"Using network analysis modularity to group health code systems and decrease dimensionality in machine learning...

Area of Science:

• Computational biology
• Health informatics
• Machine learning

Background:

• Machine learning (ML) models struggle with high-dimensional Healthcare Coding Systems (HCSs) like ICD, ATC, and DRG.
• Encoding these codes presents a challenge: balancing dimensionality reduction with information preservation.

Purpose of the Study:

• To evaluate Network Analysis modularity for grouping HCSs.
• To improve the encoding process for ML models in healthcare and pharmacy research.

Main Methods:

• A multimorbidity network was constructed using ICD-9 codes from the MIMIC-III dataset.
• Modularity detection algorithms grouped codes, and performance was compared across four strategies (modularity, hierarchy, CCS, binary encoding) for predicting ICU readmissions.
• Logistic Regression, SVM, and Gradient Boosting Machines were used to assess model performance.

Main Results:

• Modularity encoding significantly outperformed binary encoding in ML models, improving accuracy, AUC, recall, and precision.
• Modularity-based grouping generally showed superior performance compared to other methods, particularly in AUC and precision.
• Performance improvements were observed across various ML algorithms.

Conclusions:

• Modularity encoding enhances ML model performance in pharmacy research by reducing dimensionality while retaining crucial information.
• This approach is versatile, applicable to hierarchical and non-hierarchical HCSs, clinically relevant, and improves model interpretability.
• A Python package is available to support the application of modularity encoding in future research.