On-device clinical de-identification

A browser-based demonstration that removes common identifiers from clinical text without sending the pasted text to a server.

Problem

Teams need safer ways to explore text workflows without unnecessarily moving identifiable clinical information into remote services.

Users

Healthtech and healthcare teams evaluating privacy-conscious clinical text workflows.

Context

A public technical demonstration on ayothedoc.com. It is not a certified de-identification product or a substitute for an organisation's privacy review.

Role

Designed, built and published the browser-based demonstration.

Requirements

• Process pasted text locally in the browser
• Detect common structured identifiers immediately
• Offer optional in-browser name detection
• Show the transformed text for user review

Approach

Combined local pattern matching with an optional browser AI model so the source text can remain on the user's device.

Workflow

The user pastes text, local detection identifies candidate information, the browser transforms it, and the user reviews the result.

Technology

TypeScript, Transformers.js, Browser-based inference, Pattern matching

Safety and risk considerations

The page states its demonstration boundaries and keeps review with the user. Production use would require broader identifier coverage, validation, governance and monitoring.

Outcome

Published a working, inspectable demonstration of an on-device approach to clinical text de-identification.

Metrics

Not measured

Lessons

Data minimisation can be an architectural choice. Some useful healthcare AI processing can happen locally before a remote service is considered.

Next steps

Validate against representative documents, expand identifier coverage and define confidence thresholds and review rules for a specific production context.