|Home | About | Journals | Submit | Contact Us | Français|
Thorough QT (TQT) studies are designed to evaluate potential effect of a novel drug on the ventricular repolarization process of the heart using QTc prolongation as a surrogate marker for torsades de pointes. The current process to measure the QT intervals from the thousands of electrocardiograms is lengthy and expensive. In this study, we propose a validation of a highly-automatic QT interval measurement (HA-QT) method. We applied a HA-QT measurement method to the data from seven TQT studies. We investigated both the placebo and baseline-adjusted QTc interval prolongation induced by moxifloxacin (positive control drug) at the time of expected peak concentration. The comparative analysis evaluated the time course of moxifloxacin-induced QTc prolongation in one study as well. The absolute HA-QT data were longer than the FDA-approved QTc data. This trend was not different between ECGs from the moxifloxacin and placebo arms: 9.6±24msec on drug and 9.8±25msec on placebo. The difference between methods vanished when comparing the placebo-baseline-adjusted QTc prolongation (1.4±2.8msec, p=0.4). The differences in precision between the HA-QT and the FDA-approved measurements were not statistically different from zero: 0.1±0.1msec (p=0.7). Also, the time course of the moxifloxacin-induced QTc prolongation adjusted for placebo was not statistically different between measurements methods.
The US and European regulatory agencies recommend the evaluation of the propensity of a new drug to induce prolongation of the rate-corrected QT interval (QTc). In United States, the FDA requires Thorough QT (TQT) studies to be implemented in order to assess the propensity of a drug to have proarrhythmic effects by delaying the ventricular repolarization process of the heart. Today, the design and measurements implemented in TQT studies are well documented and understood even if the value of drug-induced QTc prolongation as a surrogate marker for risk of TdP remains controversial.1
To evaluate the sensitivity of clinical studies to detect small but significant increases in QT/QTc interval measurements, regulatory authorities expect the TQT study design to include a positive control drug. The common drug used is moxifloxacin, a synthetic fluoroquinolone antibiotic agent. As consistently shown in preclinical 2, 3 and human 4–11 studies, moxifloxacin slightly blocks the rapid component of delayed rectifier repolarizing current (Ikr) of myocardial cells. This induces a prolongation of the QTc interval, mainly located in the early part of the T-wave of the ECG.12 While being accepted as a safe molecule, moxifloxacin has been shown to trigger TdPs in individuals with underlying risk factors.13, 14 However, in healthy individuals moxifloxacin produces a “safe” QTc interval prolongation between 5 and 15 msec at peak plasma concentration for a single dose of 400 mg.
Recommendations for the implementation of TQT studies are described in the E14 documents of the International Conference on Harmonization (ICH) and its associated Q&A document.15 Interestingly, the E14 document does not include limitation about the definition of the technologies used to measure the QT interval. Three classes of techniques are described: the automatic, manual, and the manual adjudication (manual over-read, computer assisted and semi-automatic). Only the two latter methods are recommended in the E14 document and were reported in published TQT studies.16 ICH E14 document does not recommend the fully automatic approach because of the reported reduced performance of most algorithms when either the electrocardiogram (ECG) is polluted with noise (high and low-frequency components),17 the T-wave has a low amplitude or a U-wave merges with the T-wave.18, 19 Evidently, the manual adjudication represents a logical trade-off between fully-manual and fully-automatic methods. Yet, it raises questions about the balance between human and computer-based contributions to minimize resources while maximizing measurement precision. Also, it is noteworthy that QT measurements technologies are evolving rapidly 20 in both clinical 21, 22 and drug safety 23–26 arenas.
In this study, our group has investigated data from a set of seven TQT studies submitted to the FDA. First, we conducted a comparative study to assess the validity of highly-automatic QT measurement method in which manual review is minimized while computerized-process for quality assessment and QT measurements are “fully” exploited. It is important to note that the highly-automatic QT measurements are compared to the QT measurements submitted to the FDA which are also based on automatic QT method but all these measurements have been manually reviewed and adjudicated (so-called semi-automatic method). Therefore, we implemented this study in order to address two methodological aspects of TQT studies. First, can a highly-automatic technique measure and qualify moxifloxacin QT signal in TQT studies? Second, what are the levels of variation of QT/QTc interval measurements across studies one may expect when using such QT measurement strategy? In addition, we will provide a description of sex-related differences on the QTc interval measured in these TQT studies.
Data included in the analyses are de-identified ECG waveforms for moxifloxacin and placebo arms, as well as demographic information from 351 healthy individuals enrolled in seven TQT studies with crossover design. The studies were selected from the FDA warehouse based on willingness from pharmaceutical companies to share their data for research purpose. The pharmaceutical companies agreed on sharing their data with the Agency but requested to remain anonymous. Then, the data shared by the FDA enabled us to evaluate the placebo-controlled moxifloxacin-induced QT/QTc interval prolongation at expected peak concentration of the drug (2 hours post dosing) in six studies, and the time profile of the QTc interval within one study, the so-called study 7.
The ECGs from the baseline and the placebo arms were recorded at the same time of the day than in the moxifloxacin arm in order to control for circadian effects. Intake time of the placebo was synchronized with intake time of moxifloxacin. The digital ECG files and the respective information (age, gender, and study treatment) were provided by the Center for Drug Evaluation and Research of the FDA to our organization. The HL7 XML ECGs files were sent to the University of Rochester by Mortara Instrument (Milwaukee, MN) following FDA request and were extracted from the ECG warehouse after FDA ensured that pharmaceutical companies agreed on having their ECG used for research purpose. Any information that could help identifying the source of the data or reveal the name of organizations involved in the study (pharmaceutical companies, ECG core laboratory, etc.) was removed from files send to University of Rochester. Information about the methods adopted by ECG core laboratory to measure the QT intervals were not available. However, the FDA (CG, ML) provided a code to identify the five different ECG core laboratories (ECL) involved in the QT measurements of these seven studies (ECL A to E).
The descriptions of the methods used by ECL to measure QT interval were not available, consequently our HA-QT measurement method was compared to “current ECL methods” rather than a specific method. We expect that the ECLs involved in these seven TQT studies used semi-automatic QT method i.e. a computer algorithm identifies the end of the T-wave and a cardiologist or ECG technician visually assesses the end of T-wave and provides adjudication if needed. Most likely for each tracing, the QT measurements were based on three beats with adjudication based on on-screen tools such as the tangent method or on method using a global lead superimposition.
The HA-QT and RR intervals measurements were based on the technology developed at the University of Rochester Medical Center, NY. The COMPAS software provided the identification of the end of the T-wave based on a technique identifying the crossing-point between the baseline and the descending slope of the T-wave (least-squares technique).27 QT intervals were determined to be unreadable, preventing the measurements from being done, when the T-wave amplitude was inferior to 50 μV (flat T-wave). The QT interval measurements were computed in all available sinus beats. The median QT value from all measured cardiac beats was reported after applying rate correction formula described below.
The QT algorithm included a set of pre-processing steps such as: baseline estimation and removal based on spline interpolation, a low-pass filter with finite-impulse response of 20th order (FC=25 Hz) was designed. Finally, we developed a least-square method based on a moving-window seeking for the maximum fitting slope within the terminal part of the T-wave (independently from the identification of the end of the T-wave).
The QT measurements from University of Rochester were sent to the FDA where they were matched with the QT measurements submitted by pharmaceutical companies. If a missing QT measurement was found in one of the dataset (UR vs. FDA) then the corresponding data point was removed from the other dataset.
A quality assessment process involving a two-level filtering process was designed to identify QT and RR interval measurements outside of their “expected” range. The identification of these outlying values was based on the following two criteria:
All ECGs associated with measurements fitting at least one of these two criteria were flagged and visually reviewed and then accepted, corrected or rejected in all studies but study 7.
Outlying measurements from the study 7 were identified using the criteria describe above (similar to other studies). However, we did not include manual adjudication of QT measurements rather we rejected the outlying measurements from the analysis when the end of the T-wave was flawed. We had to change this review process because of the limited resources we had to analyze this second set of 8,911 ECGs initially not included in our research plan. Data from the study 7 were sent by the FDA to the University of Rochester at later time in order to investigate the moxifloxacin-induced QT time profile.
In summary, the QT measurements compared in this work are called the HA-QT, and the FDA-submitted QT. The HA-QT measurements are automatic QT measurement (based on COMPAS method) on which limited manual review based on quality metrics (described above) were done. All available beats in each ECGs are measured and median value from all beats is reported for each ECG tracing. The FDA-submitted QT measurements are expected to be based on automatic QT measurements with 100% manual review of the measurements. Most likely, different methods were used to measure automatically QT interval before manual review but these methods were not documented. The number of measurement manually adjudicated was not available either. Most likely, the QT measurements were based on three selected beats in each ECG tracing. Eventually, median value from the medians QT from the triplicate ECGs (when available) was computed.
Heart rate correction of the QT interval duration remains a challenging aspect of the assessment of QT-prolonging effect of drugs. First, it is known that QT/RR profiles are subject dependent and can be modified by drugs. Moxifloxacin does not significantly affect the heart rate in human; therefore, the issue of bias due to QT dependency to heart rate is not of strong concern in this analysis. Nevertheless, the Bazett’s formula is known to generate large biased and reduced QTc measurement precision so we have excluded this correction method from our analysis.29, 30 We opted for comparing QTc using Fridericia’s formula 31.
The assay sensitivity for a QT measurement technique in thorough QT studies is demonstrated when two criteria are met: 1) the lower two-sided 90% interval of the baseline adjusted and placebo controlled population-based mean of QTc changes induced by moxifloxacin is above a 5 msec threshold (at the time of expected peak concentration) and 2) the QTc profile across time follows the expected plasma concentration of the drug (moxifloxacin).
In our study, the QT interval was measured in all available beats and the median value was computed to yield one value for each ECG. Since most individuals had triplicate ECGs, the median of the available replicates was calculated for each time point. Single time-matched delta QTc durations were computed as follow:
To account for the placebo effect, ΔQTcplacebo was subtracted from ΔQTcmoxi ΔΔQTc = ΔQTcmoxi–ΔQTcplacebo for each individual. The mean Δ ΔQTc was calculated for each time point. Importantly, the QTc at baseline was the time-match value when assessing the QTc prolongation at expected peak concentration whereas the QTc at baseline was measured from the pre-dose ECG when evaluating the QTc profile in study 7. These methods follow the recommended analytic strategies adopted by the FDA.32
The population-based mean Δ ΔQTc prolongations between methods at expected maximum concentration of moxifloxacin and their two-sided 90% confidence interval are reported. T-test were used to compare the population-based average of QT interval between genders, a level of statistical significance (p-value) was set to 0.05.
When comparing the curves describing the QT time profile between the QT methods (for study 7 exclusively), the primary analysis was based on a mixed-effects linear model for repeated measures of the differences from baseline in ΔQTc between the two measurement techniques. The model had no intercept and used an unstructured covariance structure as this model showed the best goodness of fit.
The QT measurements realized at University of Rochester were sent to the FDA (to ML and CG) to implement the comparative analysis between HA-QT and the measurements received by the Agency. Thus, most of the statistical analysis was implemented by ML and GC independently from the University of Rochester group.
All results described in this section are based on highly-automatic QT measurement method realized from Lead II; if otherwise, relevant information is provided. Table 1 provides a brief description of the seven TQT studies. All studies were single-center, randomized, double-blind, placebo-controlled, active-comparator, four-way crossover studies. As shown in Table 1, all studies were single dose but study #2 and #6 which used multiple doses, 2 and 4, respectively. . The QT interval measurements from studies 3, 5, and 7 were from a similar ECG core laboratory (ECL B) while all others were from different ECL (A, C to E). No precise information related to the QT method used in these TQT studies were reported to the Agency. The overall studies included 381 healthy individuals. The ages of subjects across studies were different (p<0.001). Study 6 was based on multiple dose (n=4). Study 2 had only one single ECG for each time point while all other studies included triplicate ECGs. This study included two doses given at different days of the protocol and the data from these two days were merged in our analysis. All ECGs were recorded before dosing and information about time interval between replicate ECGs was not provided. Our investigation includes the data from a total of 13,425 ECG tracings amongst those, 8,911 were from study 7 in which we investigated the moxifloxacin-induced Δ ΔQTc time profile.
The quality assessment processes has led to identifying 216 ECG recordings with outlying QT or RR values in studies 1 to 6 (6.1% of ECGs) and 254 were identified in study 7 (5.2%). From the first set, 50 ECGs were manually adjusted (1.4%), 20 (0.6%) were rejected and the rest was not modified. In the study 7, no ECGs were manually adjusted and 200 ECGs were rejected (4.1%).
Figure 1 illustrates the mean absolute values of QTc interval measurements (rate corrected using the Fridericia formula) and their standard error for the seven TQT studies. When pooling all studies, the mean QTc durations were 408±19 msec in the baseline placebo arm, 404±20 msec in the placebo treatment, 407±19 in baseline moxifloxacin arm and 419±21 msec in the moxifloxacin treatment arm (417±20 msec when excluding the study at drug steady state, study 6).
The difference in absolute QTc interval duration between the moxifloxacin treatment and baseline arm is 23 msec for the study including only females (study 5: 442±21 vs. 419±21 msec), 21 msec for the study at drug steady state (study 6: 427±24 vs. 406±21 msec) and 8 msec in average for all other studies. In study 6, the variation from treatment to baseline arms for the moxifloxacin arm is almost three fold larger than in the other studies including both male and female subjects.
Figure 1 revealed a longer QTc in study 5 including only females than in the other studies for all arms. We investigated the gender-related difference for the absolute QTc values, for ΔQTc by treatment, and for Δ ΔQTc by merging all ECGs from the seven studies. Mean values are provided in Table 2. Females show larger moxifloxacin-induced QTc prolongation before (434±21 vs. 401±19, p<0.01) and after adjustment for baseline (16±14 vs. 7±13 msec, p<0.01), as well as after controlling for placebo effect (17±16 vs. 12±15, p=0.04).
The QT interval measurements from computerized techniques are known to provide different results depending on the approaches used to process the ECG signal, to remove the baseline wander, to increase signal-to-noise and to identify the end of the T-wave.33, 34 Today there is no “gold standard” for QT measurement methods. We used the measurements submitted to the FDA as reference even though each study used a different method to measure the QT interval. All these methods have commonly been validated (appropriate assay sensitivity) during the regulatory submission of the study. We will call these measurements the “FDA-approved” QT intervals.
First, we used Bland-Altman plots35 to investigate the level of agreements between HA versus FDA-approved QT measurements. Figure 2 demonstrates that the HA method is associated with an approximately 10 msec longer QT interval than FDA-approved data. The level of agreement between HA and the FDA-approved measurements are consistent for the placebo and moxifloxacin arms: mean difference is 9.6 and 9.8 msec, and the limits of agreement (95% CI) are 25 and 24 msec, respectively. Figure 2 was based on 4,019 ECG tracings. The figure suggests that HA-QT method does report consistent difference with FDA-approved methods on and off the drug (moxifloxacin).
The differences in absolute values of QTc describe above means that the identification of the end of the T-wave is different between methods. The Δ ΔQTc at two hours post-dosing were evaluated in the six studies and mean results are reported in Table 3. The largest difference in ΔQTc between methods was in study 5 with a 6.6 msec larger effect for the HA method compared to the FDA-approved results. This difference was peculiar. In the other studies, differences between methods ranged between −1.2 to 3.4 msec. Two studies revealed shorter Δ ΔQTc in HA method (studies 2 and 4) while this difference was positive in the remaining four studies. The average difference in population-based mean Δ ΔQTc was equal to −1.4±2.8 msec. The average difference was not statistically different from zero (p=0.4).
In term of precision, the standard error (SE) of Δ ΔQTc between HA and FDA-approved QT measurements were similar too. The averaged paired difference of SE was equal to −0.1±0.1 msec. The average difference was not different from zero (p=0.7).
The results from Table 3 are illustrated in the panel C of Figure 3. The panel A and B show the mean (90% confidence interval) for ΔQTc for the placebo and moxifloxacin arms, respectively. There is no statistical difference between FDA-related ΔQTc values and the ones computed at University of Rochester using the highly automatic techniques.
The analysis of the data from Study 7 focused on the time profiles of ΔQTc for placebo and moxifloxacin arms as well as the time profile for Δ ΔQTc. The results are provided in Figure 4. The ΔQTc values were computed using pre-dosing QTcF values instead of time-matched baseline values.36 The plots provide consistent results with the analysis of the previous section. No statistical difference was found between the profiles of ΔQTc and Δ ΔQTc based on mixed-effects linear models (p=0.7 when testing if the profiles were different). The ΔQTc between FDA-approved and HA QTcF at T=12 in the placebo arm is peculiar, more than six msec difference between methods is found for this specific time point. We did not have access to the annotated QT waveforms from the FDA and were not able to identify the source of this discrepancy. Therefore, we checked the measurements from the HA-QT method for this specific time point but we could not identify any obvious failure of the algorithm or erroneous measurements.
Our study investigates the ability of a highly automatic QT measurement method to detect the moxifloxacin signal under TQT study design. We implemented a thorough comparison between the results from the HA method developed at University of Rochester Medical Center and the QT measurements submitted by pharmaceutical companies to the FDA.
We reviewed the list of TQT studies reported in the literature and we identified the ones including results about their positive control arm. We identified nine reports describing ICH E14-based TQT studies. They were TQT studies investigating levoceretizine 37, tolterodine 38, brivaracetam 39, levetiracetam 40, rivaroxaban 41, and maraviroc 42. The studies by Davis et al. 43, Malhotra et al. .44 and Darpo et al.45 included (or focused) on assay sensitivity between fully-automatic and semi-automatic QT measurements. In E14 TQT studies, assay sensitivity is established if the ΔΔQTc at one time point is greater than 5 msec, as measured by the lower 1-sided 95% confidence bound. Most studies found in the literature were crossover design (XO) and based on a single moxifloxacin dose of 400 mg. The reported QTcF is the Δ ΔQTcF for moxifloxacin in comparison to placebo after baseline adjustment (time-matched) measured at either maximum drug concentration or its expected time (2 hours after dosing). All published studies reported a lower one-sided 95% CI above the 5 msec threshold. The range of mean double delta QTcF varies between 7.7 and 14.0 msec (mean=11.9 msec). The average size of the 2-sided 90% CI is equal to 6.0 msec ranging from 3.8 to 7.6 msec. Within the studies evaluated in our investigation, the mean ΔΔQTcF across studies is equal to 11.9 msec, 10.0 msec being the smallest and 14.3 msec the largest (for similar type of studies). The average 90% CI is 3.9 msec (2.5 to 6.0 msec). Consequently, the results we report in our investigation are consistent with current description. As a comment, we obtained a slightly lower two-sided 90% confidence interval than the ones reported in the literature (6.0 msec vs. 3.9 msec).
Interestingly when we investigated QTc profile, we observed a ΔΔQTc profile from the HA method slightly higher than the FDA-approved data in a several time points and mainly located in the descending portion of the profile. The explanation for this difference remains speculative but one would emphasize that moxifloxacin could be associated with the small but significant reduction of T-wave amplitude reported by our group in one TQT study.12 Such changes should lead to a flattening of the slope used for the identification of the end of the T-wave leading to a larger QT interval. Such explanation could not be checked but represent a likely explanation. One would note that the QT interval in study 7 were measured using a semi-automatic method that could have been based on the tangent method too. As a reminder, such trend was not observed in the data from the other studies (studies 1, 2 and 3).
As a note, one would expect the 90% confidence interval for a given method to be decreasing when the number of enrolled individuals increases. Based on Table 2, this expectation is neither met in the FDA-approved data nor in the highly-automatic method. This observation reveals that other confounding factors have an impact of the precision of QT interval measurements, amongst them ECG quality might play an important role and population selection bias could be involved as well.
As we discussed earlier the use of computerized QT measurement techniques in TQT studies is of paramount importance because it would greatly simplify and streamline the processes around ECG analysis in drug safety trial that could cascade to earlier evaluation phase 1 or later post-market evaluations (patients home monitoring). Specifically in TQT studies, the consequences of the use of more computerized methods are two-fold. First, there is an obvious impact linked to the reduction of resources and time needed to measure the thousands of tracings recorded in these safety studies, highly-automatic method represents a very cost-effective process. Second, automatic or highly automatic methods can process more cardiac beats. With more than 80% of TQT studies relying on continuous telemetric or Holter recorders, automatic QT offers an opportunity for thoroughly investigating the ECGs from TQT studies. The use of beat-to-beat QT assessment of long-term recordings in drug safety trials could help addressing the current challenges related to drug that modifies cardiac regulation.46–50
Our analysis and review of literature of TQT studies reveals that moxifloxacin-induced QT prolongation is easily captured using manual, semi-manual and, as reported by several authors, fully-automatic methods.51–54 The highly automatic method is presented in this work as a new alternative relying primarily on the performances of an algorithm while directing human interaction toward the most suspicious measurements. Furthermore, we demonstrate that with a very limited human review (6.1% of the ECGs) and adjudication (1.4% of the ECGs), we reach the same precision than the methods currently used in FDA-approved trials. Based on study 7, the assay sensitivity of our QT method is fully established. Finally, increased acceptance of computerized methods in TQT studies will create an opportunity for evaluating additional ECG markers that can easily be included in these algorithms. For instance, several groups are currently investigating the role of T-wave morphology in drug safety trials.55, 56 Our group evaluated the interest of looking at T-loop morphology in thorough QT studies and a preliminary investigation described T-loop morphology indices as more precise than QT interval measurements. Therefore, one may expect that the combination of QTc interval duration and other ECG markers might in the future help drug-safety assessment.12, 57
Finally, our gender-based analysis showed that all QTc measurements are significantly different between males and females. First, the baseline ECGs present longer QTc interval duration but also the response to moxifloxacin is stronger in women. It was expected that moxifloxacin-prolongation of the QTc interval is longer in women than in male based on current clinical evidence. The reason being two fold: women have longer baseline QTc interval duration 58 and their smaller body size might be associated with higher drug concentration (same drug intake).
In our analysis, we did not compare single beat QT interval measurements between methods rather we compared the average QT interval measurement for a given ECG tracing. One would emphasize that the mean QT interval measurement for a given ECG could have been realized on a different number of beats and may generate noise in the comparative analysis.
The criteria used to identify “abnormal” values for the QT and RR interval measurements were not evaluated to optimize the precision of the QT interval while minimizing the human interaction. One may emphasize that a sensitivity analysis would be required to provide optimal thresholds for these criteria. In addition, we did not have access to the QT measurements used as a reference namely, the FDA-approved QT. Access to this data would have help us to better understand some of the minor differences we have pointed out in our analysis. Also, our statistical approach for the identification of outlying values is based on data from the overall seven studies consequently the reported percentage of measurements to correct may not be representative of the percentage for a given study, the quality of the data may not be the same in all studies. Finally, the validation of the QT method describe in this report is valid in the context of ECGs recorded during TQT studies in which ECGs were recorded in healthy individuals with normal T-wave shapes. The ability of the method to provide same level of precision in ECG from cardiac patients with abnormal T-wave morphology is not demonstrated in this report.
In conclusion, our analysis is a unique validation study in which a highly-automatic QT method was compared to a large set of FDA-approved QT measurements from seven TQT studies. Our investigation suggest that automatic methods can be used in TQT studies. Interestingly, the FDA does not exclude the use of highly automatic or fully automatic methods to measure the QT interval from surface ECGs in TQT studies as long as the method used can demonstrate assay sensitivity. Yet, these methods are not broadly used even if their precision has been increasingly demonstrated to be appropriate in various independent studies. As far as the authors know, our analysis is the largest validation ever conducted in collaboration with FDA for evaluating the validity of a QT measurement technique in TQT studies.
This work was funded by the National Institute for Health through the R01HL084402 award.
We would like to thank Dr. Norman Stockbridge, Director of the Cardiovascular and Renal Products of the Center for Drug Evaluation and Research of the Food and Drug Administration for his continuous support.