The use of cluster analysis in clinical chemical diagnosis of liver diseases.

Abstract

Diagnostic judgement is usually based on recognition of patterns. Unfortunately more than three quantitative data cannot be judged simultaneously without help of mathematical methods. Working on laboratory reports, a clinician usually goes linearly through the columns and reduces quantitative to qualitative data. Therefore the medical decision process should be improved if data reduction is performed with the aid of mathematical methods for pattern recognition. A total of 191 consecutive outpatients with a tentative or proven diagnosis of hepatobiliary disease were examined clinically, clinically chemically and partly histologically. Nineteen clinical chemical parameters were determined. Prior to pattern cognition, a principal component analysis was performed. Using six factors, accounting for 72.4% of total variance, cluster analysis was done, applying a hierarchical algorithm for ascertaining a starting partition, followed by the k-means algorithm. The validity of the solution was scrutinized, and a stable structure was found with nine clusters. Patients with a rejected suspect of liver disease were mainly located in clusters 1, 6 and 7. Cluster 1 also contains patients with compensated cirrhosis without inflammation, idiopathic hyperbilirubinaemia, focal nodular hyperplasia and haemangioma of the liver. In contrast, one third of cirrhoses, all with inflammatory activity were assigned to cluster 5. Patients with primary biliary disease were distributed among clusters 2, 3 and 4. All malignant neoplasias were assigned to cluster 9. More than 50% of fatty livers were classified to cluster 7. Cluster 2 and 8 contain only one patient with primary biliary cirrhosis (cluster 2) and fatty liver hepatitis (cluster 8). The follow-up of 66 patients also showed clinically meaningful changes of cluster assignment.