Recall that the best case detection system imaginable would be one in which every individual in a community is examined every morning by the best diagnostician in the world. Since she would be examining everyone every day, she would be aware of patterns of illness in a community and this awareness would appropriately influence her diagnostic thinking (and treatment) of individual patients.3 This diagnostician would also never fail to report immediately each fever, early syn-dromic presentation, or reportable disease to governmental public health.

Diagnostic expert systems are the key to building such a system. The research we reviewed in this chapter has already solved most, if not all, of the technical problems. What is needed is the will to create such a system.

3 Physicians are taught (and reminded incessantly) "when you hear hoof beats, don't think of zebras". This adage is an informal statement that when the evidence available about a particular patient supports equally a diagnosis of either influenza or SARS (e.g., the patient has constitutional symptoms and no history of exposure to SARS), they should conclude that the diagnosis of influenza is far more likely than SARS.

If and when diagnostic expert systems are embedded in the clinical information systems of every hospital (animal and human), long-term care facility, clinic, and laboratory in a region, they will be able to notify a health department or other biosurveillance organization of every fever, syndromic presentation, and reportable disease in individuals receiving medical or veterinary care. If diagnostic expert systems are made available to the public or to selected high-risk populations (e.g., postal employees or patients with preexisting conditions such as asthma or diabetes), case finding would be extended to an even larger fraction of the population, approximating the "every-patient-every-day" capabilities of an ideal case detection system.

If a biosurveillance system located in a region's health department were to receive the differential diagnosis for each individual in the region in real time (anonymously, of course, and perhaps selectively based on diseases and probability thresholds), it could compute the current incidences of these conditions. It could monitor the region for increases in incidence of findings, syndromes, and diseases of concern.

Note that the outputs of a probabilistic diagnostic expert system are posterior probabilities of diseases for one patient, so the central monitoring would be monitoring of the sums of the probabilities for all reported patients. Figure 13.13 illustrates this summation. It plots the daily sums of the posterior probabilities of "flu-like" illness of all patients seen in emergency departments on each day. If the diagnostic expert systems are well calibrated, the sums of the posterior probabilities should equal the actual number of patients with the disease in the population being evaluated by the system.

Finally, if the biosurveillance system would then communicate the current fever, syndrome, and disease incidences back to the diagnostic expert systems being used by clinicians and citizens, we would realize the "diagnosticians-are-aware-of-patterns-of-illness-in-the-community" capability of an ideal case detection system.

figure 13.13 Daily sum of syndrome probabilities produced by SyCO2. SyCO2 computes the posterior probability that a patient has a flu-like illness from his chief complaint. (From Espino, J., Dara, J., Dowling J, et al. (2005). SyCo2: A Bayesian Machine Learning Method for Extracting Symptoms from Chief Complaints And Combining Them Using Probabilistic Case Definitions. Pittsburgh, PA: Center for Biomedical Informatics, University of Pittsburgh, with permission.) An outbreak-detection system would sum the posterior probabilities of flu-like illness from all patients seen in 24-hour periods to form a time series of expected daily counts of patients with respiratory illness. Readers familiar with Bayesian statistics will recognize this sum as the expectation for the number of individuals with a given diagnosis.

figure 13.13 Daily sum of syndrome probabilities produced by SyCO2. SyCO2 computes the posterior probability that a patient has a flu-like illness from his chief complaint. (From Espino, J., Dara, J., Dowling J, et al. (2005). SyCo2: A Bayesian Machine Learning Method for Extracting Symptoms from Chief Complaints And Combining Them Using Probabilistic Case Definitions. Pittsburgh, PA: Center for Biomedical Informatics, University of Pittsburgh, with permission.) An outbreak-detection system would sum the posterior probabilities of flu-like illness from all patients seen in 24-hour periods to form a time series of expected daily counts of patients with respiratory illness. Readers familiar with Bayesian statistics will recognize this sum as the expectation for the number of individuals with a given diagnosis.

This ideal approach underlies the Bayesian approach to outbreak detection described in Chapter 18 Bayesian Methods for Diagnosing Outbreaks. PANDA (Population-wide ANomaly Detection and Assessment)â€”the research system described in that chapterâ€”actually merges many individual diagnostic Bayesian networks into a very large network that also includes a subnetwork that draws inferences about the presence or absence of an outbreak. PANDA is pursuing this idea on a citywide scale. Conceptually, a PANDA network comprises millions of person-specific diagnostic Bayesian networks, each of whose probabilities of disease (e.g., anthrax) are being influenced by population-level observations (e.g., aggregate sales of over-the-counter medications) and population-level inferences (e.g., the likelihood that other individuals who may have inhalational anthrax are present in a population). PANDA also includes a prior probability distribution over outbreak diseases (e.g, the prior probability of inhalational anthrax, based on the national terror alert level). By integrating person-specific diagnostic submodels into a population-wide super model, approaches like PANDA are able to make inferences about the probability of a disease outbreak in the population as a whole, as well as the probability of disease in individual people (or subgroups of people) within the population.

Was this article helpful?

## Post a comment