Assessment of Visual Function in Mice Using Light/Dark Box and Multi-Feature Machine Learning

The light/dark box test can be used to assess visual function in rodents based on their spontaneous behavior in response to light. Commonly used assay relies on a single behavioral metric, dwell time in the light or dark compartment, which may be influenced by factors other than vision, leading to unreliable assessment results. To overcome this, we developed a multi-feature machine learning paradigm by extracting multiple mouse behavioral metrics, standardizing them as features to train machine learning models, thereby achieving reliable and automated vision assessment. We systematically compared the classification performance of single-metric versus multi-feature machine learning approaches in sighted and blind mice, using wild-type and rhodopsin-deficient mice, with a subset further subjected to double optic nerve crush. We found that the multi-feature method can improve classification performance and exhibit great robustness to different experimental settings. Additionally, we further improved model performance by applying feature importance analysis and constructing an optimized feature subset. These findings suggest that the reliability of commonly used single dwell time measure for vision assessment could become unreliable, as shown in our experiment, probably because in addition to vision other factors also impact dwell time. Our study demonstrated an improved assessment method based on a combination of multiple behavior features through machine learning. Author SummaryAssessing visual function in mice is essential for studying eye diseases and drug development. The light/dark box test evaluates visual function by measuring the spontaneous behavioral response of mice to light, providing a training-free behavioral approach that helps simplify the assessment process and improve research efficiency. However, traditional light/dark box tests rely on a single behavioral metric, dwell time in the light or dark compartment, to assess visual function, which may be influenced by factors other than vision, such as anxiety and exploratory behavior, leading to limited reliability of assessment results. Here, we demonstrate that integrating multiple behavioral features through machine learning can improve the reliability and stability of vision assessment. By automatically tracking and analyzing various behavioral metrics of mice, such as movement patterns, speed, and spatial preferences, the proposed method can more reliably distinguish between sighted and blind mice. Furthermore, the method demonstrates stable performance across different experimental settings, showing good applicability. This automated, reliable, and easily generalizable method can provide a convenient and efficient means for visual assessment in preclinical research, facilitating vision disease research and drug development.