\documentclass[fleqn,twocolumn]{article}

\usepackage{aime03}
\usepackage{psfig}

\title{\bf Prognostic Models in Medicine: \\
       Artificial Intelligence and Decision Analytic Approaches}

\author{Peter Lucas \\
        Department of Computer Science \\
        Utrecht University, PO Box 80089 \\ 3508~TB Utrecht,
        The Netherlands \\
        E-mail: lucas@cs.uu.nl
        \And
        Ameen Abu-Hanna \\
        Department of Medical Informatics \\
        AMC-UvA, Meibergdreef 15 \\ 1105~AZ Amsterdam,
        The Netherlands \\
        E-mail: a.abu-hanna@amc.uva.nl}
\date{}

\begin{document}
\maketitle

\begin{abstract}
\noindent
This paper is meant as an introduction to
the workshop on \emph{Prognostic Models in
  Medicine: Artificial Intelligence and Decision Analytic Approaches}
held during {\sc aimdm'99}. Prognosis -- the prediction of the course
and outcome of disease processes, either or not changed due to
interventions -- is an important aspect of medical tasks like
diagnosis and treatment management. Techniques for building prognostic
models vary from traditional probabilistic approaches, originating
from the field of statistics, as used in decision analysis, to more
qualitative and model-based approaches originating from the field of
artificial intelligence. The workshop brings these two fields of
research together in the hope that a fruitful exchange in ideas will
take place.
\end{abstract}

\section{Introduction}

Prognosis, the prediction of the course and outcome of disease, is a
subject that lies at the heart of patient management. There is little
sense in delving into the cause of particular symptoms and signs in a
patient, and to initiate elaborate diagnostic procedures, if it is
known beforehand that no effective treatment of the considered disease
exists.  Furthermore, also treatment selection invariably involves
taking possible future beneficial and harmful effects into account,
i.e.\ prognostic information \cite{lucas99}.

Of course, the process of patient management concerns issues other
than prognosis as well. The primary role of the physician is to guide
the patient through the disease process, which involves much more than
prognostication. Even the processes of diagnosis and treatment
selection may be seen in this light of guidance of patients.  This
view may explain why prognosis, despite its central role in medicine,
is not clearly recognised as such in typical medical textbooks, like
\emph{Harrison's Principles of Internal Medicine} \cite{harrisons}.
The subject of prognosis is only paid attention to when it
is obviously important, such as in cancer treatment.

It is likely that this situation will change in the near future, and
that the role of prognostic models in medicine will increase.
Medicine as a field is becoming increasingly complex, as is reflected
by the annually increasing number of different diagnostic tests and
therapies from which a clinician must choose.  Prognostic models are
required to guide clinicians in this selection process to ensure that
the patient will benefit from further progress in medical science.

\section{Prognostic models}

As has been said above, there are a number of fields in medicine, where
prognostic models are of particular importance. Examples of such
fields are: oncology, transplantation medicine, and trauma medicine.
Usually, prognostic models focus either on long-term or short-term
effects. For example, long-term effects dominate in treatment
considerations in oncology, whereas short-term effects are more
significant in trauma medicine.  Adequate prognostic information is of
major importance in these fields so that prognostic models of various
kinds, not necessarily mathematical in nature, have been in use for
quite some time.  Often these models are coarse and lack detail. The TNM
staging system that is used to assess a primary malignant tumour in
terms of its size (indicated by $\mbox{T}_0$ to
$\mbox{T}_4$, where an increase in subscript corresponds to an
increase in tumour size, as defined for a particular type of tumour),
regional lymph node involvement (N, also supplied with a subscript),
and presence of distant metastatis (M) is an example of a simple
qualitative tool to assess prognosis in cancer patients. Another
example is the Apache III scoring system, which is based on a logistic
regression model, and that has been shown to have a good predictive
ability for patients with severe illness, and for a large variety of
diseases \cite{knaus91}.

As one may expect, clinicians are
only prepared to accept prognostic models when it is obviously
that they will contribute to quality of care~\cite{wyatt95}.
Prognostic models are not only used in a clinical setting. They are
also used, and may even have had a larger impact, in the design of
clinical trials, counselling patients and in medical technology
assessment.

In general, and independent of particular applications of prognostic
models, the problem of the design of accurate prognostic models is the
capturing of the many possible subtle interactions among variables
that exist. It is largely determined by the (mathematical) modelling
tools used to what extent such interactions can be represented, and
learnt from data, possibly augmented with background knowledge.

\section{Artificial intelligence and decision analysis}

Medical artificial intelligence is generally concerned with the
development of medical models for various purposes, but usually the
aim is to assist clinicians in the processes of diagnosis, treatment
or prognosis of diseases in patients. A key characteristic is the
\emph{explicit} representation of the medical knowledge involved,
i.e.\ the explicit representation of meaningful interactions among the
factors that play a role in a particular medical problem is
favoured~\cite{lucas95}. However, there are a number fields in
artificial intelligence, such as neural networks, where the goal of
explicit representation is less dominant. There is now an entire array
of different techniques from which medical AI practitioners may
choose.  One of the difficult problems has been the representation of
temporal patterns, which is now addressed by a number of different
formalisms.  Progress in the field has yielded new, flexible
techniques, like Bayesian networks, neural networks and genetic
algorithms; these offer new opportunities for dealing with the issue of
prognostication.

Medical decision analysis offers a systematic approach to medical
decision making under conditions of uncertainty
\cite{sox88,weinstein80}.  It has studied the use of prognostic models
in the process of decision making for more than two decades. An
enormous amount of practical experience in building medical models has
been built up during these years. However, there has been little
progress in the field with respect to new techniques and tools that
may be used to carry out a decision analysis.

Until recently the fields of medical artificial intelligence and
decision analysis appeared to have only in common that in both model
building is of crucial importance.  In the field of artificial
intelligence there has been a revival of interest in numerical methods
stemming from probability and decision theory, and from the field of
neural networks. The new ideas and techniques that have come out of
this, has not passed by unnoticed by the medical decision analysis
community.  There currently seem to be much interest in that field
with respect to applicability of these technique. At the same time,
medical artificial-intelligence researchers realise that much can be
learnt from the more mature field of medical decision analysis.  This
workshop is therefore a timely opportunity to exchange ideas and
hopefully to learn from each other.

\section{Road-map to the workshop papers}

To conclude this introductory paper, we shall briefly summarise
the contents of the papers in the working notes.

The paper by S.S. Anand, P.W. Hamilton, J.G. Hughes and D.A. Bell,
titled \emph{Utilising censored neighbours in prognostication},
discusses an extended version of the $k$-nearest neighbour algorithm,
which is applied to the problem of prediction of the survival of
patients with colorectal cancer. Novel is the possibility of dealing
with censored patients, which is typically required in survival
analysis in medicine.  The paper by S. Antel, L.M. Li, F. Cendes, Z.
Caramanos, A.  Olivier, F. Andermann, F. Dubeau, R.E. Kearney, R.
Shinghai and D.L.  Arnold, with title \emph{A naive Bayesian
  classifier for the prediction of surgical outcome in patients with
  temporal lobe epilepsy}, focusses on a number of important issues
that arise when one wants to develop prognostic models the are
clinically useful.  The development of a Bayesian classifier for the
prediction of the outcome of patients with temporal lobe epilepsy that
undergo surgery is reported. Bayesian classifiers are also the topic
of the paper \emph{Robust outcome prediction for intensive-care
  patients} by M.  Ramoni, P.  Sebastian and R. Dybowski, but here the
main issue is how to deal with missing values in clinical data. A
comparison is made between logistic regresssion augmented with an
imputation mechanism and what is called a \emph{robust} Bayesian
classifier in which no assumptions are made with respect to the
mechanisms underlying missing data.

In the paper by H. Dreau, I. Colombet, P. Degoulet, G. Chatellieri,
titled \emph{Identification of patients at high cardiovascular risk
  using a critical appraisal of statistical risk prediction models}
not techniques, but different statistical risk-prediction models are
compared. This paper sheds light on the assumptions underlying
statistical models, and on the question to which extent assumptions
are valid and may affect the conclusions that may be drawn.

There are a number of papers in which statistical or decision-analytic
techniques are compared or combined with AI techniques.  For example
in the paper by L. Ohno-Machado and S. Vinterbo, \emph{Influential
  case detection in medical prognosis} it is studied whether a genetic
algorithm offers advantages over conventional techniques for the
selection of cases in the construction of prognostic logistic
regression models. The prediction of the prognosis of trauma patients
has been chosen as an example domain.  I. Zeli\v{c}, N. Lavra\v{c}, P.
Najdenov and Z. Rener-Prime in their paper \emph{Impact of machine
  learning to the diagnosis and prognosis of first cerebral paroxysm}
compare ID3-like decision-tree induction with naive Bayesian
classifiers from the perspective of machine learning. The comparison
is carried out in the medical domain of epilepsy.

The remaining two papers focus on medical applications of techniques
from the areas of artificial intelligence.  In the paper by N. Peek,
\emph{A specialised POMDP form and algorithm for clinical patient
  management} the formalism of partially observable Markov decision
problems (POMDPs) is studied.  This formalism has originally been
introduced in artificial intelligence as a means to handle planning
problems under conditions of uncertainty.  POMDPs, however, have also
been suggested as a suitable formalism for medical treatment planning.
Since the formalism is known to be intractable in general, this papers
proposes the use of Monte Carlo simulation to render the formalism
practically more useful. The paper by R.  Schmidt, B.  Pollwein and L.
Gierl, titled \emph{Prognoses for multiparametric time course of the
  kidney function} also discusses the suitability of a technique from
the field of artificial intelligence to the development of prognostic
models, namely the application of case-based reasoning to the
prediction of kidney function. Advantages and limitations of
case-based reasoning are clearly discussed.

We may conclude that the papers in the workshop, although all dealing
with the issue of prognostic mocels, are indeed varied; both methods
and techniques from the fields of artificial intelligence, decision
analysis and statistics are covered by the papers. Sometimes these
techniques are dealt with separately, sometimes they are combined and
in some papers they are compared to each other. It may therefore be
concluded that the title of the workshop does indeed reflect
the contents of the papers in the workshop notes.

\begin{thebibliography}{99}

\bibitem {harrisons}
K.J. Isselbacher, et al., \emph{Harrison's Principles of Internal
Medicine}, 13th edition, McGraw-Hill, New-York, 1994.

\bibitem{knaus91}
W.A. Knaus, E.A. Draper, J. Lynn, Short-term morbidity
predictions for critically ill hospitalised patients -- science
and ethics, \emph{Science} 254 (1991) 389--94.

\bibitem{lucas95}
P.J.F. Lucas, Logic engineering in medicine,
The Knowledge Engineering Review, Vol.\ 10, No.\ 2, 1995,
pp.\ 153--179.

\bibitem{lucas99} 
P.J.F. Lucas and A. Abu-Hanna, Prognostic methods in
medicine.  \emph{Artificial Intelligence in Medicine}, vol.\ 15,
1999, pp.\ 105--119.

\bibitem{sox88}
H.C. Sox, M.A. Blatt, M.C. Higgins, K.I. Marton,
\emph{Medical Decision Making},
Butterworths, Boston, 1988.

\bibitem{weinstein80}
M.C. Weinstein and H.V. Fineberg,
\emph{Clinical Decision Analysis},
W.B. Saunders, Philadelphia, 1980.

\bibitem{wyatt95}
J.C. Wyatt and D.G. Altman, Commentary -- prognostic models: 
clinically useful or quickly forgotten, BMJ, Vol.\ 311, 1995,
pp.\ 1539--1541.

\end{thebibliography}

\newpage
\mbox{}
\newpage
\mbox{}

\end{document}