Paper Number
ICIS2026-1753
Paper Type
Short
Abstract
Mental health assessment in social media is challenging because psychological expressions are implicit, heterogeneous, and difficult to interpret using purely data-driven models. While integrating domain knowledge has emerged as a primary strategy, existing approaches largely rely on static knowledge bases that fragment holistic clinical reasoning into rigid fact entries, limiting their ability to foster domain-specific reasoning. To address this challenge, we propose InfoPsych-MoE, an information-theoretic case-based reasoning mixture-of-experts framework that enables LLMs to acquire and coordinate domain-specific reasoning expertise. Specifically, we develop a Case-Based Reasoning Internalized Prompting (CBRIP) pipeline to construct transferable reasoning patterns through the Retrieve–Reuse–Revise–Retain cycle. We further introduce a PMI-based dual-level utility assessment mechanism that evaluates the contribution of reasoning patterns, guiding reasoning refinement and expert routing. Experiments on three mental health assessment tasks demonstrate that InfoPsych-MoE achieves superior performance and provides an effective framework for adaptive mental health assessment.
Recommended Citation
yao, chuang; Yan, Zhijun; and Cui, Shaoze, "InfoPsych-MoE: An Information-Theoretic Case-Based Reasoning Mixture-of-Experts Framework for Mental Health Detection" (2026). ICIS 2026 Proceedings. 3.
https://aisel.aisnet.org/icis2026/social/social/3
InfoPsych-MoE: An Information-Theoretic Case-Based Reasoning Mixture-of-Experts Framework for Mental Health Detection
Mental health assessment in social media is challenging because psychological expressions are implicit, heterogeneous, and difficult to interpret using purely data-driven models. While integrating domain knowledge has emerged as a primary strategy, existing approaches largely rely on static knowledge bases that fragment holistic clinical reasoning into rigid fact entries, limiting their ability to foster domain-specific reasoning. To address this challenge, we propose InfoPsych-MoE, an information-theoretic case-based reasoning mixture-of-experts framework that enables LLMs to acquire and coordinate domain-specific reasoning expertise. Specifically, we develop a Case-Based Reasoning Internalized Prompting (CBRIP) pipeline to construct transferable reasoning patterns through the Retrieve–Reuse–Revise–Retain cycle. We further introduce a PMI-based dual-level utility assessment mechanism that evaluates the contribution of reasoning patterns, guiding reasoning refinement and expert routing. Experiments on three mental health assessment tasks demonstrate that InfoPsych-MoE achieves superior performance and provides an effective framework for adaptive mental health assessment.
Comments
24_Social_media