Expert Summary
- The FDA has cleared over 950 AI/ML-enabled medical devices as of June 2026 — the majority in radiology (image analysis) and cardiology, with rapid expansion into pathology and ophthalmology.
- AI-assisted diagnosis has demonstrated clinically meaningful improvements in specific domains — particularly dermatology (melanoma detection), radiology (chest X-ray reading), and retinal disease screening.
- The most transformative near-term impact of AI in healthcare is operational and administrative — AI scribes, prior authorization automation, and clinical documentation tools are reducing physician burnout and saving billions in administrative costs.
Artificial intelligence is no longer an experimental technology in healthcare — it is an operational reality. Over 950 FDA-cleared AI medical devices are in clinical use, AI scribes are installed in major health systems nationwide, and the first AI-assisted drug discovery programs have reached Phase II clinical trials. Here is what is actually happening and what it means for patients and providers.
FDA-Approved AI Medical Devices: The Scale
The FDA's Center for Devices and Radiological Health (CDRH) maintains a public database of AI/ML-enabled medical devices. As of June 2026, 952 devices are cleared or approved. The breakdown by specialty:
| Specialty | Number of Approved AI Devices |
|---|---|
| Radiology | 614 |
| Cardiology | 82 |
| Pathology | 56 |
| Ophthalmology | 44 |
| Neurology | 38 |
| Other specialties | 118 |
Radiology dominates because imaging interpretation is highly amenable to computer vision approaches — there is abundant labeled training data (X-rays, CTs, MRIs with radiologist annotations) and clear quantitative performance metrics.
Where AI Actually Works: Diagnostic Evidence
Radiology: Chest X-Ray Analysis
AI chest X-ray reading tools have reached commercial deployment at hundreds of hospitals. Studies on Qure.ai's qXR and similar tools show:
- Sensitivity for pneumonia: AI systems detect pneumonia at 87–92% sensitivity (comparable to general radiologists)
- Triaging priority cases: AI can flag time-sensitive findings (pneumothorax, aortic dissection) for immediate review, reducing triage time from hours to minutes in high-volume settings
- Low-resource settings: AI has shown particular value in settings without 24/7 radiologist coverage
Ophthalmology: Diabetic Retinopathy Screening
IDx-DR (now marketed as LumineticsCore) was the first autonomous AI diagnostic device FDA-cleared for point-of-care use without a specialist's involvement (2018). By 2026, AI retinal screening is deployed in primary care settings nationwide.
Performance: 87% sensitivity, 90% specificity for diabetic retinopathy — comparable to ophthalmologist screening, at a fraction of the cost and at primary care locations where patients already present.
Dermatology: Melanoma Detection
Multiple FDA-cleared AI dermoscopy analysis tools show performance comparable to dermatologists for melanoma vs. benign lesion classification:
- A 2020 Nature Medicine study (benchmark for the field): AI achieved 84% sensitivity at 89% specificity; dermatologists averaged 82% sensitivity at 82% specificity
Important caveat: All of these comparisons are in controlled study conditions on curated image datasets. Real-world performance with variable image quality and diverse populations sometimes diverges.
Cardiology: ECG Analysis
Apple Watch Series 4 and later received FDA clearance for afib detection — the first consumer wearable cardiovascular diagnostic. By 2026, Apple reports over 10,000 users have shared ECG data showing afib with their physicians.
Algorithm-based ECG interpretation (12-lead ECG interpretation AI) is now standard in hospital ECG systems — these tools perform first-pass reading on all ECGs, flagging critical findings for physician review.
Clinical Decision Support: Beyond Imaging
Sepsis Prediction
Sepsis claims ~270,000 US lives annually — and early detection dramatically improves outcomes. Epic Systems' Sepsis Prediction Model and similar tools analyze vital signs, lab values, and clinical notes in real-time to alert providers when a patient's probability of sepsis is rising.
A 2019 study in NEJM AI found Epic's sepsis predictor had 63% sensitivity at 67% specificity in a large hospital system — meaningful but not transformative; newer 2025 models show sensitivity of 75–80%.
Drug Interaction Checking
AI has largely replaced rule-based systems for drug interaction alerts. The improvement: AI-based systems reduce alert fatigue (too many irrelevant alerts causing physicians to ignore them) by contextualizing patient factors — a drug interaction that matters for a patient with renal failure is flagged; the same interaction for a healthy patient at normal dosage may be suppressed.
AI Medical Scribes: The Highest Near-Term ROI
The highest-ROI AI application in healthcare in 2026 is clinical documentation, not diagnostics.
The problem: physicians spend 50–70% of their time on documentation — typing notes, entering orders, completing prior authorizations. Burnout rates are at record levels.
AI scribe systems (Ambience Healthcare, Nuance DAX Copilot, Nabla, Suki) listen to the patient-physician conversation and generate structured clinical notes automatically. The physician reviews and signs.
Results from early 2026 deployments:
- Nuance DAX Copilot (deployed at 500+ health systems): Reduces documentation time by 50–70% per encounter; physicians report 70% reduction in after-hours documentation burden
- Physician satisfaction in trials: 90%+ of physicians using AI scribes report improvement in work quality and reduction in burnout-related feelings
Healthcare systems are investing heavily in AI administrative tools — not because they are the most visible application, but because they deliver measurable ROI in staff retention and productivity.
AI-enabled administrative automation (prior authorization, clinical documentation, billing) is projected to reduce US healthcare administrative costs by $250–350 billion annually when fully deployed — more than the projected savings from AI diagnostics by a factor of 3–4.
Source: McKinsey Health Institute, 2026
Challenges and Limitations
Algorithmic bias: AI trained on historical health data inherits historical disparities. Pulse oximeters are less accurate for darker-skinned patients — AI trained on their outputs perpetuates this inaccuracy. Dermatology AI trained primarily on light-skinned individuals performs worse on melanoma detection in darker skin tones.
Real-world generalization: AI performance in controlled trials often does not replicate in real-world deployment. Variable image quality, different patient demographics, and clinical workflow differences all affect performance.
Liability: Who is responsible when AI-assisted diagnosis is wrong? FDA cleared the device, the physician used it, the hospital deployed it. Legal frameworks for AI malpractice liability are still developing.
Regulatory overhead: The FDA's 510(k) pathway for AI medical devices has a 12–18 month typical clearance timeline — this slows commercial deployment of innovations.
Biotechnology trends 2026: AI drug discovery and multi-omics advances →
Is AI better than doctors at diagnosis?
In specific narrow tasks with well-defined imaging criteria, AI matches or exceeds average physician performance — particularly in melanoma detection, diabetic retinopathy screening, and chest pathology. However, AI performs these narrow tasks in isolation; physicians integrate clinical history, physical examination, and comorbidities that current AI cannot replicate. AI is most valuable as a second opinion and screening augmentation tool.
What are the most common uses of AI in healthcare today?
The largest current uses are clinical documentation (AI scribes), medical imaging analysis (radiology AI for chest X-rays, mammograms, CT scans), and clinical decision support (drug interaction checking, sepsis prediction). Operational AI (scheduling, billing, prior authorization) represents the largest share of healthcare AI investment by dollar value.
Are AI medical diagnoses covered by insurance?
AI diagnostic tools used as part of a covered clinical encounter are generally reimbursable through existing CPT codes. For AI-assisted imaging reads, CMS established coverage frameworks in 2024. Standalone AI consumer diagnostic apps are generally not covered and are not FDA-cleared medical devices.