How niyaz05/early-sepsis-detection-using-explainable-ai works
This repository is a reproducible ICU sepsis early-warning research system, not a deployed clinical service. Its user workflow runs a documented sequence of notebooks: explore and prepare patient-hour data, train and evaluate four model families, then compare performance and inspect clinical lead time, explainability, uncertainty, and fairness. Shared utilities support schema, metrics, and sequence handling; saved models and result tables capture outputs. The graph follows README-described notebook stages and sampled generator/helper references; the raw data directory is documented but absent from the supplied tree.
Main entry points: 02_preprocessing_and_splits.ipynb Patient splits and features · models Saved models · results Predictions and metrics · 03_model_xgboost.ipynb XGBoost training
Components 15
Data preparation
- Clinical EDA
notebooks/01_eda_overview.ipynb - Patient splits and features
notebooks/02_preprocessing_and_splits.ipynb
Model training
- XGBoost training
notebooks/03_model_xgboost.ipynb - LSTM training
notebooks/04_model_lstm.ipynb - TCN training
notebooks/05_model_tcn.ipynb - Transformer training
notebooks/06_model_transformer.ipynb - Sequence architectures
src/torch_models.py - Saved models
models
Evaluation and analysis
- Model comparison
notebooks/07_model_comparison.ipynb - Clinical lead time
notebooks/07_model_comparison.ipynb - XAI and fairness
notebooks/08_innovation_xai_uncertainty_fairness.ipynb - Predictions and metrics
results
Shared runtime
- Shared data and metrics helpers
src/utils.py
Other components
- Researcher
- ICU dataset
Connections 27
- Researcher → Clinical EDA: runs workflow
README.md - ICU dataset → Clinical EDA: profiles
README.md - Clinical EDA → Patient splits and features: provides schema
README.md - Patient splits and features → XGBoost training: provides tabular data
README.md - Patient splits and features → LSTM training: provides sequences
README.md - Patient splits and features → TCN training: provides sequences
README.md - Patient splits and features → Transformer training: provides sequences
README.md - LSTM training → Sequence architectures: uses architecture
- TCN training → Sequence architectures: uses architecture
- Transformer training → Sequence architectures: uses architecture
- XGBoost training → Saved models: saves model
README.md - LSTM training → Saved models: saves model
- TCN training → Saved models: saves model
- Transformer training → Saved models: saves model
- XGBoost training → Predictions and metrics: writes outputs
README.md - LSTM training → Predictions and metrics: writes outputs
- TCN training → Predictions and metrics: writes outputs
- Transformer training → Predictions and metrics: writes outputs
- Saved models → Model comparison: compares saved predictions
README.md - Predictions and metrics → Model comparison: loads predictions
- Model comparison → Clinical lead time: evaluates lead time
- Saved models → XAI and fairness: loads deep checkpoints
- Predictions and metrics → XAI and fairness: loads metrics
- Clinical EDA → Shared data and metrics helpers: uses helpers
- Patient splits and features → Shared data and metrics helpers: uses helpers
- XGBoost training → Shared data and metrics helpers: uses helpers
- XAI and fairness → Shared data and metrics helpers: uses sequence helpers