AI Medical Devices

Cleared and approved AI software that reads images, flags conditions and alerts patients.

The FDA's list of AI-enabled devices reached 1,614 in September 2026, up from 950 two years earlier, and 76% are radiology tools. China's NMPA had approved 154 AI medical devices by mid-2025, and almost all Class III approvals went through clinical trials.

AI medical devices authorized, cumulative: FDA, Sep 2026 1,614; FDA, Jul 2025 1,247; FDA, Aug 2024 950; China NMPA, Jun 2025 154 AI MEDICAL DEVICES AUTHORIZED, CUMULATIVE FDA, Sep 2026 1,614 FDA, Jul 2025 1,247 FDA, Aug 2024 950 China NMPA, Jun 2025 154 DEVICES
FDA radiology share
76%
FDA AI devices with 2025 decisions
335
NMPA approvals, yearly growth
49.5%
Registry incidents
0

Deployments

Most AI medical devices support a clinician; a few decide alone

LumineticsCore, authorized in 2018, screens for diabetic retinopathy with no clinician reading the image. Yukun, United Imaging, Infervision and Deepwise hold 59 NMPA approvals between them, mostly for CT. Apple Watch hypertension alerts reach users in more than 150 countries and regions.

Flagship medical devices deployments
DeploymentDeveloper / operatorWhereScaleAutonomyStatusControls declared
LumineticsCore (formerly IDx-DR)Digital Diagnostics / US primary-care clinicsUSNot disclosed (2018)Autonomous diagnosisFDA De NovoNot disclosed
Chinese radiology AI (Yukun, United Imaging, Infervision, Deepwise)Several / Chinese hospitalsChina; United Imaging also in Europe59 NMPA approvals across the four (Jun 2025)Decision supportApprovedNot disclosed

Recent events

  • The FDA warned WHOOP in July 2025 over blood-pressure insights, then loosened wellness guidance in January 2026.
  • Radiology accounts for 106 of China's 154 NMPA AI devices.
  • Beijing, Shanghai, Shenzhen and Hangzhou hold about 78% of NMPA AI approvals.
  • The FDA list counts 335 AI devices with 2025 decisions, the most of any year.

Attack surface

Four of ten attack-surface dimensions rate very significant or critical

No dimension rates critical. Ratings use five levels, from limited to critical, and cover the paths an attacker or a failure can take into the agent.

Attack surface by dimension: Physical access significant; Identity and authentication significant; Command and control channels significant; Perception and sensors very significant; Connectivity surface significant; OTA and update pipeline very significant; Data capture and retention very significant; Integrations and permissions significant; Behavioral and policy boundary very significant; Multi-agent coordination moderate, growing Physical access SIGNIFICANT Identity SIGNIFICANT Command and control SIGNIFICANT Perception VERY SIGNIFICANT Connectivity SIGNIFICANT OTA updates VERY SIGNIFICANT Data capture VERY SIGNIFICANT Integrations SIGNIFICANT Behavior boundary VERY SIGNIFICANT Multi-agent MODERATE, GROWING
Attack surface rating by dimension. Longer, darker bars are rated higher.
Medical devices attack surface by dimension
DimensionRatingWhy
Physical accessSignificantDevice hardware in clinical and home environments
Identity and authenticationSignificantClinician credentials, patient identification, EHR integration accounts
Command and control channelsSignificantClinical workflow integration, vendor backends, remote configuration
Perception and sensorsVery significantMedical imaging, biosensors, ECG, glucose, vital signs
Connectivity surfaceSignificantHospital network, EHR integration, cloud services for AI processing, vendor backends
OTA and update pipelineVery significantFDA SaMD updates flow through manufacturer infrastructure
Data capture and retentionVery significantProtected health information under HIPAA
Integrations and permissionsSignificantEHR integration via HL7/FHIR, PACS integration, billing systems, clinical workflows
Behavioral and policy boundaryVery significantClinical decision support policy, intended use bounds, escalation thresholds
Multi-agent coordinationModerate, growingMultiple AI systems integrated into clinical workflows produce interaction surface

Incidents

No registry incident involves medical devices yet

The 137AI Incident Registry logs an incident when an agent's failure or misuse caused harm or a regulator acted, and public sources name the developer, the failed control and how it was found. No medical devices case meets that bar yet. Candidates go on the review list in the registry methodology.

Rules

The FDA lets AI devices update within pre-approved limits; the EU pushed its deadline to 2028

Rules that apply to medical devices
RuleJurisdictionWhat it requiresStatus
FDA PCCP final guidanceUSPre-authorized boundaries for AI model updatesFinal Dec 4, 2024
FDA AI-enabled device lifecycle guidanceUSLifecycle and submission recommendationsDraft Jan 6, 2025
FDA CDS and general wellness guidancesUSEnforcement discretion for some decision support; wider wellness scopeJan 2026
EU AI Act Digital OmnibusEUMedical-device AI high-risk duties moved to Aug 2, 2028Political deal May 2026
NMPA Class III registrationChinaClinical trials for high-risk AI devicesIn force

Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.

Record

Page Record

Agent class
Personal and ambient
Datasets
ai_deployments.csv (2 rows for this page); ai_attack_surface.csv (10 rows); ai_incidents.csv (0 linked)
Key sources
FDA AI device list tally (Complizen)
NMPA study (JMIR)
FDA PCCP (Ropes and Gray)
Last verified
October 2026