PbFG: Physique-based fuzzy granular modeling for non-invasive blood glucose monitoring

Weijie Liu, Anpeng Huang, Ping Wang, Chao Hisen Chu

Research output: Contribution to journalArticlepeer-review

1 Scopus citations

Abstract

Objective: Continuous monitoring of blood glucose concentrations (BGC) is a crucial and integrated part of diabetes management. The most widely used glucose monitoring devices are blood glucose meters based on minimally-invasive finger-stick tests. The recent trend, however, is shifting towards non-invasive glucose monitoring (NGM) technology, as it alleviates the pain of frequently prick on patients’ skin. In NGM technology, it is difficult to establish a universal model that meets the clinical requirements in accuracy due to the individual differences in granularity, e.g., skin thickness, blood volume, body fat, etc. How to reduce the influence of individual differences on NGM is a technical difficulty. This study aims to propose a framework and a model-level solution for NGM while improving accuracy. Methods: An optical NGM prototype is developed, and 4012 samples from 89 patients are collected by clinical trials. By clustering these samples using a fuzzy c-means algorithm, we found that human pulse waveforms can be roughly divided into four types, corresponding to four different physique characteristics, which reflect the individual differences to some extent. Based on this discovery, we propose a Physique-based Fuzzy Granular modeling (PbFG) framework, in which four granular glucose estimators are used to predict BGC values, and a fuzzy physique classifier is applied to classify user's physique. The final BGC value is then obtained by fusing the four granular BGC values with the fuzzy physique classification results. Results: Using four practical machine learning algorithms as the BGC estimators, the PbFG framework is clinically evaluated and compared to the universal modeling framework. Our experimental results show that the PbFG raises the squared correlation coefficient (R2) between the NGM prototype and other invasive reference devices to 0.851 from 0.812. Following Clarke Error Grid Analysis (EGA), more than 97.9% of the measurement points are in region A (74.1%) and B (23.8%). In individual-customized modeling analysis, the PbFG can make the R2 reach to 0.9 after merely 30 times of calibration. Conclusions: Both the accuracy and the EGA experimental results indicate that our proposed PbFG solution can reduce the influence of individual differences to a certain extent, and improve the performance of NGM to a clinically acceptable level.

Original languageEnglish (US)
Pages (from-to)56-76
Number of pages21
JournalInformation Sciences
Volume497
DOIs
StatePublished - Sep 2019

All Science Journal Classification (ASJC) codes

  • Software
  • Control and Systems Engineering
  • Theoretical Computer Science
  • Computer Science Applications
  • Information Systems and Management
  • Artificial Intelligence

Fingerprint

Dive into the research topics of 'PbFG: Physique-based fuzzy granular modeling for non-invasive blood glucose monitoring'. Together they form a unique fingerprint.

Cite this