An AI workflow scanning gastric biopsy reports for H. pylori infection matched clinician-level accuracy in a small Singapore pilot — and the efficiency gap it exposed is hard to ignore.
Researchers tested the Nimblemind Multi-Agent System (nMAS) on 54 de-identified gastric biopsy pathology reports from a Singapore healthcare system. The system evaluated four binary fields per report — including whether H. pylori was present and whether it had caused associated gastritis. Across 216 individual classification decisions, nMAS got 213 right, a 98.61% accuracy rate. A separate comparator model using MiniMax M2.5 produced similar numbers, so the paper's real claim isn't that nMAS is uniquely accurate — it's that it produces traceable, evidence-linked outputs that slot into clinical workflows more cleanly than a raw classifier would.
That distinction matters because H. pylori affects roughly 31% of Singapore's population and is a known driver of gastric cancer if left untreated. The challenge isn't diagnosing it once a biopsy is taken — it's systematically finding every positive case buried across free-text fields, coded entries, and negation language that trips up keyword search. The nMAS approach attaches source sentences to each decision, giving a clinician something to verify rather than a bare prediction to trust or reject.
The time estimate in the paper — 83 staff-hours of manual review reduced to 1.4 hours for 1,000 reports — is labeled illustrative and unmeasured, which is worth flagging: it assumes five minutes of manual review per report versus five seconds of AI-assisted verification, numbers the authors did not validate in this study. With 54 reports and no multi-institutional data yet, the scalability case remains a projection, not a finding.