Dynamic modelling of glycaemic prediction in people with Type 1 diabetes based on genetic algorithm optimisation

Abstract

Type 1 diabetes is caused by the destruction of insulin-producing mechanisms in the pancreas, resulting in uncontrolled blood glucose concentration (BGC). Previous modelling techniques have generally been marginally successful in producing accurate predictions, in part due to the lack of real physiological data from patients. This research considers a Diabetes UK study, which collected free-living data from individuals with Type 1 diabetes using a continuous glucose monitor, activity armband and food and insulin diaries to investigate methods to predict BGC over several hours. A further aim was to improve understanding of the role of physical activity in BGC variations.

Initial tests were performed to directly find invisible patterns within the raw meal, insulin, sleep and activity data but no significant correlations that could be beneficial to help develop better predictive models was found. A number of methods of modelling BGC in Type 1 diabetes are investigated and it was found out that every diabetic subject has his/her own parameter variations, different to the one-size-fits-all reference parameters, hindering improved predictions. As a resolution, an optimization-driven technique is developed for the linear time-invariant ARX class of models, aiming to refine reference constant parameters in glucose and insulin sub-models to suit the very distinct features of each set of Type 1 BGC data. The outcome of this work showed that optimising these values for parameters of digestion in the gut and insulin diffusion in subcutaneous tissue could improve the predictive properties of the ARX models. However, although the prediction was improved over previous research, it was still far from satisfactory due to the non-linearities existing in the BG data. Thus, non-stationary Gaussian Processes was investigated to address the BGC’s non-stationarity and autocorrelation with variations present both intra and inter-patient. This approach was not only able to better describe local blood glucose dynamics in response to carbohydrate and insulin intake, but also permitted the inclusion of physical activity energy expenditure data, an advantage over the previous linear time-invariant ARX.

Physiological based-models were then investigated for parameter learning, especially the physical activity data for BGC prediction over several hours. These models showed some predicting capability, highlighting periods of low and high BGC. However, physiological variables analysed were found to differ from estimates in the literature derived from healthy individuals. Therefore, a new semi-empirical compartmental model was developed to better represent the underlying physiology, however resulted in only minimal improvement. Finally, a multi-objective optimizationbased approach, constrained by published upper and lower bounds for the variables, was employed to refine these parameter estimates. This cohort-driven technique was shown to perform the best, exceeding those of the previous physiological-based modelling methods and while it is not limited to only short-term predictions, it achieved better prediction on most of the volunteers’ studied.

More information

Identifiers

ORCID for Andrew Chipperfield: orcid.org/0000-0002-3026-9890

Catalogue record

Export record