Abstract—A multiple order model migration (MOMM)
algorithm and optimal model selection strategy are proposed here
for rapid model development and online glucose prediction. First,
the optimal model order is determined for each input and a
multiple order prediction model is used. Then a MOMM
algorithm is developed based on particle swarm optimization to
simultaneously revise multiple parameters. The multiple order
parameters of each input in the old model are quickly customized
so that the revised model can be used for new subjects with
desirable prediction accuracy. In particular, the influences of the
amount of modelling samples are analyzed to check the
applicability of different methods in order to suggest the selection
guideline of optimal model in response to different data sizes. The
above issues are investigated for two types of analysis and thirty in
silico subjects. For same-case analysis, three regions are
considered. In Region I, first order model migration (FOMM)
achieves the best performance. In Region II, MOMM algorithm
should be used and the prediction accuracy is superior. With
enough samples (Region III), subject-dependent model (SM)
algorithm can be chosen. In contrast, for cross-case analysis,
MOMM can reveal more powerful generalization ability than SM,
so that only two regions are considered. FOMM achieves the best
performance in Region I and MOMM algorithm is superior in
Region II when the number of samples is larger than 4h. The
MOMM algorithm is demonstrated to be able to transfer model
for new subjects with more reasonable model structure. Besides,
each algorithm has its applicability regarding the size of
modelling samples.
I. INTRODUCTION
LUCOSE is the primary source of energy for the body's
cells and the normal body's homeostatic mechanism keeps
blood glucose levels within the normal range thresholds
(70-180 mg/dl). Diabetes mellitus (DM), is a group of
metabolic diseases in which there are high or low blood sugar
levels over a prolonged period. People with Type I diabetes can
not successfully produce the necessary hormone insulin for any
This work was supported by the Program for the National Natural Science
Foundation of China (No. 61422306 and No. 61433005).
Chunhui Zhao, Wei Wang and Chengxia Yu are with the State Key
Laboratory of Industrial Control Technology, College of Control Science and
Engineering, Zhejiang University, Hangzhou 310027 China (C. Zhao is the
corresponding author, e-mail: chhzhao@zju.edu.cn).
proper regulation of glucose level. It is important to make
glucose prediction for the proactive generation of
hyper/hypoglycemia alerts. Considering the inherent
complexity of the glucose-insulin dynamics, data-driven (or
empirical) models[1-6] can be used to explore the information
hidden in the data. Despite of the success, they, as a data-driven
algorithm, in general requires sufficient modelling data and the
work of model identification has to be repeated for different
individuals. This may cause repetitive cost and burden for
patients and clinicians and require a lot of modelling efforts.
Besides, it may not be always practical to wait time for
collection of sufficient samples before model development. It
motivates the study of modelling method based on a small
number of data in particular in an emergency.
The literature on rapid and economical modelling for glucose
prediction is rare. Previous works[1,3] have mentioned and
verified the concept of a universal or global autoregressive (AR)
model. However, it is not useful for glucose control since the
information of exogenous inputs are not employed. Besides,
Zhao et al.[3] also have pointed out that ARX models with two
exogenous inputs, insulin delivery and meal carbohydrate
(CHO), were not global. Recently, the idea of model
migration[4] was proposed for rapid modelling in glucose
prediction considering the similar model structure but different
model parameters between different individuals. It revealed
that a prediction model developed for data from one subject can
be made valid for a new subject with proper model parameter
adjustment and generate comparable performance in
comparison with subject-dependent standard model. However,
it only simply and arbitrarily considers first order for each
exogenous input which may not be the optimal model structure
for glucose prediction. It is termed first order model migration
(FOMM) for comparison here. Moreover, the parameter
adjustment strategy may converge to a suboptimal solution
since the FOMM algorithm only regulates one parameter at one
time while keeping the others invariable in particular when it
applies to multiple parameters. A proper model migration
strategy would be necessary for regulating multiple parameters
at the same time. To the best of our knowledge, model
migration approaches able to deal with multiple order of input
parameters in glucose prediction model are not available.
Moreover, the applicability of different algorithms has not been
Predicting glucose concentration for
hyper/hypoglycemia alert: multiple order model
migration and optimal model selection via
analysis of sample size
Chunhui Zhao*, Senior Member, IEEE, Wei Wang, Chengxia Yu
2017 American Control Conference
Sheraton Seattle Hotel
May 24–26, 2017, Seattle, USA
978-1-5090-5992-8/$31.00 ©2017 AACC 1468