A digital twin methodology using retrospective patient data for sample size reduction in Alzheimer’s disease clinical trials

Recruitment for Alzheimer’s disease randomized controlled trials (RCTs) is difficult and expensive. To reduce RCT sample sizes while maintaining high statistical power, our Digital Twin Trial (DTT) methodology combines an interpretable cognitive decline prediction model with prediction-powered inference. Unlike RCT sample size reduction techniques that maintain power by recruiting only patients likely to decline, prediction-powered inference is used within the data analysis stage of the trial and does not impose additional restrictions on participant eligibility. For DTT participants, our model identifies similar individuals (“Digital Twins”) from a retrospective trial-matched database and uses their cognitive scores to predict decline. Predictions adjust observed scores, reducing variance within treatment groups. We simulated 18-month DTTs and standard RCTs using mixed effects models of decline in Alzheimer’s Disease Neuroimaging Initiative subjects meeting lecanemab’s Phase 3 inclusion criteria. Predicted and observed change in Clinical Dementia Rating Sum-of-Boxes correlated at r = 0.437. DTTs required 9.5–19.0% fewer subjects than standard RCTs to detect a simulated 25% decline-slowing drug effect at 0.9 power. DTT Type 1 error was consistent with 0.05. DTTs could reduce clinical trial recruitment and cost burdens without imposing additional inclusion criteria, enabling efficient testing of new treatments in all populations who might benefit. Model interpretability will foster clinician and regulator trust in individualized prognoses. This transparency could help smooth the model’s translation to real trials and eventually the clinic, where it could help make prognoses and better evaluate treatment effects in individual patients.