Cumulative Exposure, Latency, and Delayed Toxicity

Why Exposure History May Matter

A safety event observed today may reflect exposure from earlier days or weeks rather than the concentration measured today. For example, toxicity might follow a delay after a brief high exposure, develop as exposure accumulates, or depend on exposure sustained over a recent period. A model using only current concentration can miss such patterns or attribute risk to the wrong time.

The central modeling decision is therefore how to summarize the exposure history available before each time \(t\). A lagged concentration, a moving-window area under the concentration-time curve (AUC), and cumulative dose encode different scientific hypotheses. None is automatically superior: the chosen metric should match a plausible toxicity process and be calculable without using information from after the risk time.

For a time-to-first adverse event, the exposure history can enter a hazard model. The hazard is the instantaneous event rate among individuals who have not yet experienced the event. A useful illustrative specification is

where \(\lambda_0(t)\) is the baseline hazard, \(X_i(t)\) is one selected exposure-history predictor, and \(\mathbf{Z}_i(t)\) contains other modeled covariates. This formulation assumes that the exposure predictor has a log-linear association with the hazard, conditional on the other modeled variables; the proportional-hazards form also assumes that its hazard ratio is constant over time unless the model explicitly allows otherwise. The model describes an association conditional on the specification. It does not, by itself, establish that exposure causes the event.

Constructing Exposure-History Predictors

A lagged concentration shifts the exposure measurement back in time. It is suitable when a toxicity mechanism suggests a fairly specific delay: if the relevant delay is \(L\), the model uses concentration at \(t-L\), not the concentration after that time. In practice, concentrations may be observed only at sparse sampling times. The value at \(t-L\) must then be obtained using a prespecified, defensible rule, such as interpolation from a concentration-time model. The rule and its uncertainty matter, particularly when measurements are far apart.

A moving-window AUC accumulates exposure over a period rather than selecting a single point. It can represent recent burden when exposure over multiple days is thought to matter. Its units are concentration multiplied by time, such as mg·h/L. Dividing AUC by the window length gives the average concentration over that window, which has concentration units. These two versions are proportional for a fixed window length, but their numerical coefficients have different units.

Cumulative dose is the sum or integral of administered dose up to the risk time. It can be useful when the relevant biological burden is related to total administered amount and pharmacokinetic concentration data are unavailable or not the preferred predictor. Dose and concentration are not interchangeable: the same administered dose can yield different concentrations across individuals, and concentration depends on the dosing and sampling history. A cumulative-dose predictor also depends on the chosen start time and whether the window includes all prior treatment or only a recent period.

The construction must respect temporal ordering. For an event at time \(t\), a predictor should use exposure up to, but not beyond, \(t\). Including post-event concentration or dose creates look-ahead bias: the predictor then contains information unavailable when the individual was at risk. For a first-event analysis, exposure history ends at the event or censoring time for each individual.

Why These Metrics Represent Different Hypotheses

The distinction can be made explicit with a general exposure-history summary. Suppose a latent toxic burden \(B_i(t)\) is formed by weighting past concentrations according to how strongly they are thought to contribute:

Here \(w(u)\) is a nonnegative weighting function of exposure age \(u\). If \(w(u)\) gives nearly all its weight to one delay \(L\), then \(B_i(t)\) is approximately proportional to \(C_i(t-L)\). If the weight is constant over the preceding \(W\) time units and zero earlier, the summary is proportional to a moving-window AUC. These forms are useful because they show that lagged concentration and windowed AUC are alternative assumptions about which past exposure matters, not merely different labels for the same predictor.

A cumulative-dose metric instead sums administered amounts. It corresponds to a burden hypothesis in which past doses contribute to the total, subject to the selected start time or window. A simple cumulative measure treats each included dose equally; it does not represent elimination or the changing concentration produced by each dose. More elaborate weighting can represent those features, but requires additional assumptions and data.

Comparing History-Based Predictors With Current Concentration

Current concentration is a natural comparator, but it represents a different hypothesis: that the concentration at the risk time is relevant to the event rate. A lagged concentration instead proposes a delayed relationship. A moving-window AUC proposes that accumulated recent concentration matters, while cumulative dose proposes that the amount administered over a defined history is informative. The table summarizes their main distinctions.

Because these predictors can be strongly correlated, including several together changes the question being answered. A coefficient for moving-window AUC in a model that also includes current concentration describes the AUC association conditional on current concentration, not the overall association of AUC with the event. High correlation can make those conditional estimates unstable and difficult to interpret. Compare scientifically motivated alternatives, use consistent risk-time definitions and covariates, and consider whether the data contain enough independent variation to distinguish the predictors. A better statistical fit alone does not establish that the selected history summary is the correct biological mechanism.

Assumptions, Diagnostics, and Limitations

Interpretation depends on both the time-to-event model and the exposure-history construction. The event times and risk periods must be defined consistently; exposure records must be aligned to those periods; and censoring must be handled under assumptions appropriate to the analysis. The log-linear exposure effect and proportional-hazards assumptions in the illustrative model may not fit. For example, the association could be nonlinear, or the exposure effect could change over time. Earlier material on diagnostics for exposure-response model fit provides a framework for assessing model fit; here, particular attention is needed for the exposure-history variable itself.

• Check temporal alignment. Verify that each predictor uses only information available by the risk time. Inspect how the metric is calculated around event, treatment start, interruption, and censoring times.
• Check sensitivity to construction. Assess whether plausible lag or window choices materially change estimates and interpretation. A narrow lag and a broad AUC window answer different questions; trying many choices and reporting only the most favorable one can exaggerate evidence.
• Check the exposure data. Sparse concentration measurements, uncertain interpolation, missed doses, and inaccurate treatment records can distort the constructed history metric. State how unobserved exposure between measurements is handled.
• Check correlation and support. Examine how much the history metric overlaps with current concentration and other predictors. Ensure that comparisons are supported by the observed data rather than extrapolation to exposure combinations rarely seen.
• Check model form. Examine whether a log-linear term is a reasonable approximation across the observed range and whether the effect appears to vary over time. A single coefficient may conceal a threshold or a time-dependent pattern.

A common limitation is that these metrics simplify biological processes. AUC over a fixed window weights all exposure within the window equally, then excludes exposure just outside it. Cumulative dose also weights included doses equally unless the definition says otherwise. A single lag assumes one relevant delay. Each may be useful as a transparent empirical summary, but none automatically captures tissue accumulation, recovery, or a complex toxic mechanism. In addition, exposure can change in response to emerging symptoms or laboratory abnormalities. In that setting, an observed exposure-history association may partly reflect clinical decisions or evolving health status, rather than an exposure effect alone.