Abstract
Discriminant analysis techniques were used to predict the histopathological findings
in liver biopsy specimens in asymptomatic patients with slightly to moderately raised
routine liver tests. Moderate to severe fibrosis and/or inflammation were treated
as indication for biopsy. Two methods were used to classify patients. One was the
dichotomous discrimination between “biopsy necessary” or “biopsy not necessary” groups
of patients. The other involved combining two discriminant functions trained separately
for recognition of fibrosis or inflammation, and then combined to predict the biopsy
necessity. Detection of outliers by standard techniques, directly available in the
SPSS-X package, was performed before starting discrimination procedures. Both “sharp”
assignment rules and continuous scoring rules were applied to the classification problem.
The correct classification rate reached over 85% for the algorithms tested. In the
majority of cases the classification was found to be “non-doubtful”. Elimination of
outliers (especially by standardized residuals) improved the global correct classification
rate, but only slightly improved assignment to the “biopsy necessary” group. Routine
and complementary laboratory findings were found to be the most discriminating; answers
to questionnaire and ultrasound examination were less important. Selection of the
most diagnostic features based on “clean” data without outliers enabled us to find
interesting medical associations, which were previously masked by extremely asymptomatic
values outlying from the main body of the “biopsy necessary” group.
Key-Words
Discriminant Analysis - Variable Selection - Outliers - Prediction - Liver Diseases
