Prior Informed Regularization of Recursively Updated Latent-Variables-Based Models with Missing Observations.

Area of Science:

• Biomedical Engineering
• Data Science
• Control Systems

Background:

• Traditional data-driven models often lack adaptability and fail to incorporate prior knowledge.
• Accurate prediction of glucose concentration (GC) is crucial for managing Type 1 diabetes (T1D).
• Existing models struggle with time-varying systems and missing data.

Purpose of the Study:

• To develop a novel regularized partial least squares (rPLS) algorithm for improved GC prediction in T1D.
• To incorporate prior knowledge and handle missing data within a latent variable-based modeling framework.
• To recursively update models as new data becomes available for adaptive predictions.

Main Methods:

• Proposed a regularized partial least squares (rPLS) algorithm with three steps: LV-based model development, missing data estimation, and future value prediction.
• Incorporated prior knowledge and handled missing independent covariates.
• Evaluated rPLS and rPLS with exogenous inputs (rPLSX) using simulated (in-silico) and clinical T1D data.
• Compared performance against recursive time series and kernel-based models.

Main Results:

• The rPLS algorithm demonstrated effectiveness in predicting GC variations in T1D patients.
• Achieved low Root Mean Squared Errors (RMSE) for GC prediction up to 60 minutes ahead in both simulated and clinical data.
• Simulated data RMSE: 2.52 mg/dL (30 min) and 5.81 mg/dL (60 min) with full information.
• Clinical data RMSE: 10.45 mg/dL (30 min) and 14.48 mg/dL (60 min).

Conclusions:

• The proposed rPLS approach offers a significant improvement over conventional recursive modeling algorithms for GC prediction in T1D.
• The algorithm's ability to integrate prior knowledge and manage missing data enhances its predictive accuracy.
• This adaptive, latent variable-based method shows promise for real-time glucose management in T1D.