Abstract Number: Esoc2026lb53 an Agentic AI Copilot for Exploratory Identification of Optimal Treatment Subgroups for Traditional Chinese Medicine Fytf-919 (Zhongfeng Xingnao Oral Prescription) in Intracerebral Hemorrhage

Abstract Background and aims In acute intracerebral hemorrhage (ICH), treatment effect heterogeneity may be missed by conventional hypothesis-driven subgroup analyses. We used an agentic AI copilot as a knowledge-discovery framework to identify potential treatment-effect heterogeneity of FYTF-919 in acute ICH. Methods We conducted a secondary analysis of the CHAIN prospective, randomized, double-blind, placebo-controlled trial (N=1642). The primary outcome was 180-day functional status (mRS 0-2 versus 3-6). Eighty-nine baseline demographic, clinical, and imaging variables were analyzed using the Evolutionary Multi-Agent Collaboration (EMAC) framework for subgroup discovery. EMAC employs a multi-agent system where a Data Agent screens variables for clinical validity, a Miner Agent generates and evolves candidate subgroup rules, and a Summary Agent scores each rule by treatment effect, statistical significance, subgroup size, and LLM-assessed clinical plausibility. This agent-driven cycle iteratively refines subgroup definitions toward clinically meaningful and statistically robust rules. Top-ranked subgroups were validated using multivariable logistic regression models to estimate adjusted odds ratios (aORs) with 95% confidence intervals across five drug and surgical treatment modules. Results The EMAC-assisted framework identified interpretable patient profiles showing significant treatment-effect heterogeneity in all five modules (adjusted p0.05). The strongest effect was observed in surgery-versus-no-surgery among drug-treated patients (aOR=4.55). Modules evaluating drug benefit showed consistent treatment-effect signals (aOR=2.80). Hematoma volume emerged as recurrent feature identifying patients with greatest relative treatment benefit. Conclusions An agentic AI copilot provides a structured and scalable approach for exploratory identification of treatment subgroups in ICH clinical trials. This approach may support hypothesis generation and inform future optimized trial designs for FYTF-919. Conflict of interest Figure 1 - belongs to Methods Figure 2 - belongs to Results