Phone: +34(93)413-7783
FAX: +34(93)413-7833
Email:
Angela@LSI.UPC.ES
As a consequence of the imprecision of knowledge available in biomedicine in general, one would expect that qualitative reasoning techniques, as they have been developed during recent decades by researchers working in the area of artificial intelligence, would be ideally suited to tackle problems stemming from biomedical fields. Yet, the application of artificial intelligence to biomedical sciences has not advanced rapidly in the past. Artificial intelligence techniques have been proliferated much more rapidly to and within other areas of science and technology.
Several difficulties inherent to the biomedical fields have restrained the progress in modeling and simulation of this type of systems in the past, problems that make these systems much more difficult to tackle than practically all other types of systems met anywhere in science and engineering.
The aim of the work developed in this dissertation was to address some of these difficulties, of soft sciences in general and of biomedical engineering in particular, and to come up with a methodology that would make optimal use of the limited knowledge available to the modeler, a modeling methodology that would not get confused by the inevitable incompleteness and even inconsistency of information generally available for these types of systems.
To this end, a qualitative modeling and simulation methodology called Fuzzy Inductive Reasoning (FIR) has been employed, a modeling technique of fairly recent vintage that looked promising for the task at hand, and for which a prototypical implementation was available to us. The methodology has been refined, and a second generation of FIR software was implemented that would allow us to work with biomedical and other soft science systems.
This thesis is structured in 10 chapters. The first chapter describes the motivation for this doctoral thesis research. It also discusses the selection of the methodology, and introduces the aims and scope of this dissertation. The state of the art of qualitative analysis applied to biomedical applications is presented in Chapter 2, providing an insight into the different qualitative methodologies used to this day for dealing with the problems inherent in biomedical applications.
Chapter 3 describes in detail the Fuzzy Inductive Reasoning methodology. The foundations of this methodology were laid by Prof. George Klir of the State University of New York (SUNY) at Binghamton. It has been further developed and brought to maturity in a joint effort by three Ph.D. students of the Universitat Politècnica de Catalunya under the guidance of Prof. François Cellier and Prof. Rafael Huber. The author of this dissertation was one among these three students.
The methodology used for mixed quantitative/qualitative modeling and simulation is presented in Chapter 4 of the dissertation, where a discussion of its usefulness for soft sciences is also given. This concept is described by means of a hydraulic control system example, where the validity of the proposed approach to mixed modeling and simulation is clearly demonstrated.
The next chapters are centered on biomedical systems, starting with Chapter 5 where the difficulties inherent in dealing with these types of systems are enumerated and explained in detail.
Chapter 6 focuses on the qualitative control of biomedical systems. Two vastly different types of biomedical system controllers are described: the control of the depth of anaesthesia and the central nervous system that controls the hemodynamical system. The former application represents a technical (external) controller of a biomedical system, whereas the latter represents an electro-chemical (internal) control mechanism built into a biomedical system. The purpose of the former application is the external control of a biomedical system, whereas the purpose of the latter is that of modeling a part of the central nervous system control, of understanding its functioning, and of predicting its future behavior. The difficulties encountered when modeling and simulating these two controllers are presented, and the solutions obtained by applying the fuzzy inductive reasoning approach to these systems demonstrate the strength of this methodology.
In Chapter 7, some limitations and weaknesses of the FIR methodology, as it had originally been devised, are presented. These problems are related to the limits of predictability of behavior in biomedical applications. To this end, a conceptual barrier limiting the predictability of future states of the systems under investigation is introduced. This conceptual barrier has been called causality horizon, and two systems, a linear state-space model and a biomedical system, served to demonstrate the concept.
Chapter 8 describes an enhancement of the FIR methodology for dealing with incomplete data records. To this end, a technique called missing data option is implemented. The missing data feature enables the researcher to work with incomplete data records and extract as much information from them as they contain. The feature makes it possible to convert incomplete quantitative data sets to reduced qualitative data sets in order to derive the best possible qualitative model for prediction of future system behavior.
Chapter 9 uses the missing data option to tackle one important problem, described already in Chapter 5, namely the elimination of patient-specific behavior. This work is a first attempt at tackling this difficult problem. The technique has been applied, until now, to a single anesthesiology system only. However, the results obtained are encouraging enough to, in a near future, apply this technique to other systems for which more patient data are available.
The research effort documented in this dissertation has focused on methodological issues. The main emphasis of the dissertation was to enhance the FIR methodology to be able to apply it to soft science systems in general and to biomedical science systems in particular. Therefore, the different biomedical applications presented in this doctoral thesis serve only to demonstrate the feasibility and validity of the chosen approach. The dissertation does not claim to have solved any medical problem in its full complexity. This task is left to future, more application-oriented, research projects that will need to be conducted in a joint effort of researchers with engineering and life science backgrounds.
An extensive performance comparison between the FIR methodology and other inductive modeling techniques, such as the neural network approach and the NARMAX method, is an important aspect of this thesis that confirms the capability of the FIR methodology to deal with this type of systems. This thesis demonstrates that the FIR methodology is indeed a very powerful tool for the identification of biomedical system models, and for predicting their future development.
Interested in reading the entire thesis? (256 pages, 3,389,170 bytes, pdf)