Clinical evidence
The research behind LungAI
We publish how LungAI performs, how we tested it, and where the evidence is still limited, so radiologists and screening programs can check our work.
LungAI: An AI Biomarker that Refines Lung-RADS v2022 Triage of Screen-Detected Pulmonary Nodules
medRxiv 2026. doi:10.64898/2026.11.09.26367662
Key findings
Same biopsy volume, better triage
Added to Lung-RADS v2022, LungAI moved patients along the existing management ladder: cancers toward earlier workup, benign nodules away from biopsy. Figures are per 1,000 screening examinations.
−46%
3.1 → 1.6 per 1,000
Cancers left on the 12-month rescan pathway
Fewer cancers waiting a full year for their next look.
−20%
32.7 → 26.2 per 1,000
Benign biopsy referrals
Fewer invasive procedures for nodules that turned out not to be cancer.
≈ Same
51.0 → 50.1 per 1,000
Total biopsy volume
The gains came from a better mix of who is referred, not more procedures.
0.92
AUROC (95% CI 0.88–0.95)
Malignancy discrimination
Higher than Sybil, nodule diameter, and nodule volume (all P<0.001).
Built to work with Lung-RADS®, not replace it. Compared with Lung-RADS v2022 alone, LungAI improved precision for finding cancers (AUPRC +0.31, P=0.003), but the difference in AUROC was not statistically significant (+0.03, P=0.20). Its value in this study came from adding risk information that size and growth don't capture.
Source: Suri H, et al. medRxiv 2026, doi:10.64898/2026.11.09.26367662. Assumes full adherence to recommended follow-up. The preprint also reports estimated lead time and cost effects, which depend on modeling assumptions; see the paper for details. Lung-RADS® is a registered trademark of the American College of Radiology.
Study design
How the study worked
Held-out screening scans
2,747 low-dose CT exams from 1,244 National Lung Screening Trial patients with a 4 to 30 mm nodule, never used to train LungAI. 89 pathologically confirmed cancers.
Head-to-head comparisons
LungAI was compared with Sybil, nodule diameter, and nodule volume. A board-certified radiologist assigned Lung-RADS v2022 categories on a random 226-exam subset.
Integration with Lung-RADS
LungAI scores shifted patients up or down the existing Lung-RADS management ladder, using score cutpoints fixed before the analysis.
Patient-level outcomes
Results were projected to the full cohort and reported per 1,000 examinations: cancers on each pathway, benign biopsies, and total biopsy volume.
Read this before you rely on it
Limitations and disclosures
Every study has limits. These are the ones the authors state in the paper.
Limitations
- One cohort. A held-out NLST validation with no external cohort yet. NLST scans date from 2002 to 2004.
- Solid nodules. Only 7 part-solid nodules entered the cohort, so conclusions apply to solid nodules. Ground-glass disease is out of scope.
- Retrospective. Stage at diagnosis reflects care delivered in NLST, so the study cannot show how reclassification would change stage.
- Preprint. The paper has not yet been certified by peer review.
Funding and conflicts of interest
- The study was funded by Oatmeal Health, which develops LungAI.
- Seven authors hold equity in Oatmeal Health. The radiologist who assigned Lung-RADS categories is a paid advisor.
- Co-author affiliations (UCLA, the VA, Harvard Medical School, Mount Sinai, and others) are listed for identification and do not imply institutional endorsement.
- NLST data are available through the NCI Cancer Data Access System. Analysis code is available on reasonable request.
The problem we're working on
Why nodule triage matters
Published research by other groups. These are not studies of Oatmeal Health products.
20%
Fewer lung cancer deaths with screening
Annual low-dose CT reduced lung cancer mortality by 20% in the National Lung Screening Trial, and by 24 to 33% in the NELSON trial.
NLST Research Team, N Engl J Med 2011 · de Koning HJ, et al., N Engl J Med 2020
22–37%
Of Lung-RADS 4B nodules are cancer
Most nodules in a high-suspicion category turn out benign, yet many are sent for invasive workup.
Gu JZ, et al. J Am Coll Radiol 2023;20(8):769–780
29%
Of reader pairs disagree on Lung-RADS
Lung-RADS rests mainly on size and attenuation, and secondary features are judged subjectively.
van Riel SJ, et al. Eur Radiol 2019;29(2):924–931
~15%
Of needle biopsies cause a collapsed lung
Transthoracic needle biopsy causes pneumothorax in about 15% of procedures, 6.6% needing a chest tube.
Wiener RS, et al. Ann Intern Med 2011;155:137–144
68–92%
Five-year survival at stage I
Survival falls below 10% at stage IV, so finding cancers earlier changes outcomes.
IASLC staging, Goldstraw P, et al. J Thorac Oncol 2016
50–80
Ages now eligible for screening
The 2021 USPSTF update widened eligibility to adults 50 to 80 with a 20 pack-year history, adding millions of people to screening programs.
US Preventive Services Task Force. JAMA 2021;325:962–970
What's next
Where the evidence goes from here
This preprint is the first step. We'll add each new study to this page as it's published.
November 2026
Preprint posted
Full methods and results public on medRxiv.
In progress
Peer review
Submission to a peer-reviewed journal.
Underway
External validation
Multi-site validation on contemporary screening data at several institutions.
Pending
FDA 510(k)
Premarket notification pending with FDA.
The team
The people behind the study
All twelve authors of the LungAI preprint, with affiliations as listed in the paper.
- Huzaifa Suri, MSOatmeal Health
- Yakov Keselman, PhDOatmeal Health
- Lance Legel, MSOatmeal Health
- Jonathan GovetteOatmeal Health
- Bradley Mullen, MDOatmeal Health
- Nirmal Choradia, MDOatmeal Health; University of Oklahoma Health Sciences Center
- Thamer Sartawi, MDOatmeal Health; Blanchard Valley Health System
- Fereidoun Abtin, MDDavid Geffen School of Medicine at UCLA
- Drew Moghanaki, MDDavid Geffen School of Medicine at UCLA; US Department of Veterans Affairs
- Carey Thomson, MD, MPHHarvard Medical School
- David Yankelevitz, MDIcahn School of Medicine at Mount Sinai
- Yuri Ahuja, MD, PhDOatmeal Health; Icahn School of Medicine at Mount Sinai
Affiliations are listed for identification and do not imply institutional endorsement. See the paper for each author's disclosures.
Selected prior research by our team
These studies were done before or outside Oatmeal Health and do not evaluate Oatmeal products.
Yuri Ahuja, MD PhD
Co-Founder and Chief Technology Officer
Corresponding author of the LungAI study. Physician-scientist whose research applies machine learning to health records and medical signals.
- Vaid A, …, Ahuja Y, et al. A foundational vision transformer improves diagnostic performance for electrocardiograms. npj Digital Medicine 2023;6:108. DOI
- Ahuja Y, et al. sureLDA: A multidisease automated phenotyping method for the electronic health record. JAMIA 2020;27(8):1235–1243. DOI
- Li Y, Nair P, …, Ahuja Y, et al. Inferring multimodal latent topics from electronic health records. Nature Communications 2020;11:2536. DOI
Yakov Keselman, PhD
Principal Machine Learning Engineer
Co-author of the LungAI study. Trained the 3D CT foundation model behind LungAI. PhD in computer science (Rutgers) and former assistant professor at DePaul University.
- Keselman Y, Dickinson S. Generic model abstraction from examples. IEEE Transactions on Pattern Analysis and Machine Intelligence 2005;27(7):1141–1156. DOI
- Demirci MF, Shokoufandeh A, Keselman Y, et al. Object recognition as many-to-many feature matching. International Journal of Computer Vision 2006;69(2):203–222. DOI
For radiology and screening teams
Walk through the evidence with our team
We'll take you through the methods and results, and talk about how a validation on your own data could work.
LungAI is investigational and not FDA 510(k) cleared, and is not available for sale in the United States. Results are from a retrospective study of held-out NLST data reported in a preprint that has not been peer reviewed.