Can AI evidence support health-spending decisions?
A 2,476-record scoping review found rapid growth but little validation—and its own audit judged only 51.3% of a 300-record sample genuinely relevant. The study's LLM reference standard makes that estimate informative, not definitive.
The short version
- What happened
- The PRISMA-ScR review searched PubMed for 2010–2025, screened 2,628 records and included 2,476, then audited a random sample of 300 against a blinded LLM classifier.
- Who it affects
- Patients can gain or lose access when uncertain AI-derived evidence influences reimbursement, so the final decision needs auditability and sensitivity analysis.
- What remains unproven
- Single-database, English-language search with limited conference and grey-literature coverage
Keep in mind
Single-database, English-language search with limited conference and grey-literature coverage
An LLM rather than dual human experts served as the reference for the 300-record audit
Large measured classification errors, especially for the economic-evaluation label
An unresolved 273-record discrepancy between PubMed date-query approaches
Record-level indexing counts do not establish risk of bias, real-world adoption or decision benefit
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: useful automation, weak evidence for the final decision
AI can accelerate identifiable parts of health-economics work, including screening references, extracting structured information, writing code and processing real-world data. A new systematic scoping review found documented gains where there is an observable answer against which output can be checked. It found much less evidence that a strong predictive score makes a reimbursement estimate, causal effect or cost-effectiveness conclusion more trustworthy.
That distinction matters to public services, insurers and patients. A model may find records quickly and still exclude the wrong study; predict an outcome accurately and still estimate a treatment effect poorly; or generate a plausible economic model whose assumptions drive the result. The review's central rule is therefore practical: predictive performance is not evidentiary performance. Decision makers need to test how the AI changes the quantity used for funding, access or reimbursement—not merely whether it produces a high AUC or fluent report.[1]
How 2,642 search results became 2,476 records
The author searched PubMed on 9 September 2026 for publications from 2010 through 2025 that combined AI terms with health-economics and outcomes-research terms. Eligible records were English-language journal articles, reviews or conference proceedings with an abstract. Fourteen duplicates and 152 ineligible records were removed, leaving 2,476. A published term dictionary assigned eight overlapping application domains and one of six AI method categories from titles, abstracts and indexing terms.
The raw count rose from 29 records in 2010 to 852 in 2025. Economic evaluation was assigned to 1,559 records, while real-world data and causal inference appeared in 496 and governance, reporting and validity in 494. Generative AI appeared in 103 records in 2025. Yet those are classification outputs, not verified practice counts. The author deliberately tested the search and dictionary because bibliometric totals can become marketing claims when their false-positive rate is unknown.[1]
The review audited its own headline number
A random sample of 300 included records was reclassified by an independent large language model that was blinded to the dictionary labels and received only the title and abstract. It judged 154 records, or 51.3%, genuinely relevant to health economics and outcomes research. Applying that proportion to the full corpus produced a precision-adjusted estimate of 1,271 relevant studies, with a 95% confidence interval from 1,131 to 1,410.
Agreement varied sharply by task. It was substantial for real-world data and causal inference, with Cohen's kappa of 0.729, but poor for economic evaluation, with kappa of 0.144. The dictionary labelled 186 of the 300 sampled records as economic evaluation; the blinded classifier labelled 69, and dictionary precision was 29.6%. In other words, words about cost, efficiency or resource use frequently appeared without the paper being an economic evaluation in the sense a reimbursement reviewer would use.[1]
The audit has an important circularity
The independent classifier was another LLM, not a panel of health economists. The paper openly acknowledges this circularity and treats the model as a reference standard rather than truth. It may be too strict about what counts as relevant or too lenient in unfamiliar subfields. A dual human review of the same 300 records is the necessary next step. Until then, 51.3% should be read as a measured warning about noisy counting, not a definitive correction factor.
There was also an unresolved difference of 273 records between a range query and the sum of per-year PubMed queries, probably related to ambiguous Medline dates. PubMed omits much conference and grey literature, the search was English-only, and title-and-abstract classification cannot recover every method or deployment detail. The review is stronger for exposing these errors than a study that hides them, but transparent uncertainty remains uncertainty.[1]
Validation was rare relative to publication volume
Of the 2,476 records, 674 were indexed as reviews, 121 as clinical studies, 38 as randomised trials and only 17 as validation studies. Those categories are based on PubMed indexing rather than an individual risk-of-bias assessment of every paper, but the imbalance is striking. The literature expanded far faster than evidence that tools travel between settings, preserve calibration or improve a real funding decision.
The review's narrative evidence map found the clearest measured value in tasks with checkable ground truth. Some literature-screening systems reached high sensitivity and throughput, although performance varied widely and false positives could be substantial. In causal estimation, synthetic data and deployed monitoring, apparently strong predictive summaries did not reliably guarantee accurate treatment effects, privacy protection or stable calibration. The outcome a health-spending body needs is often counterfactual: what would happen, and at what cost, under another decision. A prediction metric alone does not answer it.[1]
What commissioners, analysts and patients should require
For clerical use, organisations should measure recall, extraction accuracy, time saved and an auditable trail back to the source. Human reviewers need a way to inspect exclusions and correct structured fields. For inferential use, they should compare the final treatment effect, budget impact or cost-effectiveness conclusion with a credible non-AI analysis and vary model assumptions to show how much the decision depends on the tool. A faster pipeline is valuable only if it preserves the evidence needed for accountability.
Patients and clinicians are affected when evidence pipelines shape coverage, guidance or access. An AI-derived endpoint should not become more persuasive merely because its internal model performs well. Decision documents should disclose where AI was used, the human checks, the data population, external validation, error rates and whether the funding recommendation changes under reasonable alternatives. Procurement contracts should preserve access to prompts, code, versions and audit logs rather than accepting a vendor's performance summary.[1]
What would change the assessment
A human dual review of the archived 300-record sample would show whether the estimated false-positive problem survives expert adjudication. Wider searches across Embase, economics databases, regulatory submissions and non-English sources would test coverage. The most useful prospective studies would embed a locked AI workflow in real health-technology assessments and compare completeness, errors, time, cost, analyst workload and the final recommendation against ordinary practice.
The author reported no funding and no conflict related to the work, while disclosing an academic affiliation and a health-economics consultancy serving pharmaceutical and life-sciences clients; no client participated or reviewed the manuscript. Search strings, records, prompts, outputs and code are publicly archived. The defensible conclusion is not that AI has failed health economics. It is that automation can be faster without making a spending decision truer, and the field must measure that difference explicitly.[1]