DeepMate: Integrating Deep Learning with Ultrasonic and Video Data for Predicting Mouse Mating Behaviors. Na Zhao

Mice communicate using ultrasonic vocalizations (USVs) during various social interactions, including mating, making these vocalizations a critical area of study for behavioral and biomedical research. Current computational models struggle to analyze mating-specific USVs, especially in drug-influenced contexts. This paper introduces DeepMate, a deep learning model that integrates video-based behavior recognition with USV analysis to predict mating behavior in mice. This model dynamically merges and processes ultrasonic waves to form time series data of unequal lengths. Meanwhile, it adopts the bidirectional long Short-Term Memory network (BiLSTM) with attention mechanism to classify the behavioral data in the ultrasonic signals, and uses the yolo enhanced video tracking module for visual behavior recognition. Experimental results reveal that DeepMate achieves a prediction accuracy of 98.57%, outperforming conventional models. By identifying unique patterns in USVs during mating and non-mating periods, DeepMate provides insights into mouse behavior under normal and pharmacological conditions, particularly in carbidopa treated mice. This study highlights the potential of DeepMate as a tool for advancing research on mating-related behaviors and drug effects.