A retrospective cohort study was conducted to compare programmatically assigned and bacteriologically defined MDR-TB outcomes among adult patients in Peru who initiated treatment between August 1996 and March 2002. Eligible patients were identified from an electronic database with information abstracted from patient medical charts. Individuals who started treatment during the study period were examined for treatment outcomes.

The subset of patients who had both programmatic and laboratory-based outcomes was analyzed to determine concordance between the two methods, and to identify any socio-demographic, social or clinical characteristics associated with concordance.

The study was reviewed and approved by the ethics review boards at the Instituto Nacional de Salud del Perú, the Harvard Medical School, and the US Centers for Disease Control and Prevention.

All medical and social history variables were categorized as ‘Known history of ___’ or ‘No known history of ___’. A count variable of social risk factors tallied the number of the following: homelessness, smoking, alcoholism, drug addiction, commercial sex work, criminal activity, imprisonment, unemployment, institutionalization, and military service. Body mass index (BMI) and age were examined as both continuous and categorical variables. Urban was defined as residing in Lima or Callao, the metropolitan Lima area.

A simple κ statistic and 95% confidence interval (CI) was calculated to assess overall agreement between programmatic and laboratory outcomes. As the laboratory outcome variable had only three categories (cured, completed, failed), the programmatic outcome variable was also limited to these three categories.

Next, cured and completed outcomes were combined to form a single successful outcome, and contrasted against failure. For the purposes of this analysis, performance characteristics (sensitivity, specificity, positive predictive value [PPV], and negative predictive value [NPV]) were calculated using the laboratory outcome as the gold standard.

Finally, logistic regression was used to examine possible factors associated with outcome concordance. Available socio-demographic, clinical, and social history variables were examined for univariate association with outcome agreement, defined as 1 if laboratory and programmatic outcomes matched and 0 if they were different. Any variable with a univariate association P ≤ 0.2 was considered for inclusion in a multivariate model. All models included age and sex as potential confounders. Effect modification by treatment strategy (individualized or standardized) was assessed. A final model predicting concordance was constructed retaining only variables with P < 0.05. Analyses were performed using SAS, version 9.2 (SAS Institute Inc, Cary, NC, USA).

Of the 1152 patients declared cured by a clinician, 123 (10.7%) were bacteriologically deemed to be treatment failures (Table 2). Similarly, 27 (9.4%) of the 287 individuals clinically categorized as completed were bacteriologic failures. Together, these misclassified categories totaled 150 patients (9.0% of all patients) who were programmatically considered treatment successes, yet had laboratory evidence of persistent MDR-TB.

At the same time, 9 (6.4%) patients deemed treatment failures by their health care providers were cured, and an additional 5 (3.6%) completed treatment based on their laboratory results. Of the 71 individuals still in treatment, 31 (43.7%) had sufficient laboratory results to be considered cured, and 20 (28.2%) had laboratory results establishing treatment completion.

The overall κ statistic for percentage agreement was 0.30 (95%CI 0.25–0.34).

For the dichotomized outcome–successful (cured or completed) vs. unsuccessful (failed)–performance characteristics of programmatically determined outcomes were evaluated using the outcome based on bacteriologic laboratory results as the gold standard. The sensitivity of the clinicians’ determinations was 98.9%, but specificity was only 45.7% (Table 3). The PPV and NPV were respectively 89.6% and 90.1%.

In univariate analysis, advanced education, employment as a health care worker, previous contact with a TB case, HIV, and receiving individualized treatment were the only predictors significantly associated with concordance between programmatic and laboratory-based outcomes, based on a cut-off of P < 0.2 (Table 4). Only treatment strategy was retained in the final multivariate model as a significant predictor of the outcome match variable (individualized vs. standardized treatment odds ratio 2.00, 95%CI 1.48– 2.70, P < 0.0001). In sub-analyses using stratified models and interaction terms, there was no detectable effect modification of other covariates by treatment strategy.

This study had several limitations due to the constraints of available data. Nearly half of this patient cohort was excluded due to either a missing programmatic outcome or a lack of sufficient bacteriologic evidence to establish a laboratory outcome of cure, completion, or failure. The analysis population was statistically different from the full patient population on a number of characteristics. The necessary exclusion of this subset of patients may therefore have distorted the final results, if those without both outcome determinations were either more or less likely to have concordant outcomes.

The lack of validation of either bacteriologic or programmatic outcomes in this study makes the outcome discordance difficult to interpret. Our laboratory outcome of treatment failure was based on international consensus definitions;⁸ however, we recognize the constraints in deeming patients as failed based on bacteriologic findings at the conclusion of the treatment course without post-treatment followup.¹² Prospective studies that include 6- or 12-month post-treatment cultures and clinical monitoring to reexamine treatment failures and relapse or recurrent disease rates of MDR-TB patients would provide valuable data on the accuracy of each of these outcome determination methods.

Data for this analysis were abstracted from medical charts and program data, which were not intended for research purposes. Many of the social and medical history variables were therefore missing for the majority of patients, and there was no mechanism for verifying the abstracted information. The missing data might be one reason why no social factors were identified as significant predictors of concordance.

Another shortcoming was the lack of information on physician characteristics. A physician’s training, experience, background, or location (e.g., in a high or low TB prevalence area) could conceivably affect his or her ability to correctly judge MDR-TB treatment outcomes. Some providers or facilities may have better track records than others at correctly determining outcomes, and future research could look at provider characteristics that may predict outcome concordance.