Mixed effect machine learning: A framework for predicting longitudinal change in hemoglobin A1c.

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Longitudinal data analysis is crucial for understanding chronic disease progression.
• Traditional generalized linear mixed-effect models (GLMMs) rely on unverifiable assumptions.
• Standard machine learning models often fail with the correlated, non-identically distributed nature of longitudinal data.

Purpose of the Study:

• To develop an analytic framework integrating GLMM random-effects into non-linear machine learning models.
• To address the challenges of temporal heterogeneity, sparsity, and varying lengths in patient data.
• To predict longitudinal changes in glycemic control (HbA1c) in type 2 diabetes patients.

Main Methods:

• Formulation of a mixed-effect machine learning (MEml) framework.
• Integration of random-effects structures from GLMMs into non-linear ML models.
• Application of MEml to predict HbA1c changes in adults with type 2 diabetes.

Main Results:

• MEml demonstrated competitive performance against traditional GLMMs.
• MEml significantly outperformed standard machine learning models lacking random-effects accounting.
• MEml showed superior accuracy in predicting glycemic changes across multiple future clinical visits.

Conclusions:

• Modeling random-effects is essential for effective machine learning on longitudinal data.
• The MEml framework offers high resistance to correlated data and accounts for random-effects.
• MEml accurately predicts longitudinal clinical outcomes in real-world settings.