Drug Development

With U.S. Food and Drug Administration (FDA)-approved anti-amyloid, partially disease-modifying treatments now available, the ethical justification for randomly assigning patients to placebo has become controversial. The idea of twin-controls (or virtual-controls) has gained significant attention in recent years, with the aim of creating twin patient cohorts that can be used as a surrogate to evaluate the effects of treatment on a personalized level. While promising, the feasibility of digital twins techniques remains largely untested.

I-CONECT is a multi-site, single-blind, randomized controlled trial (RCT) examining the effects of conversational interactions on cognition among socially isolated subjects aged ≥ 75 years (normal cognition; mild cognitive impairment). 186 participants were randomized into experimental or control groups. The experimental group engaged in video chats with study staff 4 times/week for 6 months, while the control groups received weekly 10-minute phone calls. The current analysis focused on the efficacy-shown Montreal Cognitive Assessment (MoCA; global cognition) and category fluency animals (CFA; language-based executive function) at 6-month follow-up. Data from the National Alzheimer's Coordinating Center-Uniform Data Set (NACC-UDS) were used to create digital twins for treatment participants through two methods. Method 1 involved twin mapping, matching participants with 1 to 20 twins who had similar demographic, biological, and social factors, and comparing change scores between each participant and their twins. Method 2 used direct modeling by building random forest models to predict change scores as if participants were assigned to a "usual care" control group. Effect sizes were compared between original and twin-controls trials, as well as between the two methods.

Approximately 10% of NACC-UDS participants (5,332 out of 50,259) were eligible for I-CONECT. For parallel-group designs, treatment effect sizes on MoCA closely aligned between original (β=1.67) and twin-control (β=1.46-1.97) trials when the Euclidean distance mapping was applied. Similar findings were found in CFA (original trial β=2.56; twin-control trial β=2.31-3.34). For single-case, n-of-1 designs, methods 1 and 2 showed substantial agreement in identifying treatment responders (Cohen's Kappa=1 for MoCA; 0.68 for CFA).

Digital twins from publicly available datasets enhance the rigor of RCTs by providing mapped twins as controls for early-phase dementia trials.