How jtan1102/nla-mmr_cikm_2024 works
Project type and purpose
• This is a research codebase (“CIKM2024 submission 53”) implementing a drug-recommendation/multi-medication risk modeling pipeline on the MIMIC-III and MIMIC-IV datasets.
• It is not a full-stack app but an open-source research tool: ingesting raw patient data, constructing datasets, training a GNN-based model, and reporting metrics.
File structure
• Top-level:
• data/ — raw Pickle files for MIMIC-III and MIMIC-IV (ATC→SMILES mappings, drug-drug interaction matrices, patient visit records, vocabularies)
• datasets/ — Python scripts that read raw data and produce final, model-ready data splits (e.g., Drug_Rec_pt_new_final.py)
• scripts/ — core training/inference code:
– common_layers.py, util.py (shared utilities)
– gnn/ (custom GNNConv, GNNs, utils)
– main.py (entry point for training & validation)
• No frontend or deployment; focus is on data pipeline, model architecture, and training loop.
README
• Confirms the pipeline steps: data download & preparation → code clone → data placement → python main.py --dataset MIMIC-III|IV --Train
• Lists environment and dependencies (PyTorch, PyG, Transformers, etc.)
• Logs show training/inference loops with metrics.
a. Main components:
1. Raw Data Store (MIMIC-III/IV raw Pickles)
2. Data Preparation Module (datasets/*.py)
3. DataLoader & Batch Generator (scripts/util.py)
4. Model Definition (scripts/common_layers.py + scripts/gnn/)
5. Training Loop & Loss Computation (scripts/main.py)
6. Checkpointing & Evaluation (scripts/main.py)
7. Output & Metrics Logging
b. Relationships & interactions:
• Raw Data → Data Preparation → Processed Dataset → DataLoader
• DataLoader + Model → Forward Pass → Loss → Backprop → Model Update
• After each epoch: Model → Validation DataLoader → Evaluation Metrics → Logging
c. Architectural patterns & design principles:
• Modular code separation: data prep, model layers, GNN definitions, training script
• Use of custom PyG layers (GNNConv) and PyTorch training loop
• Single-process GPU training with configurable hyperparameters
d. Key technologies/frameworks:
• Python 3.8, PyTorch 2.0.1, torch-geometric (GNNs), Transformers/tokenizers (for possible pre-trained bio_pt embeddings), NumPy, Pandas, Scikit-learn.
Elements
• Data flow (→) from raw data to preprocessing to training to evaluation
• Rectangles for code modules
• Rounded rectangles for executables (main.py)
• Parallelograms for outputs/logs