ICU In-Hospital Mortality Prediction — MIMIC-III

Machine Learning & Deep Learning Project

---

⚠️ Data Access Notice MIMIC-III data is not included in this repository. Access requires:

1. Completing CITI training at citiprogram.org
2. Submitting a credentialing application at physionet.org
3. Downloading MIMIC-III v1.4 from physionet.org/content/mimiciii/1.4

Place the downloaded .csv.gz files in the data/ directory before running the notebook. A demo subset (~100 patients) is available at physionet.org/content/mimiciii-demo/1.4 for testing without full access.

---

Task

Predict in-hospital mortality using the first 24 hours of an ICU stay.

• Why this task? Different from readmission and length-of-stay. Mortality prediction within the first 24h is clinically actionable — it guides care escalation decisions, ICU resource allocation, and palliative care discussions.
• Data: MIMIC-III v1.4 (Beth Israel Deaconess Medical Center ICU data)
• Cohort: First ICU stay, age ≥ 18, ICU LOS ≥ 24 hours (~46K patients)
• Outcome: Hospital expire flag (binary: 0 = survived, 1 = died in hospital)

---

Models

#	Model	Type	Notes
1	SVM (RBF kernel)	Classical ML	Baseline; PCA-reduced features
2	Decision Tree (pruned)	Classical ML	Cost-complexity pruning via CV
3	Cox Proportional Hazards	Survival	Time-to-event; interpretable HRs
4	LSTM (Bidirectional + Attention)	Deep Learning	Temporal vital sign sequences
5	Transformer Encoder (CLS token)	Deep Learning	Self-attention on 12 time steps

---

Project Structure

icu-mortality-prediction/
├── mortality_prediction.ipynb   ← Main notebook (all sections)
├── utils/
│   ├── mimic_utils.py           ← Feature engineering utilities
│   └── __init__.py
├── data/                        ← Place MIMIC-III CSV files here
├── outputs/
│   ├── figures/                 ← All saved plots
│   └── models/                  ← Saved model checkpoints
├── requirements.txt
└── README.md

---

Setup

# 1. Create virtual environment
python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate

# 2. Install dependencies
pip install -r requirements.txt

# 3. Place MIMIC-III CSV files in data/
#    Required files:
#      PATIENTS.csv
#      ADMISSIONS.csv
#      ICUSTAYS.csv
#      CHARTEVENTS.csv    (large — ~33GB uncompressed)
#      LABEVENTS.csv      (large — ~10GB uncompressed)
#      DIAGNOSES_ICD.csv

# 4. Launch Jupyter
jupyter notebook mortality_prediction.ipynb

---

Features Engineered

Static (per admission)

• Demographics: age, gender
• Admission type: EMERGENCY / ELECTIVE / URGENT
• Ethnicity (5 categories)
• ICU unit type (MICU, SICU, CCU, CSRU, etc.)

Temporal — First 24 Hours (vital signs)

Vital	Item IDs	Stats
Heart rate	211, 220045	min, max, mean, std
Systolic BP	51, 442, 220179...	min, max, mean, std
Diastolic BP	8368, 220180...	min, max, mean, std
Temperature (°C)	223762, 676, 223761, 678	min, max, mean, std
SpO2	646, 220277	min, max, mean, std
Respiratory rate	615, 618, 220210...	min, max, mean, std
GCS total	198, 226755	min, max, mean, std

Lab Values — First 24 Hours

Creatinine, BUN, WBC, Hemoglobin, Sodium, Glucose, Bicarbonate, Lactate, Potassium, Bilirubin (min, max, mean per stay)

Diagnosis Codes

Top-50 ICD-9 diagnosis codes as binary bag-of-codes features

---

Key Notebook Sections

Section	Content
1	Environment setup, imports, reproducibility
2	MIMIC-III data loading, cohort selection
3	Feature engineering (static, vitals, labs, diagnoses)
4	EDA — outcome distribution, feature distributions, heatmap
5	PCA explained variance + scatter, UMAP projection
6	Train/test split, SMOTE class balancing
7	SVM (RBF) baseline
8	Decision Tree with cost-complexity pruning
9	Cox Proportional Hazards + KM curves
10	Bidirectional LSTM with temporal attention
11	Transformer Encoder with CLS token
12	ROC curves, PR curves, calibration, confusion matrices, results table
15	Key takeaways, feature importance, hazard ratios

---

Evaluation Metrics

• AUROC — primary metric (area under ROC curve); threshold-independent
• AUPRC — especially important given class imbalance (~10–15% mortality)
• F1 Score — at default 0.5 threshold
• Calibration — reliability diagram; well-calibrated models are safer clinically
• Concordance Index (C-statistic) — for Cox PH model
• Confusion Matrix — false negatives (missed deaths) are clinically costly

---

Results

Cohort: 32,575 ICU stays · 10.9% mortality rate · 123 features

Model	AUROC	AUPRC	F1	Accuracy
Transformer Encoder	0.7683	0.3309	0.3891	0.8195
SVM (RBF)	0.7502	0.2458	0.3687	0.7450
LSTM	0.7479	0.3067	0.3721	0.8196
Cox PH (C-index)	0.7238	—	—	—
Decision Tree	0.6955	0.2565	0.3036	0.7050

---

References

1. Johnson et al. (2016). MIMIC-III, a freely accessible critical care database. Scientific Data.
2. Wang et al. (2020). MIMIC-Extract: A data extraction, preprocessing, and representation pipeline for MIMIC-III. CHIL.
3. Harutyunyan et al. (2019). Multitask learning and benchmarking with clinical time series data. Scientific Data.
4. Rajpurkar et al. (2017). CheXNet: Radiologist-level pneumonia detection on chest X-rays. arXiv.