Computer Vision in Healthcare

The tasks that work

Clinical computer vision performs best on narrow, well-defined tasks:

• Fall detection and bed-exit prediction.
• Presence detection — is anyone in the room, is the patient alone.
• Hand-hygiene compliance at door and sink.
• Skin-color-aware vital-signs estimation (remote photoplethysmography).
• Restraint monitoring and suicide watch in behavioral units.

General "understanding" of a patient room remains far outside the current state of the art; programs that ask for it end up disappointed.

Sensor choices

Privacy by design

The most-adopted systems process video locally, discard raw frames within seconds, and expose only event-level metadata to the network. Some use depth or radar sensors that never capture an identifiable image at all. Both approaches let the system function without a stored recording of the patient.

Design controls to look for:

1. Physical shutter or a hardware disable button.
2. Signage at the door and in the admission packet.
3. Documented data-retention policy — raw frames should be measured in seconds, not days.
4. Role-based access to any archived event clips, with audit logs.

Trust with patients

Signage, consent, and a clear off-switch matter more than model accuracy for adoption. Programs that treat the camera as clinical equipment — introduced by the nurse, explained during admission, easy to disable — tend to see minimal patient objection. Programs that install cameras silently do not.

Standards

• HIPAA Privacy Rule and NIST SP 800-66r2 — safeguards for identifiable video and audio.
• IEC 62304 — software lifecycle for the vision pipeline as a medical device.
• FDA 21 CFR Part 892 and Part 880 — imaging and general hospital device classifications.
• ISO/IEC 42001 — AI management system for governance.