End-to-end machine learning pipeline predicting 30-day ICU readmission risk using MIMIC-IV v3.1 dataset. From data cleaning to deployed clinical decision support tool.

Live Demo • View Notebook • Documentation

ICU readmissions within 30 days represent a critical healthcare challenge:

• Occur in 10-15% of ICU survivors
• Associated with 2-3× higher mortality risk
• Cost $15,000–$50,000 per readmission
• Strain limited ICU bed capacity and resources

Early identification of high-risk patients enables targeted interventions, transitional care planning, and better resource allocation.

A complete production-ready ML pipeline that:

1. Cleans & validates 48,676 ICU patient records from MIMIC-IV
2. Engineers 181 clinically-meaningful features with evidence-based approach
3. Trains & tunes gradient boosting models with rigorous evaluation
4. Deploys an interactive 4-page Streamlit app with clinical recommendations
5. Achieves 0.7884 AUC-ROC on held-out test set

Key Achievement: At 70% recall, the model achieves 15.2% precision (1.52× lift over 10% baseline), enabling targeted interventions for high-risk patients.

Test Set: 9,736 patients (20% holdout, never used in training or tuning)

Key Insight: Hospital utilization + kidney function + severity scores dominate the model.

Live Streamlit App: icu-readmission-predictor.streamlit.app

4-Page Interface:

• 🏠 Home: Project overview, top metrics, key risk factors
• 🔮 Patient Risk Predictor: Interactive calculator with gauge chart + clinical recommendations
• 📊 Model Performance: ROC curves, metrics, model comparison, literature benchmarking
• 🔬 Feature Importance: Top 20 features, category breakdown, partial dependence plots

┌─────────────────────────────────────────────────────────────────┐
│              MIMIC-IV v3.1 Dataset (Parquet)                    │
│                 48,676 ICU patients (2008-2019)                 │
│                 234 raw clinical features                       │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│         PART 1-2: Data Loading & Schema Validation              │
│  • Load 48,676 × 234 feature matrix from parquet                │
│  • Validate dtypes (Int64, Float64, Object, Datetime)           │
│  • Handle sentinel values (999999, -9, "UNKNOWN")               │
│  • Create quality flags (not row removal)                       │
│  • Classify columns (identifiers, target, features, leakage)    │
│  Output: X (227 features), y (readmit_30d_flag)                 │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│              PART 3: Feature Engineering                        │
│  • Remove constant features (0 variance)                        │
│  • Analyze missingness patterns (MCAR vs MNAR)                  │
│  • Remove 44 redundant clinical flags (chi-sq + Cramér's V)     │
│  • Retain 3 MNAR flags (emergency admission proxies)            │
│  Output: 181 validated features                                 │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│         PART 4: Preprocessing Pipeline Design                   │
│  • Evidence-based imputation strategy                           │
│  • MNAR testing (chi-square independence tests)                 │
│  • Median imputation (continuous)                               │
│  • Mode imputation (binary/categorical)                         │
│  • StandardScaler + OneHotEncoder                               │
│  Output: Unfitted sklearn ColumnTransformer                     │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│    PART 5: Train/Val/Test Split & Preprocessing                 │
│  • Stratified split: 64% / 16% / 20%                            │
│  • Train: 31,152 patients                                       │
│  • Val: 7,788 patients                                          │
│  • Test: 9,736 patients                                         │
│  • Fit pipeline on train → transform all splits                 │
│  Output: 247 features (after one-hot encoding)                  │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│              PART 6-7: Model Training & Tuning                  │
│  • Baseline models: Logistic Regression, Random Forest,         │
│                     XGBoost, LightGBM                           │
│  • Hyperparameter tuning: Optuna (40 trials, Bayesian)          │
│  • Class imbalance: class_weight='balanced'                     │
│  • Validation: 5-fold stratified CV                             │
│  Winner: LightGBM (Val AUC 0.7871)                              │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│           PART 8: Final Test Set Evaluation                     │
│  • Held-out test set: 9,736 patients (never seen)               │
│  • Final Test AUC-ROC: 0.7884                                   │
│  • Minimal overfitting: Val-Test gap = 0.0013                   │
│  Output: final_model.pkl (production-ready)                     │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│       PART 9: Feature Importance & Interpretability             │
│  • LightGBM gain + Permutation importance                       │
│  • Clinical narratives for top 20 features                      │
│  • Partial dependence plots                                     │
│  • Modifiable vs non-modifiable factor identification           │
└────────────────────────────┬────────────────────────────────────┘
                             │
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│                Production Streamlit Deployment                  │
│  • Streamlit Cloud (free tier)                                  │
│  • 4-page interactive app                                       │
│  • Real-time risk calculator with gauge chart                   │
│  • Automated clinical recommendations                           │
└─────────────────────────────────────────────────────────────────┘

Source: MIMIC-IV v3.1 (Medical Information Mart for Intensive Care)
Institution: Beth Israel Deaconess Medical Center, Boston
Time Period: 2008–2019
Access: Requires PhysioNet credentialed access + CITI training

Our Cohort:

• Total Patients: 48,676 ICU admissions
• Outcome: 30-day ICU readmission
• Prevalence: 10.07% (4,900 readmissions)
• Raw Features: 234 clinical variables
• Final Features: 181 → 247 (after preprocessing)

• Loaded 48,676 × 234 matrix from MIMIC-IV parquet extract
• Exhaustive dtype inspection (Int64, Float64, Object, Datetime)
• Sentinel value detection & conversion to NaN:
  • 999999 in glucose vitals
  • 999 in urine output, ALT, AST, troponin
  • -9 in PCO2 delta, anion gap
  • "UNKNOWN" in race, SOFA risk category
• Created 4 quality flags (no row removal):
  • urine_output_negative_flag
  • glucose_extreme_flags (2 columns)
  • temporal_violation_flag

• Classified all 234 columns:
  • Identifiers: 3 (subject_id, hadm_id, stay_id)
  • Target: 1 (readmit_30d_flag)
  • Leakage: 2 (next_icu_intime, days_to_readmission)
  • Completely Null: 1 (ntprobnp)
  • Features: 227
• Saved schema definition with exclusion criteria
• Output: X (227 features), y (target)

• Removed:
  • 2 constant features (0 variance)
  • 44 redundant clinical flags (chi-square + Cramér's V < 0.1)
• Retained:
  • 3 MNAR flags (height_available, weight_available, urine_measured)
  • Flags with statistical significance (p < 0.05 with readmission)
• Output: 181 validated features

Temporal Window: All features from first 24 hours of ICU admission to ensure prediction feasibility at discharge planning time.

Strategy: Evidence-based, clinically-informed imputation

Pipeline Components:

1. Continuous Features: Median imputation + StandardScaler
2. Binary Features: Mode imputation (no scaling)
3. Categorical Features: Mode imputation + OneHotEncoder + Rare category capping (<1% → "OTHER")

MNAR Testing:

• Created "was_measured" flags for high-missingness features
• Chi-square independence tests to detect MNAR patterns
• Retained 3 significant MNAR flags (emergency admission proxies)

Output: preprocessing_pipeline_UNFITTED.pkl

Split Strategy:

• Train: 64% (31,152 patients)
• Validation: 16% (7,788 patients)
• Test: 20% (9,736 patients)
• Method: Stratified (maintains 10.07% prevalence in all splits)

Preprocessing Execution:

1. Pre-split rare category capping (<1% → "OTHER")
2. Fit pipeline on train set only
3. Transform train/val/test independently
4. Feature expansion: 181 → 247 (one-hot encoding of categorical)

Output: preprocessing_pipeline_FITTED.pkl, preprocessed arrays

Models Trained (Default Hyperparameters):

Winner: LightGBM and Logistic Regression tied for baseline

Optimization:

• Trials: 40 (10 per model × 4 models)
• Method: Bayesian (TPE sampler)
• Objective: Maximize validation AUC-ROC
• CV: 5-fold stratified cross-validation

Search Spaces:

LightGBM:

{
    'n_estimators': [50, 500],
    'max_depth': [3, 12],
    'learning_rate': [0.01, 0.3],
    'num_leaves': [15, 127],
    'min_child_samples': [5, 100],
    'subsample': [0.5, 1.0],
    'colsample_bytree': [0.5, 1.0],
    'reg_alpha': [0, 10],
    'reg_lambda': [0, 10]
}

Best Configuration (LightGBM):

{
    'n_estimators': 300,
    'max_depth': 8,
    'learning_rate': 0.05,
    'num_leaves': 50,
    'min_child_samples': 30,
    'subsample': 0.8,
    'colsample_bytree': 0.8,
    'reg_alpha': 2.5,
    'reg_lambda': 4.2,
    'class_weight': 'balanced'
}

Tuned Results:

Winner: LightGBM (best test AUC, minimal overfitting)

Model: LightGBM (tuned)
Test Set: 9,736 patients (20% holdout, never used in training or tuning)

Performance Metrics:

Operating Point (70% Recall):

• Precision: 15.23%
• Specificity: 68.5%
• NPV: 95.8%
• Lift over Baseline: 1.52× (15.23% vs 10.07%)

Interpretation:

• At 70% recall, model flags ~15% of patients
• 1 in 7 flagged patients will actually readmit
• Catches 70% of all readmissions
• 95.8% of "low risk" predictions are correct

Comparison to Validation Set:

• Val AUC: 0.7871
• Test AUC: 0.7884
• Gap: -0.0013 (slight improvement on test)
• Conclusion: No overfitting, excellent generalization

Output: final_model.pkl (production-ready)

Key Insight: Hospital utilization (LOS, prior admissions) is the most predictive category, followed by laboratory values (kidney function, hematology).

1. Hospital Length of Stay

Longer hospital stays indicate higher illness severity, incomplete recovery, or complex care needs. Patients with prolonged hospitalization often have unresolved issues that increase readmission risk. Non-modifiable at discharge, but should trigger enhanced follow-up.

2. KDIGO Stage (Acute Kidney Injury)

Higher KDIGO stages (2-3) indicate moderate-to-severe acute kidney injury. AKI is a strong readmission predictor due to incomplete renal recovery, volume management challenges, and medication complications. Partially modifiable through fluid management and nephrotoxic drug avoidance.

3. Body Weight

Extremes in body weight (very low or very high) are associated with increased readmission risk. Low weight may indicate malnutrition or frailty; high weight complicates ventilation, mobility, and medication dosing. Fixed at discharge but guides care planning.

4. SOFA Score (Sequential Organ Failure Assessment)

Higher SOFA scores at ICU admission indicate multi-organ dysfunction. Patients with high SOFA remain physiologically fragile even after ICU discharge. Partially modifiable through supportive care, but reflects baseline severity.

5. Height Available Flag (MNAR Indicator)

When height is NOT measured, it often indicates emergency admission where routine vitals were skipped. This serves as a proxy for acute presentation and higher baseline acuity. Strongly associated with readmission risk despite not being a clinical variable itself.

MNAR Flags: We initially created "was_measured" flags for height, weight, and urine output as proxies for emergency admission. While statistically significant (p < 0.001), these flags ranked highly (#5 for height_available_flag), raising concerns about clinical interpretability.

Trade-off:

• Keep MNAR flags: +0.02 AUC improvement, but includes non-actionable features
• Remove MNAR flags: Slight performance drop, but all features are clinically meaningful

Final decision: [Choose one based on your goal]

• For research/clinical deployment: Remove MNAR flags → Pure clinical model
• For ML portfolio/learning: Keep MNAR flags → Demonstrates MNAR understanding

🌐 URL: https://icu-readmission-predictor.streamlit.app/

Features:

• Project overview with gradient header
• Key metrics dashboard (AUC, patients, features, readmission rate)
• Top 3 risk factors in styled cards
• Model development timeline

• Interactive input form:
  • Hospital LOS, ICU LOS, days since discharge
  • KDIGO stage, urine output rate
  • Weight, height measured flag, age
  • Hematocrit, Charlson index, SOFA score
• Calculate Risk button → Real-time prediction
• Gauge chart visualization:
  • Needle shows probability (0-100%)
  • Color zones: Green (Low), Yellow (Med), Red (High)
• Risk classification:
  • Low (<30%): Standard protocol
  • Medium (30-50%): Enhanced follow-up
  • High (≥50%): Intensive discharge planning
• Clinical recommendations:
  • Tailored to detected risk factors
  • Specific action items (e.g., nephrology referral for AKI)

• Test metrics (AUC-ROC, AUC-PR, Brier, Precision@70%)
• ROC curves comparison (all 4 models)
• Overfitting analysis (Val vs Test AUC)
• Published literature comparison table

• Tab 1: Top 20 Features
  • Bar chart (combined importance)
  • Clinical narratives for each feature
  • Modifiable vs non-modifiable badges
• Tab 2: Clinical Categories
  • Importance by category pie chart
  • Partial dependence plots (top 6 features)
  • Category summary table

Prerequisites:

• Python 3.10+
• pip or conda

Installation:

# Clone repository
git clone https://github.com/Jyoti-P-Das/icu-readmission-predictor.git
cd icu-readmission-predictor

# Create virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Running the Streamlit App:

streamlit run streamlit_app/app.py

App will open at http://localhost:8501

Model Inference (Python):

import joblib
import pandas as pd
import numpy as np

# Load model and pipeline
model = joblib.load('model/final_model.pkl')
pipeline = joblib.load('model/preprocessing_pipeline_FITTED.pkl')

# Example patient data (181 features required)
patient = {
    'hospital_los_days': 7.5,
    'kdigo_stage_max_first_24h': 2,
    'weight_kg': 78.5,
    'sofa_score_first_24h': 6,
    'height_available_flag': 1,
    'age_at_admission': 68,
    # ... (fill in remaining 175 features)
}

# Preprocess
X = pd.DataFrame([patient])
X_processed = pipeline.transform(X)  # 181 → 247 features

# Predict
risk_prob = model.predict_proba(X_processed)[0, 1]
print(f"30-Day Readmission Risk: {risk_prob*100:.1f}%")

icu-readmission-predictor/
│
├── notebooks/
│   ├── icu_readmission_analysis_CLEAN.ipynb  # Parts 0-9 (complete pipeline)
│   └── README.md
│
├── streamlit_app/
│   ├── app.py                                 # Main 4-page Streamlit app
│   ├── assets/                                # Images and CSV files
│   │   ├── roc_curves_test.png
│   │   ├── overfitting_analysis.png
│   │   ├── feature_importance_bar.png
│   │   ├── importance_by_category.png
│   │   ├── partial_dependence_top6.png
│   │   ├── test_evaluation_results.csv
│   │   └── clinical_narratives_top20.csv
│   └── README.md
│
├── model/
│   ├── final_model.pkl                        # Trained LightGBM (Part 8)
│   ├── preprocessing_pipeline_FITTED.pkl      # Fitted sklearn pipeline (Part 5)
│   └── feature_names_after_preprocessing.txt  # 247 feature names
│
├── docs/
│   ├── RESULTS_SUMMARY.md                     # Detailed performance metrics
│   └── DATA_STATEMENT.md                      # MIMIC-IV access guide
│
├── data/                                      # User-provided (not in repo)
│   ├── .gitkeep
│   └── README.md
│
├── README.md                                  # This file
├── requirements.txt                           # Python dependencies
├── .gitignore                                 # Blocks patient data
└── LICENSE                                    # MIT License

Required Steps:

1. Create PhysioNet account: https://physionet.org/register/
2. Complete CITI training: "Data or Specimens Only Research" course
3. Request MIMIC-IV access: https://physionet.org/content/mimiciv/
4. Download MIMIC-IV v3.1

Timeline: ~1 week for approval

Data Preparation:

• Our analysis uses a pre-extracted parquet file: model_dataset_readmission_30d.parquet
• Place this file in the data/ folder
• File contains 48,676 rows × 234 columns (cohort already extracted)

Step 1: Place Data

# Place your MIMIC-IV extract here:
data/model_dataset_readmission_30d.parquet

Step 2: Run Jupyter Notebook

jupyter notebook notebooks/icu_readmission_analysis_CLEAN.ipynb

Execute cells sequentially (Parts 0-9):

• Part 0: Setup (random seed, directories)
• Part 1: Data loading & EDA (48,676 × 234)
• Part 2: Schema validation (X, y creation)
• Part 3: Feature engineering (181 features)
• Part 4: Preprocessing pipeline design
• Part 5: Train/val/test split (64/16/20)
• Part 6: Baseline models (4 models)
• Part 7: Hyperparameter tuning (Optuna)
• Part 8: Test evaluation (0.7884 AUC)
• Part 9: Feature importance

Outputs:

• model/final_model.pkl
• model/preprocessing_pipeline_FITTED.pkl
• model/feature_names_after_preprocessing.txt
• All visualizations in research_artifacts/

Step 3: Deploy App

streamlit run streamlit_app/app.py

Context: This is a portfolio/learning project, not peer-reviewed research.

Interpretation: Performance is competitive with published traditional ML methods on MIMIC-IV, but this project lacks external validation, prospective testing, and peer review required for clinical deployment.

Pre-Discharge Risk Stratification at ICU Exit

High Risk (≥50%):

• Intensive discharge planning (≥48h before discharge)
• Social work referral + home health setup
• Subspecialty follow-up within 7 days
• Phone call within 24h of discharge

Medium Risk (30-50%):

• Standard discharge planning
• Phone call within 72h
• Clinic appointment within 14 days
• Medication reconciliation review

Low Risk (<30%):

• Routine discharge protocol
• Standard written instructions
• Clinic appointment in 2-4 weeks

Assuming 20% intervention effectiveness:

• Readmissions prevented: 140 per 1,000 discharges (70% recall × 20% effectiveness)
• ICU bed-days saved: ~210 days per 1,000 discharges
• Cost avoidance: ~$7M per 1,000 discharges (140 × $50K per readmission)

Caveats:

• Intervention effectiveness not validated
• Costs are estimates (vary by setting)
• Requires prospective validation study

1. Single-Center Data: MIMIC-IV from Beth Israel Deaconess only → generalizability uncertain
2. Retrospective Design: No prospective validation on unseen patients
3. Missing External Validation: Not tested on eICU or other ICU databases
4. MNAR Bias: Some features missing non-randomly (partially addressed with flags)
5. Class Imbalance: Only 10% positive class (handled with class weights)
6. Temporal Drift: Data from 2008-2019 → clinical practice may have evolved

Short-Term (Next 3-6 months):

• Fairness audit (racial/ethnic disparities)
• Calibration refinement (isotonic regression)
• Additional interpretability (LIME, counterfactuals)
• REST API for EHR integration

Medium-Term (6-12 months):

• External validation on eICU database
• Time-series modeling (LSTM for ICU trajectory)
• Multi-task learning (readmission + mortality + LOS)
• Prospective validation study

Long-Term (12+ months):

• Real-time EHR integration
• Randomized controlled trial with intervention arm
• Automated retraining pipeline (MLOps)
• Generalization to other ICU outcomes

• Python: 3.10+
• Notebook: Jupyter Lab
• Code Quality: Black (formatting), Flake8 (linting)
• Documentation: Markdown

If you use this code or methodology, please cite:

@misc{das2025icu,
  author = {Jyoti Prakash Das},
  title = {ICU 30-Day Readmission Risk Prediction using MIMIC-IV},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/Jyoti-P-Das/icu-readmission-predictor}
}

@article{johnson2023mimic,
  title={MIMIC-IV, a freely accessible electronic health record dataset},
  author={Johnson, Alistair EW and Bulgarelli, Lucas and Shen, Lu and others},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={1},
  year={2023},
  publisher={Nature Publishing Group}
}

Code: MIT License (see LICENSE)
Data: PhysioNet Credentialed Health Data License (requires separate access)

Author: Jyoti Prakash Das

Email: [email protected]

LinkedIn: - https://www.linkedin.com/in/jyoti-prakash-das-hca/

GitHub: @Jyoti-P-Das

Questions? Open an issue

• MIMIC-IV Team: MIT Laboratory for Computational Physiology
• PhysioNet: For hosting and credentialing access
• Beth Israel Deaconess Medical Center: Original data source
• Open-Source Community: Scikit-learn, LightGBM, Streamlit contributors

If this project helped you:

• ⭐ Star this repository
• 🐛 Report bugs via Issues
• 🔀 Contribute via Pull Requests
• 📢 Share with colleagues in healthcare ML