A common focus during rehabilitation after spinal cord injury (SCI) is on promoting improvements in functional walking capacity. Fast and effective enhancement of gait function can enhance a patient's independence, life satisfaction, and subsequent reintegration into society as a fully-participating member. Body weight support treadmill training (BWSTT) has been considered a primary rehabilitative strategy for improving locomotion in patients with incomplete SCI [
1,
2]. BWSTT provides intensive, task-specific repetitive walking training, promotes supraspinal neuroplasticity involved in locomotion, and consequently helps patients to regain their walking skills [
3‐
5]. While conventionally BWSTT is supplemented by manual assistance from therapists, robotic-assisted BWSTT has become widely used over the past decade to provide symmetric walking training to support lower extremity rehabilitation in patients with neurological movement impairment. The most widely-used automated locomotor training system is the Lokomat® (Hocoma AG, Volketswil, Switzerland) [
6,
7]. Recent studies have shown that patients can receive positive physical and psychological benefits from Lokomat training [
1,
2,
8‐
17], such as improved walking capacity [
1,
8‐
10], improved metabolic performance [
2,
12], and increased activity in the cerebellum [
15]. However, a recent review from Swinnen [
18] has shown that the effectiveness of Lokomat training remains controversial, with some studies showing significant improvements in motor and walking performance after Lokomat trainings, while others showing minimal changes [
18]. The authors and previous researchers have pointed out that difficulty arises when translating these research findings to the clinical setting because of the lack of a control group in those studies [
19]. The authors also observe that the advantages of Lokomat training over other types of locomotor trainings have yet to be fully demonstrated.
In addition, the high variability in recovery pattern among patients should be explicitly considered in the analysis. Locomotor training, like any other intervention, is not expected to affect all patients equally; instead, different treatment responses are expected for different patient subgroups. Many research studies, however, oversimplify by assuming that all individuals are drawn from a single population with a similar recovery pattern; pre-training and post-training measures are then compared. This approach loses the heterogeneous response of patients to the therapy program as well as the longitudinal information describing the timeline of improvement—information that should be considered when evaluating the efficacy of a particular intervention. In fact, many studies involved with locomotor training have reported results with a standard deviation even larger than the sample mean, indicating that the high inter-individual variance within their sample population results in a less meaningful average value. Thus, it is necessary to address these deficits by first identifying distinct latent classes within each sample and then modeling the recovery patterns for each class separately. In addition, pre-vs. post-training comparisons usually have a lowered effect size and statistical power than longitudinal tracking studies, because the error term has a larger standard deviation than in other statistical models [
20].
Recent statistical techniques such as Growth Mixture Modeling (GMM) allow researchers to effectively recognize this high inter-individual variance, and to model the growth pattern of a longitudinal recovery procress. The inclusion of additional time points between the pre-and post-training evaluations can help elucidate the training effect on gait impairment, and further can dramatically improve the statistical power in situations where increasing sample size is not feasible due to costs and patient availability [
20].
In the present study, rather than performing a pre-vs. post-training comparison within one experimental treatment group, we examined in detail the growth trajectory of three aspects of gait impairment (speed, functional mobility, and endurance, as evaluated by typical clinical measures) for subjects in an intervention group receiving Lokomat training and those in a control group who received no intervention. Instead of assuming that all subjects are drawn from a single population, we objectively classified the subjects into distinct subgroups based on their growth trajectory of walking capacity, and then modeled the effect of Lokomat training for each subgroup. Finally, we evaluated whether baseline measurements of neuromuscular and clinical performance could serve as statistical predictors for these recovery patterns. In particular, we hypothesized that ankle strength—measured in terms of maximum voluntary dorsiflexion and plantarflexion torque—could predict the recovery trend for these clinical evaluations.