The FDA has published a new list of approved artificial intelligence/machine learning products. This compilation is not intended to provide an exhaustive or all-encompassing reference for medical devices incorporating AI/ML technology. Rather, it serves as a catalog of such devices spanning various medical disciplines, primarily drawn from information found in the summary descriptions of their marketing authorization documents.
Key observations from this compilation: The insights we’re about to delve into have been derived from both publicly available information contained in the summary descriptions of marketing authorization documents and aggregated data from internal sources. It’s important to note that the publicly accessible summaries don’t encompass the entirety of the information often included in the full submissions, which can run into thousands, if not tens of thousands, of pages. Additionally, these summaries are typically composed by the application submitters, reflecting what they consider significant, which may not consistently encompass details about AI/ML techniques and other insights regarding broader trends.
Growth trends: The year-over-year expansion of AI/ML-enabled medical devices showed a deceleration in 2021 (15%) and 2022 (14%) following a substantial increase of 39% in 2020 compared to the previous year (2019). Projections for 2023 anticipate a significant growth of over 30% when compared to 2022.
Distribution across medical disciplines: In the calendar year 2022, a notable 87% of the devices listed in this compilation were authorized in the field of radiology (122 devices), followed by 7% in cardiovascular (10 devices), and 1% each in neurology (2 devices), hematology (1 device), gastroenterology/urology (1 device), ophthalmic (2 devices), clinical chemistry (1 device), and ear, nose, and throat (1 device).
2023 authorizations: Up until the end of July 2023, 79% of the devices authorized this year fall under the category of radiology (85 devices), with 9% in cardiovascular (10 devices), 5% in neurology (5 devices), 4% in gastroenterology/urology (4 devices), 2% in anesthesiology (2 devices), and 1% each in ear, nose, and throat (1 device) and ophthalmic (1 device).
Radiology dominance: Radiology stands out not only for having the highest number of submissions but also for exhibiting a consistent and steady rise in AI/ML-enabled device submissions compared to other specialties.
Machine learning models: Machine learning models featured in these devices exhibit a wide spectrum of complexity, ranging from shallow models with fewer than two hidden layers to more intricate deep learning models.
Hybrid approaches: A notable trend is the increasing adoption of hybrid approaches in model design. This approach combines various algorithmic methods to achieve the desired outcome of a safe and effective medical device. For example, some devices utilize one model for feature generation and another for classification.
A few that I find of interest in radiology are:
- iCAC is a software device intended for use in estimating presence and quantity of coronary artery calcium for patients aged 30 years and above during routine care. The device automatically analyzes non-gated, non-contrast chest computed tomography (CT) images collected during routine care and outputs a visual representation of estimated coronary artery calcium segmentation (intended for informational purposes only) and both exact and four-category quantitative estimates of the patient’s coronary artery calcium burden in Agatston units. The output of the subject device is made available to the physician on-demand as part of his or her standard workflow. The device-generated calcium score or score group can be viewed in the patient report at the discretion of the physician, and the physician also has the option of viewing the device-generated calcium segmentation in a diagnostic image viewer. The subject device output in no way replaces the original patient report or the original chest CT scan; both are still available to be viewed and used at the discretion of the physician. The device is intended to provide information to the physician to provide assistance during review of the patient’s case. Results of the subject device are not intended to be used on a stand-alone basis and are solely intended to aid and provide information to the physician. In all cases, further action taken on a patient should only come at the recommendation of the physician after further reviewing the patient’s results.
- The Radiation Planning Assistant (RPA) is used to plan radiotherapy treatments for patients with cancers of the head and neck, cervix, breast, and metastases to the brain. The RPA is used to plan external beam irradiation with photon beams using CT images. The RPA is used to create contours and treatment plans that the user imports into their own Treatment Planning System (TPS) for review, editing and re-calculation of the dose. Some functions of the RPA use Eclipse 15.6. The RPA is not intended to be used as a primary treatment planning system. All automatically generated contours and plans must be imported into the user’s own treatment planning system for review, edit and final dose calculation.
- Irregular Rhythm Notification Feature (IRNF) is a software-only mobile medical application that is intended to be used with the Apple Watch. The feature analyzes pulse rate data to identify episodes of irregular heart rhythms suggestive of atrial fibrillation (AFib) and provides a notification to the user. The feature is intended for over-the-counter (OTC) use. It is not intended to provide a notification on every episode of irregular rhythm suggestive of AFib and the absence of a notification is not intended to indicate no disease process is present; rather the feature is intended to opportunistically surface a notification of possible AFib when sufficient data are available for analysis. These data are only captured when the user is still. Along with the user’s risk factors the feature can be used to supplement the decision for AFib screening. The feature is not intended to replace traditional methods of diagnosis or treatment. The feature has not been tested for and is not intended for use in people under 22 years of age. It is also not intended for use in individuals previously diagnosed with AFib.
- Gleamer device is intended to aid in the detection, localization and characterization of fractures on acquired medical images (per 21 CFR 892.2090 Radiological Computer Assisted Detection and Diagnosis Software For Fracture).
- The intended users of BoneView are clinicians with the authority to diagnose fractures in various settings including primary care (e. g., family practice, internal medicine), emergency medicine, urgent care, and specialty care (e. g. orthopedics), as well as radiologists who review radiographs across settings.
In summary, this compilation provides valuable insights into the landscape of medical devices incorporating AI/ML technology across diverse medical disciplines. While the growth rate of such devices has varied over the years, the field continues to evolve, and radiology remains a frontrunner in terms of submissions and growth. Additionally, the complexity of machine learning models and the use of hybrid approaches are notable aspects of this evolving field.
Mark Watts is an experienced imaging professional who founded an AI company called Zenlike.ai.
