Machine learning approach for analysis of rodent’s behaviors

Video analysis of animal behavior plays an important role in different preclinical research field, from pharmacological studies to neuroscience experiments. Unbiased quantification of animal behavior has always been an important challenge in the scientific test, and different approaches have been used to resolve it. For simpler behaviors that can be characterized by the position of the animal such as elevated plus maze or open field, automatic systems based on classical algorithms can be used effectively. On the other hand, for more complex behavior analysis human intervention is still necessary, but this task is time-consuming for the researcher and prone to human bias. The bias refers to both different perception and procedure by different researchers even in the same lab group (inter-variability), and time variation in the accuracy of the same researcher (intra-variability) influenced by the expertise acquired over the time. Therefore, it becomes clear that these problems limit the reliability and reproducibility of some behavioral experiments. Over the years, machine learning has been more and more used for human and animal automatic behavior characterization. Machine learning, similarly to other kinds of algorithms, permits to have an automatic and standardized scoring of the behaviors, freeing the researcher from the intense task of manual annotation. Moreover, machine learning techniques can extrapolate complex patterns not always discernible by classical algorithms, but necessary for the analysis of complex behaviors. The present thesis aims to develop machine learning models for the analysis of animal behaviors, in particular rodents. Two commonly used behavioral tests were considered: the forced swim test, and the free social interaction test. The forced swim test is a behavior test used to evaluate antidepressant drugs efficacy, while the free social interaction test is used to evaluate social experiences. In Chapter 1 I give an overview of the machine learning techniques used for the analysis of behaviors. Then, in the following chapters, the application of machine learning to the two behavioral tests is described in details. In Chapter 2 I describe the application of machine learning for the forced swim test. To train the model we constructed a dataset of videos. After the training, I validated the model on two pharmacological experiments, to evaluate its performance. The model is a 3D convolutional neural network to extract spatiotemporal features from a sequence of frames of the video. In Chapter 3 I present two different algorithms I build up for the social interaction test. Here, to train the two different models, I used a dataset provided by the Caltech Institute: the CalMS21 dataset. The first machine learning algorithm relies on the graph neural network to extract interesting spatiotemporal features for the classification of social behaviors from the coordinates of keypoints. The second algorithm tries to circumvent the use of keypoints for the classification of such behaviors by using a model based on transformer encoders. In conclusion, these studies together demonstrate how machine learning is a valid, human-bias-free, and consistent method to score complex behaviors. Applying different and specific algorithms to the specific experimental test considered, it is possible to standardize the scoring process and obtain more reliable data that can be reproducible and replicable.