Spinal cord injury (SCI) is characterized by damage and resulting dysfunction to the motor, sensory, and/or autonomic nervous systems associated with trauma or disease processes leading to traumatic or non-traumatic SCI, respectively. The functional consequences can therefore be wide-ranging across these systems, with varying degrees of muscle paralysis, sensory impairment, and autonomic dysfunction such as problems with cardiovascular control, thermoregulation, or bowel, bladder, or sexual function [1], [2]. In general, the more rostral (higher) the damage to the spinal cord, the more body systems that will be affected. With respect to motor function, persons with damage to the cervical (neck) area of the spinal cord will have impairments to both lower and upper limb muscles and are diagnosed as having tetraplegia, while persons with damage to the thoracic (back) or lumbar (lower back) area of the spinal cord will have impairments to the muscles of the thorax and/or the lower limbs only and are diagnosed as having paraplegia. Given the functional consequences of SCI are dependent on both the severity and level of the damage to the nervous system, in addition to a variety of other factors such as pre-morbid condition, additional secondary complications, and psychosocial influences, there is a significant degree of heterogeneity in the presentation of persons with SCI [1], [2].

We study the problem of adaptively identifying patient subpopulations that benefit from a given treatment during a confirmatory clinical trial. This type of adaptive clinical trial has been thoroughly studied in biostatistics, but has been allowed only limited adaptivity so far. Here, we aim to relax classical restrictions on such designs and investigate how to incorporate ideas from the recent machine learning literature on adaptive and online experimentation to make trials more flexible and efficient. We find that the unique characteristics of the subpopulation selection problem -- most importantly that (i) one is usually interested in finding subpopulations with any treatment benefit (and not necessarily the single subgroup with largest effect) given a limited budget and that (ii) effectiveness only has to be demonstrated across the subpopulation on average -- give rise to interesting challenges and new desiderata when designing algorithmic solutions. Building on these findings, we propose AdaGGI and AdaGCPI, two meta-algorithms for subpopulation construction. We empirically investigate their performance across a range of simulation scenarios and derive insights into their (dis)advantages across different settings.