The dimensional reduction method for identification of parameters that trade-off due to similar model roles

详细信息    查看全文

• 作者：Shaun M. Davidson^a ; ^{shaun.davidson@pg.canterbury.ac.nz} ; Paul D. Docherty^a ; ^{paul.docherty@canterbury.ac.nz} ; Rua Murray^b ; ^{rua.murray@canterbury.ac.nz}
• 关键词：Inverse problems ; System identification ; Parameter trade-off ; Hyperplane ; Gradient descent
• 刊名：Mathematical Biosciences
• 出版年：2017
• 出版时间：March 2017
• 年：2017
• 卷：285
• 期：Complete
• 页码：119-127
• 全文大小：2003 K
• 卷排序：285

文摘

Parameter identification is an important and widely used process across the field of biomedical engineering. However, it is susceptible to a number of potential difficulties, such as parameter trade-off, causing premature convergence at non-optimal parameter values. The proposed Dimensional Reduction Method (DRM) addresses this issue by iteratively reducing the dimension of hyperplanes where trade off occurs, and running subsequent identification processes within these hyperplanes. The DRM was validated using clinical data to optimize 4 parameters of the widely used Bergman Minimal Model of glucose and insulin kinetics, as well as in-silico data to optimize 5 parameters of the Pulmonary Recruitment (PR) Model. Results were compared with the popular Levenberg–Marquardt (LMQ) Algorithm using a Monte-Carlo methodology, with both methods afforded equivalent computational resources. The DRM converged to a lower or equal residual value in all tests run using the Bergman Minimal Model and actual patient data. For the PR model, the DRM attained significantly lower overall median parameter error values and lower residuals in the vast majority of tests. This shows the DRM has potential to provide better resolution of optimum parameter values for the variety of biomedical models in which significant levels of parameter trade-off occur.