In this study, the knowledge base was restricted to the 223 diseases common to humans and animals. Several disease information axes were identified for inclusion in the medical knowledge base. For each disease, nomenclature along with any recognized synonyms, etiology, epidemiologic descriptions, diagnostic procedures, diagnostic procedure findings, and clinical signs and symptoms were gathered and abstracted from recent editions of standard medical texts. In addition, occupational risks, food consumption data, travel history, and animal/insect exposure were included. Finally, basic treatment, management, and prevention protocols were also recorded (3-10). Bibliographic citations for all entries into the knowledge base were tied to each information axis to authenticate the content and provide a quick reference for additional reading. Glossaries of knowledge base entries (e.g., signs, symptoms, occupation, diagnostic procedure results, geography, animal/insect exposure) to be used in generating a differential diagnosis were created to ensure the consistency of the lexicon. Audits were performed to validate the knowledge base entries against the glossaries. The resulting knowledge base contains the entries listed in Table 1. A batch procedure indexing all findings to diseases for use by the inference engine resulted in more than 12,000 finding/disease links. The inference engine is the portion of the decision support system used to derive useful clinical information (e.g., possible diagnoses). It is composed of nine program modules containing more than 10,000 lines of C language code linked into one executable program.

The goal of the system is to remind the user of relationships between clinical signs, symptoms, and findings and the diseases themselves. Therefore, a simple, intuitive pattern-matching technique was employed. The user sets a "closeness-of-fit" parameter in the system that determines how tight the relationship between case signs/findings and diseases must be before a disease can become a candidate for the differential diagnosis list. A simple ratio is created with signs/findings entered that fit each disease as the numerator and the total number of signs/findings entered as the denominator. The ratio ranges from zero to 1.00 with 1.00 implying a perfect fit. If this value is equal to or greater than the "closeness-of-fit" parameter set by the user (also ranging from zero to 1.00), the disease is entered in the differential diagnosis list as a candidate. The ratio for each disease shows how close the clinical manifestations are to previously published descriptions of the disease. For example, if seven clinical signs and findings are collected on a case and six fit a particular disease by the pattern-matching algorithm, the case fit would be 0.86. If the case fit parameter in the system is set to 0.86 or greater, that disease would be added to the differential diagnosis list. If the case fit parameter is set to 1.00, only diseases that fit the clinical manifestations perfectly would become differential diagnosis candidates. Signs and findings are not weighted by the system.

The purpose of this trial was to determine if the intervention of the support system at the time of the first physician visit would significantly alter the mean number of days to the correct diagnosis. Investigation records of 96 cases of murine typhus occurring in Texas from 1979 to 1987 were obtained from the Texas Department of Health. All records were complete and were used in the study. Each record contained the date of the first contact with a physician (dp). The date on which the indirect fluorescence antibody test for Rickettsia typhi was ordered was defined as the date the correct diagnosis was suspected (ds). Therefore, the number of days from the first contact with a physician (dp) to the correct diagnosis without decision support system intervention (dc_ni) was defined as follows: dc_ni = ds - dp.

The number of days till the correct diagnosis was suspected (dc_ni) was calculated for all cases without support system intervention. A case was classified as missed if dc_ni >13 days. This number was derived from the worst-case scenario for a murine typhus test turnaround from the Texas Department of Health. During 1979-1987, the department had the only laboratory in Texas that used the indirect fluorescence antibody test for confirming murine typhus. On the basis of this criterion, 23 of the 96 original cases fell into the missed category.

The clinical history, symptoms, physical findings, and some preliminary laboratory findings as noted in the investigation records for the 23 missed patients were entered in the support system and processed at case fit parameter values of 0.25, 0.50, 0.75, and 1.00. Specific historical data such as occupation, travel, and animal/insect exposure, if available, were also entered in the support system. The support system was said to have suggested the correct diagnosis of murine typhus if the disease was included in a differential diagnosis list of five or fewer diseases at a case fit parameter of 0.50 or greater.

Testing for murine typhus was assumed to be ordered immediately if the disease was suggested by the support system. This effectively set the number of days to suspect the correct diagnosis with intervention (dc_wi) to 1 day under these conditions.

The distributions of dc_ni and dc_wi were found to be Gaussian (PROC UNIVARIATE, SAS). Therefore, parametric statistical techniques were used for data analysis. A paired t-test was run on µ_ni and µ_wi (mean time to suspecting murine typhus with and without the system, respectively) to see if there was a significant difference in the means (PROC MEANS, SAS) (14).

In 11 (48%) of the 23 cases defined as missed, murine typhus was clearly suggested by the support system in a differential list of five or fewer possibilities by using only clinical history, signs, symptoms, and preliminary laboratory data if available. In six (26%) of these cases, murine typhus was the only disease suggested by the support system. The support system did not suggest the correct disease in 12 (52%) of the cases classified as missed. Table 2 lists the most common diseases that also appeared in the differential diagnosis lists for murine typhus cases.

The mean number of days to suspect the correct diagnosis without support system intervention (µ_wi) was 11.5 days. The mean number of days to suspect the correct diagnosis with support system intervention (µ_wi) was 8.6 days, an improvement of 2.9 days (p = 0.001).

Investigation records of 342 cases of brucellosis occurring in Texas from 1980 to 1989 were also obtained from the Texas Department of Health. Of these records, 13 were incomplete, and 129 involved cases of recrudescence of the disease or had no specific date of onset of illness and were excluded from the study. The remaining 200 records contained the date of onset of illness. However, the day of the first physician visit was not available and had to be extrapolated from the day of onset of illness. The average number of days from onset of illness to the first physician visit was assumed to be 4 days, on the basis of a study in Texas for patients with murine typhus (12).

The day that the correct diagnosis was suspected was defined as the day on which the serum agglutination test or a bacterial culture for brucellosis was ordered. The number of days from the first contact with a physician to the day the correct diagnosis was suspected without support system intervention (dc_ni) was calculated as indicated above.

A case was classified as missed if dc_ni >11 days. This number was derived from the worst-case scenario for a brucellosis test turnaround from the Texas Department of Health. By this liberal criterion, 98 of the 200 original cases still fell into the missed category.

The clinical history, symptoms, physical findings, and some preliminary laboratory findings as noted in the investigation records for the 98 missed patients were entered in the support system and processed at case fit parameter values of 0.25, 0.50, 0.75, and 1.00. Specific historical data such as occupation, travel, and animal/insect exposure, if available, were also entered in the support system. By definition, the system gave the correct diagnosis of brucellosis if the disease was included in a differential diagnosis list of five or fewer diseases at a case fit parameter of 0.50 or greater.

It was assumed that the testing for brucellosis would be ordered immediately if the disease was suggested by the support system. This effectively sets the number of days to suspect the correct diagnosis with intervention (dc_wi) to 1 day under these conditions. The distributions of dc_ni and dc_wi were also Gaussian and analyzed by paired t-tests on µ_ni and µ_wi (14).

In 86 (88%) of the 98 cases defined as missed, brucellosis was clearly suggested by the support system in a differential list of five or fewer possibilities by using only clinical history, signs, symptoms, and preliminary laboratory data if available. In 69 (70%) of the cases, brucellosis was the only disease suggested. The support system did not suggest the correct disease in only 12 (12%) of the cases classified as missed. Table 3 lists the most common diseases that also appeared in the differential diagnosis of brucellosis.

The mean number of days to suspect the correct diagnosis without support system intervention (µ_wi) was 17.9 days. The mean number of days to suspect the correct diagnosis with support system intervention was 4.5 days, an improvement of 12.9 days (p = 0.0001). However, data were derived by using an extrapolated day of first physician visit, and that may have affected these comparisons.

The brucellosis investigation records screened showed that of nine cases, three involved apparently healthy pregnancies, one a premature delivery, one a miscarriage, one seizures, one weakness and chronic headaches, and two deaths. In eight of the nine cases, the support system correctly suggested brucellosis. In addition, 59 (30%) of the patients with brucellosis were hospitalized, some for extended periods during the diagnostic phase.

Twenty-one of 23 murine typhus patients were hospitalized, some for extended periods during the diagnostic phase of the case. Furthermore, many patients had to undergo expensive diagnostic testing, including 19 electrocardiograms, four computer-aided tomography scans, and a various other procedures such as bone scans, liver scans, barium enemas, and renal and pelvic sonograms. Earlier diagnoses of brucellosis and murine typhus could have eliminated the need for many of the diagnostic procedures, shortened hospital stays, and improved the course of treatment.

On the average, it took approximately 3 minutes of interaction with the decision support system to construct a differential diagnosis. Therefore, this type of screening could easily become part of the routine history-taking and physical exam of a patient. Furthermore, this study underlines the discriminatory power of the clinical history and physical signs and symptoms in suspecting the presence of a certain disease in a patient. Factors such as the patient's occupation, travel history, animal/insect exposure, and unusual dietary habits can be especially important in helping to diagnose many rare infectious and parasitic illnesses.

The results of this evaluation indicate that for two rare diseases (brucellosis and murine typhus), the decision support system appears to perform well with high sensitivity) in suggesting the correct diagnosis in a patient. However, the specificity of the system has not been evaluated. A prospective double-blinded study in a general clinical or hospital setting would make that determination.