Nimblemind nMAS finds H. pylori in gastric biopsy reports

Nimblemind Multi-Agent System (nMAS) processed 54 de-identified gastric biopsy pathology reports from a large Singapore healthcare system and correctly classified 213 of 216 feature-case decisions, an overall accuracy of 98.61% (arXiv:2607.06435, submitted 7 Jul 2026). The retrospective pilot evaluated four clinician-scoped binary fields and returned unified report-level outputs with supporting source sentences.

What did the pilot evaluate and find?

The pilot tested nMAS on four clinician-scoped binary fields across 54 gastric biopsy pathology reports, producing 216 feature-case decisions and correctly classifying 213, for 98.61% overall accuracy. The four fields evaluated were gastric/stomach biopsy, biopsy status, H. pylori positivity, and H. pylori-associated gastritis. The paper also notes background prevalence data from Singapore: about 31% of the population had evidence of Helicobacter pylori infection, explaining why scalable extraction from reports matters for screening and prevention. The authors frame nMAS as a field-name-driven, evidence-linked extraction workflow designed to locate evidence spread across coded and free-text fields while handling contextual assertion and negation that defeat simple keyword search.

How does nMAS compare to the UMA-style MiniMax M2.5 comparator?

nMAS produced aggregate and per-field metrics similar to a separately implemented UMA-style MiniMax M2.5 comparator, but nMAS preserved unified report-level outputs that linked each decision to supporting source sentences. The paper emphasizes that the contribution demonstrated is workflow integration and traceability rather than predictive superiority, since the comparator achieved comparable predictive performance. The report describes nMAS as delivering evidence-linked extraction that keeps the provenance of each classification visible, a feature the authors contrast with the comparator implementation.

Why it matters

If evidence-linked, field-driven extraction replaces or accelerates manual review, the potential time savings are large: the authors describe an illustrative, unmeasured scenario where reviewing 1,000 reports at five minutes per manual review versus five seconds per evidence-linked verification would cut review time from 83.3 to 1.4 staff-hours, a reduction of 81.9 staff-hours and about USD 6,100 in potential staff-time value. That example frames the practical benefits the authors expect from traceable, source-linked outputs: faster clinician verification and less manual labour when evidence is already tied to the extracted label. The paper cautions these gains are illustrative and calls for larger, multi-institutional studies to measure evidence-span correctness, clinician verification time, and generalizability before drawing operational conclusions.

What to watch

The next concrete milestones are external validations that measure evidence-span correctness and clinician verification time across institutions and report styles; those studies will test whether nMAS's traceable outputs scale beyond this pilot. The arXiv submission (arXiv:2607.06435, Yufan Wang et al., submitted 7 Jul 2026) recommends multi-institutional evaluations and explicit measurement of the verification workflow as the immediate follow-ups needed to move from pilot to deployment-ready evidence extraction.