RETURN TO ABSTRACT LISTING


Objective:
The adoption of FDA-approved AI algorithms into clinical practice depends largely on whether this technology provides significant value to radiologists in the clinical setting. The objective of this study is to understand current trends in FDA-cleared AI algorithms for neuroimaging and understand the value proposition of these algorithms as advertised by the developer.

Materials and Methods:
We extracted a list of AI algorithms for neuroimaging from the ACR Data Science Institute AI Central database from May 2008 to August 2022. The product information of each device was collected from the database. For each device, we collected information of advertised value as presented on the developer’s website.

Results:
The database included a total of 59 FDA-approved AI neuroimaging algorithms. 30/59 (50.8%) were compatible with CT images, with 23 (39.0%) being compatible with MR and 6 (10.2%) compatible with CTA or CTP. Five algorithms (8.5%) were compatible with multiple imaging modalities. Of these algorithms, 58 (98.3%) were approved with a 510(k) clearance and 1 (1.7%) received de novo clearance. Nearly 50% of these algorithms (29/59) are related to the diagnosis of stroke, 22% (13/59) for detection of intracranial hemorrhage, 15% (9/59) for stroke brain perfusion, and 14% (8/59) for large vessel occlusion (LVO) detection. Just over a third of all the algorithms were involved in delineating brain anatomy (20/59, 34%). Out of the 59 algorithms, we found websites that discussed the product for 55 algorithms. The most widely advertised value proposition was improved quality of care (38/55, 69.1%). A total of 24 algorithms (43.6%) argued they saved the user time, 9 (15.7%) decreased costs, and 6 (10.9%) increased revenue. Product websites for 26 algorithms (43.6%) showed testimonials advertising the value of the technology. Subgroup analysis was also performed to compare algorithms compatible with CT to those with MRI. While CT-compatible algorithms were designed to can be used for triage (15/30, 50.0%), image processing/quantification (IPQ) (14, 46.7%), and diagnosis of pathology (1, 3.3%), the MRI-compatible algorithms only helped in IPQ. Both types of algorithms had similar value propositions like other FDA-approved algorithms , with CT-compatible algorithms more commonly advertised to improve quality of care (77.8% (21/27) vs 68.2% (15/22), p - value = 0.666) and cost reduction (18.5% (5/27) vs 13.6% (3/22), p=0.943) and MRI-compatible algorithms more commonly advertised to save time (45.5% (10/22) vs 37.0% (10/27), p= 0.761) and increase revenue for hospital or provider (18.2% (4/22) vs 11.1% (3/27). Analysis was also performed to study the value propositions of stroke algorithms.

Conclusion:
Our results indicate a wide range of value propositions advertised by developers to indicate the value of their product. Most developers argued that their product would improve patient care. Given the rapid growth in the number, availability, and clinical application of AI algorithms across the US, it is important to understand the different value propositions provided by the software.