We searched five databases of peer-reviewed articles, (e.g., MEDLINE, National Technical Information Service, GrayLIT), Web sites of relevant government agencies (e.g., U.S. Department of Energy); and relevant nongovernmental Web sites. We included terms such as bioterrorism, biological warfare, decision support system, detection, diagnosis, radiology information systems, and public health. We also reviewed conference proceedings and reference lists of included articles.

We screened peer-reviewed articles to determine if they met inclusion criteria. Two investigators blinded to study authors independently abstracted articles onto pretested abstraction forms. Data abstracted from each report varied, depending on the type of system described. For descriptions of detection systems, we abstracted information about the system’s portability, ability to run more than one sample at a time, and ability to detect more than one bioterrorism agent. For descriptions of diagnostic decision support systems, we recorded whether bioterrorism-related illnesses were included in the system’s knowledge base, how the system enabled updates of the probability of bioterrorism-related illness as an epidemic progresses, the method of reasoning used by the inference engine, and whether the system used a standard vocabulary.

A complete description of the methods used to develop our evaluation criteria for reports of detection systems and diagnostic decision support systems can be found elsewhere (5). Briefly, we reviewed reports of naturally occurring and bioterrorism-related outbreaks and solicited information from relevant experts to describe the detection and diagnostic decisions that clinicians and public health officials would have to make while responding to bioterrorism. We then described the capabilities of detection and diagnostic systems necessary to assist these decisions. We augmented this list of system characteristics with previously published standards for evaluating information technologies and diagnostic tests to develop evaluation criteria for systems designed to facilitate detection and diagnosis during a bioterrorism response (5). We did not attempt to independently evaluate detection systems and diagnostic decision support systems; rather, we relied on information provided in the published reports about these systems.

We identified 55 detection systems including 4 systems that collect aerosol environmental samples; 14 particulate counters or biomass indicators that detect an increase in the number of particles in aerosol samples over baseline; 27 identification systems designed to rapidly detect bioterrorism agents collected from environmental, human, animal, or agricultural samples; and 10 systems that integrate the collection, identification, and communication of detection results (5). Other detection systems exist; however, we describe all of the systems for which we found publicly available information through the search methods described.

Only 8 of the 55 detection systems had published evaluations (Table 1 and Table 2). No system was evaluated for all the evaluation criteria. Timeliness was described for 33 of the 55 detection systems. Of these systems, 20 were described in specific terms such as minutes or hours, whereas 13 systems were described in nonspecific terms such as “rapid” or “real-time”. Several reports included general statements about system sensitivity or detection limits; however, studies of only 1 of the 55 detection systems specifically reported both sensitivity and specificity (Table 2) (14–18).

Of the four collection systems, we found evaluation data only for BioCapture (Meso Systems Technolgy, Inc., Albuquerque, NM), a device that has been used by fire departments in Seattle, Los Angeles, and New York among other sites to collect environmental samples for subsequent testing for bioterrorism agents (8). Although its sensitivity and specificity were not described, the BioCapture system had a collection efficiency reported to be 50% to 125% relative to reference standards (8). Reports on three other systems also included a comparison of the system under evaluation to a reference standard (Anthrax Sensor [7], MiniFlo [12], and the Fluorescence-based array [10]).

Most identification systems are limited in that they can evaluate a sample for only a single bioterrorism agent in each test cycle, they often run only a limited number of samples at a time, and they cannot test for many bioterrorism agents of concern (e.g., smallpox). None of the reports of the detection systems described methods for maintaining the security of the sample or test results or evaluated the systems in different clinical settings or among different populations. We found no studies that directly compared two or more systems in any given category.

In response to the 2001 anthrax cases, considerable interest was generated in the handheld antibody-based detection tests such as the Sensitive Membrane Antigen Rapid Test (SMART) (New Horizons Diagnostic Corp., Columbia, MD) and the Antibody-based Lateral Flow Economical Recognition Ticket (ALERT) (14–18). Such systems use antibodies to recognize specific targets on the toxins, antigens, or cells of interest (13,14). Limitations of these tests include nonspecific binding of the antibodies, which may lead to false-positive results and degradation of the antibodies over time, which may lead to false-negative results (13,14). Additionally, these tests are limited by the availability of antibodies. Given concerns about the diagnostic sensitivity and specificity of hand-held, antibody-based tests when used during the anthrax attacks, the Federal Bureau of Investigation and Centers for Disease Control and Prevention performed an independent evaluation of these tests (19). Although these results are not yet publicly available, the July 2002 Statement by the U.S. Department of Health and Human Services regarding hand-held assays for identification of Bacillus anthracis spores stated, “These studies confirm the low sensitivity of such assays and their potential to produce false-positive results with non-anthrax bacteria and chemicals. The performance of handheld assays for the detection of biological agents other than B. anthracis has not been evaluated and their use is also not recommended at this time” (20). Instead, law enforcement should transport samples quickly to a Laboratory Response Network facility, where cultures will be performed and preliminary results made available within 12 to 24 hours (20).

Several detection systems were designed in part, if not fully, by the military, and battlefield evaluations may have been performed. However, the paucity of publicly available information about such evaluations prevents conclusions about whether these systems will serve the detection needs of first responders and clinicians. Moreover, even if battlefield evaluation data were available, these systems would require additional study to confirm their utility for civilian users.

We identified 23 diagnostic decision support systems that may enhance clinician consideration of bioterrorism-related illness. We found six general diagnostic systems, four systems designed to improve radiologic diagnoses, four telemedicine systems, and nine systems for other diagnostic purposes (5). None has been formally evaluated with respect to a bioterrorism response; however, 15 diagnostic decision support systems had published evaluations for potentially analogous situations (Table 3).

The general diagnostic decision support systems are typically designed to assist clinicians develop a differential diagnosis list on the basis of patient-specific signs and symptoms. The included general decision support systems require manual entry of patient information by clinicians (Table 2 and Table 3). They use probabilistic or rules-based inference to compare patient information with a knowledge base to generate a differential diagnosis list that is typically ranked in descending order of likelihood. Some of the systems provide additional information about the suspected diseases and suggest appropriate diagnostic tests if clinicians choose to pursue these diagnoses.

Three diagnostic decision support systems (Columbia–Presbyterian Medical Center Natural Language Processor, Neural Network for Diagnosing Tuberculosis, and SymText) were specifically evaluated for both sensitivity and specificity and typically performed better than physicians-in-training but not as well as experienced clinicians (22,32,35,36). Also, the accuracy of the decision support systems decreased for difficult cases. The need for clinicians to manually enter patients’ data into diagnostic decision support systems, a laborious step that may be a barrier to the use of these systems and increases interuser variability, is eliminated by the few systems that automatically collect patient data from an electronic medical record (21,22,35,36). For example, diagnostic decision support systems currently available in hospitals with electronic medical records provide clinicians with an estimate of the likelihood of community-acquired pneumonia or active pulmonary tuberculosis based exclusively on data collected from the medical record (21,32,33,35,36).

Two infectious disease diagnostic decision support systems, The Computer Program for Diagnosing and Teaching Geographic Medicine and GIDEON, included most of the bioterrorism-related organisms in their knowledge bases (23,28). In an evaluation of The Computer Program for Diagnosing and Teaching Geographic Medicine, the system correctly identified 222 (75%) of 295 cases and 128 (64%) of 200 hypothetical cases (23). The clinical diagnosis was included in the computer differential diagnosis list in 95% of cases. Several cases included in this evaluation were for the causative agents of anthrax, brucellosis, cholera, Lassa fever, plague, Q fever, and tularemia.

An evaluation of GIDEON compared its diagnostic accuracy to that of medical house officers admitting 86 febrile adults to the hospital (28). The house officers listed the correct diagnosis first in their admission note 87% (75/86) of the time compared with 33% (28/86) for GIDEON (28). To limit the differential diagnosis provided by the system, users enter the geographic area where the outbreak occurred. This geographic information is compared with the known areas of natural occurrence. Adding this geographic information could falsely decrease the probability of disease if a bioterrorism agent were used in a region that had little naturally occurring disease from that organism.

Many diagnostic decision support systems use probabilistic information about the likelihood of disease. Because bioterrorism-related illness is relatively rare, in the event of bioterrorism these systems will have inappropriately low pretest probabilities for bioterrorism agents. Only Dxplain was described as being able to change the probability of disease based on information about suspected bioterrorism events to improve the system’s performance (26). Additionally, no report specifically described restricting access to the system by user type or other security measures.

For the purpose of evaluation, detection systems have much in common with diagnostic tests. Published guidelines for evaluating diagnostic tests are well established and promote study designs that provide unbiased estimates of both sensitivity and specificity (or likelihood ratios) relative to an acceptable reference standard, in the appropriate clinical population or setting.

The first important design consideration is that both sensitivity and specificity (or likelihood ratios) must be measured relative to an appropriate reference standard. Many of the studies included in our review measured only sensitivity or specificity. Because sensitivity and specificity are jointly determined by the choice of threshold for a positive (or abnormal) test, either sensitivity or specificity can be made arbitrarily high at the expense of the other. Thus, reporting one without the other is not informative. Reporting both sensitivity and specificity for a variety of thresholds for abnormal tests as a receiver operating characteristic (ROC) curve (Figure 2) is preferable. ROC curves are useful because differences in sensitivity and specificity of two tests could be due either to real differences in the accuracy of the test or to the use of a different threshold for an abnormal test. When results are reported as an ROC curve, no such confounding will occur.

To develop unbiased estimates of sensitivity and specificity, studies of detection systems should use an appropriate reference standard test, the reference standard should be applied to all samples, the tests should be interpreted while blinded to results of the reference standard, and the samples or patient population should resemble as closely as possible the populations in which the system will be used (40). The reference standard should be used for all positive and negative samples. Selective use of the reference standard, for example, using the reference standard only on samples that are positive on the test under consideration, creates so-called test referral bias which can produce overestimates of sensitivity and underestimates of specificity (40). Test-interpretation bias may occur if the result of the detection system is not determined while blinded to the reference test (and vice versa). This bias causes an artificial concordance between the detection system and reference test, which results in overestimates of both sensitivity and specificity. Finally, the detection system should be evaluated under the most realistic conditions possible, which may be difficult to implement for bioterrorism agents given the range of conditions from hoaxes with no cases to real situations with a number of cases.

Evaluations of detection systems are ongoing (19). We expect with the heightened attention to bioterrorism preparedness planning that the systems for both detection and diagnosis will improve, as will their evaluations. Evaluations that adhere to the principles for design of studies of diagnostic tests will provide substantially more information than is now available and will help users interpret the results provided by these systems. Our review of 78 detection and diagnostic systems found that many of the evaluations performed to date are critically deficient. Further evaluative studies will delineate the usefulness of these systems.