Concentration-QTc Analysis In A Thorough QT Study

Why Model Concentration and QTc Together?

A thorough QT (TQT) study collects repeated electrocardiogram measurements and drug concentrations from participants, often alongside a placebo group and a positive control. The central exposure-response question is not merely whether QTc changes after dosing. It is how the mean QTc change relates to concentration, while accounting for baseline QTc, scheduled measurement time, and repeated observations from the same participant. That relationship can then be used to predict the change at clinically relevant concentrations, including exposures that may be uncommon in the observed data.

Here, QTc denotes a heart-rate-corrected QT interval; QTcF is one common choice, using Fridericia’s correction, \(QTcF=QT/RR^{1/3}\) when QT and RR use compatible time units. The analysis below treats the QTc endpoint as measured in milliseconds. Correction does not guarantee that heart-rate effects or measurement issues have disappeared, so the endpoint and its diagnostics still matter.

Data and Model Definition

Let \(QTcF_{i,\mathrm{base}}\) be participant \(i\)'s prespecified baseline QTcF, and let \(QTcF_{ij}\) be the post-baseline QTcF at observation \(j\). Define the outcome as the change from baseline, \(Y_{ij}=QTcF_{ij}-QTcF_{i,\mathrm{base}}\), in milliseconds. Let \(C_{ij}\) be the drug concentration measured at the corresponding visit, in a stated concentration unit, and let \(t_{ij}\) identify the scheduled time. Placebo observations have drug concentration zero in this model. If concentration is centered at a reference other than zero, predictions must instead compare the target concentration with that reference.

This model uses a categorical time effect so that scheduled time-related changes need not follow a straight-line time trend. It assumes a linear mean relationship between concentration and QTc change over the concentration range being interpreted. If drug and placebo observations are collected at the same scheduled times, the time effect captures shared time-related changes, while the concentration term estimates the exposure-related component. The model does not require \(\sigma_b^2\) to be positive. When \(\sigma_b^2>0\), observations from the same participant are correlated under these assumptions; when \(\sigma_b^2=0\) and \(\sigma_\epsilon^2>0\), their correlation is zero.

The random intercept represents a participant’s persistent departure from the population mean across visits. It is not a substitute for modeling a changing individual concentration response: if participants plausibly have different concentration slopes, a random-slope structure may be considered, provided the data support its estimation. The model also assumes that the baseline and time adjustments are appropriately specified, that concentration and QTc are paired to the correct visits, and that missing observations do not create an unaddressed bias in the target estimand.

What the Concentration Coefficient Predicts

For a fixed baseline value and scheduled time, compare the model mean at concentration \(C^*\) with the model mean at placebo concentration zero. The baseline, time, and population-average random-effect terms are identical in the two predictions, so they cancel:

Thus the model’s placebo-adjusted mean QTc change at concentration \(C^*\) is \(\beta_C C^*\), in milliseconds. This is a population mean contrast at the specified exposure and time, not a prediction that every participant’s QTc changes by that amount. If the analysis includes a concentration-by-time interaction, the slope can vary by time and the contrast must use the coefficient for the relevant time. If concentration is centered at \(C_0\), the contrast is \(\beta_C(C^*-C_0)\), not \(\beta_C C^*\).

Uncertainty and a Clinically Relevant Contrast

A point prediction alone is not enough for interpretation. If \(C^*\) is treated as fixed and the fitted model is linear in concentration, the estimated placebo-adjusted mean change is \(C^*\hat\beta_C\). Its standard error is \(|C^*|SE(\hat\beta_C)\), because multiplying an estimate by a fixed constant multiplies its standard error by the absolute value of that constant. A confidence interval can then be formed using an appropriate critical value and degrees of freedom from the model-fitting method. The interval describes uncertainty in the population mean contrast under the model; it is not a range expected to contain a specified proportion of individual responses.

Repeated Measurements and Model Checks

Under the stated random-intercept model, two distinct observations from participant \(i\) share \(b_i\). Their covariance is \(\sigma_b^2\), because their residuals are independent. The variance of each observation is \(\sigma_b^2+\sigma_\epsilon^2\), so the within-participant correlation is

This expression is defined when the total variance is positive. It equals zero if \(\sigma_b^2=0\) and \(\sigma_\epsilon^2>0\); repeated observations are not necessarily correlated just because they come from the same participant. If residuals have additional serial correlation, the simple random-intercept covariance structure is inadequate and must be reconsidered.

Diagnostics should examine whether the model’s assumptions are plausible for the data and the intended prediction:

• Mean relationship. Plot residuals against concentration and compare observed summaries with the fitted concentration trend. Curvature or systematic departures can indicate that a straight-line slope is inadequate.
• Time structure. Inspect residuals and fitted changes by scheduled time. A categorical time effect is flexible across the observed visits, but unmodeled time-specific exposure effects may remain visible.
• Residual distribution and variance. Check residual-versus-fitted plots for changing spread and examine residual distributions for severe departures from normality. Large outliers or heteroscedasticity can affect standard errors and intervals.
• Participant effects. Review the estimated random-effect distribution, variance estimates, and any convergence or boundary warnings. A complex covariance structure that fails to converge is not validated merely because it has more parameters.
• Exposure support and pairing. Verify concentration units, visit matching, and overlap between observed concentrations and the target. Predictions far beyond observed exposure are extrapolations, even when the fitted equation produces a number.

Alternatives, Limitations, and Common Pitfalls

A set of timepoint-specific placebo-adjusted contrasts can estimate effects at selected visits without imposing one common concentration slope. Such contrasts may be useful when the question concerns particular times, but they do not directly use the continuous exposure-response relationship and can be less informative about predictions at a concentration between observed values. A mixed-effects concentration model instead borrows information across observations through a shared structure, but its conclusions depend on the form of that structure and the assumed concentration-response shape.

If residual patterns show systematic curvature, a prespecified nonlinear relationship, such as an Emax model, may be more suitable than a linear slope. A more flexible model is not automatically better: the concentration range and number of observations must support its additional parameters. Likewise, adding random slopes can represent between-participant variation in exposure response, but may be unstable if there are few observations per participant or little variation in concentration within participants. These alternatives should be compared by whether their assumptions and data support the estimand, not solely by which gives the most favorable result.

The analysis also has important limits. QTc correction, ECG measurement quality, timing of concentration collection, baseline definition, missingness, and concomitant factors can all affect the result. An exposure-response association is not automatically causal if concentration is related to participant characteristics or time-varying factors that also influence QTc. The prediction applies to the modeled population and concentration range; it should not be transported to a different population or extreme exposure without additional justification.