Detecting Deviation in Rodent Behaviour Through Home Cage Analysis

该项目涉及应用机器学习自动检测实验室小鼠的行为变化。通过视频记录和RFID跟踪，可以对受试者进行长期的标准化观察，从而允许研究人员事后提取高级别信息。该项目的主要目的是确定是否有可能通过分析从家庭笼中收集的自发行为来确定动物的福利状态。该项目的第一个里程碑将是产生一个分类器，该分类器在两个不同类别的小鼠中提供足够的训练集，可以高精度地确定给定的小鼠记录是否属于一个类别或另一个类别。首先，该模型将被训练以区分两种非常不同的小鼠品系，例如具有正常昼夜节律的品系和没有任何昼夜节律的品系。首先，模型将使用更容易设计的特征进行训练。这些因素包括：饮水和进食的时间，攀爬和挖洞等活动，以及睡眠模式。如果模型显示出更精细的特征，那么工作将包括工程检测更自然的行为，如梳理和社会互动。最初提出的用于表征时间序列数据的方法是将数据集划分为单个小时，并从该小时的运动活动中设计特征。然后，解释活动和睡眠的周期和模式的特征可以适合更传统的机器学习技术。探索的另一个途径是应用更深入的学习策略，试图以更无方向的方式识别活动模式。

This project is concerned with applying machine learning to automate detection of behavioural changes in laboratory mice. Through video recording and RFID tracking it is possible to perform long-term standardised observation of subjects, allowing for high-level information to be extracted by researchers afterwards. The main aim of this project is to determine whether is is possible to determine the welfare state of the animal through the analysis of spontaneous behaviour gathered from home cages.The first milestone for this project would be producing a classifier that, given a sufficient training set of mice in two distinct classes, can determine wth high accuracy whether a given mouse recording belongs to one class or the other. To start with the model will be trained to differentiate between two very distinct strains of mice, such as a strain with a normal circadian rhythm and a strain without any. At first the model will be trained using features that are more easily engineered. These would include: time spent drinking and feeding, activities such as climbing and burrowing, and sleeping patterns. If the model shows significant improvement with more elaborate features, then work would include engineering detection of more naunced behaviours such as grooming and social interactions.The initial proposed method for characterising the time series data is to split the dataset into individual hours, and engineer features from the locomotor activity over that hour. Features that interpret periods and patterns of activity and sleep can then be fitted to more conventional machine learning techniques. Another avenue of exploration is to apply deeper learning strategies to attempt to recognise patterns in activity in a more undirected fashion.