Alejandro Schuler, United States of America

Unlearn.AI Data Science

Alejandro is a data scientist and biostatistician with expertise in the intersection of causal inference, machine learning, and health data. Alejandro holds a Ph.D. in Biomedical Informatics from Stanford University, where he worked with Professors Nigam Shah, Rob Tibshirani, and others to develop methods for the groundbreaking Informatics Consult service. Before working at Unlearn, Alejandro helped direct research efforts at Kaiser Permanente to evaluate the benefit of live predictive models. Alejandro is known as one of the inventors of NGBoost, a method for probabilistic regression via gradient boosting, and as an inventor of PROCOVA, a procedure that leverages machine learning and historical data to improve the efficiency of trial analyses.

Presenter of 2 Presentations

Conference Calendar

DIGITAL TWINS ENABLE SMALLER TRIALS THAT MAINTAIN THEIR POWER: A DEMONSTRATION FOR ALZHEIMER’S DRUG TRIALS

Session Name

DIGITAL END-POINTS AND ARTIFICIAL INTELLIGENCE

Session Type

SYMPOSIUM

Date

14.03.2021, Sunday

Session Time

12:00 - 13:45

Room

On Demand Symposia B

Lecture Time

12:00 - 12:15

Presenter

Alejandro Schuler, United States of America

Session Icon

Abstract

Abstract

Aims

Smaller and faster trials could accelerate drug development in Alzheimer’s Disease (AD). We set out to demonstrate that smaller trials that maintain their power can be designed using a machine-learning model of AD progression trained on placebo subject records from past clinical trials.

Methods

Digital twins are longitudinal, patient-level placebo records with baseline characteristics and treatment duration matched to those of actual subjects randomized into a study. Because they predict outcomes for individual subjects, the outcomes of digital twins may be adjusted for as prognostic covariates to add power while preserving type-I error control (unlike many other methods of historical borrowing). We re-analyzed a placebo-controlled randomized study of docosahexaenoic acid (DHA) in 402 subjects with mild to moderate AD and compared the results to an analysis of a reduced dataset powered using digital twins.

Results

Using digital twins, the same power (80%) was attained with 18% fewer subjects than in the original trial. The estimated standard errors in the results of the digital twins trial were nearly identical to those in the original trial, indicating equal level of confidence in the results despite smaller sample size. Both analyses produced similar (not statistically significant) results on the primary endpoint of change in ADAS-Cog11 over 18 months.

Conclusions

Our retrospective analyses indicate that digital twins enable smaller trials with equal design power and can thus accelerate clinical trials without sacrificing type-1 error control. Our methodology pairs an innovative use of machine learning with proven statistical methods that easily integrate with trial protocols.

Hide

LIVE DISCUSSION