Applying Frame-Based Semantic Patterns for The Identification of Violence Against the LGBTQIAPN+ Population

Open text fields in e-medical records not rarely surpass parameterized fields in informativeness when it comes to the identification and understanding of conditions affecting patients. However, their lack of structure, reliance on conventions unique to the healthcare context, and the need to protect patients’ personal information impose challenges to their use at scale. Parameterized fields, in turn, may suffer from poor data entry quality, especially for fields such as sexual orientation and gender identity. In this paper, we extend a methodology for the identification of notifiable events in healthcare, using Frame Semantics to define patterns and detect them in the open-text fields of e-medical records. We apply the method to tackle the underreporting of gender-based violence against the LGBTQIAPN+ community in primary care settings, a scenario where parameterized registration forms fail to capture the clinical context and the social dynamics of violence. Using corpora from two Brazilian Universal Public Health systems (e-SUS APS and SINAN), we expand the FrameNet Brasil model by proposing new gender and sexuality frames and validating semantic patterns of violence. Our evaluations confirm that the semantic model outperforms traditional demographic data, revealing vulnerability trajectories marked by physical trauma, suicide risk, and institutional neglect. Designed as a transparent, interpretable, and computationally low-cost model, the system serves as a tool to support policy making and healthcare surveillance.