Nanox.AI, the medical imaging analytics arm of Nano-X Imaging Ltd (Nasdaq: NNOX), said clinical study data on its coronary artery calcium software was presented at the Society of Cardiovascular Computed Tomography annual meeting. The company’s own annual report on Form 20-F, filed with the U.S. Securities and Exchange Commission, sets out what those studies are and where they were run, and the U.S. Food and Drug Administration file on the product sets out what the software is cleared to do.

The product at issue is HealthCCSng. In the 20-F the company said its established HealthCCSng solution for opportunistic detection of coronary artery calcium was being evaluated by three university medical centers, Mass General Brigham, University of Washington and Wake Forest Health, in a study assessing the effect of calcium detection on patient treatment rates. The filing said all sites were in the 12-month follow up period and that initial results would be presented at the SCCT conference in San Diego. That is the evaluation programme the conference material draws on.

What the FDA file says the software does

HealthCCSng has been through the 510(k) route twice. The FDA cleared the original submission, K210085, on September 15, 2021, filed by Zebra Medical Vision, Ltd. and received on January 13, 2021, with HealthCCS, K172983, as the predicate. A second clearance, K241440, was granted to Nano-X AI, Ltd. on August 19, 2024, on a submission received on May 21, 2024, and was processed as a Special 510(k). Both sit under regulation 892.1750, a class 2 computed tomography device.

The indications for use are narrow and the FDA text is worth reading closely. HealthCCSng is a non-invasive post-processing software to evaluate calcified plaques in the coronary arteries. It analyses routine non-gated, non-contrast CT studies that include the entire heart of adult patients of age 30-85. It produces an exact calcium score and a four category output with preview axial images, and the clearance states that the output is meant for informational purposes only. The FDA file adds that HealthCCSng results “are not intended to be used on a stand-alone basis for risk attribution, clinical decision-making or otherwise preclude clinical assessment of CT studies”, and that the final diagnosis is made by the clinician after reviewing the scan independently of the software.

The categories are fixed thresholds. A score of 0 is reported as Zero Calcium, 1-99 as Low, 100-399 as Medium and above 400 as High. The default calcium threshold is 130 Hounsfield Units, and slices containing measured coronary calcification are annotated and returned as key images. Scans arrive from the Nanox AI Imaging Analytics Platform or another hosting application, and the software can also return a non-compliant result, a success with no category where metal artefacts suggest known cardiac disease, or a failure.

The 2024 clearance widened the product rather than remaking it. Supported slice thickness went from up to 3.1 mm in the cleared device to up to 5.1 mm. The output moved from three categories to four, separating a true zero from the low band, and the exact score was added alongside the category. The intended population gained an upper age limit of 85 where the earlier clearance covered patients aged 30 and above.

The validation behind the clearances

The performance evidence in the FDA file is the closest thing on the public record to an independent measure of the software. For K241440 the company ran a retrospective, simulated use, standalone study against a ground truth set by the majority agreement of two out of three US board certified radiologists experienced in identifying coronary calcium on non-gated CT studies. The validation dataset drew on four healthcare institutions, Intermountain Healthcare, Clalit Health System, Northwell Health and USARAD, and was engineered to contain enough cases in every calcium category.

Of 436 cases the algorithm returned a result on 427, a yield of 97.94%. Mean age was 61.89 years with a standard deviation of 13.11, and 58.31%, or 249 patients, were male. Scans came from four CT manufacturers, GE, Philips, Siemens and Toshiba, across slice thicknesses from 0.5 mm to 5 mm.

Overall agreement between the software category and the ground truth category was 89.46%, with a 95% confidence interval of 86.15% to 92.21%. By category, agreement was 86.63% for Zero, 87.65% for Low, 87.36% for Medium and 98.85% for High. The correlation between the software score and the ground truth score was r=0.959. The 2021 clearance reported overall agreement of 92.5%, with a confidence interval of 89.7% to 95.5%, on a chest CT dataset drawn from two healthcare institutions and truthed the same way.

The 2024 file also carries a paired comparison of gated and non-gated scans. On 100 paired scans read both ways, the software agreed with itself on 88.89% of gated scans scored at zero but on only 48.15% of those in the 1-99 band and 54.55% in the 100-399 band. The FDA summary states plainly that calcium scoring on non-gated CT scans is not as accurate as on gated scans.

Analysis: the clinical question is downstream of the algorithm

The measured agreement figures are not what the SCCT programme is testing. Agreement with radiologists was settled at clearance and is a property of the software. What the three university medical centers are studying, on the company’s own description in the 20-F, is the effect of calcium detection on patient treatment rates. That is a health system question rather than an imaging question: whether a flag generated on a scan ordered for something else changes what a clinician then does.

The distinction matters because the clearance forecloses the easy version of the claim. HealthCCSng is not cleared for risk attribution or clinical decision-making on a stand-alone basis, and the FDA text puts the diagnosis with the clinician reviewing the scan independently. Any finding that AI-detected calcium leads to more preventive therapy is therefore a finding about a workflow that includes a radiologist, a report and a referring physician, not about the algorithm in isolation. A reader cannot separate the software’s contribution from the reporting pathway built around it on the evidence published so far.

The one throughput figure the company publishes points the same way. On its website Nanox describes Corewell Health identifying nearly 4,000 new patients in 2023, against 268 patients having coronary artery calcium reported in the previous two years. That is a very large ratio, and it is a measure of how little incidental calcium was being reported before, not a measure of algorithm accuracy. On the paired gated and non-gated table in the FDA file, agreement is highest at the extremes and lowest in the low and intermediate bands, which are the bands where a treatment decision is most likely to be marginal.

What a careful reader would look for next is the published form of the three-centre work. The 20-F says the sites were in a 12-month follow up period, so the relevant outputs are treatment initiation rates at the follow-up point, the proportion of flagged patients who received a report action, and whether the effect holds outside the enriched category mix used for FDA validation. Conference abstracts are not peer-reviewed publications, and none of the study data has been furnished to the SEC in a filing. Until a full paper appears, the FDA file remains the only detailed, independently reviewed description of what this software does.