AI-Assisted Diagnosis: What Evidence Exists

What this page covers

"AI-assisted diagnosis" means software that analyzes medical data, most often images such as mammograms, X-rays or scans, and flags findings or a risk score for a clinician. It is different from a chatbot that talks with the public, which is covered in Patient-Facing Health Chatbots. Evidence labels: approved or cleared device (we name the regulator, product and use), established, promising but limited, mixed, insufficient evidence, and marketing claim. Rules here are for the United States unless stated, and are not legal advice.

What the FDA list shows, and what it does not

The FDA keeps a public list of AI-enabled medical devices authorized for marketing in the United States. The FDA says the devices on it have met applicable premarket requirements, but also that the list is not comprehensive, because devices were identified mainly by AI-related terms in their authorization documents. [1] When we read the list on 2026-10-22 and counted its rows, it held 1,614 entries, and 1,230 of them had radiology as the lead review panel. [1] That tally is our own count of a page that is updated over time.

"Cleared" is not the same as "shown to improve patient outcomes." A 2026 analysis of 1,357 FDA-cleared AI devices through December 2025 found that 34 (2.5%) were linked to registered prospective trials and 3 (0.2%) had evaluated patient-centered outcomes such as mortality or readmissions. [2] That analysis used registries and literature searches, so it can miss unregistered work, but the direction is clear: authorization has moved faster than outcome testing. [2]

Accuracy studies versus clinicians

• Pooled comparison. A 2019 meta-analysis of deep learning for medical imaging found that in 14 studies that compared models with clinicians on the same cases, with outside validation, pooled sensitivity was 87.0% for the models and 86.4% for health professionals. [3] The authors said few studies had external validation or used the same samples, and reporting was often poor. [3] The journal lists a published erratum for this paper.
• Study quality. A 2020 BMJ review found only 10 randomized trial records for deep learning in imaging, of which two were published. Of 81 non-randomized studies, only nine were prospective and six were tested in a real-world clinical setting. The risk of bias was high in 58 of the 81 studies. [4]
• Trials of machine learning in care. A 2022 review found 41 randomized trials with a median of 294 participants, none adhering to all CONSORT-AI reporting standards, and 15 involved endoscopy. [5]

Label: promising but limited for imaging tools generally. A tool's accuracy on a test set does not show it helps people live longer or healthier lives.

The best-tested example: AI-supported mammography screening in Sweden

The MASAI trial randomly assigned women in Swedish mammography screening to AI-supported reading or standard double reading by two radiologists. The system used was Transpara, and it triaged exams to one or two readers and marked suspicious areas. [6]

• In the first safety analysis of 80,033 women, AI-supported screening found cancer at 6.1 per 1,000 screened versus 5.1 per 1,000 for standard reading, with similar false-positive rates (1.5% in both) and a 44.3% lower reading workload. [6]
• The 2026 report on 105,934 women found interval cancers (cancers appearing between screenings) at 1.55 versus 1.76 per 1,000, meeting the trial's non-inferiority goal. Sensitivity was 80.5% with AI versus 73.8%, and specificity was 98.5% in both groups. [7]
• The sensitivity gain was seen for invasive cancer but not for in-situ cancer. [7]

How far to trust this: it is a randomized trial funded by Swedish cancer charity and government research sources, which is a strong design. [7] It tested one AI system, in one country's screening program, with radiologists who knew the AI was involved (the radiologists were not masked). [6] Results for that system do not transfer to other tools. It is also not a clinic or product endorsement.

Questions to ask

If your scan or screening mention AI, you can ask your clinician what the tool is, what it was tested on, and who makes the final reading. AI-assisted reading does not mean you need any extra test on your own initiative; screening decisions belong with your clinician.