CONSTRUCTION OF A MODEL FOR COMPUTER-ASSISTED DIAGNOSIS: APPLICATION TO THE PROBLEM OF NON-TOXIC GOITRE

Abstract

This study describes the construction of a model for the computer-assisted diagnoses of non-toxic goitre. From a study of 53 patients with Hashimoto's disease, 51 with simple goitre and 51 with thyroid cancer, information was obtained which was used in the construction of a probability matrix. The diagnosis was confirmed histologically in all of these patients. The probability matrix was constructed from 30 pieces of information relating to the results of history taking, clinical examination, and laboratory investigations and consisted of a table of the observed incidence of each piece of information in each of the three diseases. This probability matrix was then fed into an Elliott 803 automatic digital computer and formed the memory of the computer for these three diseases. A fresh series of 88 patients, 43 with Hashimoto's disease, 26 with simple goitre, and 19 with thyroid cancer were then studied. Clinical information and the results of laboratory studies were provided as data to the computer to calculate a diagnosis for these patients. Two slightly different applications of probability theory were used for each calculation. In one (Bayesian probability) the ratio of the prior probabilities that any patient would have Hashimoto's disease, simple goitre, or thyroid cancer before any tests or observations had been made, were taken to be 10:89 and 1 respectively on the basis of observations previously made on a population of patients attending a thyroid clinic. In the other application of probability theory (relative likelihood) the prior probabilities of occurrence of the three diseases were assumed to be equal. The diagnoses given by both methods were compared with each other and with the diagnosis of a clinician experienced in dealing with thyroid disorders. In all patients in whom discordence of opinion occurred a histological diagnosis was accepted as the final diagnosis. The method using relative likelihood was considered to be superior to the method using Bayesian probability for in no patient did relative likelihood miss a diagnosis of thyroid cancer when the clinician diagnosed thyroid cancer. This was not true for Bayesian probability which missed three cases of clinically obvious cancer. The possible reasons for this superiority of relative likelihood are discussed in the paper. The results given by relative likelihood agreed with the clinician's correct diagnosis of Hashimoto's disease in 26 of 28 patients. The calculated diagnosis was wrong in two of these patients in whom the clinician made the correct diagnosis. In another 12 patients in whom the clinician made a wrong diagnosis, the calculations based on relative likelihood gave the correct diagnosis. In a further three patients the clinical and calculated diagnoses were both wrong. In 24 of the 26 patients with simple goitre the clinician made the correct diagnosis. In 21 of these patients the diagnosis calculated by relative likelihood was correct and in three patients it was wrong. In two patients both the calculated and clinical diagnoses were wrong. Of the 19 patients with thyroid cancer 16 were correctly diagnosed by relative likelihood and by clinician alike. The remaining three patients in this group were incorrectly diagnosed by both. It is concluded that the results of this study support the contention that a correct diagnosis can be calculated using the theory of conditional probability. Many difficulties remain to be overcome in what still remains a highly experimental approach to the problem of diagnosis.