10decodersFrom 15–20 minutes to 3–5 minutes per document
A leading U.S. supplier of home-delivered urologic catheters and ostomy products processes a high volume of physician orders, prescriptions, and chart notes that had to be read, categorized, and verified before equipment could ship. 10Decoders built a GPU-accelerated document intelligence platform that reads, classifies, extracts clinical detail from, and validates signatures on every incoming document in a single automated pass — with human reviewers engaged only where a second look is genuinely needed.
Volume versus accuracy, resolved in one pipeline
This home medical equipment provider processes a high volume of physician orders, prescriptions, and chart notes arriving as faxes and scans. Every document has to be read, categorized, checked for completeness, and — for signed work orders — verified for a valid signature before equipment can be dispensed or a claim submitted. Done by hand, that review took 15–20 minutes per document, with inconsistent extraction between reviewers and a queue that only grew as volume did.
10Decoders engineered a document intelligence platform that runs OCR, classification, clinical data extraction, and signature validation as one unified pass per document, rather than as separate point tools. Reviewers now see a completed, audit-ready extraction in 3–5 minutes, with human judgment reserved for the documents that genuinely need it.
The stack behind a 3–5 minute review
GPU-Accelerated OCR
A three-model hybrid (PP-DocLayoutV2, PP-OCRv5, PaddleOCR-VL 1.5) with Tesseract confidence scoring reads every page.
Agentic LLM Extraction
Template-based clinical extraction over a 120B-parameter model (gpt-oss-120b) served via vLLM.
Clinical NLP
Captures permanency, frequency, catheter type, and diagnosis as consistent, structured output.
Document Classification
Sorts every fax or scan into Chart Notes, Physician Orders, Admission Forms, or Others, flagging anything out of scope.
Signature Detection & Validation
A dedicated vision model checks for a valid signature in the designated field and flags irregular or missing ones.
Evidence & Audit Trail
Every extraction is tied back to its source with PDF highlighting and coordinates, and every stage is logged.
“Miss a signature, and an unsigned order may slip through.”
The Challenge
- High-volume faxed and scanned documents needed reading, categorizing, and verifying before equipment could ship or claims could be submitted.
- Manual review took 15–20 minutes per document, with inconsistent extraction between reviewers and fatigue-driven errors.
- A missed signature could let an unsigned order slip through; a misclassified document could route to the wrong workflow; a misread clinical detail could mean the wrong product dispensed.
- Growing document volumes kept expanding the review queue without a proportional path to more reviewer headcount.
Our Approach
- Built a unified pipeline that runs OCR once per document and reuses it across classification, extraction, and signature validation.
- Cut review time to 3–5 minutes with GPU-accelerated OCR and agentic LLM extraction, standardizing output across every reviewer.
- Added a dedicated vision model that validates signatures in the designated field and flags anything unsigned or irregular as Required Review.
- Let automation absorb the routine reading and categorization work, freeing reviewer capacity to grow with volume rather than headcount.
The Solution
Every incoming fax or scan moves through one automated pass: upload, OCR, classification, clinical extraction, signature validation, and highlighting. Document classification sorts each file into Chart Notes, Physician Orders, Admission Forms, or an Others bucket, routing it to the right downstream workflow and flagging anything outside those categories for manual review instead of letting it slip through unseen.
Clinical NLP extracts details like permanency, frequency, catheter type, and diagnosis into structured output, while a dedicated vision model verifies that orders and prescriptions carry a valid signature in the designated field — routing a correctly signed document to Completed, and flagging anything irregular or missing as Required Review. Every extraction is tied back to its source with PDF highlighting and coordinates, so each review stays documented and audit-ready end to end.
Document Classification
Every scan sorted automatically into the right workflow, with out-of-scope content flagged for review.
Clinical Data Extraction
Structured, audit-ready output on permanency, frequency, catheter type, and diagnosis.
Signature Detection & Validation
Unsigned or irregularly signed orders reliably surfaced before they reach fulfillment.
24/7 Conversational Assistant
Plain-language queries over documents for reviewers, administrators, and patients alike.
One pass, from scan to sign-off
Data Sources
Physician orders, prescriptions, chart notes, and admission forms arriving as faxes and scans.
Intelligence Layer
GPU-accelerated OCR, agentic LLM extraction, Clinical NLP, and vision-based signature validation run as one pass per document.
Decisions & Actions
Documents route to Completed or Required Review with a full, auditable, evidence-linked processing trail.
Every minute saved compounds at volume
Patient Experience
Faster order confirmation and shorter wait times for patients who depend on recurring supplies.
Compliance & Reimbursement
Signature validation catches unsigned or irregular orders before they enter the fulfillment pipeline.
Review Capacity
Automation absorbs routine reading and extraction so reviewers focus on documents that genuinely need judgment.
Operational Cost
Faster, more accurate processing lowers the cost of review as document volumes grow.
Audit Readiness
Standardized, evidence-linked outputs strengthen the provider’s position with auditors and payers.
Unified Pipeline
One pass per document across OCR, classification, extraction, and signature validation closes the gaps disconnected point tools leave.
What Changed
Review Time
15–20 minutes cut to 3–5 minutes per document — roughly 75% faster.
Reviewer Productivity
Up to 3x more documents handled per reviewer per day.
Categorization Accuracy
90% accuracy classifying patient data with Clinical NLP.
Signature Detection
90–96% consistency across 70+ validated documents, with no false positives reported.
Audit Trail
Every stage from scan to sign-off is now recorded and evidence-linked.
Reviewer Focus
Routine reading and extraction automated, human judgment reserved for genuine exceptions.
Delivery, Cost & IP
Delivery
- Rolled out across DEV, UAT, STG, and PROD environments without disrupting live order processing.
- 10Decoders’ first engagement of this kind in the healthcare domain.
- Delivered as a fixed-scope platform build, tracked against a defined document-intelligence roadmap.
Cost
- Automation absorbs routine review work without added reviewer headcount.
- Cost of review stays flat as document volume grows.
- Up to 3x reviewer productivity reduces per-document processing cost.
IP
- Not patentable, but the document intelligence platform is a reusable core capability.
- Document processing framework designed for reuse across future healthcare engagements.
- Built on 10Decoders’ core document intelligence stack, not a one-off build.
Impact across the business
From reading to reviewing
Routine extraction is automated, freeing time for judgment calls on genuine exceptions.
Audit-ready, every time
Standardized, evidence-linked output strengthens the provider’s position with auditors and payers.
Faster access to supplies
Faster order confirmation and shorter wait times for recurring urologic and ostomy care needs.
