Development of a deep learning-based grid walking test for assessing post-stroke motor function in mice

Stroke is a leading cause of death and disability, affecting over 15 million people each year. Developing effective therapies depends on accurate, scalable methods to assess sensorimotor recovery in preclinical models. Standard motor tests like grid walking and ladder tests rely on manual scoring, limiting objectivity and throughput. To address this, we compared a commercially available automated ladder test with a newly developed deep learning-based grid walking test using DeepLabCut (DLC) in a photothrombotic stroke model in mice. Our results show that the automated ladder test system failed to detect significant unilateral deficits in stroke animals. It did not capture the expected increase in missteps in the affected forepaw after stroke, even though motor cortex infarcts were confirmed histologically, and the functional impairment was apparent with manual scoring in the grid walking test. In contrast, our DLC-based grid walking model provided fully automated quantification of this widely used test. It achieved millimeter-scale tracking accuracy and detected a ∼3-fold increase in foot faults in stroke-affected mice compared to controls, clearly differentiating between affected and unaffected forepaws for at least 2 weeks post-stroke. These findings expose the limitations of the commercial automated ladder test and demonstrate that the DLC-based pose-estimation-assisted misstep quantification in grid-walk test provides greater sensitivity for detecting motor deficit.