How supriya090/llm-clinical-coding-audit works
This is a code-only reference implementation for interpretable clinical ICD coding and audit. Its principal workflow prepares discharge-note evidence, predicts codes using a classifier trained with embedding-matched silver evidence, then verifies candidate codes with LLMs; audit and evaluation modules support error analysis. The README and excerpts establish the major stages, but do not show a single orchestrator or all inter-module handoffs, so some documented pipeline links are represented at subsystem level.
Main entry points: section_splitting.py Section splitting · prompts.py Shared prompts · deepseek_verifier.py DeepSeek verification · inference.py Code inference
Components 19
Note preparation
- Section splitting
src/preprocessing/section_splitting.py - Negation filtering
src/preprocessing/negation_removal.py
Evidence generation
- Silver matching
src/silver_evidence/embedding_matching.py - Shared prompts
src/prompts.py - GPT extraction
src/evidence_extraction/gpt_extractor.py - Gemini extraction
src/evidence_extraction/gemini_extractor.py - DeepSeek extraction
src/evidence_extraction/deepseek_extractor.py
Code prediction
- Classifier training
src/code_classifier/train_classifier.py - Code inference
src/code_classifier/inference.py
Code verification
- GPT verification
src/verification/gpt_verifier.py - Gemini verification
src/verification/gemini_verifier.py - DeepSeek verification
src/verification/deepseek_verifier.py
Audit and evaluation
- Ensemble crosscheck
src/audit/ensemble_crosscheck.py - Agreement analysis
src/audit/iaa_analysis.py - Audit reporting
src/audit/reporting.py - Code metrics
src/evaluation/metrics.py
Other components
- Clinical notes
- LLM services
- Audit results
Connections 31
- Clinical notes → Section splitting: provides notes
README.md - Section splitting → Negation filtering: filters phrases
src/preprocessing/section_splitting.py - Section splitting → Silver matching: supplies candidates
src/preprocessing/section_splitting.py - Section splitting → GPT extraction: supplies candidates
src/preprocessing/section_splitting.py - Section splitting → Gemini extraction: supplies candidates
src/preprocessing/section_splitting.py - Section splitting → DeepSeek extraction: supplies candidates
src/preprocessing/section_splitting.py - Silver matching → Classifier training: creates training data
src/silver_evidence/embedding_matching.py - GPT extraction → Shared prompts: uses templates
src/evidence_extraction/gpt_extractor.py - Gemini extraction → Shared prompts: uses templates
src/evidence_extraction/gemini_extractor.py - DeepSeek extraction → Shared prompts: uses templates
src/evidence_extraction/deepseek_extractor.py - GPT extraction → LLM services: requests extraction
src/evidence_extraction/gpt_extractor.py - Gemini extraction → LLM services: requests extraction
src/evidence_extraction/gemini_extractor.py - DeepSeek extraction → LLM services: requests extraction
src/evidence_extraction/deepseek_extractor.py - Classifier training → Code inference: produces model
README.md - Code inference → GPT verification: provides candidate codes
README.md - Code inference → Gemini verification: provides candidate codes
README.md - Code inference → DeepSeek verification: provides candidate codes
README.md - GPT verification → Shared prompts: uses templates
src/verification/gpt_verifier.py - Gemini verification → Shared prompts: uses templates
src/verification/gemini_verifier.py - DeepSeek verification → Shared prompts: uses templates
src/verification/deepseek_verifier.py - GPT verification → LLM services: requests verification
src/verification/gpt_verifier.py - Gemini verification → LLM services: requests verification
src/verification/gemini_verifier.py - DeepSeek verification → LLM services: requests verification
src/verification/deepseek_verifier.py - DeepSeek verification → Ensemble crosscheck: informs undercoding audit
src/verification/deepseek_verifier.py - GPT verification → Audit results: writes verification output
src/verification/gpt_verifier.py - Gemini verification → Audit results: writes verification output
src/verification/gemini_verifier.py - DeepSeek verification → Audit results: writes verification output
src/verification/deepseek_verifier.py - Ensemble crosscheck → Audit results: supports audit
README.md - Agreement analysis → Audit results: reports agreement
README.md - Audit reporting → Audit results: reports findings
README.md - Code metrics → Audit results: measures performance
README.md