Mosaic Subspace Models for CGM Glucose Sensitivity

The input signal feature space is partitioned into multiple contiguous subspaces based on sensor electrical properties, with each subspace handled by a dedicated machine learning model. This segmentation allows the system to achieve higher modeling accuracy by capturing complex relationships in different operating regions separately, while the modular structure manages the complexity through organized, independent model units.

Multiple machine learning models trained on different subspaces are merged into a unified mosaic model framework. The system selectively applies appropriate models based on the current sensor operating conditions, combining the strengths of individual models to achieve high overall accuracy. This merging approach manages system complexity by providing a structured framework for coordinating multiple models.

The patent applies local quality by training each machine learning model on data specific to its corresponding subspace of sensor electrical properties. Each model develops specialized knowledge of the local data characteristics and relationships within its designated region, rather than being trained on all data uniformly. This local specialization improves measurement precision by capturing subspace-specific patterns that would be diluted in a global model.

The system partitions the input feature space along multiple dimensions (e.g., impedance magnitude, impedance phase) to create a multi-dimensional subspace structure. This dimensional approach ensures that data is distributed across subspaces in a way that maintains adequate sample sizes in each region, preventing information loss while still capturing complex relationships. The dimensional partitioning creates a more balanced distribution of data across models.