At ESC Congress 2026, researchers from the University of Tokyo presented a study demonstrating that an AI algorithm could detect undiagnosed hypertension and diabetes from a 30-second facial video with 95% accuracy. Using pulse-wave dynamics, skin blood-flow patterns, and spectral characteristics of skin color, the algorithm performed screening comparable to clinical methods — contactlessly, from a short video of someone’s face.
This is not speculative research. The technique works and has been independently validated. Similar research has demonstrated AI-based detection of anemia from eye photos, cardiovascular risk from retinal images, and insulin resistance from facial characteristics. What has typically required laboratory tests, specialized equipment, and clinical settings now requires only a camera and the right software.
The privacy implications are direct: if AI can extract health information from facial images and video, then every photo and video you upload to a cloud service — every selfie, every video call recording, every candid shot — potentially carries health data that you never intended to create or share.
The Science Behind It
The body expresses health information visibly, and AI systems trained on medical imaging datasets are becoming capable of reading signals that aren’t apparent to human observers.
Pulse wave analysis from facial video. Blood flow through facial tissue changes with each heartbeat in ways measurable through subtle color fluctuations in video footage. This remote photoplethysmography (rPPG) technique was originally developed for contactless heart rate monitoring and has been extended to analyze blood pressure characteristics.
Skin and subcutaneous tissue analysis. Facial fat distribution, skin tone patterns, and structural characteristics of facial features carry signals associated with metabolic conditions including diabetes and insulin resistance. AI models trained on datasets linking facial images to clinical diagnoses have demonstrated statistically significant predictive accuracy.
Eye and retinal analysis. The retina is part of the central nervous system and shows systemic health conditions in ways that trained analysis can identify. AI models have detected cardiovascular disease, anemia, age, biological sex, and neurological indicators from retinal photographs. Standard smartphone cameras can now capture eye images with sufficient resolution for these models.
Voice and audio analysis. Audio recordings have been shown to carry signals associated with Parkinson’s disease, respiratory conditions, cardiac abnormalities, and various neurological conditions. The same voice memo or video audio you store in the cloud can function as medical signal data.
The research directions are moving toward fusing multiple signal types — combining facial, retinal, voice, and behavioral data — for more accurate and comprehensive health inference from everyday media.
From Research to Product
The gap between published research and deployed consumer product has been closing.
Health screening apps. Consumer applications positioning themselves as “health monitoring” tools use smartphone cameras to estimate heart rate, blood oxygen levels, and stress indicators. These apps are early versions of a technology trajectory that points toward disease screening.
Insurance and financial applications. A small number of insurance companies have explored using AI analysis of facial photos — submitted by customers for identity verification — to infer health risk, raising questions about whether standard insurance application photos are being processed for health signal extraction.
Employer and HR applications. Remote work monitoring software has incorporated facial analysis for engagement and alertness assessment. The same analysis pipeline that detects fatigue or distraction may incidentally process health-relevant signals.
Social media and content platforms. Platforms running AI analysis on uploaded content for tagging, recommendation, and safety filtering process facial images through analysis pipelines that may extract health-adjacent signals as a side effect of other analytical goals, even when health inference is not the stated purpose.
What This Means for Your Cloud Photo Library
A photo library stored in a cloud service is, from a health data standpoint, a longitudinal health record — one you didn’t intend to create.
Consider what a collection of several years of photos contains:
- Selfies and portrait photos tracking facial characteristics over time
- Video recordings with audio
- Photos in medical contexts (at a hospital, wearing a bandage, visibly ill)
- Photos showing physical changes — weight, mobility, visible skin conditions — over time
An AI model applied to this dataset could extract health trajectory information, including changes suggesting onset of conditions, recovery patterns, or treatment responses. This is not hypothetical inference — it follows directly from the research capabilities demonstrated in published medical AI literature.
The practical question is whether this is currently happening. The answer is: partially, inconsistently, and mostly as a side effect rather than an explicit goal.
Photo services running facial recognition for people identification, scene analysis for searching, and content moderation for policy compliance all process facial images through AI systems. Whether those systems extract health-adjacent signals is a question most services have not publicly addressed. The research capabilities exist; the deployment in consumer platforms is less clear but not impossible.
The Regulatory Gap
Health data is among the most strictly protected categories of personal information in most jurisdictions. The challenge is that AI health inference from photos doesn’t fit neatly into existing health data frameworks.
HIPAA (US). The Health Insurance Portability and Accountability Act protects health information held by covered entities — hospitals, insurers, healthcare providers. A cloud storage service that incidentally extracts health signals from photos is not a covered entity, and HIPAA does not apply. HIPAA was written for a world of clinical records, not for AI systems that generate health inferences from personal photos.
GDPR (EU). The GDPR classifies health data as a special category requiring explicit consent for processing. However, whether an AI-extracted health inference from a photo constitutes “health data” under GDPR — as distinct from the photo itself, which may not be health data in isolation — is an unresolved question. Several DPAs have begun investigating AI health inference, but no comprehensive guidance specific to this technology has been issued.
Biometric privacy laws (US states). Illinois, Texas, and Washington state laws governing biometric data include provisions relevant to facial analysis. If an AI system analyzes a facial image to extract health-relevant biometric characteristics, these laws may apply — but the specific question of health inference has not been adjudicated.
The regulatory environment is playing catch-up to research that has moved quickly. For the near-term, the primary protection for most people is their own choices about where they store photos and what platforms have access to them.
Practical Considerations
Review the AI analysis terms for services hosting your photos. Services that run content analysis, facial recognition, or AI-powered features on uploaded images are processing your photos through systems that may extract health-adjacent information as a consequence of those features. Understanding what analysis is applied to your photos — and what data is retained from that analysis — is worth knowing.
Treat video files as higher-risk than still images. Facial video contains significantly more health-inferrable data than still photos due to the pulse-wave analysis dimension. Video stored in cloud services where AI analysis is applied to content carries a higher health data exposure than photos alone.
Consider the audience for your photos. Photos shared broadly — on social media, in public albums — are accessible to any system that can scrape or access them. Health inference from publicly accessible photos is technically feasible with currently available AI tools. Photos stored in private cloud storage under a service that does not run AI analysis on your content are less exposed.
Be cautious with health and wellness apps requesting photo access. Apps that request access to your camera or photo library and are positioned around health monitoring are the most direct path from your photos to health inference. Reviewing what these apps do with photos — specifically whether images are uploaded and what happens to analysis results — is worthwhile before granting access.
What to Look For in Storage
If health inference from your stored photos is a concern, the relevant characteristics of a storage service are:
- Does the service run AI analysis on uploaded file contents, or does it store files without examining them?
- What data is retained from any analysis that is run?
- Is the service using your files as training data for AI systems?
A service that stores your files without examining their contents — no face recognition pipelines, no scene analysis, no content modeling — does not create health inference data from your photos as a side effect of its operations.
daftei stores files using TLS 1.3 in transit and AES-256 at rest, does not run AI analysis on uploaded content, and does not use your files for AI training. Your files are stored and returned to you. No organization, tagging, or facial recognition is applied to your content by the service.
The Trajectory
The research on AI health inference from everyday media is moving quickly, toward higher accuracy, more conditions, and more modalities. Within a few years, the technical capability to derive a fairly detailed health picture from a photo library will be straightforward to deploy, not just to demonstrate in research settings.
The regulatory frameworks are moving more slowly. The window where thoughtful choices about where and how you store personal photos shape what can be inferred from them is now, while the technology is emerging rather than ubiquitous.
A photo taken for a memory is, from a sufficiently capable AI system’s perspective, a biometric scan. Understanding that doesn’t mean stop taking photos. It means being more deliberate about which systems those photos pass through, and what those systems are permitted to do with what they find.