(XAI) Explaining Readmission of Diabetes Patients
Project Overview
This project aims to identify and analyze factors contributing to the readmission of diabetes patients using various explainable machine learning methods. The study focuses on two main questions:
- What factors contribute most over time to a diabetes patient being readmitted?
- What general features influence readmission risk in diabetic patients?
Project Poster
Key Components
- Data Preprocessing:
- Removal of features with >50% missing values
- Sorting datapoints by patient number and encounter ID
- Mapping of diagnostic codes to ICD9 descriptions
- Feature filtering based on importance scores
- Machine Learning Models:
- RuleFit
- Bayesian Networks (Tree Augmented Naive Bayes and Forest Augmented Naive Bayes)
- Explainable Boosting Machine (EBM)
- Neural Network with SHAP analysis
- Explainability Techniques:
- Rule extraction (RuleFit)
- Causal tree construction (Bayesian Networks)
- Global and local explanations (EBM)
- SHAP (SHapley Additive exPlanations) analysis
Model Performance Comparison
| Model | Accuracy | F1-score (<30 days) | F1-score (>30 days) | F1-score (No Readmission) |
|---|---|---|---|---|
| RuleFit (Filtered) | 0.58 | 0.01 | 0.31 | 0.71 |
| RuleFit (Unfiltered) | 0.57 | 0.01 | 0.33 | 0.71 |
| TANBS | 0.555 | - | - | - |
| FANBS | 0.616 | - | - | - |
| EBM | 0.58 | 0.06 | 0.40 | 0.70 |
| Neural Network (Filtered) | 0.573 | 0.03 | 0.36 | 0.71 |
| Neural Network (Unfiltered) | 0.573 | 0.02 | 0.44 | 0.71 |
Key Findings
- The most important features for predicting readmission include:
- Number of inpatient visits
- Discharge disposition ID
- Number of diagnoses
- Admission source ID
- Diabetes medication
All models struggled with accurately predicting readmissions within 30 days, highlighting the complexity of short-term risk factors.
- The integration of multiple explainable AI techniques provided comprehensive insights into both temporal and causal relationships affecting readmission risk.
Ethical Considerations
- Ensuring patient privacy and data security
- Obtaining informed consent for data usage
- Evaluating models for fairness across patient groups
- Clinical validation before deployment
Resources
- Data Source: Diabetes 130-US Hospitals for Years 1999-2008 dataset
- Libraries Used: TensorFlow, Keras, SHAP, H2O, pgmpy, interpret.glassbox
- Code Repository: GitHub - XAI DIabetes Readmission