AGENT-ATOM-GRADIENT: A NEXT-GEN AI ARCHITECTURE FOR EVIDENCE DRIVEN OUTCOMES RESEARCH
Author(s)
Achilleas Livieratos, PhD1, Maria Kudela, PhD2, Yuxi Zhao, PhD2, All-shine Chen, PhD2, Junjing Lin, PhD3, Di Zhang, PhD4, Xin Luo, PhD2, Paula Angelica Ramos, MSc2, Sai Dharmarajan, PhD5, Chinyu Su, PhD2, Mina Gaga, MD6, Margaret Gamalo, PhD2.
1SPAIML Scientific Working Group, New York, NY, USA, 2Pfizer, New York, NY, USA, 3Takeda Pharmaceuticals, Cambridge, MA, USA, 4Teva Pharmacieticals, New York, NY, USA, 5Sarepta Therapeutics, New York, NY, USA, 6Hygeia Hospital, Athens, Greece.
1SPAIML Scientific Working Group, New York, NY, USA, 2Pfizer, New York, NY, USA, 3Takeda Pharmaceuticals, Cambridge, MA, USA, 4Teva Pharmacieticals, New York, NY, USA, 5Sarepta Therapeutics, New York, NY, USA, 6Hygeia Hospital, Athens, Greece.
OBJECTIVES: Building on prior work with Mixture-of-Agents (MoA) architectures for data extraction in HEOR, this study introduces Agent-Atom-Gradient (AAG), a next-generation AI framework designed for treatment sequence optimization in complex, multi-stage diseases such as ulcerative colitis (UC) and chronic obstructive pulmonary disease (COPD). Traditional single-model LLMs struggle with the heterogeneity of patient pathways and integration of real-world evidence (RWE). Our objective was to develop and evaluate a layered AI pipeline—combining MoA, Atom-of-Thought (AoT) reasoning, and TextGrad refinement—to deliver clinically transparent, evidence-anchored treatment sequences for HEOR decision-making.
METHODS: The AAG pipeline proceeds in three stages. First, a MoA framework deploys multiple LLM proposer models and an aggregator (DeepSeek-R1) to synthesize candidate treatment sequences from RCTs, guidelines, and RWE. Second, AoT reasoning (GPT-4.1) decomposes recommendations into atomic clinical sub-questions (e.g., “After TNF inhibitor failure, which agent demonstrates superior durability in RWE cohorts?”), then recontracts them into coherent pathways. Finally, TextGrad applies iterative, gradient-style optimization, aligning outputs with regulatory trial data, guideline mandates, and RWE metrics under explicit structural constraints (first-, second-, third-line). Benchmarking was performed across UC and COPD use cases.
RESULTS: Successive reasoning layers yielded progressively richer outputs. MoA provided rapid synthesis of guideline and trial evidence but remained conceptual. AoT exposed reasoning transparency, linking each clinical decision to a specific sub-question. TextGrad produced the most actionable recommendations, explicitly pairing therapies with trial outcomes, real-world utilization metrics, and patient-subgroup guidance. This multi-stage refinement increased both interpretability and clinical utility, supporting treatment plans for UC and COPD.
CONCLUSIONS: The Agent-Atom-Gradient model illustrates the way multi-layered LLM complexes can turn conceptual AI frameworks into transparent, evidence-tethered decision support for HEOR. Combining MoA, AoT and TextGrad in a single pipeline improves scalability, personalization and payer/regulator interactions. This tool is applicable beyond UC and COPD to oncology and other multidisciplinary therapeutic areas where we can benefit from evidence-based decisions.
METHODS: The AAG pipeline proceeds in three stages. First, a MoA framework deploys multiple LLM proposer models and an aggregator (DeepSeek-R1) to synthesize candidate treatment sequences from RCTs, guidelines, and RWE. Second, AoT reasoning (GPT-4.1) decomposes recommendations into atomic clinical sub-questions (e.g., “After TNF inhibitor failure, which agent demonstrates superior durability in RWE cohorts?”), then recontracts them into coherent pathways. Finally, TextGrad applies iterative, gradient-style optimization, aligning outputs with regulatory trial data, guideline mandates, and RWE metrics under explicit structural constraints (first-, second-, third-line). Benchmarking was performed across UC and COPD use cases.
RESULTS: Successive reasoning layers yielded progressively richer outputs. MoA provided rapid synthesis of guideline and trial evidence but remained conceptual. AoT exposed reasoning transparency, linking each clinical decision to a specific sub-question. TextGrad produced the most actionable recommendations, explicitly pairing therapies with trial outcomes, real-world utilization metrics, and patient-subgroup guidance. This multi-stage refinement increased both interpretability and clinical utility, supporting treatment plans for UC and COPD.
CONCLUSIONS: The Agent-Atom-Gradient model illustrates the way multi-layered LLM complexes can turn conceptual AI frameworks into transparent, evidence-tethered decision support for HEOR. Combining MoA, AoT and TextGrad in a single pipeline improves scalability, personalization and payer/regulator interactions. This tool is applicable beyond UC and COPD to oncology and other multidisciplinary therapeutic areas where we can benefit from evidence-based decisions.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR124
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Respiratory-Related Disorders (Allergy, Asthma, Smoking, Other Respiratory), SDC: Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)