FAQS
About RosettaMD
-
RosettaMD is a free AI-powered tool that instantly translates complex medical notes — including MyChart entries, lab results, and discharge summaries — into plain English you can understand.
-
After helping his brother with confusing hospital records, Dr. Chris Wixon saw that most people don’t understand medical language. RosettaMD was created to fix that and to standardize data for population health tools.
-
Dr Chris Wixon is a physician with deep NLP experience built RosettaMD. It’s part of Archimedes Medical, a health-tech company that puts people first with AI.
-
No. It’s a web-based translation tool — think of it as Google Translate for doctor notes. You upload your document, and RosettaMD returns a plain-language summary, instantly.
-
Yes. RosettaMD does not store or share your health information. Nothing is saved. Your data is processed securely and immediately discarded.
How It Works
-
No accounts. No Logins. No emails. Just clarity.
-
RosettaMD can translate any medical note — charts, blood tests, scans, discharge or visit summaries, and EHR records.
-
Yes — RosettaMD is fully mobile-optimized. You can use it from any phone, tablet, or desktop browser.
-
No. RosettaMD helps you understand your records, but it is not a substitute for medical advice. Always consult your doctor for decisions.
-
Can be added to any desktop Chrome browser.
Once installed, it can process medical text found on any website or any PDF that you open within Chrome.
You use it directly in the browser: open a webpage or PDF, then activate the extension to translate clinical language into clear, patient-friendly explanations.
It works only on desktop Chrome (not on mobile browsers or other browsers).
Installation and use happen within your Chrome browser; no separate app is required.
🟦 Still have questions? Try RosettaMD and see for yourself.
No account. No ads. Just clarity.
💡 Technical FAQs (for Developers)
This FAQ gives developers, architects, and data scientists detailed answers about our NLP engine—covering model performance, security and compliance, architectural best practices, and commercial terms—to support smooth, confident API integration for clinical narrative structuring.
Architecture and Performance
-
To ensure high reliability and predictable performance, our API imposes a maximum input size limit of 100,000 characters per single request.
For extremely long narratives that exceed this limit, we recommend a sharding or chunking architecture.
-
Latency scales linearly with input length: ~0.6–1.0 ms per input token. Processing may vary with entity count (NER resolution to codes). We optimize for low P90 latency to meet real‑time needs. Latency includes network RTT plus model inference.
Note Types and Targets
Short Note — 450 tokens / 1,800 chars / ~8 entities (e.g., Radiology Impression, Progress Note)
Target P90 latency: 0.2–0.4 sMedium Note — 1,000 tokens / 4,000 chars / ~30 entities (Operative note, H&P)
Target P90 latency: 0.7–0.9 sLong Note — 1,600 tokens / 7,500 chars / ~50 entities (Discharge Summary)
Target P90 latency: 1.6–2.0 sExtreme Note — 24,000 tokens / 80,000 chars / ~1,200 entities (Medical journal review)
Target P90 latency: 25–30 s
-
Our API is built for fast, modern web use and easy developer work.
Input:
Plain text: Send the clinical note as a text string in a JSON request — fastest and most efficient.
Document images: Upload image-based files (PDF, TIFF, JPEG). We run OCR to extract text before NLP.
Output:
JSON: Primary output is JSON with extracted entities, codes, negation status, and human-readable format.
FHIR (future) : Convert extracted data into FHIR R4 resources (e.g., Condition, Observation, Medication) for direct import into EHRs, HIEs, and population health systems.
-
For large-scale population health needs, we offer asynchronous batch processing.
Model &Extraction Capability
-
SNOMED
ICD-10
CPT (requires third-party license)
RxNorm
NPPES
Proprietary dictionary of lay descriptions
Customized terminology capable
-
Our mission: API payload includes Patient-Friendly Language (PFL) output for each clinical concept (e.g., "Heart attack: blockage of blood flow to the heart muscle.") This PFL is produced by our proprietary model, built for clear medical communication.
Developers can use PFL to make NLP-structured clinical data immediately actionable and easy to understand in patient portals, mobile apps, and other consumer tools.
-
"All" means the NER model extracts every supported clinical entity type and resolves them across all terminologies.
Entity scope — If no types are specified, the API labels each NER into comprehensive classes (e.g., Disorders, Medications, Procedures, Findings, Person, Occupation, Religion — 14 classes modeled).
Terminology resolution — Selecting "all" maps each entity to ICD-10-CM, SNOMED CT, CPT, RxNorm, UMLS, and NPPES.
Note: To reduce latency and compute, specify needed terminologies instead of using "all."
-
No — we don’t use general-purpose LLMs. We use specialized, encoder-based Small Language Models built for high‑precision, low‑latency clinical classification and entity extraction. That design gives production advantages for healthcare:
Faster, low-latency performance for real-time clinical workflows.
Deterministic, stable outputs for reliable, auditable data and consistent code selection.
Lower compute costs, enabling scalable, competitive pricing.
Improved PHI security: private models keep data in our HIPAA‑compliant environment, avoiding public or multi‑tenant LLM exposure.
For classification and structured extraction, these domain-specific encoders outperform general LLMs on accuracy, cost, latency, and reliability—aligning with our physician‑led focus on clarity, safety, and real-world usability.
-
Ensuring fairness is crucial for population-health solutions because algorithmic bias can worsen healthcare disparities. Our bias-mitigation spans the model lifecycle:
Data-focused: Train on large, diverse de-identified clinical notes. Use oversampling and augmentation for underrepresented cases and demographics to avoid spurious correlations.
Transparency & oversight: Development and audit teams review models to detect biases and reduce “black box” risks.
We commit to equitable outcomes and continually update mitigation strategies to follow best practices in clinical AI fairness.
-
Not yet, but we're working on it. After NER and entity resolution, the next step is Relationship Extraction (RE) to map connections—essential for knowledge graphs and clinical reasoning. RE is a high-priority roadmap item.
We aim for the API to return structured triples so developers can build knowledge graphs from clinical notes.
Developers interested in piloting can contact our Partnerships team.
Model Accuracy and Robustness
-
NER: 92% (F1)
Entity Resolver
SNOMED
ICD10:
CPT
RxNorm
-
Our engine, trained on large, diverse clinical texts, uses deep learning to capture nuanced context around extracted entities.
Negation (Assertion Status): Assigns Assertion Status to concepts so population-health models don’t mislabel diagnoses from phrases that rule them out. An Assertion Detection module finds negation cues (e.g., “denies,” “no evidence of,” “without”) and their scope.
Temporality (Time and Sequence): Recognizes and timestamps clinical events, distinguishing historical vs. current concepts (e.g., “past history of,” “status post appendectomy”).
-
The model leverages the surrounding context to identify ambiguous medical documentation with a high degree of precision.
The engine is trained on "noisy" real-world clinical data, making it inherently resilient to common typographical errors.
We employ semantic matching layer which intelligently maps misspelled or non-standard terms to the closest standardized clinical concept, preventing extraction failure due to minor errors. For example, 'diabites' will be correctly mapped to Diabetes.
We maintain and regularly update a proprietary lexicon of common clinical abbreviations and their possible long-form expansions.
-
The Entity Resolver confidence score measures semantic similarity between the extracted phrase and the final standardized code (e.g., SNOMED CT, ICD-10).
Meaning: It’s the probability the assigned code is the best contextual/semantic match. Scores near 1.0 indicate near-perfect matches.
Scale: Float from 0 to 1.0.
Pre-filtering (minimum): Our API suppresses any entity with confidence < 0.95 to reduce false positives and improve data reliability. We still return the exact score for transparency so developers can set audit thresholds and monitor score distributions in logs.
Deployment and Versioning (MLOps)
-
The stability and predictability of our engine are core requirements for production environments, and we manage this through strict versioning and a predictable update schedule.
Version Pinning: Developers are required to specify the API version in their request header (e.g., API-Version: v2.1). This pins their integration to a specific, immutable model and processing pipeline, guaranteeing stability even as newer versions are released.
The underlying model is not static, as it must evolve to keep pace with new medical knowledge, emerging terminology changes (e.g., new ICD codes), and shifts in clinical language.
Update Cadence: We anticipate model updates on an annual basis. These updates incorporate the latest terminology, fix edge cases, and enhance accuracy across specific domains.
Minimal Impact: Our rigorous testing process ensures that new versions are highly compatible. We project that annual model updates will impact the ranking or selection of codes for less than 1% of outputted data. This low delta allows for high confidence in migration.
Version Sunset: New major versions are announced with a minimum 90-day deprecation notice for the previous version. This provides ample time for developers to test and migrate their systems to the new, enhanced version.
-
For any given version of the API (e.g., v2.1), the model behavior is highly stable and fully deterministic.
Identical Input, Identical Output: Repeated calls with the same clinical text will yield identical entity extractions, entity rankings, and code selections. This ensures predictable downstream processing and accurate data comparison across runs.
Compliance and data security
-
NO
-
For any given version of the API (e.g., v2.1), the model behavior is highly stable and fully deterministic.
Identical Input, Identical Output: Repeated calls with the same clinical text will yield identical entity extractions, entity rankings, and code selections. This ensures predictable downstream processing and accurate data comparison across runs.
Commercial and Support
-
Licensing & pricing:
Usage-based (default)
Best for variable or spiky volume; pay for actual use.
API called per document; billing = total input token count.
Tiered discounts lower per-token price as monthly volume grows.
Subscription (monthly/annual)
Best for predictable, high-volume production.
Fixed fee includes a large pre-allocated token/document allowance.
Locked low rate for included volume; predictable overage pricing.
Often includes enterprise features: support SLAs, VPC deployment options.
Custom enterprise
For very high volume or on-prem/VPC needs; custom license with annual volume commitments, dedicated infra, and specialized support.
Key terms
Standard terminology: license covers royalty-free, commonly available sets (e.g., SNOMED CT for US, ICD-10-CM).
Third-party terminology: you must obtain any required external licenses (e.g., proprietary CPT sets).
Data rights: per our BAA, you retain ownership of your data and extracted outputs.
-
Yes — two ways to start:
Free Developer Key: permanent, limited access — up to 100 API calls/day for testing and proofs of concept.
14‑Day Full Trial: unlimited API access for 14 days to evaluate performance, latency, and full extraction across your clinical notes.
-
Item description
-
Item description
-
Item description