NymTrace Local Medical De-identification

Local medical data de-identification for student research

NymTrace addresses one practical workflow: a medical research student needs to use Codex, Claude Code, or another authorized analysis tool, while the source files may contain names, medical-record numbers, phone numbers, identity documents, addresses, dates, institutions, or DICOM identity metadata. NymTrace first de-identifies or pseudonymizes those files on the local computer. The student submits only the de-identified analysis folder. NymTrace itself supplies no Agent and calls no cloud API.

The desktop application is designed for ordinary Windows PCs and Apple Silicon Macs. Its CPU-first pipeline prioritizes deterministic headers, code patterns, and contextual rules. A bundled quantized bilingual entity recognizer loads lazily only for natural-language passages that rules cannot settle. A simple medical-record-number/name table therefore does not invoke a large model for every cell, and no discrete GPU is required.

Categories come from the files you drop

NymTrace does not begin with a generic fixed checklist. After files or folders are dropped, it performs a read-only pre-analysis and lists the sensitive categories found in that batch, together with field, match, and file counts—without displaying patient values. All discovered categories are selected by default. A category should be cleared only when an approved protocol explicitly permits it and the researcher accepts the narrower review scope.

If a legacy workbook, embedded object, complex PDF, image, or medical-imaging surface cannot be enumerated fully, “could not analyze” is never treated as “contains nothing sensitive.” Surfaces that can be handled safely continue. Original content that cannot be safely transformed, independently re-scanned, or supported is withheld from the analysis folder and explained in NYMTRACE_EXCLUSIONS.csv using an anonymous SourceFileID and reason code. The run may therefore be PARTIAL, but it does not silently copy an untreated original.

Supported surfaces

• Text: TXT and Markdown.
• Tables: CSV, TSV, XLSX, and XLSM. Readable cells from legacy XLS/XLSB can be rebuilt into a safe derivative; macros, comments, external links, or hidden surfaces that cannot be guaranteed are omitted and reported.
• Documents: DOCX. Active and unsafe embedded objects are not copied.
• PDF: text layers, page imagery, and detectable regions. Embedded attachments are conservatively removed and reported.
• Images: JPG, JPEG, PNG, TIF, TIFF, and BMP, including detectable text and QR/barcode regions.
• Medical imaging: DICOM files and metadata. Head volumes require a validated defacing workflow, and pixels that cannot be decoded safely are not released.

“Supported” does not mean every file must pass. Password protection, corruption, active content, unknown encodings, complex pixel compression, and unenumerable hidden surfaces may produce PARTIAL or BLOCKED. The quality report states what was handled and what was excluded.

Two top-level deliverables reduce upload mistakes

A successful task produces only two top-level items:

1. 1_NymTrace_脱敏对应表_禁止上传.xlsx: the sensitive local correspondence table containing category, original, code, linked case identifier, and TraceID. Keep it only on an authorized device or institution-approved controlled storage.
2. 2_NymTrace_脱敏后完整数据_可分析/: de-identified derivatives, quality report, manifests, TraceIDs, and integrity checks. This is the only folder intended for a subsequent authorized analysis tool.

The split makes the safety boundary visible to beginners. Never upload the entire top-level task directory, and never send the correspondence table through an Agent, drive, email, or chat. A PASS does not grant permission for cloud use: ethics approval, data authorization, institutional policy, and cross-border rules still control that decision.

Bring analysis results back for controlled local restoration

If an Agent result preserves NymTrace_TraceID, [TraceID:...], or NT_* codes, a student can drop the analyzed Excel, CSV, TXT, DOCX, PDF, image, or DICOM result onto the Restore page and select the original local correspondence table. NymTrace restores uniquely matched codes into a new user-selected controlled directory and generates a sensitive field list.

Pixels permanently redacted in PDF, image, or medical-imaging output are not fabricated. Their associated originals may appear only in the controlled field list. When one visible code maps to several originals and no TraceID survives, NymTrace retains the code and reports ambiguity rather than guessing an identity. Every restored output is sensitive again and must never be returned to a cloud service.

How it behaves on low-spec computers

Files and worksheets are processed incrementally, and progress events from large workbooks are coalesced so that hundreds of thousands of rows do not saturate the UI queue. Model components load only when needed, are reused, then released; deterministic fields use batched rules and caches. Actual throughput still depends on format, OCR page count, DICOM pixel volume, storage speed, and available memory. Large image or imaging collections should be split into auditable batches, with the computer connected to power.

The release target is macOS 12+ on Apple Silicon and 64-bit Windows 10/11. At least 8 GB RAM is recommended; 16 GB is more suitable for large workbooks, PDF, image, and DICOM batches. No discrete GPU or separately installed Python environment is required.

Install and first run

macOS Apple Silicon

Fully extract the ZIP and move NymTrace.app to Applications. This research preview is ad-hoc signed and has not been notarized with Apple Developer ID. Keep Gatekeeper enabled; do not disable security controls or run untrusted bypass commands. If the system blocks the app, wait for a later notarized build or follow an institutionally reviewed IT process.

Windows x64

Fully extract the ZIP, keep NymTrace.exe and _internal together, and run the EXE. Do not run inside the archive or move the EXE by itself. The current package is not Authenticode-signed and has not completed comprehensive clean native Windows, Defender, and SmartScreen acceptance testing. Evaluate it first in a controlled non-production environment without patient originals.

Verify download integrity

The page displays each package SHA-256. On macOS, run shasum -a 256 filename; in Windows PowerShell, run Get-FileHash filename -Algorithm SHA256. The result must exactly match the page or SHA256SUMS.txt. A matching hash proves transfer consistency only; it is not platform signing, clinical efficacy, or compliance certification.

Current evidence and limits

The frozen 1.0.12 release record reports 96/96 cross-platform source tests passing, with engineering acceptance for packaged directories, ZIP integrity, bundled runtime, and principal delivery workflows. This website release re-verifies the byte length and SHA-256 of both ZIPs.

That is not a clinical gold-standard validation. There is no current institution-level precision/recall report on annotated sensitive entities, and no proof of complete recognition across all hospital templates, dialectal text, handwriting, complex imagery, or vendor-private DICOM fields. NymTrace is not a medical device, provides no diagnostic advice, and does not replace human sampling, ethics review, data authorization, or security approval.

Frequently asked questions

Does NymTrace need an LLM or a network connection?

No network or cloud LLM call is required. A lightweight bundled Chinese/English entity recognizer loads only when deterministic rules cannot resolve natural language. Names, medical-record numbers, phone numbers, identity documents, and structured headers are handled by local rules first.

Why does a file sometimes produce PARTIAL?

The application could not reach a sufficiently safe transform or residual-scan conclusion for one of its surfaces. It releases safe derivatives, excludes risky content, and records the reason in an anonymous manifest. PARTIAL does not mean everything passed.

May I send the correspondence table to an Agent?

No. It contains original sensitive values and exists only for controlled local restoration. Submit only the folder whose name contains 脱敏后完整数据_可分析.

Can Restore reconstruct text hidden in an image?

No, and it does not synthesize pixels. Restore replaces uniquely matched text, table fields, or DICOM metadata and places originals associated with redacted regions in a local sensitive field list for review.

Is this version ready for formal clinical-data operations?

It is a research preview. Start with synthetic data or an authorized non-production copy, then conduct institution-specific trials, manual sampling, and information-security review before deciding whether it belongs in a formal workflow.