Across analyses incorporating p-tau217, digital twin–based covariate adjustment reduces outcome

variability relative to standard adjustment methods. At the two-year timepoint, this variance reduction

translates into an estimated increase in statistical power for p-tau217 as a secondary endpoint, from

92% to 97%. These improvements were observed in settings with limited longitudinal biomarker

support, demonstrating that approaches incorporating AI-generated digital twins into analyses

can effectively leverage emerging biomarker information as it becomes available.

 

Emerging biomarkers offer new sources of

prognostic and diagnostic information in Alzheimer’s

disease (AD). In particular, the blood-based

biomarker p-tau217 is increasingly incorporated

into AD clinical trials as an exploratory or candidate

endpoint, reflecting its biological relevance

and less invasive measurement compared with

cerebrospinal fluid or PET-based biomarkers1. In

parallel, digital twin–based covariate adjustment

methods have previously been shown to reduce

outcome variability and improve statistical power

for clinical endpoints2. However, the performance

of these methods in trials incorporating emerging

AD biomarkers has not been well characterized,

particularly when such biomarkers are measured in

only a subset of participants or at sparse follow-up

visits. In later-phase AD trials, biomarkers such as

p-tau217 are typically used as secondary endpoints

to support interpretation of clinical outcomes and

biological mechanisms, where improving their

statistical power is valuable but constrained by

limited biomarker sampling.

 

We analyzed AD datasets containing longitudinal

clinical outcomes alongside p-tau217 measurements.

•

•

•

We considered a trial design powered on a clinical

primary endpoint at two years—CDR-SB powered

to 80% (yielding a fixed two-arm sample size of

approximately 1,230 participants)—and evaluated

the resulting statistical power for p-tau217 as a

secondary endpoint at that fixed sample size under

different analysis strategies: an unadjusted analysis

(ANOVA), baseline covariate adjustment (ANCOVA),

and digital twin–based adjustment (PROCOVA and

Bayesian PROCOVA).

Our AD Digital Twin Generator is a machine learning

model trained on historical data from prior clinical

trials and observational studies to forecast expected

placebo outcomes for trial participants [2]. Digital

twin predictions were included as covariates in

treatment effect estimation, and performance was

assessed via variance reduction, the corresponding

sample-size reduction, and power for p-tau217 as a

secondary endpoint at fixed sample size.

To assess robustness in realistic biomarker-limited

settings, we additionally compared CDR-SB sample-

size reduction in the p-tau217–measured cohort

against a size-matched cohort without p-tau217.

 

Across analyses incorporating p-tau217, digital

twin–based covariate adjustment reduced outcome

variability relative to standard adjustment methods.

•

•

•

At the two-year timepoint, digital twin adjustment

achieved a variance reduction relative to an

unadjusted analysis of approximately 70% for

p-tau217, compared with approximately 59% for

baseline adjustment alone. This variance reduction

translated into an estimated increase in statistical

power for p-tau217 as a secondary endpoint, from

92% to 97%, at the sample size fixed by the CDR-

SB primary endpoint.

For the CDR-SB primary endpoint itself, digital

twins yielded estimated sample-size reductions of

approximately 31% (PROCOVA) and 40% (Bayesian

PROCOVA), versus approximately 21% for baseline

adjustment, with the digital twin contributing the

majority of the achievable reduction and p-tau217

adding modest further gains as a covariate.

These improvements were observed in settings

with limited longitudinal biomarker support and

were preserved in the p-tau217–measured cohort

relative to a size-matched cohort without p-tau217,

demonstrating that approaches incorporating AI-

generated digital twins into analyses can effectively

leverage emerging biomarker information as it

becomes available.

 

As blood-based biomarkers such as p-tau217

become more prevalent in AD clinical trials, they

present an opportunity to further improve trial

efficiency. Digital twin–based covariate adjustment

provides a practical framework for integrating

emerging biomarkers into trial analyses—whether

p-tau217 serves as a secondary endpoint or as

an additional covariate—enabling improvements

in statistical efficiency even when longitudinal

biomarker data are sparse. This approach offers a

scalable path toward more informative and efficient

AD trial design as biomarker use continues to evolve.