aai540-group3/diabetes-readmission

Model Description

This model predicts 30-day hospital readmissions for diabetic patients using historical patient data and machine learning techniques. The model aims to identify high-risk individuals enabling targeted interventions and improved healthcare resource allocation.

Overview

• Task: Binary Classification (Hospital Readmission Prediction)
• Model Type: autogluon
• Framework: Python Autogluon
• License: MIT
• Last Updated: 2024-10-29

Performance Metrics

• Test Accuracy: 0.8865
• Test ROC-AUC: 0.6467

Feature Importance

Significant features and their importance scores:

Feature	Importance	p-value	99% CI
0	0.0563	3.24e-04	[0.0294, 0.0832]
1	0.0358	8.45e-06	[0.0290, 0.0426]
2	0.0080	0.0083	[-0.0013, 0.0173]
3	0.0046	1.96e-04	[0.0027, 0.0065]
4	0.0023	0.0055	[-0.0001, 0.0046]
5	0.0008	0.1840	[-0.0027, 0.0043]

Note: Only features with non-zero importance are shown. The confidence intervals (CI) are calculated at the 99% level. Features with p-value < 0.05 are considered statistically significant.

Features

Numeric Features

• Patient demographics (age)
• Hospital stay metrics (time_in_hospital, num_procedures, num_lab_procedures)
• Medication metrics (num_medications, total_medications)
• Service utilization (number_outpatient, number_emergency, number_inpatient)
• Diagnostic information (number_diagnoses)

Binary Features

• Patient characteristics (gender)
• Medication flags (diabetesmed, change, insulin_with_oral)

Interaction Features

• Time-based interactions (medications × time, procedures × time)
• Complexity indicators (age × diagnoses, medications × procedures)
• Resource utilization (lab procedures × time, medications × changes)

Ratio Features

• Resource efficiency (procedure/medication ratio, lab/procedure ratio)
• Diagnostic density (diagnosis/procedure ratio)

Intended Use

This model is designed for healthcare professionals to assess the risk of 30-day readmission for diabetic patients. It should be used as a supportive tool in conjunction with clinical judgment.

Primary Intended Uses

• Predict likelihood of 30-day hospital readmission
• Support resource allocation and intervention planning
• Aid in identifying high-risk patients
• Assist in care management decision-making

Out-of-Scope Uses

• Non-diabetic patient populations
• Predicting readmissions beyond 30 days
• Making final decisions without clinical oversight
• Use as sole determinant for patient care decisions
• Emergency or critical care decision-making

Training Data

The model was trained on the Diabetes 130-US Hospitals Dataset (1999-2008) from UCI ML Repository. This dataset includes:

• Over 100,000 hospital admissions
• 50+ features including patient demographics, diagnoses, procedures
• Binary outcome: readmission within 30 days
• Comprehensive medication tracking
• Detailed hospital utilization metrics

Training Procedure

Data Preprocessing

• Missing value imputation using mean/mode
• Outlier handling using 5-sigma clipping
• Feature scaling using StandardScaler
• Categorical encoding using one-hot encoding
• Log transformation for skewed features

Feature Engineering

• Created interaction terms between key variables
• Generated resource utilization ratios
• Aggregated medication usage metrics
• Developed time-based interaction features
• Constructed diagnostic density metrics

Model Training

• Data split: 70% training, 15% validation, 15% test
• Cross-validation for model selection
• Hyperparameter optimization via grid search
• Early stopping to prevent overfitting
• Model selection based on ROC-AUC performance

Limitations & Biases

Known Limitations

• Model performance depends on data quality and completeness
• Limited to the scope of training data timeframe (1999-2008)
• May not generalize to significantly different healthcare systems
• Requires standardized input data format

Potential Biases

• May exhibit demographic biases present in training data
• Performance may vary across different hospital systems
• Could be influenced by regional healthcare practices
• Might show temporal biases due to historical data

Recommendations

• Regular model monitoring and retraining
• Careful validation in new deployment contexts
• Assessment of performance across demographic groups
• Integration with existing clinical workflows

Monitoring & Maintenance

Monitoring Requirements

• Track prediction accuracy across different patient groups
• Monitor input data distribution shifts
• Assess feature importance stability
• Evaluate performance metrics over time

Maintenance Schedule

• Quarterly performance reviews recommended
• Annual retraining with updated data
• Regular bias assessments
• Ongoing validation against current practices

Citation

@misc{diabetes-readmission-model,
  title = {Hospital Readmission Prediction Model for Diabetic Patients},
  author = {Agustin, Jonathan and Robertson, Zack and Vo, Lisa},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/{REPO_ID}}}
}

@misc{diabetes-dataset,
  title = {Diabetes 130-US Hospitals for Years 1999-2008 Data Set},
  author = {Strack, B. and DeShazo, J. and Gennings, C. and Olmo, J. and
            Ventura, S. and Cios, K. and Clore, J.},
  year = {2014},
  publisher = {UCI Machine Learning Repository},
  doi = {10.24432/C5230J}
}

Model Card Authors

Jonathan Agustin, Zack Robertson, Lisa Vo

For Questions, Issues, or Feedback

• GitHub Issues: Repository Issues
• Email: [team contact information]

Updates and Versions

• {pd.Timestamp.now().strftime('%Y-%m-%d')}: Initial model release
• Feature engineering pipeline implemented
• Comprehensive preprocessing system added
• Model evaluation and selection completed

---

Last updated: {pd.Timestamp.now().strftime('%Y-%m-%d')}