If The Maltese Falcon taught us anything, it is that the clue everyone ignored at the start may turn out to be the whole show. Friends, that is precisely the energy of this new 2026 paper in npj Digital Medicine: researchers took ordinary routine blood tests, the clinical wallflowers nobody invites to the glamorous diagnostics ball, and used machine learning to turn them into a gastric cancer triage tool for people with chronic dyspepsia.[1]

That matters because dyspepsia is common, endoscopy is expensive and invasive, and stomach cancer is the sort of villain you would prefer to catch before Act Three. Gastric cancer is usually confirmed by biopsy during endoscopy, but most people with chronic indigestion do not have cancer, which means healthcare systems can burn through a lot of scopes to find relatively few true cases.[2][3] It is a needle-in-a-haystack problem, except the haystack has billing codes.

A Modest Blood Test, Wearing a Trench Coat

Seo and colleagues trained their model on 210,463 people from Hong Kong Hospital Authority records collected between 2000 and 2015, then validated it on an independent cohort of 90,479 people from 2016 to 2020.[1] The ingredients were not exotic. No glittery sci-fi liquid biopsy panel. No bespoke sequencing assay with a price tag that makes your finance department lie down on the floor. Just 24 routine blood-test and demographic features.

The result was a model they call RBT-GC. In the validation cohort, where gastric cancer prevalence was 2.3%, the model sorted patients into low-risk, intermediate-risk, and high-risk groups with cancer prevalences of 0.3%, 1.9%, and 14.0% respectively.[1] That high-risk bucket is the headline. It pushed the number-needed-to-scope from 44 down to 7.[1]

That is not a tiny nudge. That is the difference between "scope nearly everyone with persistent symptoms and hope for the best" and "maybe send the front of the line to the people most likely to need it." The model also detected 1,276 cases in validation, compared with 102 for CEA and 42 for CA19-9.[1] Put bluntly, the old-school tumor markers showed up like a flashlight with dying batteries.

Why This Is Clever, and Slightly Sneaky

The best part of the paper is not that it invents a magical new blood signal. It does something sneakier. It asks whether boring, already-collected lab data can reveal a pattern that humans do not reliably spot at scale. Machine learning is good at this sort of thing. Give it enough examples and it starts noticing combinations of small signals that look meaningless one by one, like the one detective in the film who actually reads the whole case file.

That idea lines up with broader trends in cancer diagnostics. A 2024 PLOS Medicine study of 477,870 patients in England found that common abnormal blood tests in people with abdominal pain or bloating could meaningfully raise cancer risk estimates and support earlier referral.[4] In gastric cancer specifically, recent work has explored ML models built from clinical data,[5][6] metabolomic panels,[7] and broader liquid-biopsy strategies.[8][9] The field is clearly trying to answer the same question in several accents: can we find cancer signals earlier, more cheaply, and with less ordeal for patients?

Before We Start Declaring Victory Over the Endoscopy Queue

Steady on.

This was a retrospective study from one healthcare system, and retrospective models have a well-known habit of looking sharp in the dressing room before real-world deployment asks them to sing live. Practice patterns, lab standards, patient mix, and referral behavior vary across countries. A model trained in Hong Kong may not travel neatly to Birmingham, Boston, or Brisbane without recalibration.

There is also a deeper issue. Blood-based cancer tools are improving, but reviews keep making the same sober point: promising biomarker studies do not automatically translate into routine clinical practice.[8][9][10] Endoscopy still gives direct visualization and biopsy. That is the gold-standard detective with the warrant. RBT-GC is more like the sharp desk sergeant saying, "You, my friend, should move to the top of the list."

And that may be enough.

If future prospective studies hold up, this kind of tool could help health systems use limited endoscopy capacity more intelligently, especially where universal gastric cancer screening is not practical. It could also make routine care do double duty. Blood already drawn for ordinary workups might quietly become a low-cost early warning layer. Not a replacement for diagnosis, but a smarter traffic cop.

That is the appealing part here. No miracle serum. No robot prophet. Just the possibility that the clues were sitting in the chart all along, waiting for somebody with enough data and enough nerve to connect them.

References

1. Seo M, Cheung KM, Lam SJL, et al. Development and validation of a novel blood-based biomarker for gastric cancer triage in chronic dyspepsia. npj Digital Medicine. Published April 17, 2026. doi:10.1038/s41746-026-02618-1

2. Esophagogastroduodenoscopy. Wikipedia. https://en.wikipedia.org/wiki/Esophagogastroduodenoscopy

3. Stomach cancer. Wikipedia. https://en.wikipedia.org/wiki/Stomach_cancer

4. Rafiq M, Renzi C, White B, et al. Predictive value of abnormal blood tests for detecting cancer in primary care patients with nonspecific abdominal symptoms: A population-based cohort study of 477,870 patients in England. PLoS Medicine. 2024;21(7):e1004426. doi:10.1371/journal.pmed.1004426. PMCID:PMC11288431

5. Ke X, Cai X, Bian B, et al. Predicting early gastric cancer risk using machine learning: A population-based retrospective study. DIGITAL HEALTH. 2024;10:20552076241240905. doi:10.1177/20552076241240905

6. Kuwayama N, Hoshino I, Mori Y, et al. Applying artificial intelligence using routine clinical data for preoperative diagnosis and prognosis evaluation of gastric cancer. Oncology Letters. 2023;26(5):499. doi:10.3892/ol.2023.14087. PMCID:PMC10579989

7. Zhang Y, Li Y, Wang X, et al. Metabolomic machine learning predictor for diagnosis and prognosis of gastric cancer. Nature Communications. 2024;15:1706. doi:10.1038/s41467-024-46043-y

8. Díaz Del Arco C, Fernández Aceñero MJ, Ortega Medina L. Liquid biopsy for gastric cancer: Techniques, applications, and future directions. World Journal of Gastroenterology. 2024;30(12):1680-1705. doi:10.3748/wjg.v30.i12.1680. PMCID:PMC11008373

9. Nozeret K, Harlé A, de la Fouchardière C, et al. Protein Biomarkers of Gastric Preneoplasia and Cancer Lesions in Blood: A Comprehensive Review. Cancers. 2024;16(17):3019. doi:10.3390/cancers16173019

10. Liatsou E, Driva TS, Vergadis C, et al. Current Role of Artificial Intelligence in the Management of Gastric Cancer. Biomedicines. 2025;13(12):2939. doi:10.3390/biomedicines13122939

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.