ABSTRACT
In Chapter 8 we developed the basic techniques required for statistical diagnosis in the simple standard situation where there is: a single clinic; complete, precise though possibly mixed feature vector v; disease type ascertained with certainty and a well-defined relationship between selection of past cases and referral of a new patient. In this chapter we consider realistic non-standard situations in diagnosis which can be regarded as extensions or deviations from the above. We consider the following situations. Diagnostic system transfer: a calibration problem. Under what circumstances is it valid to use a diagnostic data set from one clinic to assess new undiagnosed cases in another? Clinic amalgamation: a calibration problem. If diagnostic data are available from more than one clinic on a given set of diseases, how can they be combined to produce a more efficient diagnostic system, taking into account the interplay of referral and selection? In relation to this and the preceding question, how can we cope with features which are either measured in one clinic but not in the other, or features which are not identical but related, for example through calibration by an assay technique? Imprecision in the feature vector. For a single clinic how do we deal with imprecise features in v? Missing features. For a situation with a single clinic how do we deal with missing values in some feature vectors? Uncertainty in ascertainment of type. How do we arrive at diagnostic techniques in a clinic where the typings are uncertain and all the diagnostic information we have about the type u of a case is a composite diagnosis giving pr(u = j) (j = 1, . . . , k)?
