Computer Aided Diagnosis of Jaundice. A Comparison of Two Data Bases

Abstract

A computer system for probabilistic diagnosis of jaundice was tested on a patient sample from a geographical area different from that for which it was first constructed. 144 consecutive patients with jaundice seen in two Stockholm hospitals were interviewed and examined to record a total of 82 indicants from history, demographic details, physical findings and laboratory tests. Data were compared with those of 319 jaundiced patients previously interviewed and examinated at different London hospitals. It was found that disease incidences were different in the two patient samples. There were more patients with acute viral hepatitis, chronic active hepatitis and primary biliary cirrhosis in : he London data base whereas the Stockholm data base included significantly more patients with Gilbert's syndrome and alcoholic cirrhosis. Indicant frequencies, standardised for disease incidence, differed with respect to age (Stockholm patients were on average six years older), time from onset of first symptom to hospital admission (Stockholm patients had on average a two-week shorter history of disease) and a number of symptoms such as nausea, vomiting, anorexia, weight loss, itching, pale stools and dark urine which were more frequent among the London patients. Differences in hospital admission policy was regarded as an important reason for the differences in indicant frequency.

The results of probabilistic diagnosis were poor. Only 49% of the cases were correctly classified into twelve diagnostic groups. In particular the computer model was poor at separating different causes of malignant bile duct obstruction and at differentiating between malignant and benign bile duct obstruction. However, all cases of acute viral hepatitis were correctly classified and the computer model was 87% accurate in differentiating between medical and surgical jaundice. Reclassification of the 144 patients on their own data showed the computer system to be well calibrated and 97% of the cases were correctly classified according to this procedure. In conclusion, the computer system could not be directly transferred for use in a Swedish hospital but the results of reclassification were sufficiently encouraging to warrant prospective studies.

Key Words:

• Bile duct obstruction
• computer-aided diagnosis
• differential diagnosis
• liver disease
• mathematical model
• medical history
• physical examination
• physical signs
• symptoms