A Comparative Analysis of Patient Similarity Measures for Outcome Prediction

Area of Science:

• Medical Informatics
• Computational Biology
• Clinical Decision Support

Background:

• Personalized medicine tailors treatments using patient characteristics for improved outcomes.
• Accurate patient similarity measurement is vital for identifying comparable patient cohorts.
• Existing research lacks comprehensive comparisons of patient similarity measures in large electronic health record (EHR) analyses.

Purpose of the Study:

• To conduct a large-scale comparative analysis of four patient similarity measures.
• To focus on feature weighting mechanisms within these similarity measures.
• To evaluate the effectiveness of these measures using retrospective EHR data.

Main Methods:

• Analyzed EHR data from 46,968 hospitalized patients.
• Compared four distinct patient similarity measures.
• Focused on feature weighting strategies, including type-based feature combination with grid-searched weights.

Main Results:

• The method using grid-searched weights to combine features based on their types demonstrated superior performance.
• This approach outperformed other evaluated patient similarity measures.
• Effectiveness was assessed through predictions of acute kidney injury, readmission, and mortality.

Conclusions:

• Optimized feature weighting, specifically combining features by type with grid-searched weights, enhances patient similarity assessment.
• This refined approach improves the identification of similar patient cohorts for clinical decision-making.
• The findings support the advancement of personalized medicine through more accurate patient stratification.