Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Predictability of non-CGM Diabetes Data for Personalized Recommendation

Tu Nguyen, Markus Rokicki

TL;DR

The study investigates predicting blood glucose for patients without continuous CGM using sparse, home-collected data. It compares baseline, classical, and ensemble machine-learning models, introducing prediction intervals via bagging to quantify uncertainty and enable abstention. Random Forest and Extra Trees achieve the best accuracy (MdAE as low as $2.16$) and demonstrate robustness to noise, with stability and confidence filtering further improving performance, especially after about 25–30 training instances. The findings support feasible, patient-specific recommendations from non-CGM data and inform uncertainty-aware deployment in home monitoring settings.

Abstract

With continuous glucose monitoring (CGM), data-driven models on blood glucose prediction have been shown to be effective in related work. However, such (CGM) systems are not always available, e.g., for a patient at home. In this work, we conduct a study on 9 patients and examine the online predictability of data-driven (aka. machine learning) based models on patient-level blood glucose prediction; with measurements are taken only periodically (i.e., after several hours). To this end, we propose several post-prediction methods to account for the noise nature of these data, that marginally improves the performance of the end system.

On the Predictability of non-CGM Diabetes Data for Personalized Recommendation

TL;DR

The study investigates predicting blood glucose for patients without continuous CGM using sparse, home-collected data. It compares baseline, classical, and ensemble machine-learning models, introducing prediction intervals via bagging to quantify uncertainty and enable abstention. Random Forest and Extra Trees achieve the best accuracy (MdAE as low as ) and demonstrate robustness to noise, with stability and confidence filtering further improving performance, especially after about 25–30 training instances. The findings support feasible, patient-specific recommendations from non-CGM data and inform uncertainty-aware deployment in home monitoring settings.

Abstract

With continuous glucose monitoring (CGM), data-driven models on blood glucose prediction have been shown to be effective in related work. However, such (CGM) systems are not always available, e.g., for a patient at home. In this work, we conduct a study on 9 patients and examine the online predictability of data-driven (aka. machine learning) based models on patient-level blood glucose prediction; with measurements are taken only periodically (i.e., after several hours). To this end, we propose several post-prediction methods to account for the noise nature of these data, that marginally improves the performance of the end system.

Paper Structure

This paper contains 15 sections, 6 equations, 12 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Dataset Overview
  3. Prediction
  4. Setup
  5. Algorithms
  6. Overall Results
  7. Prediction confidence
  8. Prediction intervals
  9. Regression evaluation
  10. when to predict: on the training size evaluation.
  11. when to predict: on the model stability.
  12. when to predict: on the prediction confidence.
  13. when to predict: combined factors.
  14. Overall results with Filtering methods
  15. Conclusion

Figures (12)

  • Figure 1: Blood glucose distribution for each patient.
  • Figure 2: Blood glucose prediction scenario.
  • Figure 3: Glucose, carbohydrate and insulin values per hour of day for patients 13 and 14.
  • Figure 4: Error bar graphs for predicted BG using unbiased variance.
  • Figure 5: Incremental training size - error bar graphs for predicted BG using unbiased variance for patient 8.
  • ...and 7 more figures