Privacy-First Multi-Agent AI Powered Virtual Counselor for SBIRT: Integrating Large Avatar Models with Zero-Retention Architectures

Abstract

The Screening, Brief Intervention, and Referral to Treatment (SBIRT) framework is critical for early detection of substance use problems, yet its clinical adoption is impeded by resource constraints and patient reluctance to disclose sensitive behaviors. While Large Language Models (LLMs) offer conversational capabilities, their clinical deployment is limited by unreliable protocol adherence and privacy concerns. We introduce a Privacy-First Multi-Agent Virtual Counselor that automates SBIRT with high fidelity by treating LLMs as protocol-driven generators governed by an explicit state machine, rather than autonomous decision-makers. Our blackboard-based architecture orchestrates specialized agents for screening like AUDIT (Alcohol Use Disorders Identification Tests) and DAST (Drug Abuse Screening Test), motivational interviewing, and crisis response, achieving 97.2% protocol accuracy and 98.8% crisis detection recall on 500 simulated profiles. We address data security through a Hybrid Privacy Architecture combining server-side Zero-Retention (RAM-only processing, no persistent storage) with optional client-side encrypted state, supported by an explicit threat model and verifiable engineering controls. To enhance therapeutic alliance, we integrate a Large Avatar Model providing synchronized non-verbal cues. This work demonstrates a scalable, auditable approach to automated behavioral health intervention.

Ruiheng Yu

PhD Student of Biomedical Engineering

My research interests include Computer vision, explainable AI, LLM in health area.

Xinyu Qin

PhD Student of Biomedical Engineering

My research interests include Causal Inference, Survival Analysis, and Digital Twin.

Lu Wang

Assistant Professor of Biomedical Engineering

My research interests are developing and applying Machine Learning, Data Mining and Statistical methods (e.g., Multi-task Learning, Survival Analysis, Clustering, Risk Factor Analysis and Causal Discovery) on various data including gene expression, electronic health/medical records, and DNA sequencing reads for both cognitive disorders (e.g., delirium, Alzheimer’s disease, dementia, major depressive disorder) and chronic diseases (e.g., cancer, obesity, hypertension). Inspired by the human factors approach, she also designs and develops Human-Centered Artificial Intelligence tools for users to integrate, visualize, analyze, and interpret health data in order to improve the interoperability and accessibility of AI-assisted healthcare decision support.