Wireless AI-Powered IoT Sensors for Laboratory Mice Behavior Recognition

According to the U.S. Department of Agriculture in 2018, there are more than 100 million animals used in research, education, and testing per year. Of the laboratory animals used for research, 95 percent are mice and rats as reported by the Foundation for Biomedical Research (FBR). We present here our work in developing wireless Artificial Intelligent (AI)-powered IoT Sensors (AIIS) for laboratory mice motion recognition utilizing embedded micro-inertial measurement units (uIMUs). Based on the AIIS, we have demonstrated a small-animal motion tracking and recognition system that could recognize 5 common behaviors of mice in cages with accuracy of ~76.23%. The key advantage of this AIIS-based system is to enable high throughput behavioral monitoring of multiple to a large group of laboratory animals, in contrast to traditional video tracking systems that usually track only single or a few animals at a time. The system collects motion data (i.e., three axes linear accelerations and three axes angular velocities) from the IoT sensors attached to different mice, and classifies these data into different behaviors using machine learning algorithms. One of the challenging problems for data analysis is that the distribution of behavior samples is extremely imbalanced. Behaviors such as sleeping and walking dominate the entire sample set from different mice. However, machine learning algorithms often require a balanced sample set to achieve optimal performance. Thus, several methods are proposed to solve the imbalanced sample problem. Data processing methods for data segmentation, feature extraction, feature selection, imbalanced learning, and machine learning are explored to process motion data including sleeping, walking, rearing, digging, shaking, grooming, drinking and scratching. For example, by tuning the parameters of a machine-learning algorithm (i.e., Support Vector Machine (SVM)), the average accuracy of classifying five behaviors (i.e., sleeping, walking, rearing, digging and shaking) is 48.07% before solving the imbalance sample issue. To address this problem, an iteration of sample and feature selection is applied to improve the SVM performance. A combination of oversampling and undersampling is used to handle imbalanced classes, and feature selection provides the optimal number of features. The accuracy increases from 48.07% to 76.23% when the optimized combination is used. We further obtained an average accuracy of 86.46% by removing shaking, which is proved to have a negative effect on the overall performance, out of these five behaviors. Furthermore, we were able to classify less frequent behaviors including rearing, digging, grooming, drinking and scratching at an average accuracy of 96.35%.