Case TS-32818B9D15 Aug 2026researchCompound claim

AI

“Bengaluru-based startup Dognosis reported about 90% sensitivity in a Phase 2 study using cancer-sniffing dogs combined with AI to detect early-stage cancers across seven broad cancer groups covering more than 20 cancer types, and has begun a Phase 3 trial involving around 10,000 people across 10 Indian hospitals.”

Plain restatementA published Phase 2 diagnostic-accuracy study by Dognosis reported approximately 90% sensitivity for early-stage cancer detection from breath samples using trained dogs plus computational modeling, across seven cancer groups, and the company has started a larger prospective study targeting roughly 10,000 participants at 10 Indian hospitals.

Mostly accurateConfidence Medium
What this verdict means →

Distortion code this site does not recognise yet: unreleased_as_released. Not collectible until the field guide has an entry.

This claim mostly holds up. Dognosis, a Bengaluru startup, did publish a study in the Journal of Clinical Oncology, a leading cancer journal, reporting 90.8% sensitivity and 91.3% specificity for detecting cancer from breath samples using trained dogs plus a statistical model, with 90.6% sensitivity for early-stage disease across seven cancer groups. The numbers in the post are accurate. Two things are missing, though. First, this was a case-control study, meaning the dogs sorted samples from people already diagnosed with cancer against separately recruited healthy people, and the study authors themselves say it only establishes that the signal is real and that it must now be tested in an actual screening population. In a general population where cancer is rare, a test with 91% specificity would produce far more false alarms than true finds. Second, the study was funded and run by Dognosis, whose CEO is the corresponding author and a shareholder, and no independent group has reproduced the result. The "more than 20 cancer types" figure is the company's description of what the system targets, not something the study validated type by type. The larger 10,000-person trial at 10 hospitals did begin in April and is a recruitment target, not a completed study.

The drift / as claimed vs as evidenced

[drifted from the evidence:] Bengaluru-based startup Dognosis reported about 90% sensitivity in a Phase 2 study using [drifted from the evidence:] cancer-sniffing dogs [drifted from the evidence:] combined with AI to detect early-stage cancers across seven [drifted from the evidence:] broad cancer groups [drifted from the evidence:] covering more than 20 cancer types, and has [drifted from the evidence:] begun a [drifted from the evidence:] Phase 3 trial involving around 10,000 [drifted from the evidence:] people across 10 Indian hospitals.


A [added by the neutral restatement:] published Phase 2 [added by the neutral restatement:] diagnostic-accuracy study [added by the neutral restatement:] by Dognosis reported approximately 90% sensitivity for early-stage cancer detection from breath samples using [added by the neutral restatement:] trained dogs [added by the neutral restatement:] plus computational modeling, across seven cancer groups, and [added by the neutral restatement:] the company has [added by the neutral restatement:] started a [added by the neutral restatement:] larger prospective study targeting roughly 10,000 [added by the neutral restatement:] participants at 10 Indian hospitals.

Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.

The trace / claim to source

Where it appeared
⌿ Omitted qualifier
A load-bearing condition from the source quietly disappears from the claim.
$ Marketing as evidence
Promotional material dressed up as independent proof.
▲ Exaggeration
A real finding gets inflated: stronger, bigger, faster, or more certain than the evidence supports.
unreleased_as_released
Tertiary sourcedata aggregators
Tracxn, PitchBook, Crunchbase company profiles
Secondary sourcenamed-outlet accountable journalism, originating report
Bloomberg, "Indian Startup Leans on Cancer-Sniffing Dogs and AI for Early Detection," Satviki Sanjay, 14 Aug 2026
Secondary sourcenamed outlet
Business Standard syndication of the Bloomberg report, 14 Aug 2026
Secondary sourcetech press, downstream of the Bloomberg story
TheNextWeb and Digital Trends coverage, 14-15 Aug 2026
Primary sourcerefereed venue, top-tier oncology journal
Kulgod S. et al., "Canine Olfaction Combined With Bayesian Modeling for Multicancer Detection From Breath Samples: A Phase II Study in India," *Journal of Clinical Oncology* 2026;44(19):1774-1783
Primary sourcepreprint, unrefereed version of the refereed paper
medRxiv preprint of the same study, 2025.09.21.25336259, full text and PDF publicly available
Primary sourcevendor channel
Dognosis newsroom, "Dognosis publishes Phase-2 Trial in Journal of Clinical Oncology," 24 Apr 2026, and dognosis.tech/breatheasy
Primary sourcerefereed, but authored entirely by Dognosis staff and explicitly a perspective/hypothesis piece
Dakhave M., Bitan I., Moitra P., Kulgod A., "Digitized canine olfaction and multimodal biosensing...: a hypothesis-driven perspective," *J. Breath Research* 20(2), Feb 2026
Primary sourceconference abstract, interim, not full peer review
AACR abstract P93, "Multi-Cancer Screening Using AI-Enhanced Canine Olfactory Detection: Interim Results," *Cancer Res* 2026;86(13_Suppl)
● Primary source found
What is true
  • Dognosis is a Bengaluru-based startup working on breath-based cancer detection using trained dogs plus computational modeling. Confirmed by the paper's affiliations and by multiple named outlets.
  • A Phase 2 study exists, is real, and is published in the Journal of Clinical Oncology, a leading refereed oncology journal. This is not a preprint-only or press-release-only claim.
  • The "about 90% sensitivity" figure is accurate and, if anything, slightly understated: 90.8% overall sensitivity and 91.3% specificity.
  • The early-stage specificity of the claim holds. Stage I to II sensitivity was 90.6%, which is genuinely notable because early-stage sensitivity is where blood-based multi-cancer tests typically weaken.
  • "Seven broad cancer groups" is exactly what the paper reports for the test cohort.
  • The study is substantial in scale: 3,275 participants enrolled, 1,502 in the test cohort, six hospitals, assessor-masked, with training and test sets partitioned at the participant level so no test sample or person appeared in training.
  • A Phase 3 prospective study at 10 hospitals targeting roughly 10,000 participants, begun in April, is reported by Bloomberg and corroborated by the company's own description of its prospective "FirstAlert" study.
  • The post's attribution to Economic Times is plausible. The originating report is a Bloomberg story dated 14 Aug 2026, widely syndicated to Indian outlets including Business Standard.
What is misleading
  • Omitted qualifier: The single most important missing piece of context is that this was a case-control study, not a screening study. Cases were people already diagnosed by biopsy, and controls were separately recruited. The authors themselves declined to report positive predictive value precisely because the case-control design artificially inflates prevalence, and they state the results establish analytical validity and require confirmation in prospective population studies. A reader of the post has no way to know that "90% sensitivity" describes classification of stored, frozen breath samples in a cohort roughly 19% of whom had confirmed cancer, rather than performance in a general population.
  • Marketing as evidence: The post presents the result without noting that the study was funded by Dognosis, that the corresponding author is the company's CEO and a shareholder, that site investigators received institutional funding from the sponsor, and that a patent is disclosed. Peer review at JCO is a meaningful check, but peer review is not independent replication. No third party has run this system.
  • Exaggeration (scope): "covering more than 20 cancer types" attached to the study result. The paper reports seven cancer groups. The ">20 cancer types" figure is a company descriptor of what the approach targets, which TheNextWeb attributes to the company rather than to the study. The study provides no validated per-type breakdown across 20-plus types, and where the paper did break results down, the small groups had very wide confidence intervals, including a pooled rare-cancer category at 76.9% with a CI spanning 53.8 to 100.0. The phrasing implies validated 90% performance across 20-plus individual cancers, which the study does not demonstrate.
  • Unreleased as released (mild): "a Phase 3 trial involving around 10,000 people" describes a recruitment target over 12 months, not current enrollment. As of mid-August 2026 the study began in April and is enrolling. The post's phrasing reads as though 10,000 people are already involved.
  • Demo to product conflation (mild, in the post's image text rather than the caption): The overlay text states the startup "is using cancer sniffing dogs and AI to detect cancer from human breath" in the present tense as an operating capability. Coverage notes BreathEasy is a prescreening tool in trials, not a proven diagnostic, and that a positive result would direct someone toward further testing rather than provide a diagnosis. The caption is more careful than the image.
What is uncertain
  • How much the "AI" contributed to the 90.8% figure. This matters because the claim's headline pairing is "dogs combined with AI." The JCO abstract describes the fusion as incorporating historical dog performance and participant-level covariates. The preprint abstract says the study also assessed whether sensor-derived behavioral signals improved accuracy. The preprint's discussion describes EEG-based dynamic weighting of dog-specific likelihood terms in the future tense, saying such refinement "is expected to" narrow uncertainty and improve performance, which suggests the brain-computer-interface component described in the post's caption was not the driver of the published result. I could not retrieve the paywalled full text to settle the sensor contribution, so I am not asserting this. It is flagged as unresolved.
  • The role of participant-level covariates in the model. A Bayesian fusion that incorporates participant covariates is partly using non-olfactory information. Without the full text and the model specification, I cannot quantify how much of the reported accuracy derives from the dogs' reads versus from covariates.
  • Phase 3 registration. I could not locate a CTRI registration entry for the prospective 10,000-person study. Its existence rests on company statements relayed by Bloomberg and on the company's own site.
  • The CTRI number discrepancy in the preprint (CTRI/2024/10/075938 in the abstract versus CTRI/2024/03/061847 in the methods) is unexplained.
  • The commissioned JCO editorial. I could not retrieve it. The quotations circulating come from the company's own Substack post, which is an interested party quoting its reviewers.
  • Whether the post's stated source, Economic Times, carried these exact figures. I could not retrieve the ET article directly. The figures match the Bloomberg original and its syndications.
Evidence summary

The underlying study is real, is published in a top-tier refereed oncology journal, and the headline number in the claim matches it closely. The JCO abstract states the fusion system achieved 90.8% sensitivity (95% CI, 87.2 to 94.5) and 91.3% specificity (95% CI, 89.7 to 92.9), with an ROC AUC of 0.962. Sensitivity for stage I to II disease was 90.6%. The design was an assessor-masked, multicenter case-control study across six hospitals in Karnataka, India, registered as CTRI/2024/10/075938. A total of 3,275 participants were enrolled, 1,773 for training and 1,502 for testing. The test cohort comprised 283 treatment-naive, biopsy-confirmed cancer cases spanning seven major cancer groups, and 1,219 controls consisting of healthy volunteers, people with non-oncologic chronic disease, and people with benign biopsy results. Breath was collected on cotton surgical masks, stored under cold-chain conditions at minus 20 degrees Celsius for up to six months, and later evaluated by seven trained dogs at a central laboratory. The JCO abstract describes the modeling step as integrating individual dog indications "using a Bayesian fusion framework incorporating historical dog performance and participant-level covariates." The authors are explicit about what the study does and does not establish. The paper frames the result as analytical validity and states that these findings "support prospective evaluation in true screening populations." The preprint's discussion states that positive and negative predictive values are not reported "given the case-control design and artificially inflated" prevalence, and that the results "require confirmation in prospective, population-based studies to determine the clinical utility of true screening prevalence." On the Phase 3 element, Bloomberg reports that Dognosis began the trial in April with 10 hospitals across India and plans to recruit about 10,000 people over 12 months, testing asymptomatic higher-risk people and cancer survivors at risk of recurrence. Bloomberg also reports Kulgod's statement that the test does not need Indian regulatory approval because it is positioned as a prescreening rather than a diagnostic device.

Complete reasoning
Every load-bearing number in the claim traces to a retrievable primary artifact and matches it. The 90% sensitivity, the early-stage figure, the seven cancer groups, and the published Phase 2 status all check out against JCO 2026;44:1774-1783 and its openly available medRxiv version, and the Phase 3 details are corroborated by named-outlet reporting. I considered and rejected "Source exists but framing is misleading" because the claim uses appropriately hedged language throughout: it says "reported," it names the study phase, and it says the next trial is "to further validate the technology," which signals the technology is not yet validated. I also considered and rejected "Accurate" because two gaps are real: the ">20 cancer types" figure is a company descriptor presented as a study finding, and the omission of the case-control design means a reader will over-read what 90% sensitivity implies for real-world screening. Confidence is Medium rather than High because the JCO full text is paywalled, leaving the actual contribution of the AI and sensor components unresolved, because no independent party has replicated this system, and because the Phase 3 details rest on company statements relayed through a single originating wire report with no trial registration located. As of 2026-08-15.
Use this case

The reply is formatted for pasting into the thread where the claim is circulating.

Compact share page: ai.trueseeker.com/s/32818b9dfccd/gwYdspq-W_JNheUJn7J6hnBb3AY

Ask this case

Answers come only from the case file above; nothing is added.

Did Dognosis really report about 90% sensitivity for detecting cancer using dogs and AI?

Yes. The Phase 2 study published in the Journal of Clinical Oncology reported 90.8% overall sensitivity and 91.3% specificity, with 90.6% sensitivity for early-stage disease. The claim's headline figure matches the published paper closely.

Does this mean the test can reliably screen the general population for cancer?

The investigation did not establish this. The study was a case-control design, comparing people already diagnosed with cancer to separately recruited healthy people, and the authors themselves say positive and negative predictive values were not reported because the case-control design artificially inflates prevalence. They state the results need confirmation in prospective, population-based studies before clinical utility can be judged.

Was the study covering more than 20 types of cancer?

No. The published study covered seven broad cancer groups, not more than 20 individual cancer types. The '20-plus cancer types' figure is a company description of what the approach targets generally, not something the study validated type by type.

Is the study result independent, or is it tied to the company?

The study was funded and run by Dognosis, its CEO is the corresponding author and a shareholder, and a patent is disclosed. It was peer reviewed and published in a leading journal, but no independent group has reproduced the result.

Has the Phase 3 trial with 10,000 people already happened?

No. The Phase 3 study began in April at 10 Indian hospitals and is targeting roughly 10,000 participants over 12 months, according to Bloomberg. It is a recruitment target and ongoing enrollment, not a completed study with 10,000 people already involved.

Similar cases on record