Presentation (CTI)

OBJECTIVES: To evaluate whether an external control arm (ECA) constructed using latent similarity matching reproduces the placebo recovery trajectory of the phase 2b Nogo inhibition (NISCI) trial.
METHODS: The placebo arm was emulated using the European Multicenter Study about Spinal Cord Injury registry (EMSCI). An autoencoder trained exclusively on registry baseline covariates learned a latent representation. Placebo patients were projected into this space, and registry controls were identified by nearest-neighbour matching under a severity grade, level of injury, and motor score caliper. Hyperparameters were selected through an outcome-blind, reproducible search. The primary estimand is the difference in mean upper extremity motor score (UEMS) recovery slopes between matched cohorts, reproducing the trial model. Equivalence was evaluated with Two One-Sided Tests at a 4-point UEMS margin. The primary analysis is contextualized by results using covariate-balancing propensity score, Mahalanobis distance, and coarsened exact matching. The preregistration included the target trial protocol, ICH E9(R1) estimand, and directed acyclic graph. The preregistered pipeline was executed under a reduced hyperparameter search.
RESULTS: Outcome-blind selection produced an ECA of 111 controls for all 37 placebo patients, with a post-match standardized mean difference below 0.3 on clinical matching covariates. The difference in mean UEMS was −0.44 [90% CI: −3.2, 2.4], within the pre-specified margin and robust under bootstrap resampling. Under comparable conditions, only covariate-balancing propensity score matching achieved equivalence at 0.65 [90% CI: −3.5, 2.2], whereas Mahalanobis distance matching and coarsened exact matching did not.
CONCLUSIONS: Our proposed approach constructed an external control arm that reproduced the randomized placebo recovery trajectory in this initial evaluation. These findings support the potential of machine learning-derived external controls, combined with transparent target trial emulation, to reduce reliance on concurrent control arms in spinal cord injury trials.