AltRAG+: Retrieval-Augmented Graphs for Dynamic Mortality Trajectories
This repository contains the implementation of AltRAG+, a retrieval-augmented clinical decision-support system designed for ICU patients using MIMIC-IV.
AltRAG+ integrates:
Structured EHR processing (MIMIC-IV)
Graph construction using NetworkX
Graph storage & querying using Neo4j
Semantic patient retrieval using SBERT + FAISS
Causal modeling with IPTW-weighted Cox PH
ML prediction models (Logistic Regression, XGBoost, LSTM)
LLM-based biomedical embeddings (BioGPT prototype)
The goal is to identify similar past ICU patients, analyze their treatment pathways, and generate K alternative, clinically-equivalent treatment strategies with predicted outcomes. AltRAG-Plus/ ├── Group9_Capstone.ipynb # Main notebook (end-to-end workflow) ├── INFO 5082 Final PPT.pptx # Final presentation ├── flowchart.png # Workflow diagram ├── README.md └── .gitignore Key Features
- Data Integration & Preprocessing
Combined 7+ MIMIC-IV tables
Added vitals, labs, treatment intensity features
Handled missing values (median, domain-based rules)
- Graph Construction
Created heterogeneous graph:
Patients (50920 nodes)
Diagnoses (16k)
Drugs (1423)
Procedures (146)
Exported to CSV and stored in Neo4j
- Patient Retrieval (SBERT + FAISS)
Encoded patient profiles
Built FAISS ANN index (~73k vectors)
Retrieved top-K similar “digital twins”
- Causal Modeling & Counterfactuals
IPTW weighting
High vs low treatment intensity
Cox PH survival comparison
- Prediction Models
Logistic Regression (AUC ~0.83)
XGBoost (AUC 0.87)
LSTM Time-Series model (Accuracy 0.910)
- Treatment Recommendation Engine
For a new ICU patient:
Encode profile
Retrieve similar historical patients
Extract treatment subgraphs from Neo4j
Predict mortality trajectory using LSTM
Generate alternative treatment paths How to Reproduce
To reproduce this work:
Request MIMIC-IV access from PhysioNet.
Download the dataset locally.
Update paths inside the notebook.
Install dependencies (PyTorch, Sentence-Transformers, FAISS, Neo4j, scikit-learn, xgboost, lifelines, tensorflow).
If referencing this project:
Rajesh Sathuri et al., “AltRAG+: Retrieval-Augmented Graphs for Dynamic Mortality Trajectories,” UNT Graduate Capstone, 2025.