MetaNet,
A Computer Program
for Neural Network Aided
Diagnosis of Inherited
Metabolic Diseases
J. Pepper, ServiceWare Inc., Pittsburgh,
Principal Investigator
C. E. Wyatt, Applied Analytic Systems Inc.,
Pittsburgh, PA, MetaNet Technical Developer
D.C. Lehotay and J.T.R. Clarke, Medical Consultants, The University
of Toronto Hospital for Sick Children, Ontario,
Canada
We have developed a
prototype computer program, MetaNet, that uses a
combination of artificial neural networks and
knowledge-based expert systems to assist in the
diagnosis of inborn errors of metabolism in children.
Results of amino
acid analysis data of normal children, and of
patients diagnosed with a number of amino acid and
organic acid abnormalities were used as inputs to
train the neural network component of the program. To
diagnose new cases, plasma or urinary amino acid
results are entered. The knowledge-based expert
system then asks questions of the user regarding the
presence or absence of common clinical and/or
biochemical abnormalities.
Using both the amino
acid data and the answers to the questions, the
MetaNet program integrates the output of the neural
network and the results of the knowledge-based expert
system to yield a provisional diagnosis.
The diagnostic
output is accompanied by a numerical *belief vector*,
which indicates the degree of confidence of the
program in the diagnosis. Altering any of the input
variables followed by reprocessing of the data
generates a new diagnostic output and a revised
belief vector. This allows analysis of the importance
of any input variable to the proposed diagnosis. The
knowledge-based expert system also includes a section
entitled *Independent Metabolic Disease Reference
Documents*, which provides additional information
about a suspected metabolic disease when requested by
the user. The neural network component consists of
eight, three-layer neural networks that are trained
using a back-propagation approach. Analysis of the
hidden layers following training of the neural
network revealed both expected and novel, unexpected
connections between specific diagnoses and clusters
of amino acids. Such data may be used as a guide for
future investigation of the contribution of the
metabolism of specific amino acids to amino acid
disorders.
The program runs
under Windows 3.1 or Windows 95, and promises to be
useful both as a model for computer assisted
diagnosis of inborn errors and as a research tool.
Additional...
MetaNet Screen
Shot.
|
