ICU Mortality Prediction with Fairness-Aware Machine Learning

This project uses real-world data from the MIMIC-III clinical database to build machine learning models that predict in-hospital mortality for ICU patients. A core focus is placed on identifying and mitigating bias in model predictions across sensitive attributes like gender, age group, and insurance type.

Project Goals

• Build predictive models for ICU patient mortality.
• Analyze and audit fairness across key demographic features.
• Apply bias mitigation techniques to reduce disparity without sacrificing accuracy.

Key Highlights

• Dataset: MIMIC-III (via PhysioNet, de-identified patient data).
• Models: Logistic Regression, SVM, Random Forest, Gradient Boosting.
• Fairness Audits:
  • Disparate Treatment
  • Disparate Impact
  • Disparate Mistreatment
• Bias Mitigation:
  • Reweighing (pre-processing)
  • Group-Aware SMOTE (resampling)
• Best Model:
  • Random Forest (Group-Aware SMOTE)
    • Accuracy: 85.8%
    • ROC AUC: 0.739
    • F1 Score (class 1): 0.285
    • Reduced fairness disparity across gender, age, and insurance

---

Author

Lohith Basavanahalli Anjinappa