Archived

This content is available here for research, reference, and/or recordkeeping.

Author ORCID Identifier

https://orcid.org/0009-0003-8667-0098

Date Available

10-1-2028

Year of Publication

2026

Document Type

Thesis

Degree Name

Master of Computer Engineering (MCompE)

College

Engineering

Department/School/Program

Electrical and Computer Engineering

Faculty

Robert Adams

Xiaohua Douglas Zhang

Simone Silvestri

Abstract

Background:

Continuous Glucose Monitoring (CGM) is vital for modern diabetes management, yet missing data gaps caused by sensor failures or user behavior frequently bias clinical metrics. While simple interpolation is common, it often fails to capture complex glucose dynamics. We evaluated a principled Multiple Imputation by Chained Equations (MICE) framework using various Machine Learning engines to optimize missing value recovery.

Methods:

Using data from the Type 1 Diabetes and Exercise Initiative (T1DEXI, n = 497), we followed random and group-based marking and simulated five missingness scenarios (5%–40%). Six models were integrated into the MICE and GAIN framework, including K-Nearest Neighbor (KNN), Random Forest (RF), XGBoost, LightGBM, Long Short-Term Memory (LSTM), and Autoregressive Integrated Moving Average (ARIMA). Performance was assessed via Mean Absolute Percentage Error (MAPE), Mean Relative Difference (MRD), and coefficient of determination (R2). We also evaluated the impact of a 1% missing gap reduction rule (excluding gaps >100 minutes).

Results:

For the random-missingness approach, the statistical method ARIMA emerged as the strongest overall engine. It achieved an MAPE of 0.097% and an MRD of 0.000975 at 5% missingness, and it maintained the highest accuracy at 40% missingness, with an MAPE of 1.816%. Random Forest was the second-best performer, while the gradient-boosted models also demonstrated strong stability.

In contrast, under the group-based missingness approach, ARIMA did not consistently outperform the other models across all missing-data levels. At 5% missingness, ARIMA achieved the best performance. However, from 10% to 40% missingness, machine learning models produced significantly better results. Among these models, KNN and XGBoost performed the best, although only by a small margin compared with the other machine learning methods.

Although GAIN is a popular framework to handle missing values, it could not produce a consistent result in this work.

Conclusion:

The statistical method ARIMA and the ensemble boosting method XGBoost within a MICE framework provide a robust, uncertainty-aware strategy for repairing CGM data. To facilitate clinical and research application, we have released these optimized models via the open-source imputeCGM Python and R packages.

Lastly, we developed a web-based platform that allows technical and non-technical users to impute missing CGM data by simply uploading a dataset.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2026.437

Archival?

Archival

Download

Available for download on Sunday, October 01, 2028

Contact Author

Share

COinS