Innovative statistical method for longitudinal and hierarchical data modeling: the GMEXGBoost method

Area of Science:

• * Computational statistics and machine learning.
• * Biostatistics and health data science.

Background:

• * Exponential data growth in healthcare necessitates advanced analytical methods beyond conventional machine learning.
• * Traditional algorithms struggle with correlated, longitudinal, and hierarchical data common in healthcare.
• * Generalized mixed-effects models (GLMMs) and XGBoost have limitations when used independently for complex data structures.

Purpose of the Study:

• * Introduce GMEXGBoost, a novel algorithm extending GLMMs with XGBoost's boosting framework.
• * Develop a method that explicitly incorporates data correlations while retaining predictive power.
• * Evaluate GMEXGBoost's performance against existing models in simulations and real-world data.

Main Methods:

• * Developed GMEXGBoost by integrating GLMM fixed-effect estimation with XGBoost's boosting and random-effect accounting.
• * Evaluated performance using simulations with varying effect structures and a real-world cohort study.
• * Benchmarked against GLMM, GLMMTree, GMERF, and XGBoost using metrics like PMAD, PMCR, sensitivity, specificity, accuracy, and AUC in RStudio.

Main Results:

• * XGBoost showed the lowest average errors in most scenarios.
• * GMEXGBoost demonstrated superior stability and accuracy with large random-effect variances or strong correlations.
• * GMEXGBoost outperformed other models on real-world data across key performance metrics.

Conclusions:

• * GMEXGBoost effectively combines GLMM and XGBoost capabilities for improved performance on complex problems.
• * The algorithm offers clear advantages for analyzing hierarchical and longitudinal datasets with strong correlations.
• * GMEXGBoost is a valuable tool for decision-making in healthcare and other fields with structured data.