Derivation and validation of a pulmonary tuberculosis prediction model.

OBJECTIVE

To describe the derivation and validation of a pulmonary tuberculosis (TB) prediction model that would enable early discontinuation of unnecessary respiratory isolation.

DESIGN

Patients placed in isolation for suspected pulmonary TB were studied retrospectively (derivation cohort) and prospectively (validation cohort). Independent predictors of pulmonary TB in the derivation cohort (January 1992-March 1994) were identified by retrospective analysis. Predictors in the model were assigned weights on the basis of the results of the multivariate analysis in order to quantitate the risk of TB in an individual patient. The prospective validation consisted of application of the model to patients placed in isolation during the period April 1994 to June 1995. The predictability of the model in the derivation and validation cohorts was evaluated using receiver operating characteristics (ROC), curve analysis, and calculation of the area under the ROC curve (AUC).

SETTING

A university-affiliated, urban, public hospital with a large population of prison inmates and patients with human immunodeficiency virus infection.

INTERVENTIONS

Prospective application of the prediction model to patients placed in isolation during the validation period.

RESULTS

Four factors were found to be independent predictors of pulmonary TB among 296 isolation episodes in the derivation cohort; positive acid-fast sputum smear (odds ratio [OR], 5.8; 95% confidence interval [CI95], 3.0-11.0; weight = 3 points), localized chest radiograph findings (OR, 2.5; CI95, 1.3-4.9; weight = 2 points), residence in a correctional facility (OR, 2.3; CI95, 1.2-4.4; weight = 2 points), and history of weight loss (OR, 1.8; CI95, 1.0-3.2; weight = 1 points). Infection control practitioners applied the model prospectively to 220 isolation episodes. The mean (+/-SE) AUCs of the ROC curve for the derivation and validation cohorts were not significantly different (.86 +/- .04 vs .86 +/- .07; P = .90). There was a significant decline in the mean duration of isolation from the onset of an automatic TB isolation policy in August 1992 to the end of the study (P = .045 by analysis of variance).

CONCLUSIONS

A pulmonary TB prediction model was derived and validated prospectively in a hospital with a moderately high prevalence of TB. The model quantitated the risk of TB in an individual patient and aided infection control practitioners and primary-care physicians in their decisions to discontinue isolation during the validation period. Utilization of the model was responsible, in part, for a decrease in the mean duration of isolation during the study period. Although the model may not have general applicability due to the uniqueness of the patient population studied, this study illustrates how prediction models can be developed and used effectively to deal with a clinical problem.