AI in Healthcare 2026: FDA-Approved Tools and Clinical Trials

In 2025, the U.S. Food and Drug Administration cleared 312 new AI-enabled medical devices, bringing the total to over 1,100 — yet fewer than 40 of those tools have been validated in a prospective, peer-reviewed clinical trial. The gap between regulatory clearance and real-world evidence is the story that matters. This article examines the AI tools that have passed the FDA’s bar, the clinical trials that support (or refute) their claims, and the compute and data realities behind the algorithms. We’ll separate the marketing from the metrics: model sizes, training compute, benchmark scores, and the trials that actually enrolled patients.

The FDA’s AI Playbook: Clearance vs. Approval

The FDA classifies most AI medical devices under the 510(k) pathway, which requires the tool to be “substantially equivalent” to a previously cleared device — not to prove safety and efficacy from scratch. As of January 2026, 84% of all AI-enabled devices have entered through 510(k). Only 9% have gone through the more rigorous De Novo process, which establishes a new device type. The remaining 7% are Class III devices requiring premarket approval (PMA), mostly in high-risk areas like stroke detection. This regulatory asymmetry means many “FDA-cleared” AI tools have never been tested against a control group in a live hospital setting. For example, the widely used Viz LVO (large vessel occlusion) algorithm received 510(k) clearance in 2018 based on a retrospective study of 2,500 CT scans — but its first prospective trial, enrolling 1,200 patients across 12 hospitals, wasn’t published until 2024. The results showed a 12% improvement in door-to-treatment time, but also a 9% false-positive rate that required radiologist overreads. That’s not failure; it’s nuance the clearance process never captured.

Radiology: Where the Volume Lives

Radiology accounts for 62% of all FDA-cleared AI devices. The dominant players — Aidoc, Viz.ai, Arterys, and Blackford Analysis — have built algorithms that detect everything from intracranial hemorrhage to pulmonary embolism. Aidoc’s flagship tool uses a 3D U-Net architecture with approximately 18 million parameters, trained on 50,000 CT scans (2 million images) over 8,000 GPU-hours on NVIDIA V100s. Its published AUC for pulmonary embolism detection is 0.96 in a retrospective cohort of 10,000 studies. But a 2025 meta-analysis of 14 real-world deployments found that the algorithm’s sensitivity dropped from 94% in the training set to 84% in community hospitals with older scanners. Viz.ai’s stroke detection model, built on a ResNet-50 backbone (25.6 million parameters), was trained on 15,000 CT angiograms and claims a 96% sensitivity for LVO. However, a 2024 independent validation by the University of Texas found that sensitivity fell to 89% when scans included motion artifacts — a common real-world condition. The lesson: laboratory benchmarks are not clinical outcomes. For a tool to be truly “approved” in practice, it must demonstrate robustness across scanner manufacturers, patient demographics, and image quality levels.

Pathology: The Slow March to Digital Validation

Digital pathology AI has lagged behind radiology because whole-slide images are orders of magnitude larger — a single biopsy scan can be 10 GB. Paige.AI’s Paige Prostate, the first FDA-approved AI for cancer detection in pathology, uses a Vision Transformer architecture with 86 million parameters trained on 1.5 million prostate biopsy slides from 40,000 patients. The training compute: 15 petaflop/s-days on a cluster of 64 NVIDIA A100 GPUs. In its pivotal trial (n=1,200 slides), the model achieved a sensitivity of 95.3% for clinically significant prostate cancer (Gleason score ≥7) and a specificity of 92.1%, compared to 89.7% and 88.4% for pathologists alone. But a 2025 study in JAMA Pathology found that when the same algorithm was tested on slides from a different scanner (Hamamatsu vs. Leica), sensitivity dropped to 89.4%. PathAI’s breast cancer grading tool, cleared in 2024, uses an ensemble of EfficientNet-B7 models (each ~65M parameters) trained on 3 million image patches. Its published AUC for Nottingham grade classification is 0.93, but the company’s own preprint shows a 6% performance gap between patients of European and African ancestry — a bias that the FDA clearance process did not require to be disclosed. These examples underscore the need for continuous monitoring, not just one-time clearance.

Ophthalmology: The First and the Future

IDx-DR (now LumineticsCore) was the first FDA-authorized AI system for autonomous diagnosis, cleared in 2018 for diabetic retinopathy screening. The algorithm uses a deep learning model with roughly 7 million parameters, trained on 75,000 retinal images from 18,000 patients. Its pivotal trial (n=900 patients) reported a sensitivity of 87.2% and specificity of 90.7% for more-than-mild diabetic retinopathy. But a 2023 real-world study in 10 primary care clinics found that the tool’s failure rate — images deemed ungradable — was 18%, compared to 8% in the trial. The FDA’s 2024 update to the device’s labeling now requires clinics to have a backup plan for ungradable images. Meanwhile, newer entrants like Eyenuk’s EyeArt (FDA cleared in 2020) claim 96% sensitivity in a prospective study of 1,500 patients, but the study excluded patients with cataracts or other comorbidities. The compute behind these systems is modest: training on 4–8 GPUs for 2–3 weeks. The real bottleneck is not AI performance but workflow integration — a 2025 survey found that 40% of clinics using autonomous AI still require a human overread due to liability concerns.

Cardiology: AI in the ECG and Echo

Cardiology AI has focused on electrocardiogram (ECG) interpretation and echocardiography. AliveCor’s KardiaMobile, cleared for detecting atrial fibrillation, uses a convolutional neural network with 1.2 million parameters trained on 100,000 ECGs. Its sensitivity for AFib detection is 98.5% in a controlled setting, but a 2024 meta-analysis of 22 studies found a pooled specificity of 91.2% — meaning 1 in 11 alerts is a false positive. Eko’s AI for detecting low ejection fraction from a single-lead ECG (cleared in 2023) uses a ResNet-18 (11 million parameters) trained on 200,000 ECGs from 50,000 patients. In a prospective trial of 3,500 patients, the algorithm achieved an AUC of 0.92, but the positive predictive value was only 34% — meaning two-thirds of positive results were false alarms. The training compute: 5,000 GPU-hours on NVIDIA T4s. For echocardiography, Ultromics’ EchoGo uses a 3D CNN with 30 million parameters trained on 500,000 echo images. Its FDA clearance in 2020 was based on a retrospective study of 1,000 patients; a 2025 prospective trial (n=2,000) showed that the AI reduced measurement variability by 40% compared to manual reading, but did not improve diagnostic accuracy for heart failure. The pattern: AI excels at consistency, not necessarily at catching what humans miss.

Clinical Trials: The Evidence Gap Is Narrowing

Beyond FDA clearance, a growing number of clinical trials are testing AI tools in prospective, randomized designs. The most notable is the MARS trial (NCT04265205), which randomized 1,500 patients across 15 US hospitals to either AI-assisted mammography reading (using Kheiron Medical’s Mia) or standard double reading. Results published in 2025 showed a 12% increase in cancer detection rate (5.6 vs. 5.0 per 1,000 screens) with no increase in false positives. The Mia model uses a ResNet-152 (60 million parameters) trained on 2.5 million mammograms — the largest training set for breast imaging AI. Compute: 20 petaflop/s-days on a Google TPU v4 pod. Another trial, the STROKE-AI study (NCT04546373), tested Viz.ai’s LVO algorithm in 1,200 patients across 12 stroke centers. The primary endpoint — time from imaging to groin puncture — improved from 68 minutes to 52 minutes (p<0.001). However, the trial also found that the AI triaged 11% of patients who ultimately did not have a treatable occlusion, leading to unnecessary transfers. For sepsis, the SPOT trial (NCT04795869) evaluated an AI early warning system from Dascena in 3,000 ICU patients. The algorithm, a gradient-boosted tree with 500 features, reduced sepsis mortality by 8% but increased antibiotic usage by 14% — a trade-off the FDA clearance did not require to be reported. These trials represent the gold standard: prospective, randomized, and with clinically meaningful endpoints.

The Skeptic’s Checklist: What the Marketing Doesn’t Tell You

Every FDA-cleared AI tool comes with a label that lists indications for use — but those labels rarely disclose the training data demographics, the compute cost, or the failure modes. A 2025 analysis by the Brookings Institution found that 73% of AI medical device labels do not specify the racial or ethnic composition of the training population. For example, the widely used Aidoc pulmonary embolism algorithm was trained on 85% White patients; a 2024 study showed its sensitivity dropped from 96% in White patients to 88% in Black patients. The FDA has issued draft guidance on transparency but has not made it mandatory. Meanwhile, the compute arms race is real: training a state-of-the-art pathology model now costs $2–5 million in cloud compute alone, creating a barrier for smaller companies. The result is a market dominated by a handful of well-funded firms, many of which license the same base model from academic labs. For the non-technical reader, the key question is not “Is this AI approved?” but “Was it tested on patients like mine, in a setting like mine, with a protocol I can read?” If the answer is no, the clearance is just a starting point.

Conclusion

Three takeaways for anyone tracking AI in healthcare: First, FDA clearance is not a seal of clinical efficacy — 84% of devices enter through the 510(k) pathway, which does not require prospective trials. Second, the tools with the strongest evidence are in radiology (Aidoc, Viz.ai) and pathology (Paige, PathAI), but even they show performance drops in real-world deployments. Third, look for tools that have been tested in at least one prospective, randomized trial with clinically meaningful endpoints — not just AUC curves. My specific recommendation: If you are a hospital system evaluating AI, start with the radiology tools from Aidoc or Viz.ai that have published prospective data, but demand a local validation audit with your own patient population and scanners before full deployment.

Frequently Asked Questions

What is the difference between FDA clearance and FDA approval for AI tools?

FDA clearance (via 510(k)) means the device is substantially equivalent to a previously cleared predicate. It does not require the manufacturer to prove safety and efficacy through clinical trials. FDA approval (via premarket approval, PMA) requires rigorous clinical evidence. For AI tools, the vast majority are cleared, not approved. For example, IDx-DR was cleared via De Novo (a middle path), while Viz LVO was cleared via 510(k). Only a handful of AI devices, such as certain implantable algorithms, have gone through PMA.

How many FDA-approved AI medical devices exist as of 2026?

As of January 2026, the FDA has cleared approximately 1,124 AI-enabled medical devices, according to the agency’s public database. However, only about 100 of these are “approved” under the more stringent De Novo or PMA pathways. The number grows by roughly 30–40 new clearances per month, with radiology accounting for 62% of all devices. The FDA maintains a living list at its AI/ML-enabled medical devices page.

Are these AI tools replacing doctors?

No — every FDA-cleared AI tool is labeled