Frequently Asked Questions
Find answers to common questions about using DelPHI for clinical document de-identification.
Getting Started3 questions
What is DelPHI and how does it work?
DelPHI (Deidentify and Label PHI) is a purpose-built de-identification platform for clinical documents. It runs a three-stage pipeline — a clinical NER transformer detects candidate PHI, a proprietary clinical-code shield removes false positives, and a redactor rewrites the surviving spans — so the residual text can feed downstream NLP, coding, and analytics without exposing patient identity.
What file formats are supported?
PDF (with OCR for scanned documents), DOCX, and TXT. DelPHI extracts the text from each file automatically, deciding page by page whether to use the digital text layer or OCR. If a scanned page comes back with low OCR confidence it is flagged for review, and the Results step opens on a Review tab showing that page beside its editable text — correct it and re-run detection. Layout and embedded images are dropped at extraction time (in-place redaction is on the roadmap).
How do I process my first document?
DelPHI is a three-step wizard: Upload & Mode → Process → Results. Drop a file (or click "Try sample note"), choose a redaction mode, optionally narrow the PHI types, then click "De-identify". When processing finishes you land on the Results step with a side-by-side comparison. See the How-To Guide for a step-by-step walkthrough.
The De-identification Engine3 questions
What engine does DelPHI use?
A single engine: the OpenMed-PII clinical NER transformer (a fine-tuned DeBERTa-v3-large, ~434M parameters) running on a hosted inference endpoint. There is no engine picker — DelPHI deliberately ships one well-characterised model wrapped in its own clinical safeguards.
What is the clinical-code shield?
It is DelPHI’s proprietary value-add: a post-filter that knows what a generic NER model does not. It detects ICD-10-CM, HCC, CPT, and NDC code ranges in the source text and drops any model-proposed PHI span that overlaps them, and it drops spans matching a curated 151-term clinical vocabulary — so an MRN-shaped diagnosis code or a medication name is never mistaken for PHI and scrubbed.
How accurate is the PHI detection?
The transformer benchmarks at 95.8% overall recall on a 17-category synthetic clinical-note set, with 14 of 17 categories at 100% and one known weak spot (device identifiers at ~32%). Every detected span carries a confidence score, and the Analysis view summarises how many cleared the high-confidence (≥90%) bar.
PHI Detection & Privacy4 questions
What types of PHI can DelPHI detect?
All 17 HIPAA Safe Harbor identifiers — names, dates, ages, addresses, ZIP, phone, fax, email, SSN, MRN, health-plan ID, account, license, URL, IP, device ID, organization — plus a catch-all "Other ID". You choose which types to enable per run from the PHI-types checklist.
What are the redaction modes?
Three: Anonymise replaces each span with a category token (John Doe → [PERSON]); Pseudonymise swaps in a realistic synthetic equivalent (John Doe → "Mark Reed"); Delete removes the span entirely. The same detected-span data is returned either way for audit and review.
Is DelPHI HIPAA-compliant, and is my data safe?
DelPHI is an assistive de-identification control, not a sole compliance guarantee — pair it with a human QA checkpoint or a second independent scan for Safe Harbor sign-off. For production live-PHI workloads the inference runs on a BAA-covered endpoint; the public demo path uses a HuggingFace endpoint that has no BAA and is for synthetic data only.
What happens to false positives?
The clinical-code shield removes the most common medical false positives before redaction. Beyond that, every span and its confidence are surfaced on the Results step, so you can review before exporting — and you can narrow the enabled PHI types or add custom always-redact terms to tune behaviour.
Reviewing Results3 questions
How do I see what was changed?
The Results step offers six views of the same document: Text (plain side-by-side), Highlighted (color-coded PHI spans), Diff (a unified diff of exactly what changed), Reflowed (readable prose), Document (your original file in its native format), and Analysis (a stats dashboard). Panes can be maximized, scrolled in sync, and copied.
What is in the Analysis view?
Detection-quality meters (average confidence and the share of spans ≥90% confidence), a row of metric tiles (PHI entities, entity types, medical terms preserved, processing time, throughput, text length), and a "PHI removed by type" breakdown you can read as a Chart, a flat List of every span, or raw JSON.
How do I interpret the confidence scores?
Confidence runs 0–100% per span. Above ~90% is reliable, 70–90% may warrant a glance, and below 70% is worth manual verification. The Analysis view’s high-confidence meter (e.g. 19 / 26) is a fast signal for how much manual review a document needs.
Batch & Export3 questions
How many files can I process at once?
Use the Batch or Folder upload tab to queue many files under one shared mode and PHI-type configuration; they process in parallel. On the Process step a progress bar and per-file status let you Pause, Resume, or Cancel, and a "Retry N failed" button re-runs only the files that failed.
How do I navigate batch results?
When more than one file is processed, the Results step shows a row of file tabs. Click any tab to switch documents; the stat tiles, the PHI inspector, and all six views follow your selection.
What export formats are available?
Four: .txt (plain de-identified text), .doc (Word), PDF (print-ready), and a JSON report (a full audit trail of every span with type, position, and confidence). Exports use opaque <hash>.<mode>.<ext> filenames so the same source in different modes shares a hash prefix; only the JSON report carries the original filename.
Troubleshooting3 questions
Why are my results showing no changes?
Either no PHI was detected in the text, or the backend could not reach the NER endpoint. The Process step translates known backend errors into a plain-English headline and fix (for example "Delphi-OM is not configured" → set DELPHI_NER_ENDPOINT and DELPHI_NER_TOKEN, or "Cannot reach the API" → confirm the backend URL).
The first request is slow — is that normal?
Yes. The inference endpoint scales to zero when idle, so the first request after a quiet period can take ~30–60s to warm up; subsequent requests are fast. Very large documents also take proportionally longer. If a file times out on a cold start, use "Retry failed".
I’m getting upload errors — how do I fix this?
Make sure the file is a supported format (PDF, DOCX, or TXT) and within the size limit. The wizard validates files as they are added; if a file is rejected, check its type and try re-adding it, or use the sample note to confirm the rest of the flow works.
Still have questions?
Can't find the answer you're looking for? Check our How-To Guide or contact our support team.