WASHINGTON, July 8 — The history of artificial intelligence in medicine is not a sudden breakthrough but a slow, cumulative arc, one that began with rule-based systems in the 1970s and has now reached a point where the U.S. Food and Drug Administration has approved over 500 AI-based medical devices as of 2023. The record shows that the earliest applications were narrow and logic-driven. Stanford University developed MYCIN in 1976, an expert system designed to diagnose bacterial infections.
It operated on a set of if-then rules, a method that worked in controlled settings but struggled to scale. For decades, the field remained largely academic, constrained by the limits of available computing power.
That began to change in the 1990s, when machine learning algorithms started to improve. But the real turning point came in 2012. That year, the ImageNet competition was won by a deep learning model developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton.
Their model demonstrated that deep learning could process visual data with an accuracy that far surpassed traditional computer vision methods. The victory signaled a shift in the entire field of artificial intelligence.
By 2017, researchers had applied deep learning to medical imaging. Studies published that year showed AI systems matching or exceeding human performance in detecting diabetic retinopathy and skin cancer. The results were striking enough to draw the attention of regulators.
In 2018, the FDA approved the first AI-based medical device for clinical use: IDx-DR, a system for screening diabetic retinopathy. It was a milestone, opening the door for a wave of subsequent approvals. As of 2023, the FDA has approved over 500 AI medical devices.
The pace has accelerated rapidly. The agency now reviews these products through a variety of pathways, including the traditional 510(k) clearance and the newer De Novo classification process.
The devices cover a broad range of specialties: radiology, cardiology, neurology, ophthalmology, and pathology. Many focus on image analysis — detecting tumors, measuring organ volumes, flagging abnormalities in X-rays and CT scans. The trajectory is clear.
What began with a single rule-based system at Stanford has grown into a regulatory and clinical ecosystem that now processes millions of patient scans each year. The 2012 ImageNet moment catalyzed a wave of investment and research that transformed the field.
The 2018 approval of IDx-DR established a regulatory precedent. And the number of approved devices has climbed steadily since.
The Shift from Rules to Data
The difference between MYCIN and today’s deep learning systems is fundamental. MYCIN encoded the knowledge of human experts into explicit rules. It could explain its reasoning, but it could not learn from new data on its own.
Modern deep learning models, by contrast, learn patterns directly from thousands or millions of labeled images. They do not require a human to define every rule.
But they also operate as black boxes: it can be difficult to determine exactly why a model reached a particular conclusion. That trade-off has been a central tension in the FDA’s approach to regulation. The agency requires evidence of safety and effectiveness, but it has adapted its review processes to accommodate the unique characteristics of AI-based devices.
Some devices are locked at the time of approval, meaning their algorithms do not change after deployment. Others are designed to learn and adapt over time, raising questions about how to ensure continued safety and performance.
On the timeline, the period from 2018 to 2023 saw the most rapid growth. The FDA has published guidance documents and held public workshops to address the challenges of adaptive AI. The agency has also established a framework for reviewing modifications to AI-based devices, acknowledging that the technology evolves after initial approval.
What the Numbers Show
The figure of over 500 FDA-approved AI medical devices as of 2023 represents a substantial ecosystem. It includes both large, well-funded companies and smaller startups. The majority of approved devices are in radiology, where AI can assist in reading mammograms, chest X-rays, and CT scans.
But approvals have also been granted for dermatology, cardiology, and pathology. The IDx-DR approval in 2018 was a landmark.
It was the first time the FDA had authorized a device that could make a screening decision without the need for a specialist to interpret the results. The system runs on a retinal camera and provides a yes-or-no result for diabetic retinopathy, a leading cause of blindness. The approval established a regulatory pathway for autonomous AI devices that could be used in primary care settings, expanding access to screening for patients who might not see an eye doctor regularly.
By 2023, the list of approved devices had grown to include tools for detecting strokes on CT scans, measuring cardiac function on echocardiograms, and identifying skin lesions suspicious for melanoma. The pace of approvals has drawn attention from lawmakers and patient advocacy groups, who see both promise and risk in the rapid deployment of AI in clinical settings.
Looking Forward
The record shows that the history of AI in medicine is still being written. The number of FDA-approved devices is likely to continue growing, as more companies seek clearance for their products and as the technology matures. But questions remain.
How will the FDA handle devices that learn and change after approval? How will clinicians integrate AI recommendations into their decision-making?
And how will the system ensure that AI does not amplify existing disparities in healthcare access and outcomes?
