Background:
Adverse drug reactions (ADR) can influence treatment completion rates and the effectiveness of tuberculosis (TB) treatment. The rate of ADR related to anti-TB treatment (ATT) can be affected by HIV and diabetes mellitus. Given the paucity of prediction models for ATT-associated ADR, we developed a prediction model of ADR on TB treatment, incorporating important clinical variables.
Methods:
We included culture-confirmed, drug-susceptible, pulmonary TB participants from the RePORT-Brazil cohort, who received standard ATT between 2015-2019. ADR was defined based on physician-assigned attribution of relation to ATT and described according to the affected organ system, HIV status, severity, timing, and duration. Diabetes was categorized as HbA1c < 5.7% (no diabetes); ≥5.7% to < 6.5% (pre-diabetes); and ≥6.5% (diabetes). The predictive model of ADR risk used a bootstrapped backward selection approach. Of 13 candidate predictors (e.g., HIV status, HbA1c, ancestry markers), the variables retained in at least 70% of prediction models across 500 bootstrap replications were included in the final model. Model discrimination was evaluated by c-statistic and calibration with a calibration plot.
Results:
Of 945 participants, 102 (11%) experienced ADR. Among 156 ADR occurrences, most (78%) were of moderate severity and occurred during the first two months of ATT (77%). Hepatic ADR were the most frequent (n=82, 53%). ADR occurred in 38 (21%) people living with HIV/AIDS (PLWHA) and in 64 (8%) HIV-seronegative. Overall, 35 (10%) normoglycemic participants had ADR, while 47 (13%) and 19 (9%) participants with pre-diabetes and diabetes, respectively, experienced ADR. Variables included in the final prediction model for ADR were HIV status, HbA1c, age, ancestry markers, and concomitant medication use. Use of concomitant medication (mainly other antibiotics) and HIV status were highly predictive of ADR; they were included in 100% and 82% of all prediction models, respectively. The final prediction model demonstrated reasonable accuracy (c-statistic=0.75 [95%CI=0.70-0.80]) (Figure 1A) and suitable calibration (Figure 1B). Bootstrap internal validation indicated that the model was robust (optimism-corrected c-statistic of 0.73 [95% CI: 0.68-0.78]).
Conclusions:
We developed a robust, accurate, and reliable prediction model of ATT-related ADR. Knowledge and intervention based on important factors such as use of concomitant medication at the time of ATT initiation and HIV status could improve treatment tolerability.
Figure 1. Performance of prognostic model for predicting ADR in TB treatment. (A) the receiver operating characteristic curve indicates good fit, with c-statistic of 0.75 (95% CI: 0.70-0.80) indicating good discriminatory ability. (B) The calibration curve also indicated good fit with an optimism-corrected intercept and slope of -0.22 and 0.87, respectively.