🏥 Inference Trace: GenAI’s 2nd Opinion

An Ethical, Modular Clinical Decision Support Pipeline for ICU Patient Summarization

---

📌 Overview

Inference Trace: GenAI’s 2nd Opinion is a modular, end-to-end Clinical Decision Support (CDS) pipeline built on real ICU data from the MIMIC-IV database. It combines explainable ML, FHIR compatibility, and GenAI-based clinical summarization to simulate how modern, ethical, and transparent CDS systems can function in real-world hospital environments.

This system is designed to:

• ✅ Ingest & normalize patient data as FHIR Bundles
• 🔍 Predict risks using structured ML models (e.g., Random Forest)
• 🧠 Provide explainability via SHAP + Permutation Importance
• 🩺 Summarize patient status using ClinicalBERT QA
• 📜 Log all model inferences & expose results for audit

Built with fairness, traceability, and modularity at its core.

---

Core Modules in src/

Module	Purpose
etl_mimic.py	ETL pipeline: transforms MIMIC-IV CSVs into clean, FHIR-compatible Bundles
icd9_decode.py	Decodes ICD-9 codes using local dictionaries for interpretability
train_los_model.py	Trains a classifier to predict patient Length of Stay
shap_model.py	Generates SHAP + permutation importance for any classifier
shap_model_iv.py	Special SHAP workflow for ICU-focused risk modeling
genai_infer.py	Prompts GenAI (LLM) over patient Bundles for explainable summarization

---

▶️ Usage Example

Below is a minimal example of how a single patient journey moves through the Inference Trace pipeline:

# Step 1: Extract & preprocess from MIMIC-IV CSVs to FHIR JSON
python src/etl_mimic.py --input core/ --output output/patients/

# Step 2: Predict Length of Stay (LOS) and generate SHAP explanations
python src/train_los_model.py --input output/patients/ --model output/models/los.pkl
python src/shap_model_iv.py --input output/patients/ --model output/models/los.pkl --output output/shap/

# Step 3: Prompt GenAI for a natural language summary
python src/genai_infer.py --input output/patients/ --output output/genai/

# Step 4: Review GenAI response
cat output/genai/case_10109128_bundle_answer.txt

---

🧠 **Explainability by Design**  
✅ **SHAP**: Patient-specific, feature-level model attributions  
✅ **Permutation Importance**: Global, model-agnostic insight into key drivers  
✅ **Bundle Injection**: Explanations embedded directly into FHIR-compatible patient JSON  
✅ **FHIR-Safe Integration**: Traceability preserved without violating schema structure

---

## 🤖 GenAI Summarization

- 🧬 Uses `emilyalsentzer/Bio_ClinicalBERT` for embeddings
- 💬 Prompts `NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO` via Hugging Face Inference API
- 🛡️ Clinical reasoning prompts written for **precision + interpretability**

```text
"You are a clinical reasoning assistant. Given the patient bundle, summarize risks and reasoning."