Predicting in-hospital mortality for ICU patients with Sepsis-Associated Acute Kidney Injury using MIMIC-IV.

SA-AKI (Sepsis-Associated Acute Kidney Injury) is one of the deadliest complications in critical care. This project builds ML models to identify high-risk patients early, using 356 clinical features extracted from the first 24 hours of ICU stay:

• Vital signs (HR, BP, SpO₂, temperature, respiratory rate)
• Laboratory panels (creatinine, lactate, bilirubin, CBC, coagulation, ABG)
• Severity scores (APACHE III, SOFA — 6 organ-specific components)
• Comorbidities (17 Charlson flags)
• Therapies (mechanical ventilation, vasopressors, RRT)
• Fluid balance & urine output

Each time-series feature includes 9 statistical aggregations (first, last, median, IQR, range, delta, AUC, slope, count) over the 24-hour window.

Binary classification on event_observed (1 = in-hospital death, 0 = survived/censored), with time_to_event_hrs available for future survival analysis.

Pipeline:

1. Drop ID columns (stay_id, subject_id, hadm_id) to prevent data leakage
2. Drop features with >99% missing values
3. Categorical encoding — CatBoost native handling / OneHot for sklearn pipelines
4. Median imputation for numeric features
5. Stratified train/test split (80/20, seed=42)
6. Cross-validated AUROC evaluation

git clone https://github.com/stabgan/saaki.git
cd saaki
pip install -r requirements.txt
python saaki_model.py

Runs logistic regression (cross-validated) followed by CatBoost training, reports AUROC on a stratified 80/20 test split.

Note: Requires the MIMIC-IV dataset in data/. See Dataset below.

saaki/
├── saaki_model.py          # Training & evaluation pipeline
├── data/                   # MIMIC-IV SA-AKI cohort (PhysioNet access required)
│   ├── mimic_saaki_final.csv
│   └── mimic_saaki_final.xlsx
├── doc/                    # Data dictionary (356 columns)
├── requirements.txt        # Python dependencies
├── AGENTS.md               # Project context & methodology
├── plan.md                 # Roadmap
└── changelog.md            # Version history

Uses MIMIC-IV v3.1, which requires credentialed access through PhysioNet:

1. Complete CITI training for human research data
2. Sign the MIMIC-IV data use agreement
3. Place the processed cohort file in data/

Data files are not included in this repository.

• AUROC plateaus around 0.80 — feature engineering (missingness indicators, interaction terms) and ensemble stacking are planned
• requirements.txt includes libraries for planned future work (survival analysis, explainability) not yet used in the main pipeline
• Survival modelling (Cox PH, DeepSurv) not yet implemented

MIT — see LICENSE for details.

Built by Kaustabh Ganguly