Mixed-effects neural network modelling to predict longitudinal trends in fasting plasma glucose

Area of Science:

• Biostatistics
• Machine Learning
• Endocrinology

Background:

• Accurate prediction of fasting plasma glucose (FPG) trends is crucial for managing type 2 diabetes mellitus (T2DM).
• Traditional linear mixed-effects (LME) models struggle with complex, nonlinear data, while machine learning (ML) models often overlook random effects.
• A need exists for models that integrate the strengths of both LME and ML for improved FPG prediction.

Purpose of the Study:

• To develop and evaluate novel models for predicting FPG levels in T2DM patients.
• To combine LME and back propagation neural network (BPNN) approaches into a hybrid model (LMENN).
• To compare the predictive performance of the LMENN model against standalone LME and BPNN models.

Main Methods:

• Data from 779 T2DM patients were utilized, split into training (80%) and testing (20%) sets.
• Random Forest (RF) screening identified the top 10 predictive features (HOMA-β, HbA1c, HOMA-IR, urinary sugar, insulin, BMI, waist circumference, weight, age, group).
• An LME model was built, followed by multiple BPNN models, and finally, an integrated LMENN model using stacking. Performance was evaluated using 10-fold cross-validation and a test set.

Main Results:

• The best-fitting LME model indicated that baseline glucose levels influenced subsequent measurements, with consistent trends over time.
• The LMENN model effectively combined LME and BPNN, accounting for random effects and improving prediction accuracy.
• LMENN achieved Root Mean Square Error (RMSE) ranges of 0.447-0.471 (training), 0.525-0.552 (validation), and 0.511-0.565 (test sets), outperforming single models.

Conclusions:

• The integrated LMENN model offers a promising approach for analyzing longitudinal FPG monitoring data in T2DM.
• This study presents innovative methods for enhancing blood glucose prediction accuracy.
• The LMENN model has potential applications in clinical practice for personalized diabetes management.