Abstract
In many preclinical spinal cord injury (SCI) studies, assessment of locomotion recovery is key to understanding the effectiveness of the experimental intervention. In such rat SCI studies, the most basic locomotor recovery scoring system is a subjective observation of animals freely roaming in an open field, the Basso Beattie Bresnahan (BBB) score. In comparison, CatWalk is an automated gait analysis system, providing further parameter specifications. Although together the CatWalk parameters encompass gait, studies consistently report single parameters, which differ in significance from other behavioral assessments. Therefore, we believe no single parameter produced by the CatWalk can represent the fully-coordinated motion of gait. Typically, other locomotor assessments, such as the BBB score, combine several locomotor characteristics into a representative score. For this reason, we ranked the most distinctive CatWalk parameters between uninjured and SC injured rats. Subsequently, we combined nine of the topmost parameters into an SCI gait index score based on linear discriminant analysis (LDA). The resulting combination was applied to assess gait recovery in SCI experiments comprising of three thoracic contusions, a thoracic dorsal hemisection, and a cervical dorsal column lesion model. For thoracic lesions, our unbiased machine learning model revealed gait differences in lesion type and severity. In some instances, our LDA was found to be more sensitive in differentiating recovery than the BBB score alone. We believe the newly developed gait parameter combination presented here should be used in CatWalk gait recovery work with preclinical thoracic rat SCI models.
Introduction
In preclinical spinal cord injury (SCI) research, measuring locomotor recovery is the primary metric for evaluating therapeutic efficacy. For decades, the field has relied on the Basso Beattie Bresnahan (BBB) score, a subjective 21-point scale based on visual observation of rats in an open field. While foundational, the BBB score is prone to observer bias and often lacks the sensitivity to detect subtle changes in gait coordination.
Automated systems like CatWalk address these limitations by recording paw prints on a glass walkway and generating hundreds of objective metrics. However, the sheer volume of data—often exceeding 200 parameters—presents a challenge: researchers frequently report only a few isolated metrics, making it difficult to achieve a holistic view of recovery or compare results across different laboratories.
Methodology
To resolve this "data dump" problem, researchers developed a single, predictive value ($p_{LDA}$) using Linear Discriminant Analysis (LDA). This machine learning approach distills complex gait dynamics into a standardized index.
- Feature Selection and Ranking: The team evaluated 22 candidate parameters, comparing uninjured rats against injured models. They selected the nine parameters with the largest effect sizes (Cohen’s $d > 1.3$) to ensure the index focused on the most distinctive injury-induced changes.
- Weighting: LDA was employed to assign specific weights to these nine parameters, including body speed, stride length, and inter-limb coordination. This "recipe" creates a final score that optimally differentiates healthy gait from injured gait.
- Validation: The index was validated across five distinct studies involving various injury types (contusions vs. hemisections), rat strains (Fischer 344 and Wistar), and therapeutic interventions (e.g., Progesterone or Epothilone D).
Key Results and Insights
Sensitivity and Discrimination
The $p_{LDA}$ index demonstrated superior sensitivity compared to traditional methods. It clearly separated uninjured rats (score ~0.12) from injured cohorts (scores 0.03–0.06). Notably, the index significantly distinguished between "moderate" and "moderate-severe" contusions ($p = 0.018$), a distinction that individual parameters like stride length often failed to make.
The "Forepaw Surprise"
Analysis revealed that four of the nine critical parameters were related to the forepaws, despite the injuries being located in the thoracic spine. This indicates that rats fundamentally alter their front-limb mechanics to compensate for hind-limb deficits—a subtle shift that human observers often overlook but which is vital for a holistic understanding of gait.
Cross-Lab Robustness
The index remained effective across different rat strains, sexes, and hardware versions (CatWalk 7.1 vs. XT). This robustness suggests that the $p_{LDA}$ could serve as a universal metric, facilitating standardized comparisons across the global research community.
Discussion and Limitations
While the $p_{LDA}$ offers a powerful tool for objective assessment, it is subject to a "floor effect." The system requires the animal to be capable of weight-supported walking across the runway, making it unsuitable for the earliest stages of recovery (typically BBB < 10). Additionally, while the index performed exceptionally for thoracic injuries, it did not distinguish deficits in cervical (C4) lesion models, suggesting that neck-level injuries may require a different parameter combination.
To encourage widespread adoption, the authors have made the $p_{LDA}$ calculation tool available as an open-access Excel download, providing a turnkey solution for researchers to enhance the objectivity of their gait analysis.
Summary updated