Predicting Time-to-Event Outcomes From Mechanistic Trajectories

From Mechanistic Trajectories to Event Times

A QSP model produces trajectories of biological states, but a clinical time-to-event outcome requires an additional rule that says when an event has occurred. The same simulated trajectory can imply different event times under different rules. For example, a change in a biomarker is not automatically a clinical event; the model must specify which biomarker value, direction of change, duration, or event intensity corresponds to the clinical endpoint.

The predictive target should therefore include the event definition, the time origin, the population being predicted, and the prediction horizon. Earlier material on defining predictive targets and mapping parameter ensembles to clinical predictions provides the general framework: here the outcome derived for each simulated virtual patient is an event time \(T\), potentially not observed within the simulated horizon.

Two common ways to derive \(T\) from a mechanistic trajectory are a threshold rule and a hazard rule. A threshold rule makes the event time a deterministic consequence of the trajectory, conditional on the model inputs. A hazard rule specifies an instantaneous event rate conditional on remaining event-free; it can represent event timing that is not determined by a single state crossing.

Threshold Rules and Crossing Times

Let \(x_i(t)\) denote a relevant state for virtual patient \(i\). A simple upward threshold rule with threshold \(c\) defines \(T_i\) as the first time \(x_i(t)\geq c\). A downward rule instead uses the first time \(x_i(t)\leq c\). The direction, threshold, and any requirements such as persistence above threshold must be specified before calculating event times. If the trajectory never satisfies the rule during the prediction horizon, the event time is not known to be the horizon: it is only known to exceed that horizon.

QSP trajectories are usually stored at discrete time points. If a crossing occurs between two output times and linear interpolation is appropriate, the crossing time can be approximated from the two bracketing state values. For an upward crossing between \(t_a\) and \(t_b\), where \(x_i(t_a)<c\leq x_i(t_b)\), the interpolated time is:

This calculation assumes that linear interpolation is adequate over that interval and that the crossing is genuine. A coarse output grid can miss a brief crossing or misplace a steep crossing. Oscillating states also require care: a first-crossing rule and a rule requiring the state to remain beyond threshold for a specified duration are different endpoints.

Hazard Rules for Events Not Fixed by a Crossing

A hazard rule can describe an event whose timing depends on both the mechanistic state and a stochastic event process. The hazard \(\lambda_i(t)\) is the instantaneous event rate at time \(t\), conditional on the patient having remained event-free just before \(t\). It has units of inverse time. A hazard model must specify how the trajectory determines this rate, for example \(\lambda_i(t)=g(x_i(t),\theta_i)\), where \(g\) is a nonnegative function and \(\theta_i\) denotes relevant model inputs.

For a specified hazard trajectory, the cumulative hazard through time \(t\) is the area under the hazard curve. The conditional survival probability is the exponential of minus this cumulative hazard:

This relationship assumes the hazard is defined conditional on being event-free and that the event process is represented by the specified hazard. To simulate an event time, draw \(E\) from an exponential distribution with rate 1 and set \(T_i=\inf\{t:\Lambda_i(t)\geq E\}\). If cumulative hazard has not reached \(E\) by the simulation horizon, the simulated event is beyond that horizon. Alternatively, for a milestone probability, it may be more precise to average conditional survival probabilities rather than add noise from individual event-time draws.

Propagating Uncertainty to Survival and Milestones

For a virtual population of \(N\) equally weighted patients with simulated event times \(T_1,\ldots,T_N\), the empirical survival estimate at time \(t\) is the fraction whose events occur after \(t\). For a threshold-derived event time or a set of simulated hazard event times, this gives a direct estimate of the population prediction:

Here \(I(\cdot)\) equals 1 when its condition is true and 0 otherwise. A milestone probability, such as the probability of an event by month 6, is \(\widehat{F}(6)\). With a hazard rule, the conditional survival probabilities can instead be averaged across the virtual population:

This estimates survival averaged over the simulated population and its hazard-model inputs. It avoids Monte Carlo variation from drawing a separate event realization for each patient, although uncertainty from a finite number of virtual patients remains. When patients have unequal population weights, each term must be weighted and the weights normalized to sum to 1.

The predicted median event time is the earliest time at which the population survival curve is at or below 0.5. If the curve remains above 0.5 through the horizon, the median is not identified within that horizon. A median calculated from a pooled predictive distribution also need not equal the median of medians calculated separately for different parameter sets: taking a median and averaging or pooling are different operations.

In a full prediction, each patient's event time is derived from a trajectory generated using uncertain parameters and patient characteristics. The distribution of those event times, or the corresponding averaged hazard survival probabilities, propagates those inputs to the clinical endpoint. If uncertainty is separated into components, variation among virtual patients describes patient variability, while changes across plausible parameter sets describe epistemic uncertainty. A single survival curve that averages over both does not by itself show how much each component contributed.

Assumptions, Diagnostics, and Alternatives

Before interpreting a prediction, check that the event rule matches the intended clinical outcome, the time origin and units are consistent, the simulated horizon is long enough, and the numerical time grid resolves relevant crossings or hazard changes. For a hazard calculation, verify that rates are nonnegative, have the intended time units, and are integrated with adequate numerical accuracy. For a threshold calculation, inspect trajectories near the crossing and check whether repeated crossings or brief excursions change the event definition.

A useful diagnostic is to compare event-time summaries after refining the output time step. Large shifts in crossing times or milestone probabilities suggest that numerical resolution is inadequate. Also check the number of patients with no event by the horizon. Treating each such patient as if an event occurred exactly at the horizon creates an artificial concentration of event times and can bias the survival curve and median downward.

These approaches are alternatives only when each is scientifically defensible for the endpoint. A threshold is not automatically preferable because it is simpler, and a hazard is not automatically realistic because it is flexible. Both can be misspecified, and both depend on the quality and scope of the underlying QSP model. Predictions beyond the simulated horizon, or from trajectories that leave the model's supported domain, require particular caution.

Common Pitfalls and Interpretation

• Calling the last simulation time an event time. If the event rule was not met, the result is \(T>\tau\) at horizon \(\tau\), not \(T=\tau\).
• Leaving the threshold rule vague. State the variable, threshold, crossing direction, time origin, and any persistence requirement.
• Reporting a conditional hazard curve as the population survival curve. Average conditional survival over the target population to obtain marginal survival.
• Equating a mechanistic state crossing with clinical validation. A crossing produces a prediction only under the assumed mapping from model state to clinical event.
• Reporting a median when fewer than half the predicted population has events within the horizon. In that case, the median is beyond the horizon or not identified by the available simulation.