Spinal cord injury (SCI) can impair sensorimotor pathways and reduce walking ability. The effects of sensorimotor impairments are complex, as they interact with other factors such as age, pain, injury level, and injury severity. Traditional regression analyses have been used to describe relationships, but they frequently assume linearity. Explainable AI methods such as SHapley Additive ex-Planations (SHAP), based on cooperative game theory, can reveal feature importance globally and locally. Gaussian Process Regression (GPR) can handle limited datasets, a common challenge in observational clinical studies with small sample sizes. In this study, we proposed and evaluated a framework applying GPR and SHAP to quantify how neurological impairments and other factors, including muscle strength, sensory function, age, pain, spasticity, etc., contribute to walking performance, specifically walking speed and net oxygen cost during a six-minute walk test. This approach estimates each factor’s contribution both on a group level and for each individual. Thirty four adults with SCI underwent a clinical assessment and the six-minute walk test with preferred walking aids if relevant. Muscle strength was the most influential factor in both walking speed and net oxygen cost. Male sex, lower age, and less pain were associated with increased walking speed. More pain, higher body mass index, and higher sensory score were associated with lower net oxygen cost. Spasticity, injury level and sensory injury levels had relatively small influence on either outcome measure. Individual SHAP analyses quantified how neurological factors influenced walking performance for each participant. We demonstrate how nonparametric regression and explainable AI provide insights into the complex neurological factors affecting walking ability in persons with SCI.

Background: Incomplete spinal cord injury (iSCI) often causes heterogeneous locomotion dysfunctions, depending on remaining sensorimotor function. Clinical tests and traditional gait analysis have limited ability to quantify the diversity of gait impairments. Unsupervised learning techniques can objectively identify common gait patterns among the overall heterogeneity. Explainable artificial intelligence approaches, when combined with machine learning models, can reveal important features often missed by traditional gait analyses. This study presents a framework to characterize gait heterogeneity among persons with iSCI based on several data-driven methods. We aimed to stratify overall gait heterogeneity by deriving clusters with similarities without a priori identification of parameters, and to assess possible clinical correlations in the derived clusters. 

Methods: A cohort of 28 adults with iSCI and control group of 21 non-disabled adults were recruited. The iSCI group underwent a standard physical assessment of overall mobility, lower extremity strength, and spasticity. Both groups underwent instrumented 3D gait analysis, walking at self-selected pace. Distinct iSCI gait pattern subgroups were identified with dependent dynamic time warping and hierarchical agglomerative clustering. Distribution of clinical descriptives and outcome measures among and between groups were evaluated.  Gait predictors that distinguish each cluster from control gait were identified with a random forest classifier and explainable AI.

Results: Six distinct gait clusters were identified among the 280 iSCI gait cycles. Clusters with relatively low walking speed exhibited shorter step lengths and less ankle plantarflexion in pre-swing than controls. Gait patterns and walking performance in  clusters with high walking speed were relatively similar to controls. Overall muscle strength, walking independence, walking speed, step length, step width, sex distribution, and types of walking aids significantly differed between all six clusters. Ankle plantarflexion angle in pre-swing correlated strongly with walking speed and step length.

Conclusion: Through a series of advanced data-driven approaches, common gait patterns can be objectively identified and comprehensively characterized within a heterogeneous iSCI population. This work represents an initial step in developing individualized rehabilitation programs for persons with iSCI.

Individuals with incomplete spinal cord injury (iSCI) exhibit diverse walking capabilities due to partial sensory and/or motor impairment below the injury level. This study examined how neuromuscular weakness patterns influence compensatory gait strategies, dynamic balance(margins of stability, MoS), and joint kinetics across iSCI subgroups compared to non-disabled controls. We analyzed gait data from 21 iSCI participants previously classified into four subgroups through dynamic time warping and hierarchical clustering in our previous study.

Temporospatial parameters, anterior-posterior (AP) and mediolateral (ML) MoS, and joint kinetics were extracted. The most functional group (mild plantarflexor weakness) walked comparably to controls but exhibited elevated peak hip flexion moments, suggesting increased risk for hip osteoarthritis or femoroacetabular impingement. The moderate plantarflexor weakness group demonstrated sagittal-plane compromises: slower walking speed, smaller magnitude of AP MoS, and lower peak ankle plantarflexion moments, power generation, and positive work during stance. Two lower-functioning groups with moderate-to-severe plantarflexor weakness combined with moderate hip extensor and hip abductor weakness showed similar sagittal-plane compensations plus frontal-plane compensations, including wider step width, higher ML MoS, lower peak hip extension and abduction moments. They also exhibited lower knee negative work during swing. Notably, the most impaired group relied primarily on hip muscles (extensors, abductors, and adductors) for stance-phase work generation, contrasting with other groups and controls who predominantly relied on plantarflexors. These findings demonstrate how varying levels of neuromuscular weakness,especially at the ankle and hip muscles, affect gait biomechanics after iSCI.

Predictive gait simulations hold promise for understanding motor control strategies in healthy individuals, but their application to pathological conditions like incomplete spinal cord injury (iSCI) remains challenging. Individuals with iSCI navigate multiple competing functional constraints rather than simply minimizing energy or fatigue, creating a multi-objective optimization problem with numerous viable solutions. Current approaches require manual fine-tuning of objective weights, preventing systematic exploration of compensatory strategies. In this study, we developed a bilevel optimization framework using Bayesian optimization to systematically identify objective weights that minimize prediction errors against experimental gait data. We tested this framework on a 69-year-old female with iSCI, asymmetric muscle weakness, and mild right knee osteoarthritis, comparing two multi-objective walking policies differing in incorporating pain avoidance at the right knee: one minimizing ground reaction force yank (Policy A) and another minimizing knee contact force (Policy B). Both policies shared six other objective functions. The framework successfully automated weight identification, with Policy A achieving superior performance across kinematic, kinetic, and ground reaction force deviation indices compared to Policy B and literature-based reference weights. These results demonstrate that bilevel optimization could systematically identify individual-specific compensatory gait strategies used in predictive gait simulations, offering a pathway toward personalized rehabilitation planning for neurological conditions. Future work should validate this approach across larger cohorts and incorporate refined biomechanical objectives.