A modified approach to objective surface generation within the Gauss-Newton parameter identification to ignore outlier data points. (September 2016)

• Record Type: Journal Article
• Authors: Gray, Rebecca A.L.; Docherty, Paul D.; Fisk, Liam M.; Murray, Rua
• Is Part Of: Biomedical signal processing and control. Volume 30(2016)
• Page Start: 162
• Page End: 169
• Publication Date: 2016-09
• Subjects: Parameter identification -- Glycemic modelling -- Gradient descent methods -- Gauss-Newton methods -- Outlier data -- Three sigma; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2016.06.009
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The Gauss Newton parameter identification method was adapted to ignore outlier data. The method is capable of determining outliers as a function of measurement variance. The method was successfully tested in two models describing glycaemic dynamics. Method application is not complicated and could thus be used in numerous scenarios. Abstract: The Gauss-Newton method is a simple iterative gradient descent method used to modify a mathematical model by minimising the least-squares residuals between the modelled response, and some observed behaviour. A common issue for parameter identification methods that optimise least-square residuals is the sporadic occurrence of outlying data in the observation data set. This research proposes an amendment to the Gauss-Newton parameter identification approach that limits the influence of outlying data by dissipating the contribution of outlying data to the objective function that drives iterations. The modified method was tested in two and three-dimensional parameter identification exercises using virtual data from the dynamic insulin sensitivity and secretion test (DISST). The data incorporated random normally distributed noise (CV = 3%) or random normally distributed noise in concert with an outlying data point. The proposed method performed similarly to the original method when no outlying data was included and found the model that fit accurately to the majority of data points when an outlying data point was present. The proposed approach provides a valuable tool for the rejection of outlier data that is operator independent, does not require multiple stages of analysis, or manual removal of data.