Abstract
Background and Aims
Poor adherence of T2D patients on basal insulin treatment can lead to life-threatening health complications. In this context, CGM technology may be a future key to improve diabetes management by positively impacting treatment adherence. The aim of this study was to explore the use of Machine Learning (ML) for automated detection of basal insulin injection adherence in a retrospective view based on CGM data.
Methods
A cohort of subjects with simulated CGM data was generated using a T2D modified Medtronic Virtual Patient model with adherence annotation based on patient-specific once-daily basal insulin injection. Simple feature-engineered (CGM consensus inspired) ML classification models (logistic regression) were compared to more advanced deep learning (DL) models based on automatic feature-extraction (convolutional neural networks). Additionally, a fusion of both expert-dependent features and automatically extracted features (acquired from raw CGM data) were investigated. All models were reported as ensemble accuracies from a leave-one-patient-out cross-validation step.
Results
The results indicate that the simulated CGM data from the classification day and the consecutive day provide close to all information on whether a CGM day is considered adherent or not with ~ 80% accuracy (50% baseline), i.e. whether the basal insulin was injected or not. In this context, the fused DL models performed similarly to the simple feature-dependent classification models.
Conclusions
Given the simulated T2D based CGM data, simple feature-engineered ML models can be used for retrospective adherence detection based on the classificarion day and the consecutive day. The study should be followed up by analysis of real CGM data.
