Health economic models serve as foundational analytical tools for evaluating the comparative cost-effectiveness of novel healthcare interventions, providing structured quantitative evidence that directly informs public health policies, national reimbursement decisions, and global market access for medical devices and pharmaceuticals. Developing these models requires synthesizing multidimensional clinical trial data, complex epidemiological registries, disease progression trajectories, and therapy sequencing conventions through rigorous multidisciplinary collaboration involving clinicians, biostatisticians, and specialized health economists. The initial stage of conceptualisation, which establishes the overarching structural architecture and defining clinical assumptions of a decision problem, remains one of the most resource-intensive bottlenecks within health economics and outcomes research. In response to rising pathway complexity across modern clinical practice, researchers have evaluated the feasibility of automating and refining health economic model conceptualisation through an agentic generative artificial intelligence architecture that combines graph-based retrieval-augmented generation, chain-of-thought prompt engineering, and human expert validation across two contrasting clinical case studies.
The traditional human-led conceptualisation process begins with systematic disease characterisation, in which analysts formalize the analytical scope by establishing core Population, Intervention, Comparator, and Outcome parameters alongside explicit institutional perspectives. Health economists subsequently decompose complex decision problems into distinct segments covering disease pathways, discrete health states, and resource utilization categories before iteratively analyzing each component to define baseline assumptions, structural uncertainties, and data requirements. Once these individual components are established, modelers synthesize the pieces into an integrated framework, ensuring that transition rules and comparative clinical assumptions remain methodologically aligned with established health technology assessment standards. However, because this manual progression is exceptionally labour-intensive and can take up to three months of continuous technical work, critical steps are occasionally expedited or restricted, leaving foundational model designs vulnerable to oversimplification or structural gaps that fail to capture real-world treatment realities.
To overcome these systemic bottlenecks, investigators engineered a multi-layered hybrid intelligence architecture grounded in human-in-the-loop methodology, operating on the principle that autonomous algorithmic reasoning must be continuously steered and validated by domain experts. The conceptual framework selected a human-in-the-loop configuration over an alternative artificial-intelligence-in-the-loop arrangement, recognizing that while the latter restricts machine involvement to narrow operational subtasks like automated data extraction or routine mathematical calculations, the former positions artificial intelligence as an active exploratory partner capable of proposing novel state transitions, simulating competing clinical pathways, and flagging internal structural contradictions. Independent human modelers possessing five to twelve years of professional conceptualisation experience provided iterative oversight, verifying that all interim proposals remained clinically justifiable and compatible with health technology assessment principles while allowing the autonomous agent to lead reasoning tasks.
At the reasoning core of the experimental framework, the team utilized OpenAI GPT-5.2 within an external Python programming environment, while noting that the underlying analytical sequence remains model-agnostic and transferable to comparable large language models. The architecture implemented chain-of-thought prompting compounded with self-consistency to mitigate the inherent stochastic instability of open-ended language model generations. Because health economic conceptualisation lacks a single absolute ground truth, self-consistency served as an empirical proxy for structural robustness by issuing identical prompts across five independent parallel reasoning trajectories. Each run generated a distinct step-by-step chain of thought, after which an external language model judge, powered by GPT-5.1, applied a majority-voting consensus protocol to extract the most recurrent health states, directional transitions, and overarching structural configurations across the five sampled iterations, thereby neutralizing idiosyncratic reasoning errors.
Because foundational language models are constrained by general pre-training corpora that lack niche health economic methodologies and granular clinical parameters, the framework incorporated an external retrieval-augmented generation database to infuse domain-specific guidelines into the reasoning loop. Rather than utilizing standard vector retrieval, which treats ingested passages as isolated document fragments, the investigators deployed Microsoft GraphRAG to organize domain knowledge into an interconnected topological network of conceptual nodes and relational edges. This relational graph mapped intricate dependencies between disease progression stages, therapeutic interventions, health utilities, and healthcare cost variables, closely mirroring structured human conceptual memory. Constructed from regional clinical guidelines, health technology assessment frameworks from the National Institute for Health and Care Excellence and Canada’s Drug Agency, and published clinical trials, the graph was refined using language model entity extraction and approved by human experts before active querying.
The operational orchestration was governed by an agentic execution layer configured to manage multi-step reasoning through an auditable reason-act-observe loop following the ReAct design pattern. Working through an externalized scratchpad maintained within a structured message buffer, the autonomous agent evaluated problem parameters, identified specific evidentiary gaps, and formulated natural-language queries executed as score-based similarity searches against the GraphRAG repository. The buffer recorded every emitted tool call, intermediate retrieved passage, and subsequent reasoning update, compiling a comprehensive, traceable audit trail accessible to human reviewers. At each discrete junction, the language model determined whether accumulated evidence was sufficient to finalize structural recommendations or whether further retrieval was necessary, cycling iteratively between internal deliberation, contextual evidence retrieval, and logical reassessment before delivering its structured conceptual output.
To evaluate adaptability across sharply contrasting evidence landscapes, the investigators conducted two proof-of-concept case studies that reflected opposing extremes of healthcare information availability. The first case study targeted advanced hormone receptor-positive, human epidermal growth factor receptor 2-negative breast cancer, representing a data-rich oncology indication with abundant clinical literature, established treatment paradigms, and mature published health economic evaluations. The model utilized published data from the phase three MONALEESA-3 clinical trial assessing ribociclib plus fulvestrant against standard care in post-menopausal women. To prevent data contamination and ensure authentic exploratory problem-solving, researchers explicitly withheld all existing health economic models and health technology assessment submissions from the GraphRAG corpus, restricting retrieval access entirely to clinical trial reports and authoritative institutional appraisal guidelines.
For the breast cancer evaluation, the agentic framework formulated an extensive disease concept diagram and recommended a four-state cohort Markov model comprising progression-free survival, first progression, second progression, and an absorbing death state. Transitions between non-absorbing states were driven by documented disease progression events corresponding to sequential therapeutic lines, with each health state retaining an active transition pathway to death. While acknowledging that partitioned survival models represent the conventional standard within historical health technology assessment submissions for advanced breast cancer, the agent explicitly selected a cohort Markov structure. The system reasoned that a Markov approach enabled distinct cost and quality-of-life utility assignments across successive treatment lines while capturing therapy sequencing dynamics that aggregate partitioned survival curves obscure, all while avoiding reliance on restricted individual patient survival data.
The second proof-of-concept study examined hereditary transthyretin amyloid polyneuropathy, an ultrarare, progressive disease characterized by severe multisystem peripheral neuropathy, representing a severely data-sparse clinical environment with minimal published economic modeling. Relying on aggregate clinical findings from the pivotal APOLLO trial evaluating patisiran against standard care, the agent formulated a disease concept map and recommended a seven-state cohort Markov structure anchored to polyneuropathy disability staging. The recommended model incorporated health states defined across stages 0, I, II, IIIA, IIIB, and IV, alongside death as an absorbing state. The agent reasoned that polyneuropathy disability stages provided clear, clinically validated boundaries with greater descriptive granularity at the critical transition between independent ambulation and wheelchair dependence compared to older familial amyloid polyneuropathy classifications, while avoiding unsupportable individual-level microsimulation complexity.
Structured external validation assessed the framework-generated architectures against published technology appraisals accepted by the National Institute for Health and Care Excellence, using four pre-specified comparative criteria. In the oncology evaluation, the agent’s four-state Markov design aligned closely with the three-state Markov model accepted in appraisal TA687, with expert health economic reviewers and oncologists confirming that partitioning progression into first and second stages constituted a clinically defensible refinement that enhanced therapy sequencing transparency. In the rare disease evaluation, the agent’s seven-state structure successfully reproduced the core neuropathy staging adopted in technology appraisal HST10. However, the agent did not reproduce the published appraisal’s secondary twelve-state stratification based on N-terminal pro B-type natriuretic peptide cardiac biomarker thresholds, a discrepancy reflecting the corpus-based framework’s lack of patient-level correlation data.
The experimental implementation demonstrated substantial procedural efficiency improvements over conventional health economic workflows. The automated conceptualisation run completed in approximately one hour, extending to between one and two days when incorporating multi-round human expert validation and iterative exploratory re-runs, contrasting sharply with the up to three months typically required for purely manual development. Financial expenditure remained modest, with OpenAI API costs totaling approximately ninety-four dollars per proof-of-concept evaluation under prevailing GPT-5.2 pricing, alongside an initial one-time engineering development requirement of three to four weeks. While these operational savings are substantial, investigators underscored that the generated outputs represent proof-of-concept demonstrations requiring subsequent Delphi panel consensus and formal expert elicitation, emphasizing that future implementations must incorporate double-blinded expert review, direct patient-level analytical integration, and stringent compliance with regional healthcare data privacy regulations.
Subject of Research: Generative artificial intelligence and graph retrieval-augmented generation for health economic model conceptualisation
Article Title: Conceptualising Health Economic Models Using a Graph Retrieval-Augmented, Agentic AI Framework with Human-in-the-Loop Validation: Proof-of-Concept Study
Article References: Srivastava, T., Irfan, H., & Swami, S. (2026). Conceptualising Health Economic Models Using a Graph Retrieval-Augmented, Agentic AI Framework with Human-in-the-Loop Validation: Proof-of-Concept Study. PharmacoEconomics – Open, 10(6), 987-1004. https://doi.org/10.1007/s41669-026-00687-2
Image Credits: AI Generated
DOI: 10.1007/s41669-026-00687-2
Keywords: Health Economics, Generative Artificial Intelligence, Model Conceptualisation, Graph Retrieval-Augmented Generation, Chain-of-Thought Reasoning, Self-Consistency, Markov Models, Human-in-the-Loop, Health Technology Assessment, Breast Cancer, Transthyretin Amyloid Polyneuropathy, MONALEESA-3 Trial, APOLLO Trial, Decision Problem Formulation, NICE Guidelines, ReAct Pattern, Cost-Effectiveness Analysis
News Source: Blake Davidson. (November 20, 2026). Agentic Artificial Intelligence Framework Accelerates Conceptualisation of Health Economic Decision Models. Scienmag.



