Seite - 28 - in Applied Interdisciplinary Theory in Health Informatics - Knowledge Base for Practitioners

per Kelvin. However, although it would be wrong to say that Shannon entropy is “the same thing” as “entropy”, it would be equally wrong to say they are unrelated: the two equations only differ by a constant (which defines the scale of measurement), and one can begin to reconcile the two if one relates the probabilities of the microstates of the system under consideration with the probabilities of the symbols generated by that system. Indeed, Jaynes argued in depth that the information theoretic view of entropy was a generalisation of thermodynamic entropy [3][4]. We implicitly advocate the same position in the context of medical diagnosis. Going back to our document example. If we take a new document, pick a character at random and that character turns out to be a “z”, a character with one of the lowest probabilities of occurrence in a typical English document, then that is providing us with more information about it (relative to a “normal” document) than if we had received an “e”. Two general properties are also worth noting. Firstly, if only one outcome in an ensemble M has a non-zero probability of occurring (in which case, its probability must be 1), then: Property 1: H(M) = 0 (By convention, if p(mk) = 0, then Ͳൈ݈݋ ݃ଶͲؠͲ). At the other end of the scale, the H(M) is maximized if all of the outcomes are equally likely. An expression for the value for this is quite easy to derive. Let our ensemble Me have K possible outcomes. Then we must have for all k, p(mk) = 1/K. Substituting this into Equation 2, we get: ܪሺܯ௘ ሻൌെ෍ ͳ ܭ ݈݋ ݃ଶ ͳ ܭ ௄ ௞ ୀଵ ൌ ͳ ܭ ݈݋ ݃ଶሺܭሻ෍ ͳൌ݈݋ ݃ଶሺܭሻ ௄ ௞ ୀଵ (Noting that log(1/K) = -log(K) and that log(K) is a constant and so can be factored to the outside of the summation). So, Property 2: H(Me) = log2(K) if all K outcomes are equally likely In the case of our English document example, if the characters were uniformly distributed, then we would have H(Muniform) = log2(27) = 4.76 bits. This is slightly higher than that for our representative English language document (4.1 bits). Returning to the application of this to medical diagnosis, we can interpret these two situations as follows: H( ) = 0 if only one message/positive test result is possible. That is, a specific diagnosis has been confirmed. H is at its maximum when all messages are equally possible. That is, we are at a state of complete ignorance about the patient’s internal state. From this we can see that the challenge of diagnosis is to reduce the entropy to as close to zero as possible, and to select tests so that the result of each test (what we are calling “messages” here) maximises the reduction of entropy. Two points should be emphasised here before we move on: 1. We are equating the probability of occurrence of messages with the probability of microstates of the patient under examination, to justify the usage of the term “entropy”; P.Krause / InformationTheoryandMedicalDecisionMaking28