Elicit for Health Economics & Outcomes Research

Health economics and outcomes research often hinges on turning scattered clinical evidence into usable estimates of burden, costs, and intervention impact. A practical workflow demonstrated here uses Elicit to search for relevant studies, extract standardized metrics across papers, and sanity-check extracted numbers—aiming to speed up inputs needed for cost-benefit or ROI-style analyses in areas like drug pricing, payer decisions, and policy budget allocation.

The session frames HEOR as an impact-assessment problem: quantify how many people face a condition, how severe the outcomes are (including quality-of-life losses), and what medical interventions change in terms of both direct costs and downstream complications. The example query targets a concrete question—“the frequency of biopsies and complications among lung cancer patients in the US”—to illustrate how Elicit can approximate market need and clinical burden by aggregating evidence from multiple studies.

Elicit’s “Find papers” workflow starts by searching a database of public papers and selecting a small set (eight) of studies most relevant to the query. To keep the evidence aligned with a specific decision context, the workflow adds predefined columns such as Region (to focus on the US), Data set (to gauge sample sizes and data sources), and Methodology (to identify study designs). The results skew toward retrospective studies, with prospective studies treated as generally more compelling and small case studies treated more cautiously.

Next comes the extraction step: Elicit pulls quantitative details like the number of patients, the number of biopsies, and the frequency of complications. Each extracted figure is tied to supporting quotes from the paper, enabling quick verification. The walkthrough highlights how default extraction can occasionally mis-handle relationships between related quantities (for example, an average biopsies-per-patient figure that appears inconsistent with the total number of biopsies and the number of patients). High accuracy mode is then used as a targeted upgrade: it relies on more advanced (and more expensive) models, returns answers in a more structured format (often bullet points), and can better handle table-derived values and arithmetic consistency checks.

The example also shows how uncertainty is surfaced via confidence flags; when Elicit is not confident, users are encouraged to double-check the underlying text. Another key constraint is access to full text: if a paper is open access, Elicit can extract from the full text; otherwise, it relies on the abstract, which can omit critical details. For deeper extraction from non-open-access PDFs, the workflow points to an “extract data from PDFs” approach.

Finally, the session demonstrates how to scale beyond a handful of studies by creating custom columns—such as a “study type” classifier—and filtering papers based on methodology keywords (e.g., retrospective vs. prospective). The takeaway is a repeatable HEOR research pipeline: search systematically, extract standardized burden and procedure/complication metrics with traceable citations, validate with high accuracy mode when stakes are high, and export results (CSV) for downstream analysis.