Development and Validation of a Nomogram for Prediction of QT Interval Prolongation in Patients Administered Bedaquiline-Containing Regimens in China: a Modeling Study

ABSTRACT

Corrected QT duration (QTc) interval prolongation is the most frequent adverse event associated with bedaquiline (BDQ) use. It may affect the safety of antituberculosis therapy, which leads to the consequent demands of needing to monitor during therapy. Our objective was to establish and validate a prediction model for estimating the risk of QTc prolongation after initiation of BDQ-containing regimens to multidrug-resistant tuberculosis (MDR-TB) patients. We constructed an individualized nomogram model based on baseline demographic and clinical characteristics of each patient within a Chinese cohort during BDQ treatment. The generalizability of this model was further validated through use of externally acquired data obtained from Beijing Chest Hospital from 2019 to 2020. Overall, 1,215 and 165 patients were included in training and external validation cohorts, respectively, whereby during anti-TB drug treatment, QTc prolongation was observed in 273 (22.5%) and 29 (17.6%) patients within these respective cohorts, for whom QTc values were >500 ms in 86 (31.5%) and 10 (34.7%) patients, respectively. Next, a total of four Cox proportional hazards models were created and assessed; then, nomograms derived from the models were plotted based on independent predictors of clofazimine, baseline QTc interval, creatinine, extensive drug-resistance (XDR), moxifloxacin, levofloxacin, and sex. Nomogram analysis revealed concordance index values of 0.723 (95% confidence interval [CI], 0.695 to 0.750) for the training cohort and 0.710 (95% CI, 0.627 to 0.821) for the external validation cohort, thus indicating relatively fair agreement between predicted and observed probabilities of QTc prolongation occurrence based on data obtained during 8-week, 16-week, and 24-week anti-TB treatment of both cohorts. Taken together, results obtained using these models demonstrated that coadministration of clofazimine and abnormal baseline QTc interval significantly contributed to QTc prolongation development during MDR-TB patient treatment with a BDQ-containing anti-TB treatment regimen.

INTRODUCTION

The emergence of drug resistance poses a major threat to global tuberculosis (TB) control (1, 2). The World Health Organization (WHO) has estimated that approximately 4.7% of the 10 million incident TB cases in 2019 were caused by the Mycobacterium tuberculosis complex (MTB) strains with resistance to rifampicin (RR-TB), of which 78% of cases also harbored resistance to isoniazid (INH) and thus were classified as multidrug-resistant TB (MDR-TB) cases (2). Notably, only 57% of individuals afflicted with MDR-TB achieve successful treatment outcomes (e.g., microbiologically cured, completed treatment), whereas thousands of patients die or discontinue treatment prematurely (default) (2). Given the high risk of morbidity and mortality among MDR-TB patients, the development of novel drugs is essential for improving clinical outcomes and reducing mortality (3).

Bedaquiline (BDQ), a member of the diarylquinoline chemical drug class, was the first medication approved during the past half century by the United States Food and Drug Administration (FDA) for the treatment of adult pulmonary drug-resistant TB (4). BDQ inhibits mycobacterial ATP synthase by targeting subunit c of the enzyme to significantly decrease intracellular ATP levels (5, 6). Due to its novel mechanism of action, BDQ exhibits potent efficacy against drug-resistant tubercle bacilli in vitro (7), while results of numerous clinical trials have demonstrated that BDQ administration leads to a high rate of culture conversion when used to treat MDR-TB patients (8–10). In view of abundant strong evidence supporting BDQ benefits, the WHO has endorsed BDQ use for the treatment of MDR-TB by including it in the new WHO consolidated drug-resistant TB treatment guidelines as a prioritized group A drug (11).

Despite impressive clinical benefits, corrected QT duration (QTc) interval prolongation, the most frequent adverse event associated with the administration of anti-TB drugs, has been noted in diverse populations receiving BDQ-containing regimens (8, 12, 13). Thus, continual clinical monitoring to detect adverse events is of great importance to ensure effective therapeutic management of MDR-TB cases. Importantly, results obtained from previous trials of BDQ-treated MDR-TB patients suggest that a high baseline QTc interval value of ≥400 ms is probably the most powerful predictor of QTc prolongation occurrence (14). Meanwhile, results of other studies have shown that QTc prolongation by 500 ms may be triggered by several other clinical factors, including concomitant treatment with other anti-TB drugs (in addition to BDQ) and underlying diabetes (12, 15). Therefore, prior to the clinical use of BDQ to treat MDR-TB cases, clinicians should thoroughly review patient medical histories and assess patients for comorbidities to determine the risk of cardiac events before implementing screening to detect QTc prolongation, as serial electrocardiogram (ECG) monitoring places an extra burden on patients and TB treatment providers. Toward this goal, a prognostic model can be used to focus screening efforts on the patient group at highest risk of developing QTc prolongation during treatment with BDQ-containing anti-TB drug regimens.

Nomograms have been accepted as reliable and pragmatic prediction tools in many academic fields of study for use in quantifying risks associated with a variety of key factors, including adverse event risks associated with clinical characteristics of individual patients (16). They have been used widely in constructing prognostic models in many fields. Nevertheless, to our knowledge, these tools have not yet been used to predict QTc prolongation risk in TB patients, prompting this study. Here, we collected baseline demographic and clinical characteristics of a cohort of patients with MDR-TB during treatment for TB with the aim to develop and validate a prediction model for estimating patient QTc prolongation risk after initiation of BDQ-containing anti-TB therapeutic regimens.

RESULTS

Characteristics of patients.

For the training cohort, which initially included 1,306 patients who received BDQ-containing regimens, 1,215 patients who met study inclusion criteria were enrolled. For the validation cohort, we recruited an additional 165 consecutive patients. Details of the study population and missing data are shown in the appendix (see Fig. S1 and Table S1 and S6 in the supplemental material). Overall, QTc prolongation occurred in 273 (22.5%) and 29 (17.5%) patients in training and validation cohorts, respectively. Among the 273 patients experiencing QTc prolongation, 216 (79.1%) exhibited QTc increases of ≥60 ms over baseline and 86 (31.5%) exhibited QTc prolongation of >500 ms. No statistically significant difference was found in Kaplan-Meier QTc prolongation curves between training and validation cohorts (Fig. 1). The median age of patients in the training cohort was 37 years (range, 18 to 74 years) and that in the validation cohort was 34 years (range, 19 to 70 years). Subjects in the training cohort were more likely to be male (70.5%), pre-XDR cases (41.6%), and recurrent TB cases (75.1%). The most common comorbidity was diabetes (25.1%). In addition to BDQ, most anti-TB treatment background regimens included linezolid (87.7%), cycloserine (84.0%), and clofazimine (55.6%). Characteristics of patients in validation and training cohorts were similar except that a lower proportion of patients were recorded as MDR-TB in the validation cohort. Details regarding patient characteristics are shown in Table 1.

FIG 1.

Kaplan-Meier curves of training and validation cohorts for QTc prolongation.

TABLE 1.

Demographics and laboratory results of multidrug-resistant tuberculosis patients at cohort entry

Demographic or characteristic	Results for:
	Primary cohort (n = 1,215)		Validation cohort (n = 165)
	n or median	% or range	n or median	% or range
Outcome
QTc prolongation	273	22.5	29	17.6
QTc increase of >60 ms	216	17.8	23	13.9
QTc of >500 ms	86	7.1	10	6.1
Age, yrs	37	18–74	34	19–70
Sex
Male	856	70.5	121	72.5
Female	359	29.5	46	27.5
BMI (kg/cm2)	20.3	12.5–41.2	21.3	14.9–24
Drug resistance
MDR TB	458	37.7	77	46.1
Pre-XDR	505	41.6	45	26.9
XDR	252	20.7	45	26.9
Baseline QTc (ms)	414	370–448	404.5	375–440
Serum potassium (mmol/L)	4.0	2.3–5.8	4.1	2.9–5.2
Serum calcium (mmol/L)	2.3	1.0–3.4	2.3	2–2.9
Creatinine (μmoI/L)	67.9	22.3–173.2	66.6	34.4–116.2
History
Initial	303	24.9	11	6.6
Recurrent	912	75.1	156	93.4
Background disease
Diabetes	305	25.1	29	17.4
Hepatitis	12	0.9	0	0
Hypertension	7	0.5	0	0
HIV positivity	3	0.2	0	0
Background regimen
Moxifloxacin	240	19.8	21	12.6
Levofloxacin	400	32.9	86	51.5
Linezolid	1,066	87.7	133	79.6
Clofazimine	676	55.6	104	62.3
Cycloserine	1,020	84.0	122	73.1

Univariate analysis and multivariate analysis.

Results of the univariate analysis of all candidate variables revealed that statistically significant factors for QTc prolongation were as follows: sex, body mass index (BMI), QTc interval at baseline, XDR, creatinine, previous TB, diabetes, and creatinine. After backward stepwise selection, significant risk predictors for the occurrence of QTc prolongation in models C and D included older age, elevated or decreased QTc interval at baseline, female sex, XDR, decreased creatinine, previous TB, nondiabetic status, elevated or decreased serum potassium, decreased BMI, and a personalized background regimen containing clofazimine. Penalized maximum likelihood estimation based on a cross-validated optimal penalty factor (model A, log λ = −3.57; model B, log λ = −3.38) (see Fig. S3 in the supplemental material) reduced the coefficients for age, BMI, previous TB, diabetes, and serum potassium in models A or B to zero, while the following variables remained independently predictive: sex, baseline QT interval, XDR, creatinine, and coadministered clofazimine. The detailed results of the multivariate analyses are shown in Table 2. Based on two transformations of baseline QTc values for models A and D, we found that the QTc values increased by ≥60 ms in a significantly greater proportion of patients in the group with QTc prolongation values of <395 ms (88/95, 92.63%) than those in the group with QTc prolongation values of >425 ms (82/121, 67.77%; chi-square, P < 0.001).

TABLE 2.

Cox proportional hazards regression model showing the association of variables with the occurrence of QTc prolongation^a in the 24-week treatment

Variable	n (%)	Results for:b
		Univariable		Multivariable
		HR (95% CI)	P value	Model A		Model B		Model C		Model D
				HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)	P value
Factors selected
Sex
Male	856 (70.5)	1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]
Female	359 (29.5)	1.53 (1.15–2.02)	0.003	1.53 (1.15–2.04)	0.003	1.52 (1.17–2.05)	0.003	1.63 (1.23–2.16)	<0.001	1.52 (1.15–2.03)	0.003
QTc baseline numeric, ms
QTc squared						0.32 (0.27–0.39)	<0.001	0.77 (0.73–0.82)	<0.001
QTc						1.00 (1.00–1.00)	<0.001	1.00 (1.00–1.00)	0.052
QTc baseline, categorical
395-424. ms	569 (46.8)	1 [Reference]		1 [Reference]						1 [Reference]
<395 ms	267 (22.0)	4.12 (2.97–5.7.3)	<0.001	4.48 (3.21–6.23)	<0.001					4.64 (3.33–6.48)	<0.001
>425 ms	379 (31.2)	3.65 (2.67–5.00)	<0.001	3.43 (2.50–4.73)	<0.001					3.48 (2.53–4.78)	<0.001
Drug resistance
Pre-XDR	963 (79.3)	1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]
XDR	252 (20.7)	1.90 (1.42–2.54)	<0.001	1.64 (1.26–2.14)	<0.001	1.53 (1.17–2.01)	<0.001	1.62 (1.24–2.11)	<0.001	1.60 (1.23–2.09)	<0.001
Creatinine
≥60 mmol/L	816 (67.2)	1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]
<60 mmol/L	399 (32.8)	1.49 (1.17–1.89)	<0.001	1.43 (1.08–1.89)	0.011	1.42 (1.08–1.86)	0.017	1.29 (0.97–1.70)	0.073	1.44 (1.10–1.90)	0.009
Clofazimine
No	539 (44.4)	1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]
Yes	676 (55.6)	2.93 (2.21–3.89)	<0.001	2.93 (2.19–3.92)	<0.001	3.23 (2.42–4.32)	<0.001	2.83 (2.11–3.80)	<0.001	2.74 (2.04–3.67)	<0.001
Levofloxacin
No	815 (67.1)	1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]
Yes	400 (32.9)	1.02 (0.79–1.31)	0.889	1.29 (0.99–1.69)	0.061	1.35 (1.03–1.76)	0.062	1.13 (0.97–1.70)	0.138	1.26 (0.96–1.65)	0.092
Moxifloxacin
No	975 (80.2)	1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]		1 [Reference]
Yes	240 (19.8)	0.87 (0.64–1.19)	0.378	1.15 (0.82–1.61)	0.412	1.18 (0.84–1.65)	0.257	1.18 (0.85–1.67)	0.445	1.13 (0.81–1.59)	0.461
Age, yrs		1.04 (1.01–1.07)	0.006					1.03 (1.00–1.06)	0.035	1.03 (1.01–1.07)	0.009
BMI (kg/m2)		0.95 (0.90–0.99)	0.049					0.96 (0.92–1.00)	0.098
Previous TB
No	303 (24.9)	1 [Reference]								1 [Reference]
Yes	912 (75.1)	1.37 (1.02–1.85)	<0.036							1.28 (0.94–1.73)	0.071
Diabetes
No	910 (74.9)	1 [Reference]						1 [Reference]		1 [Reference]
Yes	305 (25.1)	0.48 (0.27–0.83)	0.030					0.69 (0.43–1.09)	0.102	0.59 (0.38–0.94)	0.059
Blood potassium (mmol/L)
3.8–4.2	526 (43.3)	1 [Reference]
<3.8	291 (24.0)	1.17 (0.83–1.65)	0.360					1.39 (1.03–1.87)	0.033
>4.2	398 (32.8)	1.24 (0.82–1.55)	0.451					1.27 (0.95–1.70)	0.101
Factors not selected
Blood calcium (mmol/L)
<2.2	422 (34.7)	1 [Reference]
>2.2	793 (32.8)	1.13 (0.85–1.50)	0.418
Linezolid
No	149 (12.3)	1 [Reference]
Yes	1066 (87.7)	1.36 (0.91–2.02)	0.133
Cycloserine
No	195 (16.0)	1 [Reference]
Yes	1020 (84.0)	0.99 (0.72–1.36)	0.938

^aQTc prolongation, the first occurrence of a QTc prolongation event, defined as an QTc interval of ≥500 ms or an increase in an QTc interval of ≥60 ms compared with the baseline QTc at admission.

^bReference represents the reference group to be compared.

Model performance.

Based on training cohort data, after correcting the concordance index (c-index) values of nomograms via bootstrap optimism analysis based on resampling of 1,000 bootstrapped samples, differences in c-index values between nomograms were apparently trivial (Table 3). Model C produced the highest c-index, whereas model D produced the lowest c-index. Results of calibration plots exhibited excellent agreement between predicted and actual observations for 8-week, 16-week, and 24-week QTc prolongation probabilities calculated using all four models. The overall precision of the models, as reflected by Brier scores, varied between 0.129 and 0.133.

TABLE 3.

The bootstrap-corrected c-index of QTc prolongation occurrence for multivariate model performance in the training set and the validation set

Model for survival prediction	Results for:
	Training cohort			Validation cohort
	c-indexa	95% CI	Precision Brier score	c-index	95% CI	Precision Brier score
Model Ab	0.732	0.707–0.760	0.133	0.683	0.605–0.762	0.091
Model Bc	0.723	0.695–0.750	0.129	0.724	0.627–0.821	0.088
Model Cd	0.733	0.707–0.761	0.129	0.675	0.587–0.765	0.091
Model De	0.721	0.693–0.748	0.129	0.679	0.603–0.755	0.092

^ac-index, the drop-off in the c-index was considered when comparing performance of fits from 1,000 optimism bootstrap samples to the performance of the bootstrap-derived model on the original training data sets.

^bModel A, LASSO method with baseline QTc categorized in quartiles (<395 ms, 295 ms–424.9 ms, and ≥425 ms).

^cModel B, LASSO selection with baseline QTc squared.

^dModel C, backward stepwise selection with baseline QTc squared.

^eModel D, backward stepwise selection with baseline QTc categorized in quartiles (<395 ms, 395 ms–424.9 ms, and ≥425 ms).

In the external validation cohort, the bootstrap-corrected c-index was 0.683 (95% CI, 0.605 to 0.762) for model A, 0.724 (95% CI, 0.627 to 0.821) for model B, 0.675 (95% CI, 0.587 to 0.765) for model C, and 0.679 (95% CI, 0.603 to 0.755) for model D. Notably, only model B results exhibited relatively fair agreement between predicted and observed probabilities of QTc prolongation at 8, 16, and 24 weeks of anti-TB treatment (Fig. 2). However, model B precision was lowest (Brier score, 0.088) relative to Brier scores obtained for model A (0.091), C (0.091), and D (0.092).

FIG 2.

Calibration plots of models A, B, C, and D for multivariate Cox models in the training and validation cohort. Capital letters stand for training cohort. Lowercase letters stand for validation cohort. (A, a) Calibration curves of the model A at 8, 16, and 24 weeks in the training and validation cohort. The LASSO selection method with baseline QT categorized in quartiles (<395 ms, 395 ms to 424.9 ms, and ≥425 ms) was used in model A; (B, b) model B, LASSO selection with baseline QTc squared; (C, c) model C, Backward stepwise selection with baseline QTc squared; (D, d) model D, backward stepwise selection with baseline QTc categorized in quartiles (<395 ms, 395 ms to 424.9 ms, and ≥425 ms).

Nomograms.

In the present study, the predictive nomogram was applicable to patients with a baseline QTc of ≤450 ms and QTc prolongation (as defined as a QTc interval of ≥500 ms or an increase in QTc interval of ≥60 ms over the baseline QTc value). To obtain the predicted QTc prolongation probability from the nomogram, we located patient values at each axis and then drew a vertical line to the “point” axis to determine how many points were attributed to each variable value. The resulting coefficients obtained from the Cox model B were used to create the final nomograms for use in predicting the first occurrence of QTc prolongation (Fig. 3). To obtain a predicted probability of time of first QTc prolongation occurrence from the nomograms, we further located patient values at each axis and then drew a vertical line to the point axis to determine how many points were attributed to each variable value. Finally, we located the sum on the “total points” line and then drew a vertical line to meet the 8-week, 16-week, and 24-week QTc prolongation probability axes to determine the probable time of first QTc prolongation occurrence. In the nomogram, except for the baseline QTc interval, administration of clofazimine was the most important contributing predictive factor of QTc prolongation occurrence.

FIG 3.

Predictive nomogram of QTc prolongation in multidrug-resistant tuberculosis patients with bedaquiline-containing regimens. XDR, extensive drug-resistance; Cfz, clofazimine; Lfx, levofloxacin; Mfx, moxifloxacin; Scr, serum creatinine.

Use of the nomogram to stratify patient QTc prolongation risk.

In order to evenly stratify patients in the training cohort into two risk groups, we determined the median probability from the nomogram as the boundary between the two risk groups. Each group represented a distinct level of QTc prolongation risk (incidence of QTc prolongation in training cohort: high risk group, 217/605 [35.9%], low risk group, 56/610 [9.2%]; that in validation cohort: high risk group, 23/83 (27.7%); low risk group, 6/82 [7.3%]), such that stratification could effectively discriminate between outcomes associated with the two proposed risk groups based on data obtained from both training and validation cohorts (Fig. 4).

FIG 4.

Kaplan-Meier curves of risk group stratification for QTc prolongation. Nomogram risk group stratifications for the 50 percentile are shown for the training cohort (A) and for the validation cohort (B). P value was calculated based on the log-rank test.

DISCUSSION

Clinical management of MDR-TB patients remains challenging in view of very limited available treatment options (1, 17). The introduction of BDQ has fueled renewed hope for improved treatment outcomes and for mitigating ongoing community transmission of MDR-TB (17). Nevertheless, the occurrence of adverse events has greatly undermined the achievement of favorable patient outcomes due to treatment interruptions. In this study of MDR-TB patients receiving BDQ-containing regimens in China, we derived and externally validated a prediction model that effectively stratified patients with significantly different risks of QTc prolongation (QTc interval of ≥500 ms or an increase in QTc interval of ≥60 ms compared with the baseline QTc at admission). This simple model allowed us to predict (prior to anti-TB treatment initiation) whether a given patient was at high risk of developing QTc prolongation during anti-TB treatment with BDQ. The use of this approach is based on targeted monitoring to potentially help clinicians assess the benefit-risk balance of BDQ administration. In addition, the use of this prediction model could accelerate the formulation of less toxic regimens for MDR-TB patients to reduce the incidence of QTc prolongation resulting from toxic synergy between BDQ and other QT interval-prolonging drugs.

The variables that make up our prediction model included several established risk factors for QTc prolongation that highlighted complex interactions between coadministration of administered QT interval-prolonging drugs, baseline QTc interval, and sex. In our model, the coadministration of clofazimine was associated with the greatest risk for QTc prolongation risk among MDR-TB patients (apart from baseline QTc), as consistent with previous studies conducted on TB patients receiving BDQ and clofazimine (18). In contrast, results of another reported study on the safety of clofazimine-based treatment regimens indicate that administration of this agent did not lead to significant QTc prolongation (19). These conflicting results may be explained by the fact that BDQ is metabolized primarily by the cytochrome P450 isoenzyme 3A4 (CYP3A4), which converts the drug to a less-active compound (20). Notably, significant ethnic differences in single nucleotide polymorphism (SNP) distribution of CYP3A4 have been reported that can lead to different plasma BDQ concentrations and incidence rates of QTc prolongation (21). In spite of its status as a group B drug used for the treatment of MDR-TB (11), significant QTc prolongation risk associated with clofazimine use may weaken the role of this drug in clinical management of Chinese MDR-TB patients.

Of note, having a QTc of <395 ms at baseline was a strong predictor of future QTc prolongation manifesting as an increase in the QTc interval of ≥60 ms over baseline (>425 ms). This observation may reflect the fact that the majority of patients (88/95, 92.63%) in the group with QTc of <395 ms were classified as having long QTc intervals based on QTc increases of ≥60 ms, a significantly greater proportion than was observed in the group with QTc of >425 ms. In the real world, a prolonged QTc is often a concern when the QTc value exceeds 450 ms and 470 ms for men and women, respectively (22). Therefore, the high potential for a QTc interval increase of ≥60 ms for a patient with a QTc value of <395 ms at baseline may deprive the patient of an opportunity to benefit from BDQ treatment. As an alternative, a QTc of >500 ms may be more clinically relevant than a >60-ms increase over baseline, such that for patients with >60-ms QTc increases, more intensive monitoring would likely improve clinical outcomes. Notably, the predictive model for QTc of >500 ms was based mainly on contributing factors of coadministration of clofazimine, baseline QTc interval, and initial XDR status (see Table S9 to S12 and Fig. S6 and S7 in the supplemental material).

In our patient population, female sex was a significant independent predictor of QTc interval prolongation. Previous pharmacokinetics research has revealed that BDQ preferably accumulates in adipose tissue (20), such that BDQ stored in adipose droplets may serve as a dead-end pool that maintains an effective drug concentration in extracellular fluids. Based on the well-known fact that women generally possess a higher percentage of body fat than men (23), one plausible explanation for higher QTc prolongation risk for females versus males may be attributed to relatively higher body fat associated with female patients that supports BDQ lipid partitioning. In turn, lipid partitioning leads to slow and extended BDQ release from adipose tissue into plasma, thus causing sustained high BDQ plasma concentration. However, elevated BDQ plasma concentration is a two-edged sword, whereby it improves culture conversion rates in MDR-TB patients but also increases the risk of adverse events associated with BDQ administration that may vary across populations (24, 25). Thus, additional experimental studies are needed urgently to validate pharmacokinetic differences among diverse populations.

Fluoroquinolones, BDQ, and linezolid are viewed as core agents for use in treating MDR-TB cases. Given that moxifloxacin use prolongs the QTc interval more frequently than levofloxacin, WHO guidelines recommend switching from moxifloxacin to levofloxacin to reduce additive QTc prolongation effects (26). Interestingly, our model revealed that levofloxacin use was more likely to be associated with QTc prolongation than the use of moxifloxacin when either drug was used in combination with BDQ. However, this result may be attributed mainly to the fact that a greater proportion of patients received clofazimine in the levofloxacin group (data not shown), thus emphasizing the predominant role of clofazimine in QTc prolongation.

In the present study, we developed and validated four Cox models for use in estimating the absolute risk of a first occurrence of QTc prolongation during anti-TB treatment based on baseline clinical parameters. We found that the overall performance of model B, which included seven predictors, was superior to the overall performance of the other three models. Independent predictors in model B included treatment with clofazimine, baseline QTc interval, creatinine, XDR, treatment with moxifloxacin, treatment with levofloxacin, and sex. Based on results obtained using the Cox regression model, nomograms used for predicting QTc prolongation showed excellent discrimination ability (c-index in the training cohort, 0.711; c-index in the validation cohort, 0.710). More importantly, calibration curves indicated optimal agreement between predicted and actual outcomes, thus confirming the reliability of the established nomograms. In addition, this model was based on clinical parameters without the need for sophisticated laboratory tests and thus was easy to use. We believe that the establishment of this prediction model for use in determining the risk of QTc prolongation is of clinical significance, due to its potential to help clinicians develop more individualized anti-TB regimens and to assess the risk of future severe QTc prolongation in patients treated with drug regimens that include BDQ.

In present study, we developed and validated four models for estimating the absolute risk for the first occurrence of QTc prolongation based on baseline clinical parameters. We found that the overall performance of model B including seven predictors was superior to the overall performance of other models. The independent predictors in model B included clofazimine, baseline QTc interval, creatinine, XDR, receiving moxifloxacin, levofloxacin, and sex. Based on the results of the Cox regression model, the respective nomograms for predicting QTc prolongation showed excellent discrimination ability (c-index in training cohort, 0.711; c-index in validation cohort, 0.710). More importantly, calibration curves show optimal agreements between prediction and actual observation of the studied outcomes, which guarantees the reliability of the established nomograms. In addition, this model based on clinical parameters without the need for sophisticated laboratory tests is easy to use. We believe that the establishment of this prediction model for the risk of QTc prolongation is of clinical significance, which could help clinicians develop more individualized regimens and assess the patient’s risk of severe QTc prolongation in advance when the patients’ regimens include BDQ.

Our study had several obvious limitations. The first major limitation stemmed from potential selection bias related to our patient cohort, although the likelihood of bias was reduced through the recruitment of study participants who had been enrolled in a nationwide prospective multicenter study in China. Second, we did not collect pharmacokinetic data from patients and thus were unable assess how risk factors changed with BDQ plasma concentration. Nevertheless, these data were not needed to meet our objectives, as we aimed to build a model to predict risks of adverse events prior to the initiation of anti-TB treatment. Third, the training cohort data set that was used for model development excluded individuals with baseline QTc values of >450 ms, which may have limited the generalizability of our conclusion. Fourth, the relatively small sample size of the validation cohort and the existence of other potential predictors may have adversely impacted the performance of our predictive model. In fact, in addition to the baseline variables that were incorporated in our predictive model, many other unknown factors acting during treatment may have influenced QTc prolongation occurrence. Fifth, QT baseline values varied significantly among different population groups as a reflection of heterogeneity across populations of genetic determinants associated with QT intervals (27). As an example, frequencies of genetic polymorphisms of the CYP3A4 gene have been shown to vary across populations (28, 29), with resulting diversity in BDQ metabolism possibly influencing the occurrence of prolonged QT intervals. Therefore, the nomogram developed in this work, which is tailored to the Chinese population, may require slight modifications for use in clinical applications involving other populations. Sixth, coadministration of fluroquinolone with BDQ is known to cause QTc prolongation (30), thus suggesting that the model developed here may not be useful for predicting QTc prolongation risk in approximately 40% of patients who do not receive oral fluroquinolone treatment. Finally, HIV data were not available, due to low HIV prevalence in our patient population. Therefore, risk scores obtained using the model developed here may not reflect QTc prolongation risk in this special population. Nevertheless, the model developed here should be used to identify BDQ-treated patients at high risk for QTc prolongation in order to ensure that these patients receive more frequent QTc interval testing followed by implementation of early interventions that can be used to prevent adverse cardiac events. In addition, our results revealed that concurrent administration of clofazimine with BDQ is a major risk factor for QTc prolongation. Thus, clinicians should carefully evaluate risks and benefits associated with the inclusion of clofazimine in BDQ-containing anti-TB regimens before this drug is used to treat patients at high risk of QTc prolongation.

In conclusion, this is a first modeling study to quantify the risk of QTc prolongation in patients treated with BDQ-containing regimens. The model developed here demonstrated that coadministered clofazimine and abnormal baseline QTc interval significantly contributed to the development of QTc prolongation, especially for the group with QTc values of >500 ms. Ultimately, our model offers a practical tool for use in evaluating MDR-TB patients to identify those at high risk of developing QTc prolongation, while also providing important information to guide the optimization of MDR-TB patient treatment regimens.

MATERIALS AND METHODS

Training cohort.

The training cohort for nomogram development was assembled using patients who were enrolled in the New Drug Introduction and Protection Program (NDIP) in China between January 2018 and December 2019. Briefly, patients recruited by the NDIP who were infected with MDR-TB were identified based on results of 65 pilot hospitals (9). In order to optimize the BDQ-containing regimen administered to each patient, clinical tests were conducted to assess in vitro MTB drug sensitivity to BDQ, linezolid, fluroquinolones, clofazimine, and other potentially efficacious anti-TB drugs. BDQ was administered for 24 weeks in addition to other anti-TB drugs included within personalized anti-TB regimens that were administered for durations ranging between 13 and 30 months as based on initial drug susceptibility profiles and follow-up culture results. Patient inclusion criteria included the following: (i) MDR-TB diagnosis as confirmed by drug susceptibility testing, (ii) failure to respond to current MDR-TB regimens lacking BDQ, and (iii) 18 years of age or older. Reasons for exclusion included the following: (i) pregnant or breastfeeding; (ii) QTc interval greater than 450 ms; (iii) history of risk factors for prolonged QTc interval; and (iv) concomitant serious illness, including alanine aminotransferase/aspartate aminotransferase (ALT/AST) of >3× upper limit of normal (ULN) or total bilirubin of >2× ULN, creatinine clearance of <30 mL/min, hemoglobin of ≤7.0 g/dL, and/or platelets of <50 × 10⁹/L at screening. After the enrollment of patients in our training cohort, clinical visits were scheduled at baseline and at time points after treatment initiation (weeks 2 and 4 and once every 4 weeks thereafter). At each visit, each patient was subjected to a physical examination, an electrocardiogram, a routine blood count test, a biochemical test, a urinalysis test, and bacteriological assessments of sputum specimens. In addition, some patients received extra electrocardiogram testing as deemed necessary by their physicians. A standardized data collection form was designed to retrieve all relevant sociodemographic data at baseline (age, sex, smoking history, body mass index [BMI], previous TB history, and household TB registry data), baseline laboratory data (e.g., drug resistance and levels of serum potassium, serum calcium, creatinine, serum magnesium, serum phosphorus, ALT, and AST), and background anti-TB treatment regimen details (e.g., levofloxacin, moxifloxacin, linezolid, clofazimine, and cycloserine).

External validation cohort.

To examine the generalizability of the model, an external validation cohort of 165 consecutively enrolled MDR-TB patients was studied. Patients who were selected for this cohort received treatment with BDQ-containing anti-TB regimens at Beijing Chest Hospital (affiliated with Capital Medical University) between January 2019 and December 2020 and provided sufficient data to score all variables needed to establish the nomograms. The study protocol was designed in accordance with guidelines outlined in the Declaration of Helsinki and was approved by the Ethics Committee of Beijing Chest Hospital, Capital Medical University.

Follow up.

The main outcome of interest that was assessed in this study was based on the first occurrence of a QTc prolongation event, as defined as a QTc interval of ≥500 ms or an increase in QTc interval of ≥60 ms compared with the baseline value at admission (31). Here, we used Fridericia’s formula to calculate corrected QT duration (QTc). All follow-up data were uploaded to the active drug safety monitoring (aDSM) system. After the enrollment of patients in our cohort, clinical visits were scheduled at baseline and at weeks 2 and 4 after treatment initiation and once every 4 weeks thereafter. Based on results obtained from scheduled QTc interval assessments, patients were sometimes asked to submit to additional electrocardiogram testing as deemed necessary by their physicians. Follow-up care was scheduled after referring to clinic attendance records, with follow-up care duration defined as the period between enrollment and either the first QTc prolongation event, death from any cause, date of loss to follow up, or the date of completion of the 24-week BDQ treatment. Events occurring after the 24-week treatment were ignored.

Statistical analysis.

QTc prolongation time was defined as the time from the start of administration of a BDQ-containing regimen to the time of first appearance of QTc interval prolongation (QTc interval of ≥500 ms or an increase in QTc interval of ≥60 ms compared with the baseline value at admission). In the training data set, survival curves for training and validation cohorts were plotted based on Kaplan-Meier estimates and compared using a log-rank test.

We built the models in four steps. First, all baseline variables that were considered clinically relevant or that showed a univariate relationship with patient outcome were included in multivariate analyses (see Table S3 in the supplemental material). Second, to improve the degree of curve fitting obtained using the model, we used model χ² to find the best transformation for determining nonlinear continuous predictors (see Table S2 in the supplemental material). As a result, two transformations with good fit for the baseline QTc were considered. Third, in order to improve simplicity, backward stepwise selection was conducted using the Akaike information criterion (AIC), while least absolute shrinkage and selection operator (LASSO) were used to extract the most useful predictive features from the training cohort data (32, 33). During the process of variable selection, two binary variables (receiving levofloxacin: yes versus no; receiving moxifloxacin: yes versus no) were forcibly incorporated within the model, due to their clinical correlations with QTc prolongation. Finally, considering the fact that 103 of 1,215 (8.48%) patients were lost to follow-up monitoring, four Cox proportional hazards models were developed, as follows: model A, backward stepwise selection with baseline QTc categorized in quartiles (<395 ms, 295 ms to 424.9 ms, and ≥425 ms); model B, LASSO method with baseline QTc categorized in quartiles (<395 ms, 295 ms to 424.9 ms, and ≥425 ms); model C, backward stepwise selection with baseline QTc squared; and model D, LASSO selection with baseline QTc squared (see Fig. S2 in the supplemental material). Hazard ratios (HRs) were presented with 95% confidence intervals (CIs). Selected variables were incorporated within the nomograms to predict the probability of QTc prolongation at 8, 16, and 24 weeks of anti-TB treatment. For allocating points in the nomograms, regression coefficients were applied to each individual observation to define the linear predictor.

The discrimination ability of each model for predicting QTc prolongation was measured using concordance index (c-index) values as reported by Harrell et al., whereby the c-index is equivalent to the area under the curve (AUC) of the receiver operating characteristic curve of censored data (34, 35). A maximum c-index value of 1.0 indicates a perfect separation between results of groups of patients with different outcomes, while a c-index of 0.5 indicates the absence of discrimination between groups. After c-index values were determined for the four Cox models, the degree of optimism of each model was determined and then each model was validated internally and externally via bootstrapping based on 1,000 samples to correct the concordance c-index and explain overoptimism variance. Next, we estimated the drop-off of the special c-index value when comparing the performance of fits of bootstrapped samples to the performance of the bootstrap-derived model using the original training data sets (32, 34). A bootstrapped sample from the study group was used to perform calibration using a calibration plot, a graphic representation of the relationship between observed outcome frequencies and predicted probabilities. In a well-calibrated model, predictions should fall on a 45-degree diagonal line. Calibration of the nomogram for use in predicting QTc prolongation time based on 8, 16, and 24 weeks of treatment was performed by comparing the predicted survival to observed survival. Overall goodness-of-fit of each model was assessed for training and validation cohorts using the Gronnesby and Borgan test (see Table S5 in the supplemental material). Overall, the performance of the prediction model was quantified and then expressed as a scaled Brier score, which measures the average squared deviation between observed and predicted outcomes in order to achieve both calibration and discrimination, with a lower score corresponding to greater accuracy (32, 35, 36). The four models were compared using the rcorrp.cens function of Hmisc in R (see Table S4 and Fig. S4 in the supplemental material). After comparing the performance of models using both training and validation data sets, the best model was identified. During external validation of the nomograms, the total points for each patient were calculated according to the nomograms, and then Cox regression was performed using the value of total points as a predictor for the validation cohort. In addition to conducting numerical comparisons of the discrimination ability using the c-index, we also demonstrated the independent discrimination ability of each nomogram for training and validation cohorts. After grouping patients evenly into two risk groups (≥P50, high risk group; <P50, low risk group) based on nomogram scores in both the training and validation cohorts, we analyzed the predictive abilities of the models using the abovementioned risk category cutoffs as assessed using Kaplan-Meier survival curves. A two-sided P value of <0.05 was deemed significant.

Considering that missing data in covariates can prevent the robust detection of associations, a sensitivity analysis was conducted using the multiple imputation analysis. Five imputations based on results of the multivariable regression analysis using covariate and outcome data were used to provide missing values for diabetes (n = 52; 4%), serum calcium (n = 8, 0.6%), baseline QTc interval (n = 3, 0.2%), and serum potassium (n = 1, 0.1%), with the performance of the four models evaluated using multiply imputed data sets (see Table S6 and S7 and Fig. S5 in the supplemental material).

Statistical analyses of survival data were conducted using SAS 9.4. Nomograms were formulated to provide visualized risk prediction based on results of multivariate analyses obtained using R 4.0.5 (http://www.r-project.org) survival and rms packages. The calibration plot was created using the R package pec. The MICE package was used to provide to substitute missing data with plausible data.