AI tool reviews 50 national cancer plans and finds weak financing and outcome goals

Study title: Assessment of comprehensiveness of national cancer control plans using a multiagent artificial intelligence model: a global analysis

Researchers built a multiagent AI pipeline called PRISM to read and score the most recent national cancer control plans from 50 countries against a framework of 12 elements and 76 sub-elements. The framework was checked by experts through a two-stage Delphi process, and the AI was tested against masked ratings from two experts on six plans.

Read the study abstract

Background National cancer control plans (NCCPs) vary in scope and comprehensiveness, yet expert-led review is resource intensive and limits cross-country comparability. We aimed to assess NCCP comprehensiveness across 50 countries using the policy reasoning integrated sequential model (PRISM), a validated multiagent artificial intelligence (AI) system. Methods We identified the most recent national NCCP for each country from the International Cancer Control Partnership repository, the WHO Noncommunicable Diseas…

By the Truvace deskNovember 3, 20261 min read1 sourceEvidence: Emerging · 48/100 — Benchmark or lab evaluation · n = 50 · 1 independent sourcePeer-reviewedTRV-2026-2342· cite & verify

Assessment of comprehensiveness of national cancer control plans using a multiagent artificial intelligence model: a global analysis — Insurance policy issued by the Post Office of the Japanese Government-General of Korea (January 1936) by The Government-General of Korea, Empire of Japan.. Public domain

In brief

The work matters because it shows that automated appraisal could make policy review more comparable and timely across countries, while revealing that plans are weakest on financing, outcomes and threat analysis. The validation set was small, and the authors describe AI appraisal as a possibility rather than a proven replacement for expert review.

Main points

Median overall comprehensiveness across 50 plans was 2.78 on a 0-5 scale, in the moderate band.
Europe and lower-middle-income countries had the highest medians, while the western Pacific and low-income countries had the lowest.
Implementation, health services and governance scored highest; financing, outcomes and threats scored lowest.
Only five of eleven financing sub-elements had median scores of 1.0 or less, the largest gap found.

The gain

An AI system validated against expert ratings could score 50 national cancer control plans on 76 sub-elements, making cross-country comparison faster than resource-intensive expert review and exposing financing gaps.

The rundown

The team selected each country's most recent national plan from the International Cancer Control Partnership repository, the WHO noncommunicable disease document repository and government websites. Sub-national plans, broader health strategies, drafts and superseded versions were excluded.

PRISM combined document ingestion with Qwen2.5-VL-72B, policy analysis using retrieval-augmented generation with Llama 4 Scout, and a semantic quality assurance step. It was validated against masked dual-expert ratings for six purposely selected plans before being applied to all 50.

Element-level medians were 4.17 for implementation, 3.89 for health services and 3.67 for governance, versus 1.88 for financing, 1.83 for outcomes and 2.19 for threats. Low-income countries had a median of 2.05 and Europe 2.96. The study was funded by Harvard University.