The ABLE framework for attributable, blinded and leakage-controlled evaluation of generative AI in health economic model conceptualisation
Objectives. Generative artificial intelligence is used to propose health economic model (HEM) structures through systems layering structured reasoning, retrieval, graph retrieval and agentic orchestration onto a base large language model (LLM). Evaluations rarely show which layer produced the result, whether the benchmark appraisal was already in the training data, or whether assessors knew which system they were scoring. We propose a prospective design addressing these problems. Methods. We drew on published generative-AI applications to HEM conceptualisation and on evaluation methods from LLM benchmarking and clinical assessment to identify design requirements, then translated them into a protocol with rubrics, case-selection rules, an output schema and an analysis plan. Results. ABLE has four pillars: Attribution, through within-case comparison of configurations added stepwise to one frozen base LLM; Blinding, through a common output schema, deterministic rendering and independent assessors; Leakage control, through temporal eligibility, a pre-inclusion recall probe and a time-sliced corpus; and Evaluation standards, through a registered primary comparison and prespecified outcomes. Reference design is a ladder from a base LLM (A0) to a retrieval- and agent-supported system (A4), with a compute-matched base arm (A0k) separating repeated sampling from architecture. A4 versus A0k is confirmatory; other contrasts are exploratory. Structural concordance with an HTA reference model is scored separately from expert acceptability of alternative structures; critical errors, unsupported claims, stability and efficiency are secondary outcomes. Conclusions. ABLE complements reporting guidance such as ELEVATE-GenAI by specifying how configurations are compared rather than what is disclosed. Protocol for empirical application is registered (OSF 10.17605/OSF.IO/UBN5H)
Authors
- Saurabh Srivastav (ORCID: https://orcid.org/0009-0000-3699-3145)
- ABHINAV KUMAR
- Roshan Singh
Institutions
- Indian Institute of Management Ranchi (IN)
- Indian Institute of Information Technology Ranchi (IN)
Publication Details
- Published
- 2026-11-17
- DOI
- https://doi.org/10.33774/coe-2026-7bkgl
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- preprint