A Clinical Histories-Based ``Gravitational'' Decision Support System

Eugenio Roanes-Lozano (1), Basilio de la Torre (2,3), Luis M. Laita (4),
Eugenio Roanes-Macias (1)

(1) Algebra Dept., Universidad Complutense de Madrid
(2) Guadalajara Universitary Hospital 
(3) Surgery Dept., Universidad de Alcala
(4) Artificial Inteligence Dept., 
    Universidad Politecnica de Madrid

Rule based expert systems (RBES) in Medicine are frequently oriented to medical
diagnosis and are usually based upon the intrinsic knowledge of a panel of
experts.  Meanwhile, statistical methods are usually applied to experimentally
support or reject a theory or hypothesis.

We have found a need for a simple decision support system that could try to
predict the result of a certain treatment or surgery, based on a set of
clinical histories collecting data from different variables of the patients and
including a rating of the result of such surgery or treatment.  We suppose that
these variables are reasonably (or statistically) correlated to the ``result''
variable.

We have considered a ``gravitational'' distance-depending model.  It is similar
to the k-nearest neighbors algorithm for k = ``size of the sample'', but
instead of using simple majority within the neighbors to decide, all initial
data influence the value predicted (conversely to the square of its distance to
the new point).

Unlike in an expert system, there is no knowledge base or inference engine.
They are substituted by a data set and a statistical-algebraic algorithm.  The
system automatically checks the number of patients in the sample and the number
of variables of the patients in the clinical histories (i.e., the dimension of
the space).

As the systems handles lists, summations, plots... a computer algebra system
(Maple) has been chosen to implement it (for the authors' convenience).

This work is partially supported by the research project MTM2004-03175
(Ministerio de Educacion y Ciencia, Spain.)