Diagnostics of pancreatic insufficiency using multivariate statistical and pattern recognition methods

Abstract

The complex evaluation of Lundh-test data was investigated in patients with pancreatic insufficiency and in controls using multivariate statistical methods. It has been shown that isolated parameters do not give satisfactory separation. The lower limits of the normal values were calculated in advance from the data of 77 control cases with the help of distribution analysis and of 2.5 percentile value for all parameters. On the basis of these normal values, and the results of other examinations, two groups were formed from the data of 137 more patients: (1) certainly pathologic patients, and (2) control cases. The application of multivariate statistical and pattern recognition methods considering the parameters simultaneously led to the correct diagnosis in 88-80-93% of the cases using the centroid, nearest neighbour and linear discriminant analysis, respectively. Linear discriminant analysis successfully separated controls from the patients with pancreatic insufficiency also in mild cases.