FDA AI Medical Devices List: Approvals vs. Commercial Reality
The gap between FDA approval and commercial success in digital health has never been clearer. After analyzing FDA's public database of AI/ML-enabled medical devices—1,357 devices where FDA explicitly identified artificial intelligence or machine learning in the regulatory review process, spanning 1995 to September 2025—a striking pattern emerges about where AI/ML technology finds commercial traction in healthcare.
The Radiology Dominance in AI/ML Medical Devices
Within FDA's AI/ML-enabled medical devices inventory, radiology represents 77% of all approvals. The next closest specialty, cardiovascular, accounts for just 10%. Neurology represents 4.5%.
These specialty classifications reflect FDA's product code assignments and primary review panels, so some cross-cutting technologies (such as ICU decision support or sepsis prediction algorithms) are categorized where FDA placed them rather than necessarily where they're deployed clinically. Nonetheless, the concentration reveals where AI/ML technology has found both technical maturity and commercial viability. The dominance isn't arbitrary—it reflects the reality that diagnostic imaging provides the ideal substrate for machine learning algorithms: large datasets, objective endpoints, and clear clinical utility metrics.
For context, IQVIA's Digital Health Trends 2025 report notes that over 300 billing codes now exist to support digital health technologies, with 117 specifically for software-based solutions. The concentration of AI/ML device approvals in diagnostic imaging suggests that where billing infrastructure exists (radiology fee-for-service codes) and clinical workflows are established (PACS integration), AI/ML technology finds faster adoption pathways.
Growth Patterns Signal Market Maturation
Year-over-year growth in AI/ML device clearances (based on author analysis of the FDA list) reveals an important inflection:
2020 → 2021: 114 → 130 approvals (+14.0%)
2021 → 2022: 130 → 162 approvals (+24.6%)
2022 → 2023: 162 → 226 approvals (+39.5%)
2023 → 2024: 226 → 235 approvals (+4.0%)
2024 → 2025: 235 → 258 approvals (+9.8% through September)
The dramatic slowdown in 2024—from 39.5% growth to just 4.0%—likely reflects multiple factors: market saturation in the dominant radiology imaging category, maturing development pipelines, the reality that early applications (basic image classification, detection algorithms) have been largely addressed, and potentially the cadence of FDA review processes and public database updates. Companies now face higher bars for differentiation in an increasingly crowded field.
What's Actually Getting AI/ML Review
Examining recent 2025 approvals in the AI/ML-enabled devices list reveals the nature of these technologies:
Radiology (sample from September 2025):
Brain white matter hyperintensity detection (Quantib)
Aortic measurements for vascular imaging (iSchemaView)
Dental AI analysis (VideaHealth)
MRI system with AI reconstruction (Shanghai United Imaging)
Cardiovascular:
ECG-based ejection fraction analysis (Bunkerhill, Tempus AI, Anumana)
Atrial fibrillation detection (iRhythm, AliveCor, Apple)
Coronary CT analysis (HeartFlow)
Neurology:
Surgical navigation systems (Mazor Robotics, THINK Surgical)
EEG analysis platforms (Holberg, Beacon Biosignals)
The pattern is clear: these are augmentative diagnostic and decision-support tools—AI/ML that assists existing specialists in detection, quantification, or treatment planning, rather than standalone therapeutic interventions. The FDA's AI/ML device review process has primarily been applied to tools that enhance clinical decision-making within established specialties, not software that directly delivers treatment.
What's Notably Absent
Therapeutic digital health applications—digital therapeutics (DTx) —are largely absent from this AI/ML-enabled devices inventory. DTx exist via other FDA pathways, but most aren't positioned as AI/ML devices requiring algorithmic review—they're content delivery platforms (CBT modules, gamified interventions) focused on clinical outcomes, not diagnostic performance. This explains their commercial divergence: AI diagnostic tools solve objective technical problems; DTx must prove behavior change and symptom reduction.
Regulatory Pathways: Why Radiology AI Takes the "Evolution" Route
Understanding the commercial divergence requires understanding the regulatory machinery. Most radiology AI devices in the FDA AI/ML inventory reached market through the 510(k) clearance pathway, while many first-in-class digital therapeutics followed the De Novo classification route. This difference fundamentally shapes their commercialization trajectories.
510(k) Clearance: The Path of Substantial Equivalence
The 510(k) pathway requires demonstrating that a new device is "substantially equivalent" in safety and effectiveness to a device already legally marketed (the "predicate device"). For radiology AI, this means:
Intended Use Match: The AI tool must have the same intended use as an existing predicate (e.g., "aid in detection of lung nodules on CT scans")
Technology Similarity: If using different technology (deep learning vs. traditional CAD algorithms), those differences cannot raise new safety or effectiveness questions
Evidence Requirements: Primarily bench testing, software validation, and performance data comparing to the predicate
Clinical Data: Often limited or not required if performance can be demonstrated through retrospective, annotated imaging datasets
Radiology AI's advantage: diagnostic accuracy metrics (sensitivity, specificity, AUC) can typically be validated using existing annotated imaging databases, satisfying the 510(k) standard without lengthy prospective clinical trials.
De Novo Classification: The Path of Novel Evidence
The De Novo pathway is for novel devices with no legally marketed predicate but deemed low-to-moderate risk. Unlike 510(k), manufacturers cannot rely on substantial equivalence—they must establish safety and effectiveness from the ground up:
No Predicate Comparison: Must prove reasonable assurance of safety and effectiveness through valid scientific evidence
Clinical Data Requirements: Typically requires more extensive clinical studies than 510(k)
The RCT Connection: For therapeutic claims (treating depression, substance abuse, insomnia), demonstrating that software-delivered interventions change clinical outcomes often requires randomized controlled trials
Creates New Predicate: Successful De Novo grants establish new product codes that future similar devices can reference via 510(k)
Regulatory Pathway Comparison
510(k) Clearance (Radiology AI)
Risk Class: Typically Class II
Predicate Required: Yes - must prove equivalence
Key Burden: Demonstrate substantial equivalence
Typical Evidence: Bench/performance testing, retrospective data
Timeline: ~90 day FDA review, generally faster
De Novo Classification (DTx, Novel AI)
Predicate Required: No - first of its kind
Key Burden: Demonstrate safety & effectiveness
Typical Evidence: Prospective clinical trials, often RCTs
Timeline: ~150+ day review, generally slower
This regulatory architecture explains why radiology AI can follow an "evolution" strategy—building on established CAD predicates with retrospective performance data—while therapeutic digital health must pursue a "revolution" strategy, generating novel clinical evidence at pharmaceutical-grade rigor and cost.
Why Radiology AI Succeeds Commercially
The radiology dominance in FDA's AI/ML device approvals correlates directly with commercial success, but not because of regulatory ease. These solutions solve four critical commercialization problems that other AI/ML applications struggle with:
1. Integration: Radiology AI plugs into existing PACS (Picture Archiving and Communication Systems) workflows. Radiologists don't change their reading process—the AI highlights potential findings within their existing environment, often requiring only a software update to current systems.
2. Evidence Requirements: Diagnostic accuracy studies are faster and cheaper than randomized controlled trials for therapeutic interventions. A radiology AI can demonstrate clinical utility with hundreds of retrospective cases analyzed over months, not thousands of patients followed over years.
3. Existing Reimbursement Infrastructure: Radiology already operates on fee-for-service CPT codes for imaging studies. An AI tool that improves efficiency, diagnostic accuracy, or workflow throughput doesn't require creating new coverage policies—it enhances the value of existing billable services without changing the fundamental reimbursement model.
4. Professional Adoption Readiness: Radiologists have worked with computer-aided detection (CAD) systems since the 1990s, particularly in mammography screening. AI-based tools represent an evolution of familiar technology rather than a revolutionary change to practice patterns. The specialty's comfort with algorithmic assistance created receptive ground for ML/AI adoption.
Where the Market Is Moving: Cardiovascular AI
The cardiovascular category offers insight into emerging AI/ML opportunities. With 130 total approvals in the FDA inventory and accelerating momentum in ECG-based algorithms, this specialty combines several favorable factors:
Objective, quantifiable endpoints (ejection fraction, arrhythmia burden, ischemia assessment)
Clear clinical utility propositions (heart failure screening, stroke prevention through AF detection)
Consumer device integration possibilities (Apple Watch, AliveCor, other wearables)
Growing professional acceptance of algorithmic interpretation in cardiology
The addition of new CPT codes for AI-enabled cardiac diagnostics, including proposals for coronary plaque analysis and cardiac risk assessment tools, signals that cardiovascular AI is being normalized in coding and payment infrastructure—an early but important step toward established clinical practice. While implementation timelines vary, the trajectory mirrors radiology's earlier AI adoption curve.
Implications for Life Sciences Companies
For medtech and diagnostics companies evaluating AI/ML investments, the FDA device approval patterns provide clear guidance on what works commercially:
Higher probability of successful commercialization:
AI/ML integrated into existing medical devices (imaging systems, monitoring equipment, diagnostic platforms)
Augmentative tools for specialties with objective, measurable endpoints
Solutions that fit within established reimbursement categories and fee schedules
Technologies that demonstrably reduce provider workload or improve diagnostic accuracy without adding workflow steps
Lower probability despite FDA clearance:
Standalone AI/ML tools requiring creation of new clinical workflows
Solutions demanding new payer coverage policies or benefit categories
Technologies attempting to replace rather than assist clinical specialists
Applications in specialties without established comfort with algorithmic decision support
The fundamental lesson: achieving FDA clearance for an AI/ML-enabled medical device through the 510(k) pathway has become increasingly routine. Building a commercially viable business around that clearance requires solving the harder problems of clinical workflow integration, payer reimbursement alignment, and professional adoption—challenges that radiology AI has successfully navigated but that remain significant barriers in other specialties.
Looking Ahead
The 1,357 AI/ML-enabled device approvals in FDA's inventory represent more than regulatory milestones—they map the terrain where artificial intelligence and machine learning have found genuine clinical and commercial traction in healthcare.
As the market matures and growth moderates, success will increasingly depend on solving the complete commercialization equation: algorithms that satisfy FDA review, evidence that satisfies payers, integration that satisfies providers, and outcomes that satisfy patients. Radiology AI companies have demonstrated this is achievable. Cardiovascular applications are following a similar trajectory.
For companies entering this space, the lesson is unambiguous: study where AI/ML has already succeeded in gaining FDA review and market adoption. Build tools that make existing specialists measurably better at their current jobs, within their existing workflows, using their established reimbursement infrastructure.
The FDA's AI/ML device data rewards evolution, not revolution—at least for now.
Methodology Note: This analysis uses FDA's publicly available AI/ML-Enabled Medical Devices database, which represents a technology-based cross-section of the medical device universe rather than a comprehensive registry of all software as medical devices (SaMD). The dataset includes devices where FDA reviewers explicitly identified AI/ML algorithms during the clearance or approval process. Specialty classifications, year-over-year counts, and company rankings are author-derived tallies from this list, not official FDA statistics. Specialty assignments reflect FDA product code classifications and primary review panels; some cross-cutting technologies may be categorized based on regulatory assignment rather than clinical deployment. The analysis focuses on where AI/ML technology has been successfully deployed and commercialized within the subset of devices that underwent explicit AI/ML review, rather than attempting to characterize all digital health solutions.