The Validation Gap

Executive Summary

The widespread adoption of AI tools in health economics and outcomes research has created a critical but largely unexamined problem: the metrics most commonly used to evaluate these tools were designed for a fundamentally different class of task. Accuracy, sensitivity, and specificity originated in clinical classification problems where a knowable ground truth exists, errors can be counted, and performance can be expressed in compact numbers. As AI has migrated from classification into complex research assistance, such as synthesizing literature, constructing health economic models, drafting HTA submissions, those metrics have followed.

This white paper traces the validation gap across a spectrum of HEOR tasks, demonstrating precisely where traditional metrics begin to strain, where they break down structurally, and where they lose meaningful applicability altogether. The argument is not that accuracy, sensitivity, and specificity should be discarded. It is that they are necessary but no longer sufficient, and that treating them as adequate creates a form of false assurance uniquely dangerous in evidence generation contexts.

The field's leading methodological bodies have reached the similar conclusions independently. ISPOR's ELEVATE-GenAI reporting guidelines devote only two of ten evaluation domains to traditional accuracy metrics; the remaining eight address reproducibility, data provenance, fairness and bias, model transparency, human oversight, governance, and ongoing monitoring. The NICE Decision Support Unit has stated directly that technical accuracy is "an unreliable guide to real-world impact." These reflect accumulated evidence that what an algorithm does in isolation and what a human-algorithm system accomplishes in practice are different questions, and that traditional metrics answer only the first.

This whitepaper examines three use-cases to demonstrate the activities within the “evidence” spectrum, where we see more AI integration. Systematic literature review represents the most favorable case for conventional validation. Its deterministic logic, i.e. an article either meets PICO criteria or it does not, means that recall and precision do real work in auditing AI-assisted screening and extraction. Yet even here, inter-reviewer disagreement rates of 15 to 25 percent on borderline inclusion decisions reveal that binary ground truth breaks down precisely where it matters most. Endpoint selection, risk-of-bias adjudication, and the temporal decay of validated performance fall entirely outside what these metrics can measure.

Health economic modeling marks the point of structural breakdown. Model outputs are designed artifacts, not classifications. Reasonable, experienced modelers make different structural choices, select different parameter sources, and construct defensible alternatives that diverge from one another without any being wrong. Traditional accuracy scoring applied to this design space conflates genuine errors with legitimate methodological variation. That is a category error that undermines rather than supports quality assurance.

For HEOR professionals, health economists, and HTA methodologists evaluating or deploying AI research tools, the practical implication is clear: validation portfolios built primarily around benchmark accuracy scores are insufficient for the tasks these tools are now being asked to perform. Organizations need evaluation frameworks that explicitly address reproducibility, transparent reasoning, bias detection, and human oversight alongside (not instead of) traditional performance metrics. The question is no longer whether AI can achieve high accuracy on a structured test. The question is whether the human-AI system produces trustworthy, reproducible, and auditable evidence. That question requires different tools to answer.

Key Takeaways

• Insufficient Metrics: Traditional metrics (accuracy, sensitivity) originated for structured classification and are inadequate for evaluating complex AI research tools.
• False Assurance: Relying purely on accuracy scores creates a 'validation gap', missing vital failure points and creating false assurance in evidence portfolios.
• Institutional Alignment: Leading bodies like ISPOR and NICE now require expanded evaluation spanning reproducibility, transparency, bias, and continuous oversight.
• Integrated Evaluation: Validation must measure the combined human-AI workflow and function as continuous assurance, not a one-time point-in-time check.