Fighting Time in Adaptive Clinical Trials with Longitudinal and Predictive Modeling

The Constraint of Time in Adaptive Clinical Trials

In adaptive clinical trials, the timetable for learning (the lag between enrollment and outcome measurement) dictates efficiency. Most fixed designs defer all decisions until primary endpoints mature, frequently a year or more after enrollment. Trial conduct grinds forward without adjustment: no dose selection, no arm dropping, no adaptive allocation, no operational efficiency. This standard practice assumes that all key questions can be answered only after the last subject reaches the endpoint, regardless of available interim evidence.

The promise of adaptive trials is that things are learned during the trial that had they been known at the design stage there would have been a different design. Had the designers known the low dose was ineffective or that there is a patient population that doesn’t respond as well, or that the clinical effect size is larger than expected, the design would have been different. A design that can adapt and evolve to the accruing information has great promise to provide better answers, more efficiently. The challenge is to learn as efficiently as possible in the trial. Waiting for data creates a challenge of time in adaptive trials. “Fighting time” is a central statistical challenge. If adaptation is hamstrung by slow information, promised advantages are muted.

Decision-making in adaptive trials thrives on data-driven learning. Delayed knowledge translates directly into wasted patients, time, and resources.

Longitudinal and Predictive Modeling as Essential Tools

Longitudinal modeling and predictive modeling are the technical linchpins for fighting time. These models explicitly relate early markers or surrogates such as MRI readings, interim biomarker values, cognitive test results, functional scores, etc. to the ultimately delayed clinical endpoint.

For example, the I-SPY 2 neoadjuvant breast cancer platform trial models the 6-month pathologic complete response (pCR) endpoint from MRI measurements at 1 and 3 months. For each MRI a model is based on 13 ordinal categories of percent tumor reduction. Each category is linked to a calibrated probability of six-month pCR, continually estimated from prior cohorts with both MRI and pCR outcomes.

In the adaptive Alzheimer’s trial BAN 2401 (lecanemab), early cognitive scores on ADCOMS at 3, 6, 9, and 12 months were modeled to forecast 18-month outcomes. These models, supported by historical data, informed by the new patients, allowed ongoing response-adaptive randomization and early futility or graduation triggers—decisions that would be prohibitive if inhibited by waiting for the final endpoint alone.

AWARD-5 (Trulicity/dulaglutide, Eli Lilly) further highlights operational considerations. Response-adaptive allocation across seven doses was driven by models forecasting 12-month HbA1c, built on monthly data. Multiple endpoints—weight, blood pressure, heart rate, glycemic control—were integrated into a utility-based allocation strategy. Notably, Lilly deliberately controlled accrual rates, based on extensive simulations, to increase per-subject interim information; when enrollment was too rapid, forecast precision eroded and efficiency gains were lost.

In ICECAP, 30-day modified Rankin Scale (mRS) scores—an ordinal clinical measure—are used to predict 90-day outcome. Monotonic Bayesian models across the scale for 30-day outcomes are used to model 90-day outcomes, rigorously calibrated to prior data, can add months of design efficiency.

The Bayesian approach is natural for longitudinal models in adaptive trials. Models are fit with prior distributions, updated as patient data accumulate, and recalibrated at every interim. For each subject lacking a matured endpoint, models generate a distribution of likely outcomes based on observed interim data. Datasets are multiply imputed thousands of times, with adaptations based on the resulting probability spectrum. This approach formalizes uncertainty, prevents overconfidence, and guarantees that adaptations are using all available outcomes at every analysis.

Real-World Impact and Limitations

I-SPY 2’s longitudinal modeling enabled rapid arm graduation, reducing exposure to ineffective therapies and moving efficacious drugs forward months earlier. In BAN 2401, interim model projections, based on early cognitive endpoints, accurately predicted 18-month results and guided adaptive allocation as the trial unfolded. The AWARD-5 seamless phase 2/3 trial for Trulicity advanced two optimal doses, thanks in part to mathematical modeling of early glycemic and weight-loss outcomes, coupled with intentional accrual management. In ICECAP, use of 30-day mRS to predict later function delivers gains in adaption speed and resource use.

The value of adaptation can be improved depending on the predictive quality of early markers. In most trials there are early outcomes that are likely to be predictive of longer-term outcomes. Longitudinal models can learn whether an early marker is predictive – and when it is not, this is learned and the adaptations appropriately delayed. Each scenario requires empirical vetting of modeling assumptions, variable selection, and operational feasibility.

Conclusion

Fighting time is necessary in the modern adaptive trial. Bayesian longitudinal modeling and predictive modeling compress knowledge timelines, equipping teams to make earlier adaptive decisions. The ultimate promise of an adaptive design is to improve the design based on learnings in the trial. At any given time in an adaptive trial there will be different levels of exposure to patients. Longitudinal modeling can allow more information to be included in adaptive decision making, ultimately improving the designs.