Abstract:
A two-part study was designed to test the hypothesis that sufficient information is
available from a modern hematology analyzer (the Coulter STKS) to reach a reliable
intermediate conclusion which can be used as input to the next decision-making level
in the design of a high-performance expert system for hematology diagnosis. In phase
one, we analyzed the performance of three probabilistic systems (using Bayes’ rule)
which interpret STKS data: a control system which took the traditional approach of
classifying cases into specific diagnoses, and two test systems which were designed
to reach only an intermediate conclusion but not a final diagnosis. One of the test
systems classified cases into “textbook categories” of disease and the other utilized
defined diagnostic patterns. The systems were tested with 150 cases. The pattern approach
ranked the correct choice first in 141 of 150 cases (94%). In phase two, we abandoned
Bayes’ rule, reformulated the pattern approach into a heuristic classification system,
and tested its reliability on 820 cases. The algorithm of the reformulated system
was able to classify all 820 cases into the same predominant pattern as a panel of
three experienced laboratory hematologists.
Keywords
Expert System - Hematology - Pattern Analysis - Bayesian Probability - Heuristic Classification
