Casting a Wider Net: Data Driven Discovery of Proxies for Target Diagnoses

Area of Science:

• Health Services Research
• Computational Biology
• Health Informatics

Background:

• The Hospital Readmissions Reduction Program (HRRP) links hospital reimbursement to readmission rates for specific diagnoses.
• Current HRRP methods may be susceptible to manipulation due to a limited set of target diagnoses.

Purpose of the Study:

• To develop a novel computational approach for identifying proxy indicators of target diagnoses.
• To improve the stability and accuracy of hospital readmission performance metrics.

Main Methods:

• Constructed a Markov Blanket using Google's PageRank algorithm to identify co-occurring diagnoses and procedures.
• Applied the approach to Medicare and Medicaid discharge records from California and New York (2009-2011).
• Evaluated model performance using Area Under the ROC Curve (AUC) and feature stability analysis.

Main Results:

• Achieved high predictive accuracy for target diagnoses like Sepsis (AUC=0.94) and Pneumonia (AUC=0.68).
• Demonstrated consistent feature stability, improving with larger training datasets.
• The novel approach showed prediction accuracy comparable to or exceeding existing methods for discharge record data.

Conclusions:

• The proposed Markov Blanket approach effectively identifies relevant diagnoses and procedures.
• This method offers improved accuracy and interpretability for predicting hospital readmissions.
• The findings suggest a more robust alternative for hospital performance evaluation under HRRP.