Abstract:
Domestic violence impacts millions of individuals worldwide, but remainshighly underreported due to fear, shame, and coercive control. To a degree,healthcare settings are intervention points for domestic violence, or can becritical detection points. At present, existing methods for domestic violencedetection are based upon self-disclosure by patients, and overt signs ofwounds/injuries, which does not capture many instances of control, emotional,and subtle/early-stage physical abuse. This research presents a nascent artificialintelligence framework to detect potential domestic violence victims byanalyzing micro-expressions, vocal stress patterns, and inconsistencies inbehavior during everyday healthcare interactions. To achieve this goal, the studydevelops a multimodal AI system that combines computer vision for analyzingmicro-expressions, natural language processing for vocal stress patterns, andgenerally-federated learning architectures to avoid disrupting patient care andensure privacy. The study includes ethically-sourced training data usingsustained consent from the participants. The training also follows a differentialprivacy approach to minimize likelihood of reidentifying an individualparticipant residing in a multitude of countervailing confounding factors (e.g.cultural, social, economic status, race, sexual orientation, etc.), which candelineate patterns within socio- demographic groups. The anticipated outputs ofthis endeavor are a culture-sensitive tool capable of identifying cases for trainedprofessionals to investigate with care. The complete system aspires to be 85%accurate with at-risk individuals and zero inaccurate identifications based onhuman-in-the-loop manipulations. Implementation will include working withhealthcare organizations, domestic violence organizations, and an ethics boardto ensure responsible implementation. We are addressing a significant gap for AIand social good by allowing healthcare professionals the use of a non- invasivetool to identify vulnerable populations. Using the anonymity preserving protocolallows patients to maintain their confidentiality while we potentially protectthem with early intervention and connected resources to assistance.
Page(s):
117-117
DOI:
DOI not available
Published:
Journal: 4th International Conference of Sciences “Revamped Scientific Outlook of 21st Century, 2025” , November 12,2025, Volume: 1, Issue: 1, Year: 2025
Keywords:
Domestic violence detection
,
federatedlearning
,
behavioral pattern recognition
,
social impacttechnology
,
healthcare AI
,
microexpression analysis
,
privacypreserving machine learning