In approximately 1.5 million person-years, 768 persons met at least one of the HMO-based screening criteria, with a positive screening criteria rate of 0.4 per 10,000 person-years among persons with automated ambulatory-patient records and 0.7 per 10,000 person-years among those without such records. Thirty-nine (9%) incident TB cases were identified among the 415 persons with automated ambulatory-patient records who met screening criteria, and 4 (1%) incident TB cases were identified among the 353 persons without automated ambulatory records who met screening criteria. The response rate to the provider questionnaire was 100%, as was the agreement rate between classification of TB cases based on provider questionnaire results and on-site medical record review.

Thirty-five (81%) of the 43 incident TB cases detected by HMO-based screening had been identified previously by the public health department. Of these 35 cases, 32 were culture-positive, and three met the clinical case definition. Two additional TB cases, both of which were culture-positive, were known to the public health department but did not meet HMO-based screening criteria. These two patients received treatment and medication from state-funded TB clinics. Thus, 45 cases were identified through either HMO-based screening or public health department records. All 45 cases met the CDC surveillance definition. Eight (18%) of these cases were unknown to the public health system. Most cases (41 of 45) were diagnosed at one of the HMO centers with automated ambulatory-patient records, a proportion consistent with the concentration of urban regions within their catchment areas. The rates were approximately 11.7 TB cases per 100,000 population among HMO members with automated ambulatory-patient records and four TB cases per 100,000 population among those without such records.

The sensitivity of each of the screening criteria was 0% to 100%, and the positive predictive value was 0% to 52% (Table 2). Screening criteria based on pharmacy dispensing information had the best combinations of sensitivity and positive predictive value. Two or more dispensed antituberculosis drugs, combining the results for persons with and without automated ambulatory-patient records, had an overall sensitivity of 89% (95% confidence interval [CI] = 76%, 96%) and positive predictive value of 30% (95% CI = 22%, 39%). Three or more antituberculosis drugs had an overall sensitivity of 84% (95% CI = 71%, 94%) and positive predictive value of 50% (95% CI = 38%, 62%), and two or more antituberculosis drugs dispensed on the same date had an overall sensitivity of 87% (95% CI = 73%, 95%) and positive predictive value of 36% (95% CI = 27%, 50%). The differences between the performance of two or more dispensed antituberculosis drugs among persons with automated ambulatory-patient records (sensitivity = 90%, positive predictive value = 34%) and persons without automated records (sensitivity = 75%, positive predictive value = 12%) were not statistically significant, although the small number of TB cases in each group limits the power to detect a difference.

Among the 71 persons with automated ambulatory-patient records who received two or more antituberculosis drugs but did not have incident TB, 9 (13%) had active TB diagnosed outside the study period, 23 (32%) were treated for other mycobacterial infections, 11 (15%) received more than one drug during TB prophylaxis, 2 (3%) received drugs for multiple unrelated conditions (e.g., rifampin for eradication of Staphylococcus aureus; ethambutol for M. avium complex prophylaxis), and the remaining 26 (37%) were suspected of having active TB without subsequent confirmation (Table 3).

Of the 118 persons with automated ambulatory-patient records who met screening criteria involving a diagnosis code for TB but did not have incident TB, 57 received the diagnosis code as an indication of routine prenatal screening for TB, 12 had a previous history of TB, 31 were suspected of having active TB without subsequent confirmation, and 18 had the diagnosis code documented in their HMO ambulatory medical record for no apparent reason (Table 3).

Of the eight patients whose cases had not been identified by the public health department, seven were culture-negative and met the TB clinical case definition, and one did not have a microbiology culture and met the smear-positive TB case definition. Three of the patients had AFB smear-positive pathology specimens; of these, two had negative cultures for M. tuberculosis, and one did not have a culture performed. Of the eight cases, one involved pulmonary TB, and the remaining seven were extrapulmonary. One patient was 2 years old at the time of diagnosis; the remaining seven were 18 years of age or older. All cases were confirmed by review with the Massachusetts State Tuberculosis Control Officer. Of these eight cases, seven were detected by the two or more antituberculosis drug screening criterion.

Since the establishment of a national surveillance system for TB in 1953, TB surveillance has depended on laboratories, public health clinics, and reporting by private practitioners. Several retrospective studies performed by local TB programs suggest that TB cases may be underreported (5-7). Although ascertainment of culture-positive cases is likely to be nearly complete, since laboratories are required by law in most states to report isolation of M. tuberculosis to the state health department, surveillance for cases lacking positive cultures depends largely on reporting by health-care providers or referrals to public health clinics for treatment. Underreporting of TB cases without positive cultures may contribute to incomplete surveillance. A study assessing the completeness of TB case reporting in Puerto Rico (6) found that 19.5% of patients with TB were not reported, partly because of underreporting of cases without positive cultures for M. tuberculosis.

The recent shift into managed care of populations at high risk for TB, including Medicaid and Medicare recipients, has raised additional concern about the continued completeness of reporting. However, HMOs routinely collect information that can be used to identify persons likely to have TB. McCray et al. (6) noted that, according to pharmacy prescription data in Maryland, the cases of 19% of patients receiving two or more antituberculosis drugs had not been reported to the public health department; however, the patients' medical records were not reviewed to verify a diagnosis of active TB. Maggini et al. (8) evaluated the use of Italy's National Health Service pharmacy dispensing information to identify TB cases in the province of Rome and found that pharmacy screening detected seven times more new TB cases than routine passive surveillance. Hripcsak et al. (9) evaluated a number of screening rules based on automated information available at an urban medical center in New York City and found that inpatient use of antituberculosis drugs had a sensitivity of 68% and a positive predictive value of <1% for detecting TB cases based on their health department's TB registry. These investigators did not, however, have access to records of antituberculosis drugs received by ambulatory patients and did not specifically evaluate the use of more than one antituberculosis drug as a screening criterion. No previous study has compared the utility of pharmacy data with that of other automated administrative or health-care data.

Of the screening criteria evaluated in our study, dispensing of two or more antituberculosis drugs was the most useful, with an overall sensitivity of 89%. The most common reasons for dispensing of two or more antituberculosis drugs to persons without TB were the empiric use of antituberculosis drugs for suspected active TB (37%) and the use of antituberculosis drugs for treatment of mycobacterial infections other than TB (32%). In addition, 15% of patients without TB received more than one drug for TB prophylaxis, which can occur, for example, when isoniazid is switched to another antituberculosis drug because of adverse drug reactions. A possible strategy for improving the positive predictive value of screening criteria based on pharmacy dispensing information is the use of more rigorous criteria, such as dispensing of three or more antituberculosis drugs (positive predictive value = 50%), or restricting the timing of drug dispensing, such as requiring that two or more antituberculosis drugs be dispensed on the same date (positive predictive value = 36%). The improvement in positive predictive value for these more rigorous criteria, however, must be weighed against loss of sensitivity in identifying TB cases. For our HMO study population, requiring three or more antituberculosis drugs missed two TB cases, and requiring that two or more drugs be dispensed on the same date missed one case detected by the less stringent criterion. The choice of the screening criterion with the most useful balance between sensitivity and specificity depends in part on the surveillance strategy used.

Surveillance based on HMO pharmacy dispensing information can be used to identify HMO enrollees most likely to have active TB, so that efforts can be focused on additional evaluation of these persons. As with traditional TB surveillance methods, surveillance based on pharmacy dispensing information requires information from the patients' medical records to verify whether the TB case definition is satisfied. Using this surveillance strategy, screening for two or more antituberculosis drugs would require reviewing the medical records of approximately three patients to identify each case of incident active TB. We feel that the positive predictive value of 30% is sufficient to make this surveillance screening method practical if it can be applied in other managed care settings.

The positive predictive values of screening criteria that include TB diagnosis codes are limited by a number of factors. TB diagnosis codes, for example, were frequently used for patients with suspected active TB during the weeks required for diagnostic work-up or observation for clinical response to therapy. These codes were also frequently used to indicate that routine TB skin testing had been performed rather than to indicate the presence of active disease or prior history of TB.

The difference in the TB case rates between HMO members with automated ambulatory-patient records (approximately 11.7 TB cases per 100,000 population) and members without such records (approximately four TB cases per 100,000 population) in our study could either reflect a true difference in the underlying risk for TB in the two populations or case ascertainment bias resulting from differences in the methods used to identify TB cases. The former explanation is more likely for several reasons. First, the HMO health centers with automated ambulatory-patient records serve a largely urban population concentrated in the Boston area, while the HMO-affiliated practices without such records serve a largely suburban population. The difference in the rates found in our study mirrors the difference in the 1992 to 1998 TB case rate averages reported by the Massachusetts Department of Public Health for the city of Boston (17.7 TB cases per 100,000 population) compared with the rest of the state of Massachusetts (4.1 TB cases per 100,000 population). Second, the match between the health department's TB registry and the HMO membership list did not identify any TB patients who had not previously been detected through screening criteria and record review based on our modified RVCT results among HMO members without automated ambulatory patient records. This argues against inadequate case finding resulting in apparent lower TB case rates in this group.

A substantial number of TB cases in our study were unknown to the public health department (18% of cases among our HMO study population). This proportion is comparable with the fraction described in the studies cited above. Underreporting of these cases compromises the usefulness of TB surveillance. Screening for dispensing of antituberculosis drugs may be a particularly useful method for identifying cases without positive cultures for M. tuberculosis that might otherwise be missed by routine surveillance methods dependent on laboratory- and provider-based reporting.

The positive predictive value of screening criteria based on the dispensing of antituberculosis drugs may also be limited to some degree in clinical settings where many patients receive these medications for other indications, including other mycobacterial infections (e.g., in cases of HIV infection). Strategies that could be applied in such settings include excluding those persons also receiving medications frequently used for treatment of M. avium complex infections (e.g., clarithromycin). During our study period, however, more than 1,000 known HIV-infected patients were treated in HMO centers with automated ambulatory-patient records, of whom only 23 (Table 3) had false-positive cases identified by the two or more TB drug criterion. In addition, widespread implementation of new CDC recommendations for use of multidrug therapy for TB prophylaxis may require modification of the screening criteria. One possible strategy would be to require that antituberculosis drugs be dispensed over a minimum time interval (e.g., >4 months).

Although TB surveillance based on pharmacy dispensing information depends upon availability of automated pharmacy data, such data are available for most of the U.S. population, including most Medicaid and Medicare recipients. Our results indicate that pharmacy dispensing information routinely collected by many HMOs has high sensitivity and reasonable positive predictive value and is particularly useful for identifying TB cases without positive cultures, which may be missed by traditional public health surveillance.