## Info

Notes: The notation p(FMD) actually is shorthand for a probability of an event variable called FMD that can take the values true or false. So, p(FMD) actually stands for p(foot and mouth disease is present) or p(foot and mouth disease is not present). If we to write Table 13.1 out in its full form, it would have four rows corresponding to p(FMD is present) = 0.001, p(FMD is absent) = 0.999, p(MCD is present) = 0.001, p(MCD is absent) = 0.999.The laws of probability require that the sum of the probabilities for all the event variable outcomes will be equal to 1 (e.g., p[FMD is present] + p[FMD is absent] = 1, because it is a certainty that something is either present or it is absent), so knowledge engineers usually save space by only writing one row per disease. FMD, foot and mouth disease; MCD, mad cow disease.

Notes: The notation p(FMD) actually is shorthand for a probability of an event variable called FMD that can take the values true or false. So, p(FMD) actually stands for p(foot and mouth disease is present) or p(foot and mouth disease is not present). If we to write Table 13.1 out in its full form, it would have four rows corresponding to p(FMD is present) = 0.001, p(FMD is absent) = 0.999, p(MCD is present) = 0.001, p(MCD is absent) = 0.999.The laws of probability require that the sum of the probabilities for all the event variable outcomes will be equal to 1 (e.g., p[FMD is present] + p[FMD is absent] = 1, because it is a certainty that something is either present or it is absent), so knowledge engineers usually save space by only writing one row per disease. FMD, foot and mouth disease; MCD, mad cow disease.

3.3. How Diagnostic Expert Systems Work: Differential Diagnosis

Like the physician, the computer program generates a differential diagnosis for the sick individual. A differential diagnosis is a list of diseases that are most likely to account for the findings in a patient. The diagnostic expert system typically creates its differential diagnosis by computing the posterior probability for every disease given the findings. A disease's posterior probability is the chance that the patient has the disease, given the findings.

Diagnostic expert systems use Bayes rules to compute the differential diagnosis. The relevance of Bayes rules to medical diagnosis was first introduced theoretically by Ledley and Lusted (1959) and first used in a diagnostic expert system by Homer Warner in 1962. Developers of diagnostic expert systems continue to use the same methods as did Homer Warner, as well as more complex Bayesian methods (but the original technique generally works well).

3.4. How Diagnostic Expert Systems Work: Question Generation

A differential diagnosis must be "resolved"; that is, the diseases in the list must be ruled in or ruled out. Like an expert table 13.2 Conditional Probabilities for FMD and MCD

Finding

Disease p(FindinglDisease)a

Drooling of saliva present FMD present

Drooling of saliva present FMD absent

Drooling of saliva present MCD present

Drooling of saliva present MCD absent

More than one animal affected FMD present

More than one animal affected FMD absent

More than one animal affected MCD present

### More than one animal affected MCD absent

0.95 (sensitivity) 0.05 (1 - specificity) 0.001 (sensitivity) 0.05 (1 - specificity) 0.95 (sensitivity) 0.2 (1 - specificity) 0.001 (sensitivity) 0.2 (1 - specificity)

aSimilar to the previous table, the knowledge engineers have left out half of the combinations because they can be derived from those listed listed by subtraction from 1. For example, p(drooling of saliva is not presentlFMD is present)= 1 - p(drooling of saliva is presentlFMD is present) = 1 - 0.95 = 0.05. FMD, foot and mouth disease; MCD, mad cow disease. Probabilities from BOVID, a cattle diagnostic program. Courtesy of Animal Information Management Pty. Ltd, Victoria, Australia, BOVID, a cattle diagnostic program.

physician, the diagnostic expert system engages in a cyclic process often referred to as "hypothesize and test" to resolve the differential diagnosis. A diagnostic expert system uses value-of-information calculations (Weinstein, 1980) to recommend to the physician user additional findings that will resolve the differential diagnosis efficiently. Value-of-information calculations identify those findings, which, if known to be present or absent, optimally discriminate among the diseases in the differential diagnosis, where optimality takes into account not only the probability of a diagnosis, but cost-benefit considerations, such as whether a diagnosis is treatable.

As new findings become available (either as a result of following recommendations from the diagnostic expert system, consultants, or the physician's own judgment), the physician user can rerun the diagnostic expert system to recompute the differential diagnosis. The most likely diagnosis from the first run may become less or more likely as the new information acts to rule in or rule out each diagnosis. The user can rerun the program whenever new findings about the individual become available during the course of the diagnostic work-up. The net result of this cyclic process is the diagnostic certainty about the diagnoses in the differential increases over time (i.e., probabilities of the diagnoses in the differential move towards zero, indicating certainty that a disease is not present, or one, indicating certainty that a disease is present) (Figure 13.1).

4. EXAMPLES OF DIAGNOSTIC EXPERT SYSTEMS: