Objective:

The article describes ClarIA, an AI tool designed to identify possible indicators of gender-based violence (GBV) in Brazilian health records, and reports on its pilot analysis and early deployment in Recife.

Approach:

• Pilot analysis: Vital Strategies and the Federal University of Juiz de Fora analyzed 13,000 formally documented GBV cases in Recife.
• Detection: ClarIA analyzes clinical notes and medical histories for semantic patterns resembling records of confirmed GBV cases; it does not rely on isolated keywords.
• Alerts and response: Amber alerts flag patterns of signs, symptoms, and complaints associated with confirmed cases. Red alerts concern prior disclosure or identification of violence. Trained clinicians receive guidance on sensitive follow-up.

Key Findings:

• In the pilot, women’s clinic visits began an average of 92 days before GBV was recorded in their cases.
• Records showed patterns including increased headache complaints and mental health problems; many women did not explicitly disclose violence during visits.
• After deployment began in Recife in March, ClarIA operated across 39 clinics and hospitals. Health professionals submitted eight GBV notifications after receiving alerts.
• Women whose assaults were not formally notified faced an eightfold greater risk of death than women with reported assaults.
• Among femicide victims identified by linking violence and mortality records, more than 60% died within 30 days after a GBV notification.

Interpretation:

ClarIA alerts are prompts for clinicians to examine potential clinical, psychosomatic, and social indicators of violence; they are not diagnoses.

Limitations:

• The program is new, and available data remain limited.
• The article describes a pilot and early deployment; it does not report a controlled evaluation of ClarIA’s effect on violence or mortality.

Conclusion:

ClarIA is being deployed across clinics and hospitals in Recife to flag possible GBV indicators in health records and prompt trained clinicians to assess patients sensitively.

Sources:

• Article context provided by the user