Linking neurological status to functional outcomes in spinal cord injury: a multi-class, task-specific approach.

backgroundSpinal cord injury (SCI) causes long-term neurological deficits resulting in functional disabilities. While longitudinal recovery patterns of sensorimotor outcomes after SCI have been studied, few analyses have applied machine learning to systematically model the relationship between neurological impairments and functional independence at different post-injury phases.

methodsThis study compared ordinal and nominal classification models predicting functional independence from sensorimotor status cross-sectionally. Inputs included motor and sensory scores from the International Standards for Neurological Classification of SCI, age, sex, and time since injury collected in the European Multicenter Study about SCI. Models were evaluated on a task from each domain of the Spinal Cord Independence Measure, namely grooming (self-care), bladder management (respiration and sphincter management), and indoor mobility (mobility). Analyses were stratified into early (≤ 40 days), intermediate (70–100 days), and late (> 182 days) post-injury phases. Models were ranked based on five evaluation metrics, and interpretability explored using Shapley Additive Explanations (SHAP).

resultsModel accuracy improved over time (early phase: 46–71%, late phase: 50–85%), indicating that functional independence is more reliably determined from sensorimotor scores in later post-injury phases. Across all scenarios, random forest achieved the best overall performance (0.93 ± 0.03, averaged across different metrics). Ordinal models yielded fewer severe misclassifications compared to nominal models. Motor scores were stronger predictors than sensory scores, with lower limb function (L2–L4) strongly associated with mobility, voluntary anal contraction with bladder control, and upper limb function (C6, C8) with grooming ability, highlighting that models utilise known relationships.

conclusionWe show that both multiclass and ordinal models can accurately classify SCIM-based functional independence outcomes after SCI from neurological assessments at different time points post-injury. Ordinal approaches provide particular clinical value by minimizing severe misclassifications, a crucial advantage when distinguishing between functional independence classes that require fundamentally different care approaches. Interpretability analysis showed that the predictions are grounded in clinical knowledge. The developed models provide the basis for a modular prognostic framework, in which predicted ISNCSCI scores can be used to derive the most likely functional independence class, enabling a modular, computationally efficient and scalable approach to prediction in SCI care across a range of neurological and functional outcomes.