longer, healthier lives around the world.
More information Analítica Comunicação – Communication consultancy at Vital Strategies
Gabriela Scheinberg – gabi.scheinberg@analitica.inf.br, +55 11 98111-9294
Raíza Dias – raiza.dias@analitica.inf.br, +55 11 95272-3944
Wilma Loures – wilma.loures@analitica.inf.br, +55 11 96324-6565
Mauricio Esposito – mauricio.esposito@analitica.inf.br, +55 11 99915-7583
Erica Benute – erica.benute@analitica.inf.br, +55 11 99185-7078