Classifying sex and strain from mouse ultrasonic vocalizations using deep learning.

Area of Science:

• Animal Communication
• Bioacoustics
• Machine Learning in Biology

Background:

• Mice utilize a diverse range of ultrasonic vocalizations (USVs) for communication in various social contexts.
• Identifying the sex of interacting mice is crucial, but prior research was inconclusive on whether individual USVs convey this information.

Purpose of the Study:

• To investigate if ultrasonic vocalizations (USVs) contain information about the sex of the emitting mouse.
• To develop and evaluate machine learning models for automated sex classification from mouse USVs.

Main Methods:

• Deep neural networks (DNNs) were employed to classify mouse sex directly from USV spectrograms.
• Performance was compared against Support Vector Machines (SVM) and regression models.
• Network architecture and feature importance were analyzed to understand sex differentiation within USVs.

Main Results:

• DNNs achieved unprecedented performance (77%) in classifying mouse sex from USVs, significantly outperforming SVM (56%) and regression (51%).
• Higher accuracy (85%) was observed when DNNs incorporated individual mouse properties, though with limited practical application.
• The study identified spectrotemporal differences in USVs enabling partial classification and sexual recognition, even in cortexless mice.

Conclusions:

• Spectrotemporal characteristics of mouse ultrasonic vocalizations (USVs) contain sufficient information for sex classification.
• Advanced deep learning models can automate sex attribution in USVs, facilitating the analysis of mouse social behavior.
• This technology advances our understanding of animal communication and social recognition mechanisms.