Legal
How NeuraLog AI handles protected health information.
NeuraLog AI is built for de-identified research data. It is not designed or intended to receive protected health information (PHI).
Users are responsible for de-identifying clinical notes under their institution's IRB-approved process before uploading them to the platform. NeuraLog AI does not perform, certify, or substitute for that process.
As an additional safeguard, all note text passes through a server-side redaction function before it is sent to a language model. This runs inside the API route itself — the request is redacted before it leaves the server.
The function looks for identifiers that appear in a recognizable structured or labeled format and replaces each match with a fixed placeholder (e.g. [REDACTED_NAME]). It currently catches:
What this is — and isn't
This is a safeguard, not a substitute for institutional de-identification. It matches identifiers in specific structured or labeled formats. A name written in free-text narrative without a recognized label, a bare city or county name, an unformatted phone number, or an identifying label the function does not recognize will pass through unredacted. Notes must be de-identified under your institution's IRB-approved process, as stated in the Privacy Policy, before they reach this platform.
3.1 Server-side
The API routes that call the redaction function do not write note content — raw or redacted — to any database or file storage. The redacted text is sent to the model provider and the response is returned directly to the browser in the same request. When identifiers are detected, the server logs only the names of the categories found (e.g. "NAMES, DATES") for gate-monitoring purposes — it does not log the identifier values or the note text itself.
What reaches our AI provider is de-identified text only. NeuraLog AI's PHI scrubbing function runs server-side before every API call — what is sent for processing has already had all 18 HIPAA Safe Harbor identifier categories removed. The Anthropic API processes this de-identified text, not the original note. Under Anthropic's standard API terms, prompts and outputs are not retained by default.
3.2 Browser-side
De-identified extraction results — including scale scores, medication and diagnosis entries, and short quoted excerpts of redacted note text used for source verification — are saved to the browser's localStorage so that a research session can be resumed without re-uploading. This data stays on the device running the browser; it is not transmitted to or stored on NeuraLog AI servers. Clearing browser storage removes it.
NeuraLog AI uses the Anthropic Claude API (api.anthropic.com) as its AI processing layer. This is a deliberate architectural choice: Anthropic is a leading AI safety company whose models are purpose-built with safety, accuracy, and responsible deployment as core design principles. Every note processed by NeuraLog AI reaches Anthropic's infrastructure as de-identified text only — PHI scrubbing runs on NeuraLog AI's servers before any content is transmitted. Anthropic never receives identifiable patient information from NeuraLog AI workflows.
Anthropic offers a Business Associate Agreement (BAA) for its API under a HIPAA-ready configuration, covering the processing of de-identified research data under appropriate safeguards including encryption, access controls, and audit logging. NeuraLog AI's BAA with Anthropic is pending entity formation and is not yet in place. Until that point, the platform operates exclusively with de-identified research data under the researcher's own de-identification obligation per 45 CFR 164.514.
Researchers and institutional compliance teams reviewing this platform should note: NeuraLog AI's protection model is built on de-identification before transmission, not on preventing transmission to a third-party AI provider. This is the architecturally correct approach for research software of this type, and it is consistent with how clinical NLP research platforms operate across academic medical centers.