John Snow Labs Solutions for the Public Sector
Healthcare-Specific Large Language Models - Built for Clinical Accuracy, Not General Text
General-purpose LLMs were not designed for clinical language. John Snow Labs' Medical LLM is trained on clinical notes, biomedical literature, and EHR data, and benchmarked against the models your team is already considering.
Benchmark results across 13 clinical AI tasks:
- #1 on 12 of 13 benchmarks - average score 76.8 vs. 70.9 for GPT-5.4, 70.0 for Gemini-3.1-Pro, and 68.3 for Claude-Opus-4.6
- Preferred by medical practitioners in blind evaluation: 88% more often on factuality for clinical note summarization, 256% more often on conciseness for biomedical question answering
- Domain-specific 8B and 70B models outperform GPT-4o by 45% to 92% on factuality, clinical relevance, and conciseness (CLEVER, JMIR AI, 2025 - blind evaluation by practicing physicians)
- Deployed by the VA National AI Institute for clinical note summarization across 9 million patient records
Peer-reviewed safety evaluation: Evaluated against 690 clinically grounded red-team scenarios across 9 domains at ECIR 2026 - covering safety, robustness, and fairness. Publishes results against 150+ subcategories.
Clinical capabilities: note summarization, information extraction, biomedical question answering, EHR-to-SQL, risk scoring, clinical trial matching
Deployment: Runs on a single GPU inside your secure infrastructure. No external API dependency. No patient data leaves your network.
3,000+ Pretrained Clinical Models - More Accurate Than AWS, Azure, or GCP
Healthcare NLP is the most widely used NLP library in healthcare (source: gradientflow.com). It includes 3,000+ pretrained models for clinical and biomedical text, covering entity recognition, relation extraction, assertion status detection, entity resolution, de-identification, and summarization.
Peer-reviewed accuracy benchmarks:
- Makes 4-6x fewer errors than AWS Medical Comprehend, Azure Text Analytics for Health, and Google Cloud Healthcare API
- 96% F1-score on PHI detection - vs. 91% for Azure, 83% for AWS, 79% for GPT-4o (ECIR 2025, peer-reviewed)
- SOTA on 7 of 8 major NLP benchmarks, including the 2014 n2c2 de-identification challenge, 2010 i2b2 clinical relations, and 2018 n2c2 medication extraction challenge
- NER precision up to 0.989 for clinical procedures; assertion model accuracy 0.889 (JMIR AI, 2025 - 138,250 clinical notes, Mount Sinai Hospital)
- Outperforms AWS Comprehend Medical by 8.9% and Google Cloud Healthcare API by 6.7% on clinical NER (Software Impacts, Elsevier, 2022)
What's included:
- Clinical entity recognition (signs, symptoms, drugs, procedures, diagnoses, labs, vitals, social determinants, adverse effects, demographics)
- Biomedical NER (genes, variants, proteins, organisms, cell lines, chemicals, phenotypes)
- Assertion status detection, negation, relation extraction, entity linking
- Mapping to SNOMED-CT, ICD-10, RxNorm, LOINC, MedDRA, CPT, and custom terminologies
- Scales to a cluster; hardware-optimized for Intel and NVIDIA
Deployment: On-premises or private cloud. No data sharing.
Human-in-the-Loop AI Workflows for Regulated Environments - No Coding Required
Generative AI Lab enables clinical and compliance teams to annotate, validate, train, and deploy AI models on their own data without writing code. It supports regulatory-grade Human-in-the-Loop workflows with full audit trails, role-based access control, and air-gapped deployment.
Used by 500+ enterprise customers, with 7.5M+ expert annotation hours completed.
Key capabilities:
- High-throughput annotation - annotate clinical notes, PDFs, images, audio, and video with built-in QA, consensus analysis, and reviewer comments
- AI-assisted pre-labeling - pre-annotate with Healthcare NLP models, LLM prompts, or zero-shot classifiers to accelerate expert review
- Model training and tuning - fine-tune NER, classification, and relation extraction models on your data; no ML engineering required
- De-identification with human oversight - combine automated PHI detection with expert validation to meet regulatory standards
- Medical terminology coding - resolve ICD-10, LOINC, CPT, SNOMED CT during annotation, not as a post-processing step
- Responsible AI testing - automated bias, fairness, and robustness test suites; regression testing for Responsible AI compliance
- Active learning - models improve automatically as annotators correct errors
Deployment: Available on AWS Marketplace, Azure Marketplace, and on-premises. Air-gapped deployments supported for high-compliance environments.
Turn Clinical Text Into Standardized Codes - Instantly, Privately, at Scale
The Terminology Server converts free-text medical phrases to standardized codes across all major vocabularies, with sub-second latency and no external data sharing. It runs inside your environment - queries never leave your network.
Supported code systems: ICD-10-CM, ICD-10-PCS, ICD-9-CM, SNOMED-CT, LOINC, RxNorm, CPT-4, MedDRA, NDC, HCPCS, UMLS, HPO, MeSH, NCI Thesaurus, and more
Key capabilities:
- Free-text to code: Input any clinical phrase; receive ranked, contextually relevant codes with spelling correction and abbreviation handling
- Code-to-code mapping: Crosswalk between vocabularies - SNOMED-CT to ICD-10, LOINC to SNOMED, RxNorm to NDC
- Document coding: Extract and normalize clinical entities from full documents in a single pass; deterministic output with sub-second latency
- Custom value sets: Upload your own code lists, apply version control, filter searches by your defined subsets
- Prebuilt public value sets: CMS Chronic Conditions, Acute Diagnosis ICD-10-CM, Potentially Planned Procedures, and more
- MCP-native integration: Expose terminology tools as agent tools for real-time concept normalization in agentic AI pipelines
Deployment: Available on AWS Marketplace, Azure Marketplace, and on-premises. Zero PHI storage - only codes and terminology metadata are stored.
The Most Widely Deployed Clinical De-identification Solution in Large Academic Medical Centers
John Snow Labs de-identification is the only solution independently validated at 2 billion patient notes, with zero re-identifications in adversarial red team testing and published peer-reviewed methodology.
Validated performance:
- 99%+ PHI detection accuracy - surpasses triple manual review by domain experts
- 500,000 notes processed in 2.5 hours
- Deployed by Providence Health across 2 billion real-world patient notes - first system externally certified for regulatory-grade accuracy at this scale
- 96% F1-score - vs. 91% for Azure, 83% for AWS, 79% for GPT-4o (ECIR 2025, peer-reviewed)
- 50% fewer errors than AWS Medical Comprehend, 475% fewer than Azure, 575% fewer than Google Cloud Healthcare API (ML4H 2023)
- 80%+ cheaper than API-based solutions at scale
- 100% success rate for DICOM tag retention, date shifting, and UID consistency; 99.9%+ text accuracy on medical images (ICHI 2026)
Data formats supported: clinical text, PDFs (digital and scanned), DICOM images, structured tables, FHIR resources - one pipeline, zero inconsistencies
PHI entity types detected: 23+ including names, dates, locations, contact information, medical record numbers, provider information, biometric data
Compliance support: HIPAA Safe Harbor, HIPAA Expert Determination (with legal documentation), GDPR pseudonymization and anonymization, CCPA
Deployment: On-premises within Databricks, Snowflake, Azure, or AWS. Zero data movement. Fixed-cost deployment with no per-token fees. Supports air-gapped environments.
Automated Patient Registries, Cohorts, and Quality Measures From Clinical Documents
Clinical data curation automates the extraction of structured, research-ready data from unstructured clinical documents. It replaces manual chart review at scale, producing accurate patient registries, disease cohorts, quality measures, and real-world evidence datasets.
Built for the problems that matter in clinical research and quality improvement:
- 87% of clinical facts exist only in unstructured notes - structured EHR fields miss up to 40% of diagnoses, treatment responses, and functional status indicators
- Extracts 400+ medical entities - diagnoses, medications, procedures, lab results, clinical measurements, social determinants of health, severity indicators, and temporal relationships
- Maps to standard terminologies - SNOMED-CT, ICD-10, RxNorm, LOINC, and custom code systems; no post-processing required
- OMOP-ready output - extracted data is delivered in OMOP CDM format for analytics, AI training, and regulatory submission
- Full provenance - every extracted fact traces back to the source document, line number, and extraction timestamp
Peer-reviewed research: SDoH-NER outperforms GPT-4 in precision for homelessness/housing and marital status detection (AIH 2024). In a study of 17 million clinical notes across 109,076 patients, adding unstructured data doubled suicidality/self-harm event detection - events invisible in structured EHR fields alone (PHUSE/FDA CSS 2025).
Proven at scale: Extracted 60,717 SDOH factors from 13,258 oncology notes for MIBA; reduced cancer registry abstraction time from 120 minutes to under 2 minutes per case while maintaining regulatory-grade accuracy (Winner, RWE Catalyst Challenge, PHUSE US Connect 2026). Used by Roche, Optum, GE Healthcare, and Dana-Farber Cancer Institute.
Deployment: On-premises or private cloud. PHI never leaves your infrastructure.
From Raw Clinical Data to a Unified, Research-Ready OMOP Dataset
Patient Journey Intelligence ingests any multimodal patient data - clinical notes, FHIR resources, DICOM, EHR exports, claims data - and delivers a de-identified, conflict-resolved, confidence-scored OMOP dataset inside your environment. No data leaves your infrastructure.
The problem it solves: EHR structured fields miss critical clinical context. Disease severity, treatment rationale, functional status, and patient preferences exist only in unstructured notes. Without multimodal data integration, research and risk models operate on incomplete inputs.
What the platform delivers:
- Multimodal ingestion - clinical notes, lab reports, imaging studies, claims, HL7, FHIR, CSV, JSON, and structured EHR exports; one pipeline
- Clinical NLP extraction - 400+ medical entities with temporal relationships, severity indicators, negation, and patient context
- Automated conflict resolution - reconcile contradictory information across systems using clinical reasoning; supporting and contradictory evidence tracked
- 99%+ de-identification - validated by Providence Health against human expert reviewers; HIPAA-compliant
- OMOP standardization - automatic mapping to SNOMED-CT, ICD-10, RxNorm, LOINC
- Confidence scoring and provenance - every extracted fact includes confidence level and full audit trail to source document
Prebuilt AI agents: Care-Gap Co-Pilot (guideline adherence at population scale), Clinical Trial Matching (protocol-to-cohort in seconds), Intelligent Cohort Creation (natural language queries), Complete Patient Timelines, Clinical Measures (HEDIS, CMS, custom)
Trusted by: Mayo Clinic, UCLA Health, Cleveland Clinic, Optum, West Virginia University, Ohio State University, Mount Sinai, Cigna, and more.
Deployment: On-premises or private cloud. Air-gapped environments supported.
Certified AI Governance for Healthcare - 200+ Regulations, 9 Pillars, Zero Compromise
John Snow Labs provides an end-to-end AI governance framework that covers the full lifecycle of clinical AI - from development and deployment to monitoring and incident response. All products are certified through Pacific AI (pacific.ai), John Snow Labs' AI governance partner, and conform to 200+ laws, regulations, and industry standards worldwide.
The 9 governance pillars:
- Risk Management - AI risk identification, classification, and mitigation across the deployment lifecycle
- Safety - red-team evaluation, adversarial testing, and real-world safety validation at scale
- Copyright - training data provenance, IP controls, and responsible use of licensed clinical content
- System Lifecycle - version control, model deprecation, rollback procedures, and deployment documentation
- Privacy - HIPAA, GDPR, CCPA compliance; data minimization; patient data never leaves your environment
- Incident Reporting - structured response protocols, audit-ready incident logs, and regulatory notification support
- Transparency - explainable outputs, confidence scores, provenance trails, and published benchmarks
- Fairness - automated bias and equity testing across demographic subgroups using LangTest (100+ test types)
- Acceptable Use - enforceable use policies, role-based access control, and human oversight requirements
Published peer-reviewed evidence:
- De-identification: 96% F1-score vs. Azure 91%, AWS 83%, GPT-4o 79% (ECIR 2025)
- Assertion detection: 0.962 accuracy vs. GPT-4o 0.901 (ECIR 2025)
- Red-team safety evaluation: 690 scenarios across 9 clinical safety domains (ECIR 2026)
- FDA Sentinel MOSAIC-NLP: SDoH and pharmacovigilance signal extraction from 17M+ clinical notes
LangTest: Open-source library for automated bias, fairness, robustness, and toxicity testing of NLP models. 100+ test types. Compatible with Spark NLP, Hugging Face, spaCy, OpenAI, and custom models. Published in Software Impacts (Elsevier, 2024).
Certified by Pacific AI (pacific.ai) - AI governance certification covering regulatory mapping, risk documentation, audit trails, and third-party validation.
Deployment: Governance tooling runs inside your environment. Air-gapped deployments supported. No PHI transmitted externally.
Audit-Defensible HCC Risk Adjustment - Powered by Martlet AI
Accurate HCC coding requires more than extracting diagnosis mentions from clinical notes. A valid coding recommendation must be supported by current documentation, MEAT evidence (Monitored, Evaluated, Assessed, Treated), ICD/HCC mapping logic, provenance, and human review. John Snow Labs powers HCC risk adjustment through Martlet AI (martlet.ai), an AI-driven risk adjustment platform built on the same clinical NLP and medical LLM infrastructure used across the John Snow Labs product suite.
Three risk adjustment workflows:
- Prospective - point-of-care gap capture for clinicians; surfaces undocumented or undercoded chronic conditions before the encounter closes; clinician-friendly interface designed for adoption in high-volume settings
- Retrospective - high-volume chart review with MEAT-aware evidence extraction; surfaces undercoded HCCs with ranked supporting evidence from clinical notes; prioritizes gaps by revenue impact and documentation confidence
- RADV - mock audit preparation and defensible packet generation for Risk Adjustment Data Validation; both proactive (pre-submission) and reactive (post-audit) workflows; audit trail from clinical text to submitted HCC code
Clinical AI capabilities underlying Martlet:
- Medical LLM and Healthcare NLP extraction of diagnosis evidence, MEAT criteria, and supporting context from unstructured clinical notes
- ICD-10-CM to HCC mapping with regulatory-grade accuracy
- Peer-reviewed accuracy benchmark - validated methodology
- Human-in-the-loop review workflows with full audit trail
Compliance: HIPAA-compliant on-premises deployment. Clinical notes never leave your infrastructure.
Industry partners: WVU Medicine, Arkos Health, Pacific AI