Identification of Nonlinear Physiological Systems

David T Westwick and Robert E Kearney
(IEEE Press 2003 ISBN 0-471-27456-9)


This book has been produced as a successor to the classic text by Marmarelis and Marmarelis on the theme of Advanced Methods in Physiological Modeling: the White Noise Approach, which was published in 1978. Since that time, much has occurred in relation to the field of nonlinear identification particularly as a consequence of advances in computational methods.

Before commenting further, however, a caveat is needed in relation to the title of this work by Westwick and Kearney. The focus of this new text, as was the case with that of Marmarelis and Marmarelis, is on signals which are to be interpreted in terms of black box models, for instance those occurring in electrophysiology. It does not specifically address the problem of identifying systems represented by physico-chemical models, for example metabolic, endocrine and physiological organ systems; models in which the parameters correspond explicitly to physical or chemical properties of the dynamic processes being considered. There is a large distinctive literature for this separate class of problem. Readers wishing to pursue this field would be well advised to consult the writings of long-standing experts such Cobelli, DiStefano and Godfrey; leading-edge researchers of today, just as they were thirty years ago.

Returning to this new work by Westwick and Kearney, the authors assume that the readership will have a basic grounding in linear signals and systems. Background material beyond this level is summarised and there is an excellent set of references to further reading. Each chapter includes both a number of simple problems as well as more extensive computational exercises, designed to be tackled using MATLAB and the nonlinear system identification toolbox (NLID) developed by the authors. The style of writing is very accessible with many illustrations drawn from physiological systems.

After some relevant background material, there are basic chapters on models of linear and nonlinear systems, with the latter embracing Volterra and Wiener series, block structures, parallel cascades and the Wiener-Bose model. This is followed by a chapter on the identification of linear systems. This is designed to lay the foundations for the more elaborate treatment of the nonlinear identification problem, rather than to provide a comprehensive review of all available methods.

There then follows the final three substantive chapters which focus on issues and techniques relating to nonlinear identification. The first of these deals with correlation methods. Although superior methods are now largely available, they still find widespread adoption. Moreover, an appreciation of them is helpful as a precursor to the discussion of more recent approaches.

This is followed by a treatment of explicit least-squares methods. These include orthogonal algorithms, expansion bases and principal dynamic modes. These methods are appropriate where model structures are linear in their parameters, such as Wiener or Volterra series, and offer substantial improvements in model accuracy as compared to correlation methods. However, they are computationally expensive and, of course, are limited to cases where linearity in the parameters exists.

These limitations provide a logical lead in to the final chapter which features iterative least-squares methods. These can be used to tackle the real nonlinear problem, namely that where models have structures whose outputs are nonlinear functions of some or all of their parameters. The chapter begins with treatments of the full range of gradient-based optimisation methods. This is followed by consideration of the more recent parallel cascade methods. Full details of the mathematical treatments are presented in a lucid manner and the chapter ends with a case study demonstrating the power of such methods when applied to a model of the visual processing occurring in the light-adapted fly retina.

As one who many years ago reviewed that earlier text by Marmarelis and Marmarelis, I am happy to declare that this new work by Westwick and Kearney is indeed a worthy successor. The authors communicate their ideas in a clear and logical manner, covering both the underlying principles and the computational issues which need to be addressed in moving towards problem solution. It is a volume that deserves to be on the bookshelf of all researchers involved in this increasingly important interface between signals and systems theory on the one hand and physiology and medicine on the other.

Ewart Carson
February 2004
sd395@soi.city.ac.uk