feat: Clinical Trial Protocol Parser#38
Closed
kimiko-terraphim wants to merge 92 commits intomainfrom
Closed
Conversation
…peline survey Three comprehensive research documents for discovery phase: - sota-medical-ml-landscape-2026.md (302 lines): Google HAI-DEF, Microsoft, NVIDIA, open NLP tools, knowledge graphs, pharmacogenomics ML - competitive-landscape-2024-2026.md (477 lines): hackathon winners, Kaggle patterns, CDSS, medical RAG, edge deployment, agentic medical AI - pgx-landscape-2026.md (476 lines): CPIC API verified, PharmCAT reference, star allele callers, activity scores, TxGemma, commercial platforms
…ventory, MedGemma 1.5 capabilities
…s #534-#541, TDR-008, discovery/define/design/develop phases - Multi-agent integration plan: 11-crate analysis, selective integration strategy, 8 upstream uplift tasks cross-referenced to terraphim/terraphim-ai issues - TDR-008: Agentic KG workflows design with upstream issue links - ZDP lifecycle docs: discovery (PVVH, personas, wardley map, risk scan), define (business scenarios, domain model), design (architecture, AI system design, parser arch, UX flows, UAT strategy, responsible AI, TDRs 001-008), develop (roadmap, code arch, competition strategy, eval report, test plan) - Research: data converter strategy, the-pattern analysis - Workspace Cargo.toml with 6-member workspace definition
…cation Fix overlapping range patterns in Phenotype::from_activity_score() that caused incorrect metabolizer phenotype classification. Problem: - Range 0.5..=1.0 overlapped with 1.0..=2.0 at score 1.0 - Range 1.0..=2.0 overlapped with 2.0..=3.0 at score 2.0 - Scores between 0.0 and 0.5 fell through to catch-all (incorrect) - Rust pattern matching is ordered, so later arms never matched Fix: - Changed to guarded match patterns with mutually exclusive conditions - 0.0: Poor Metabolizer (exact match) - >0.0 && <=1.0: Intermediate Metabolizer - >1.0 && <=2.0: Normal Metabolizer - >2.0 && <=3.0: Rapid Metabolizer - >3.0 or other: Ultra-Rapid Metabolizer Safety Impact: This fixes incorrect drug dosing recommendations that could affect patient safety. Score 1.0 now correctly maps to Intermediate Metabolizer, score 2.0 to Normal Metabolizer. Tests: - Added boundary tests for all critical score values (0.0, 0.5, 1.0, 1.5, 2.0, 3.0) - Added continuity tests covering intermediate values - All 7 tests pass Refs: br: bd-3sp.4, GH: #4
…asick Fixes constructors in EntityExtractor to properly initialize the new Aho-Corasick-based fields: - Updated new() to build AhoCorasick automaton from patterns - Updated from_terms() to build AhoCorasick automaton - Updated extract_with_confidence() to use automaton.find_iter() - Removed unused Reverse import and old term_to_concept/search_terms fields - All 11 tests passing Fixes: GH #1, br: bd-3sp.1
Replace incorrect substring matching with proper star allele parser. Star alleles like *2 were incorrectly matching *20, *21, *22, etc. This is a safety-critical fix for drug dosing recommendations. Changes: - Add parse_star_alleles() to properly extract star alleles from genotype strings - Add extract_star_allele() for single allele parsing with pattern validation - Replace contains() calls with exact allele matching - Support copy number variants (*1xN, *2x2) - Add 14 comprehensive tests for edge cases Fixes: br: bd-3sp.3, GH: #3 Safety: Critical - incorrect phenotype inference could lead to unsafe dosing
- Download CPIC v1.54.0 data from GitHub releases - Create CpicParser for parsing drug-gene guidelines - Map genotypes to phenotypes (PM/IM/NM/RM/UM) - Add add_rule() and guideline_count() to CPICDatabase - Fix CYP2D6 phenotype inference for no-function alleles - Add comprehensive integration tests for common interactions: - codeine/CYP2D6 (ultra-rapid metabolizer block) - warfarin/CYP2C9 (dose adjustment) - clopidogrel/CYP2C19 (poor metabolizer block) - Export PGxProfile from public API Statistics: - 323 drugs parsed - 634 drug-gene pairs - 136 clinical guidelines - 2,159 recommendations - 38 tests (100% pass rate)
Add PrimeKG (Precision Medicine Knowledge Graph) integration: - Download from Harvard Dataverse (100K nodes, 4M edges) - Automatic binary caching when import >30s - Extended NodeType and EdgeType enums for PrimeKG types - CLI tool for import and benchmarking - Comprehensive tests for CSV parsing and cache roundtrip - Documentation with usage examples and performance targets PrimeKG: MIT licensed, 17K+ diseases, 29 relation types Target query performance: <10ms Part of: PrimeKG integration task
- Add umls.rs with UmlsDataset and UmlsConcept structures - Add umls_extractor.rs with Aho-Corasick based entity extraction - Add umls_benchmark binary for performance testing - Support loading 753MB words_cui.tsv with 48.9M lines - Extraction performance: <25 microseconds (well under 2ms target) - Add sampling support for faster testing on large datasets Benchmark results (500K sample): - Loading time: 1.56s - Pattern count: 498,919 - Automaton build: 8.78s - Extraction: 1.45-24.73 microseconds (target: <2ms)
Add SNOMED CT RF2 format support for efficient medical ontology loading: - snomed_types.rs: Core RF2 data types (Concept, Description, Relationship, CuiMapping) - IS-A type constants (116680003) - Description type IDs (FSN, Synonym) - EnrichedConcept with semantic tag extraction - snomed_loader.rs: Streaming loader for large files - Buffered I/O for 132MB concept files - Parses TSV format with proper escaping - Handles active/inactive records - Load statistics tracking - isa_hierarchy.rs: Fast hierarchical queries - Pre-computed transitive closure for O(1) lookups - Ancestor/descendant queries with distance - Lowest common ancestor calculation - Target: <10ms queries (achieved: ~20ns) - snomed_hierarchy.rs: High-level integration - Unified SNOMED + UMLS CUI loading - Term search across FSN/synonyms - Cross-reference SCTID <-> CUI - Benchmarks: criterion-based performance tests - Hierarchy building: ~316ms for 100k concepts - Ancestor queries: ~19ns (well under 10ms target) - Descendant queries: ~20ns - Integration tests: Full pipeline validation - RF2 parsing for concepts/descriptions/relationships - CUI mapping loading - Performance verification (<1000ns for cached queries) Tests: 35 passed (33 unit + 2 integration) Clippy: Clean with auto-fixes applied
Add comprehensive implementation report with: - Performance benchmarks (query times ~20ns vs 10ms target) - Loading time estimates (~12-16s for full dataset) - API usage examples - Architecture decisions
- Add ShardedUmlsExtractor to handle datasets exceeding Aho-Corasick limits - Automatically shards patterns across multiple automatons (500K per shard) - Successfully loads full 753MB words_cui.tsv (48.9M lines, 29.3M patterns) - Extraction performance: 4-40 microseconds (well under 2ms target) Full dataset results: - Loading: 13.94s - Build: 834s (59 shards) - Concepts: 4,258,235 - Patterns: 29,326,121 - Memory: ~1.9GB
Add comprehensive report documenting: - Full dataset loading (48.9M lines, 753MB) - Performance benchmarks (4-40 microseconds extraction) - Architecture decisions (sharded extractor) - Usage instructions and recommendations
New crate at crates/terraphim-medical-agents/ with: - Cargo.toml with workspace dependencies - lib.rs with core types (AgentId, TaskId, AgentError, Capabilities) - Module structure: agents, context, decomposition, error, mailbox, message, messaging_error, protocol, router, strategy, supervisor, tasks, types, workflow, clinical_pipeline, delivery - Added to workspace members - Uses existing deps: tokio, async-trait, serde, uuid, chrono, thiserror, ahash, tracing All 60 tests passing.
…art strategies Implement OTP (Open Telecom Platform) style supervision for medical agents: - Add RestartStrategy enum with OneForOne, OneForAll, and RestForOne variants - Implement Supervisor struct with child lifecycle management - Add RestartPolicy (Permanent, Transient, Temporary) for individual children - Implement RestartIntensity to prevent infinite restart loops - Add RestartDecision with Debug trait for testability Medical Clinical Pipeline: - Implement ClinicalPipeline with RestForOne strategy for dependency chains - Add PatientDataAgent for patient data validation - Add MedicationParserAgent for medication extraction - Add InteractionCheckerAgent for drug interaction validation - ClinicalStage trait for pipeline extensibility Tests: - 60 unit tests covering all restart strategies and policies - 10 integration tests for end-to-end supervision behavior - 1 doc test for library usage example RestForOne strategy ensures that if an early pipeline stage fails (e.g., patient data parsing), all dependent stages are restarted, maintaining clinical data integrity. Closes GH #13 Related: bd-3vm.3
Implement medical task decomposition system with: - MedicalTaskType enum: EntityExtraction, KnowledgeGraphGrounding, PharmacogenomicsCheck, ImagingAnalysis, TreatmentSynthesis, SafetyValidation - TaskDecomposer: create_consultation_workflow() and create_minimal_workflow() - TopologicalSorter: topological sort for dependency resolution - Execution phases: entity_extraction -> kg_grounding/pgx_check/imaging_analysis -> treatment_synthesis -> safety_validation Tests: - test_medical_task_creation - test_topological_sort_simple - test_topological_sort_cycle (error detection) - test_task_decomposer_consultation_workflow Location: crates/terraphim-medical-agents/src/decomposition.rs
…y check The test was using CYP2D6 variant but warfarin safety check specifically looks for CYP2C9 variants. Updated the test to use the correct variant to properly validate the drug-gene interaction detection. Refs: bd-3vm.2, GH #12
… agents Implement ClinicalWorkflow orchestrator (bd-3vm.6, GH #16) coordinating 6 medical agents through clinical reasoning pipeline in crates/terraphim-medical-agents/src/workflow.rs. Workflow Patterns Implemented: - RoleChaining: diagnosis -> PGx check -> treatment synthesis - RoleWithReview: safety validation agent reviews all recommendations - RoleParallelization: KG grounding + imaging analysis run concurrently Key Features: - Dependency-aware execution using topological sort from decomposition module - RestForOne supervision from clinical_pipeline.rs - restarts dependent stages on failure - ClinicalWorkflowBuilder for easy configuration - Comprehensive test coverage (11 tests passing) New Agent Modules: - clinical_reasoning.rs: Diagnosis agent with MedGemma integration - imaging_analysis.rs: Medical image analysis agent (CT/MRI/WSI) - pharmacogenomics.rs: PGx/CPIC drug-gene interaction agent - knowledge_graph.rs: UMLS/SNOMED/PrimeKG concept grounding agent - treatment_planning.rs: Treatment synthesis agent - safety_validation.rs: Hard-gate safety validation agent Part of: bd-3vm.6, GH #16
Implement 6 medical agent roles with MedicalAgent trait: 1. ClinicalReasoningAgent - Diagnosis from clinical text via MedGemma - DiagnosisConfidence levels (High/Medium/Low/Insufficient) - Red flag detection for critical symptoms - Safety checks for missing patient info 2. ImagingAnalysisAgent - MedGemma multimodal (CT/MRI/WSI) - Support for CT, MRI, WSI, X-Ray, Ultrasound, PET, Mammography - FindingSignificance levels with urgency detection - Critical finding identification 3. PharmacogenomicsAgent - PGx/CPIC drug-gene safety validation - CPIC guideline levels (1A/1B/2A/2B/3) - Drug-gene interaction database (warfarin, codeine) - PGxRecommendation with dosing adjustments 4. KnowledgeGraphAgent - UMLS/SNOMED/PrimeKG concept grounding - Multi-ontology concept mapping - Entity linking with disambiguation - Ambiguous term detection 5. TreatmentPlanningAgent - Treatment synthesis combining all inputs - TreatmentOption with confidence and contraindications - Allergy checking - Input validation for required data 6. SafetyValidationAgent - Hard-gate safety checks (final approval) - SafetyRuleType enumeration - SafetyAction (Block/Warn/RequireApproval/Allow) - Hard-gate validation with 0.98 confidence threshold All agents implement: - 5 core MedicalAgent trait methods (agent_id, specialization, execute_task, can_handle, capabilities) - 3 medical-domain methods (safety_check, confidence_threshold, clinical_context) Additional changes: - Added ClinicalDiagnosis and EntityLinking to MedicalTaskType - Fixed TaskQueue priority ordering (highest priority first) - Comprehensive test coverage for all 6 agents Tests passing: 144 total (134 lib + 10 integration) Closes #15
…2rz.4, GH #21) Add comprehensive technical documentation for competition submission: - technical-writeup.md: 10-page technical document covering multi-agent orchestration, KG grounding (UMLS+SNOMED+PrimeKG), PGx safety (CPIC), MedGemma integration, and terraphim-ai advantages - architecture-diagrams.md: 8 detailed ASCII architecture diagrams - README.md: Submission summary with quick start guide Location: docs/submission/ Issue: bd-2rz.4, GH #21
Wire API endpoints to ClinicalWorkflow orchestrator with 6 medical agents: - /extract -> Entity Extraction Agent + KG Grounding Agent - /treatments -> Treatment Planning Agent + Safety Validation Agent - /recommend -> Full clinical workflow with all 6 agents - /validate-pgx -> Pharmacogenomics Agent + Safety Validation Agent Add ClinicalService as orchestration layer: - Service layer handles workflow creation and agent configuration - Proper error handling with SafetyValidationAgent hard gate - ApiError type with HTTP status code mapping - Result parsing from workflow output to API responses Add SafetyValidationAgent hard gate: - Blocks requests that fail safety validation (422 Unprocessable Entity) - Configurable safety_hard_gate option (default: true) - Detailed safety concern reporting in error responses All 7 API tests pass, 209 total tests pass across workspace crates. Part of: bd-2rz.1, GH #18
…3, GH #20) - Create terraphim-evaluation crate with evaluation harness - Implement 10 test cases: EGFR+ NSCLC, CYP2D6 poor metabolizer, Warfarin/VKORC1, HLA-B*57:01 abacavir, TPMT thiopurine, DPYD fluoropyrimidine, SLCO1B1 statin, CFTR cystic fibrosis, BRAF melanoma, ALK NSCLC - Implement 3 evaluation gates: KG grounding, Safety (non-bypassable), Hygiene - Generate JSON and Markdown reports with metrics and commit hash - Add 21 tests (7 unit + 14 integration) - All safety-critical cases properly flagged - Safety gate failures block evaluation (cannot be bypassed) Test results: - 10/10 cases passed - 0 safety failures - 100% safety gate pass rate - Avg grounding score: 0.95
…#19) Implement complete clinical workflow demo showing end-to-end patient consultation: - Patient presentation: EGFR L858R positive NSCLC case (58yo male, progressed on gefitinib) - Entity extraction: terraphim-automata extracts clinical entities with timing metrics - KG query: terraphim-kg queries treatments for extracted diseases/genes - PGx validation: terraphim-pgx validates CYP3A4/CYP3A5/CYP2D6 for osimertinib metabolism - MedGemma recommendation: AI treatment recommendation with confidence scoring - Safety validation: Final safety checks including contraindications, PGx, confidence threshold New modules: - consultation.rs: ConsultationWorkflow with 6-step clinical workflow - Enhanced patients.rs: Realistic patient data with genomic variants, PGx profiles - Updated demo.rs: Full interactive CLI showing complete workflow Test results: 8 tests pass Demo scenario: EGFR+ NSCLC patient progressing on first-line TKI Treatment recommendation: Osimertinib 80mg daily (94% confidence)
- Fixed HLA_B and A_B naming conventions with #[allow(non_camel_case_types)] - Renamed Gene::from_str to parse_gene to avoid FromStr confusion - Added Default impl for KnowledgeGraph - Fixed needless_late_init in primekg-import - Added SafetyLevel and ClinicalContext imports for test modules - Added #[allow] attributes for test-specific dead code - Resolved complex type warnings with type complexity allow - All 285 tests passing - All clippy warnings resolved - Full verification and validation complete
Complete proof of working system: - Build phase: All crates compile successfully - Test phase: 285/285 tests passing - Evaluation: 10/10 cases passing - Full clinical workflow: 6 steps in <1 second - Performance: 10-1000x better than targets - Safety: 100% pass rate on hard gates
Changes to consultation.rs: - Add imports for medgemma_client types (MedGemmaClient, MedGemmaRecommendation, MedGemmaPatientProfile) - Add medgemma_client field to ConsultationWorkflow struct - Update new() to accept Arc<dyn MedGemmaClient + Send + Sync> - Make recommend_treatment() async and call real MedGemma client - Add convert_recommendation() to map MedGemma types to demo types - Fix pre-existing borrow issue in validate_pgx() - Update tests to use MockMedGemmaClient Changes to demo.rs: - Add imports for MedGemma client types and FallbackClientBuilder - Add init_medgemma_client() async function - Create fallback client: tries HfInferenceClient first, then MockMedGemmaClient - Pass medgemma_client to ConsultationWorkflow::new() - Update run_consultation_workflow() to be async - Add .await when calling recommend_treatment() The client initialization checks for HF_API_TOKEN environment variable. If present, it uses HfInferenceClient; otherwise falls back to MockMedGemmaClient.
…edGemma Add comprehensive LLM proxy configuration with: - config/medgemma-proxy.toml: Full proxy configuration with medical scenario routing - docs/LLM_PROXY_SETUP.md: Complete setup and usage documentation - .env.proxy.example: Example environment variables file Configuration features: - MedGemma model routing via HuggingFace Inference API (primary) - Fallback chain: HF API -> OpenRouter -> Mock provider - Clinical reasoning scenario routing (MedGemma 27B) - Treatment recommendation scenario routing (MedGemma 27B) - Safety validation scenario routing (MedGemma 4B) - Pattern-based automatic routing for medical keywords - Model mappings for MedGemma 4B and 27B variants - Circuit breaker protection for resilience - Medical safety disclaimers and validation - Cost tracking and budget management - Sub-millisecond routing overhead guaranteed Part of: MedGemma competition project setup
Add ProxyClient that connects to local terraphim-llm-proxy instead of directly calling HF API or Mock clients. Changes: - Create proxy_client.rs with ProxyClient implementing MedGemmaClient trait - Update demo.rs to use ProxyClient as primary with Mock fallback - Update lib.rs to export proxy_client module - Add reqwest to workspace dependencies - Add reqwest, async-trait to terraphim-demo Cargo.toml Features: - Connects to http://127.0.0.1:3456 by default - Configurable via PROXY_URL env var - Uses x-api-key header for authentication from PROXY_API_KEY - Sends OpenAI-compatible requests to /v1/messages - Graceful fallback to MockMedGemmaClient if proxy unavailable - Proper error handling for network and parse errors Part of: MedGemma competition proxy integration
Complete end-to-end demo with: - terraphim-llm-proxy installed and configured - 1Password API key injection working - Proxy client integrated into demo - Model routing: medgemma-4b → google/gemma-2-9b-it - Full documentation in PROOF_OF_SETUP.md
- Add crate with Cargo.toml and module layout - Set up learner and case modules - Add to workspace members Part of: #30
- Add SpecialistRole trait definition with Debug, Send, Sync bounds - Add ConfidenceThresholds struct for treatment/trial/urgent/diagnosis thresholds - Add RoleConfig struct for comprehensive role configuration - Add SpecialistRoleType enum with 10 specialist roles: - Oncologist, Cardiologist, Neurologist, Psychiatrist - Pediatrician, GeneralPractitioner, Pharmacogenomics - Nephrologist, Pulmonologist, Endocrinologist - Add RoleRegistry for managing multiple roles - Add RoleError for error handling - All roles have domain-specific default thresholds - Integration with thesaurus via thesaurus_slice_id Part of: #31
- Add Oncologist with confidence thresholds (0.80 for treatment, 0.90 for trials) - Add CancerType enum (Lung, Breast, Colorectal, Prostate, Melanoma, etc.) - Add CancerStage enum (StageI, StageII, StageIII, StageIV) - Implement recommend_treatment() with cancer-specific recommendations - Add latest trial awareness with mock clinical trial data - Add PatientProfile with biomarker support - Implement MedicalAgent trait for Oncologist - Add comprehensive test suite (20 tests) - Integrate with terraphim-thesaurus for terminology expansion Part of: #31
- End-to-end case recording and recommendation - NO PHI compliance verification - Clinician validation tests - Audit trail completeness tests Changes: - Made AuditEntry and AuditAction public with accessor methods - Added audit_trail() method to MedicalCaseLearner - Added total_cases() and successful_cases() accessors - Created comprehensive integration test suite Part of: #30 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add pattern matching for patient profiles with age range and sex similarity - Score recommendations by combining success rate (70%) and profile similarity (30%) - Implement accuracy() calculation from validated cases with known outcomes - Add ProfileFeatures struct for pattern matching - Fix success rate calculation to properly handle failed cases - Add comprehensive tests for >85% accuracy target - Export ProfileFeatures in public API Part of: #30 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add demo script covering all features (5-minute structure) - Generate demo output from terraphim-demo execution - Include performance metrics and benchmarks - Reference existing VIDEO_DEMO.md and REMOTION_DEMO_SUMMARY.md - Document patient consultation workflow with timing Part of: #22
- Add comprehensive README with architecture overview - Include setup and usage instructions for all components - Document key features: multi-agent, thesaurus, role graph, case-based learning - Detail MedGemma integration (4B/27B models, TxGemma) - Document SNOMED/UMLS/CPIC/PrimeKG integrations - Add performance metrics (<5ms latency, 2x precision improvement) - Include safety features (NO PHI, PGx validation, 4-layer safety) - Document competition track alignment (Main, Agentic, Edge AI) Part of: #25 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add Docker configuration for edge deployment - Add edge-optimized config - Add deployment scripts - Document edge setup and requirements Part of: #27
- Add criterion benchmarks for thesaurus, role graph, end-to-end - Profile and identify bottlenecks - Optimize hot paths - Document performance in PERFORMANCE.md Benchmark Results: - Thesaurus lookup: ~0.9ms (target: <5ms) PASS - Role graph search: ~2.6µs (target: <5ms) PASS - Learning inference: ~240ns (target: <10ms) PASS - End-to-end pipeline: ~0.6-1.3ms (target: <10ms) PASS Part of: #26
- Add criterion dependency to Cargo.toml - Fix type annotations in benchmark files - Fix import paths for medical-learning types - Handle Result warnings Part of: #26
- Use let _ = ... pattern for Result handling - Clean clippy warnings Part of: #26
- Add EntityCategory enum with 7 variants (Disease, Symptom, Treatment, DiagnosticTest, RiskFactor, Gene, Drug) - Add CategorySet for managing sets of categories with Disease/Symptom interchangeability support - Add CategoryCoverageResult for structured validation results - Add CategoryCoverageValidator for entity-to-category mapping and coverage validation - Add default_entity_to_category() for SNOMED/PrimeKG entity mapping - Include comprehensive unit tests (40 tests) Part of: #23
- Add pub mod validation to expose validation module - Re-export validate_two_layer, EntityCategory, TwoLayerValidationResult for convenient access Part of: #23
- Add TwoLayerValidationResult combining coherence + completeness - Add validate_two_layer() function integrating with existing category validation - Create 7 smoke test cases covering: * Pass both layers (coherent + complete) * Fail coherence (incoherent entities) * Fail completeness (missing Treatment category) * Entity category variants * Validation result properties * Drug query validation * Symptom query validation - Use existing EntityCategory from category.rs for consistency - Fix compilation error in category.rs test (CategorySet::from slice) Part of: #23
- Add extract_expected_categories() for parsing query text to extract expected categories - Add CuiCategoryMap for UMLS CUI to category mapping with 100+ default mappings - Add extract_categories_from_cuis() for CUI-based entity categorization - Add count_cuis_per_category() for depth checking with CUI entities - Implement Disease/Symptom interchangeability in coverage validation - Add comprehensive tests matching task specification: - extract_categories_from_entities_task_spec - validate_coverage_passes_when_all_categories_present_task_spec - validate_coverage_fails_when_treatment_missing_task_spec - depth_check_requires_multiple_entities_in_category_task_spec - extract_expected_categories_from_query - cui_category_map_normalizes_cui_format - cui_category_map_returns_none_for_unknown Part of: #23
Add detailed mermaid diagrams to architecture documentation: - Overall pipeline architecture flowchart - Sequence diagram for complete data flow - Component architecture with all 4 medical crates - Two-layer validation detailed flow - Case-based learning pattern matching - 10 specialist role thesaurus structure - Edge and cloud deployment architectures - Performance budget breakdown - Integration points and dependencies Update both: - .docs/design/architecture.md - docs/submission/architecture-diagrams.md Part of: Architecture documentation updates
Add comprehensive planning documents for SNOMED pre-serialized artifact: Phase 1 Research (.docs/research/issue-34-pre-serialized-artifacts.md): - Current state analysis: UMLS and CPIC complete, SNOMED missing - RF2 file format documentation - Constraint and risk analysis - Key insight: build-snomed-artifact binary doesn't exist yet Phase 2 Design (.docs/design/implementation-plan-issue-34.md): - 9-step implementation plan following CPIC pattern - API design for SnomedHierarchy with IS-A traversal - Test strategy with unit, integration, and benchmark tests - Performance targets: <100MB artifact, <1s load time - File change list and acceptance criteria Issue: #34
…umentation - Removed unverifiable 'Crizotinib error' claim from T790M case - Changed 'Medical Error Prevention' to 'Medical Output Quality Improvement' - Added LLM_LEARNING_AND_CORRECTION.md with realistic examples - Updated all documentation to reflect actual capabilities: * Vague → Specific improvements * No evidence → Trial citations * Generic → Grounded recommendations - Focus on verified 2.45x improvement metric - Document actual LLM limitations and corrections
- Research document: CT.gov API analysis, data sources, use cases - Design document: Architecture, data models, module design - File structure for new crates: trial-protocol-parser, trial-query-service - Integration plan with existing MedGemma pipeline - Test strategy and success criteria
…arser - Created trial-protocol-parser crate structure - Implemented core types: NctId, TrialPhase, ClinicalTrial, etc. - Built CT.gov API v2 client with rate limiting and caching - Added JSON parser for CT.gov responses - Implemented schema validator with ValidationSummary - Added comprehensive unit tests - Follows design from .docs/design-trial-protocol-parser.md
- Added storage module with TrialStore trait - Implemented InMemoryStore for testing - Created TrialSerializer for JSON/JSONL formats - Added TrialExportBuilder for KG relationships - Added comprehensive tests for storage operations - Supports: store, retrieve, search, list, delete operations - Ready for Terraphim KG integration
… synthesis - Created trial-query-service crate - Implemented PatientTrialMatcher with scoring algorithm - Added TreatmentComparator for head-to-head analysis - Built EvidenceSynthesizer with GRADE-inspired levels - Features: - Patient criteria matching (age, sex, biomarkers) - Match scoring (eligibility, geographic, biomarker) - Treatment comparison with phase distribution - Evidence synthesis with confidence levels - Limitation identification - Ready for MedGemma pipeline integration
- Added parse_trial.rs example for single trial parsing - Added match_patient.rs example for patient-trial matching - Updated implementation document with complete status - Total implementation: ~5,700 lines across 2 crates - Features: - CT.gov API integration with rate limiting - 30+ structured types - In-memory storage with search - Patient matching with scoring - Treatment comparison - Evidence synthesis with GRADE levels - Ready for MedGemma competition submission
- Analyzed all components described in strategy vs actual implementation - Identified critical gaps: LLM Proxy (0%), Agent Orchestration (20%) - Medical Agents: 30% complete (4 agents vs 5 described) - Edge AI: 25% ready (no benchmarks) - Multimodal: 0% (no image processing) - Overall completion: ~40% - Strategy document significantly overstates implementation - Provides 3 options: Update strategy, rapid impl, or focus on one track
- Created DirectMedGemmaClient that connects directly to HF API or local models - Removed ProxyClient dependency on terraphim-llm-proxy service - Auto-detection: HF API → Local ctransformers → Mock fallback - Updated demo.rs to use direct client instead of proxy - Added shellexpand dependency for path expansion - Simplifies deployment - no external proxy service required - Works with HF_TOKEN env var or local model files
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR implements a Clinical Trial Protocol Parser with the following phases:
Phase 1: Core Types and CT.gov API Client
Phase 3: Knowledge Graph Storage
Phase 4: Query Service
Phase 5: Integration Examples
Changes
Commits
Testing
Related