Diagnosis. II. Diagnostic models based on attribute clusters: A proposal and comparisons

Abstract

A few patients had class "4" history grades but no positive radio-allergosorbent test results. This may indicate an allergen yet to be discovered-possibly an occupational one-a non-IgE-mediated mechanism or perhaps an exceptional interference in the radio-allergosorbent test by antibodies of another class. Whatever the reason, in such cases referral to an allergy specialist would be recommended. Indeed, referral should be considered when either the patient's symptoms are not fully explained by the report or advice on management is required. We conclude that microcomputers are now powerful enough to operate a complex self-learning data base that may be used (a) for taking and analysing a standard history, (b) predicting the likely causes of symptoms, and (c) linking the results of IgE tests to the patient's history so that a valuable interpretation of the patient's symptoms can be used by the general practitioner as a basis for referral to an allergy specialist. This work would have been impossible without either the assistance of Josephine Merrett in processing the IgE tests or the support of the Post Office and Civil Service Sanatorium Society. Requests for reprints should be addressed to T 0 Merrett, RAST Allergy Unit, Benenden Chest Hospital, Cranbrook, Kent TN17 4AX References 'Ishizaka K, Ishizaka T, Hornbrook MH. Physico-chemical properties in reaginic antibody. J Immunol 1966;97 :840-53. 2 Johansson SGO. Raised levels of new immunoglobulin class (IgND) in asthma. Lancet 1967;ii:951-3. 3 Eriksson NE. Diagnosis of reaginic allergy with house dust, animal dander and allergens in adult patients. III. Case histories and combinations of case histories, skin tests, and RAST compared with provocation tests. Abstract Senior neuroradiologists or radiologists of 42 hospitals with computed tomography available for NHS patients in England and Wales were contacted by postal questionnaire about the use of this facility in the management of patients with acute head injuries. Replies were obtained from 39 hospitals. Requests for computed tomography from general surgeons or physicians and staff of accident and emergency departments received positive responses for scanning with only half to three-quarters the frequency of responses to requests from neurosurgeons. Continuous computed tomography facilities were available generally to neurosurgeons. The combined effect of partial responses to requests and the availability of the computed tomography service meant that only 44% of hospitals gave a continuous service for general surgeons or physicians. The percentage of hospitals giving a continuous service to accident and emergency departments was 54%. It appeared that computed tomography scanning was being used most often as a diagnostic/

Bayes Theorem and Conditional Dependence of Symptoms: Different Models Applied to Data of Upper Gastrointestinal Bleeding

“…Missing values were handled in the same way as in the independent Bayes model. Another way to account for interactions between variables in Bayes' theorem is the combination of Independence Bayes with cluster analysis [19]. It is the aim of this approach to form clusters of variables so that variables are conditionally dependent within a cluster and conditionally independent between clusters.…”

“…This model is the straightforward incorporation of pairwise interactions into the independent Bayes model. This approach is encouraging for many reasons: It is easy to understand by the clinician and simple to perform on a personal computer; it is theoretically more justified than Independence Bayes, and there have been good results in the few studies in the literature [16,18,19]. However, most of these promising models have not yet been compared with each other and with regard to standard popular models (Independence Bayes, linear logistic regression) using adequate methodology.…”

A New Procedure for Analysis of Medical Classification

“…Because of these issues we proposed the GOM ll:10del-a model which differed in disease probability func-t~on from the posterior probability forms of logistic regres-SIon and cluster analysis. Specifically, we may relate GOM to logistic regression and Bayesian and linear discriminant analysis by defining a criterion event variable Yi and a set of predictor variables :lSi' In logistic regression the relation of Yi and :lSi may be expressed as log (8) where the regression weights in b T and the statistics ofthose weights are the desired quantities. In clinical settings :lSi is typically known and the true Yi is typically not (although a post .hoc groupin.p is often sub~ti.tuted for yJ In this case the predIctor Y i = b :lSi, (the logIstic score) performs the same function as the discriminant function score.…”

“…In a similar fashion, the Bayesian discriminant analysis is concerned with the analysis of a criterion variable Yi and a set of predictor variables :lSi. In this case, however, the :lSi are viewed as being drawn from a set of populations or classes where 10g(Plje/Paje) is analogous to b in (8) and log(l'tl/Jto) is analogous to the first term in (8). Thus, the same relation holds between GaM and Bayesian categorical discriminant analysis as between Fisherian discriminant analysis and logistic regression.…”

The Relative Accuracy of a Variety of Medical Diagnostic Programs

“…Experience suggests that the strong statistical assumption of conditional independence is invalid in many medical applications; some symptoms and signs do tend to eclipse others (e. g., [11]). Indeed, numerous studies have shown the presence of interactions in a variety of applications [12][13][14]. However, it does not necessarily follow that classifiers which take these interactions into account are more accurate.…”