Quantifying ultrasonic mouse vocalizations using acoustic analysis in a supervised statistical machine learning...

Area of Science:

• Bioacoustics
• Animal Behavior
• Machine Learning in Biology

Background:

• Rodent vocalizations, particularly mouse ultrasonic vocalizations (USVs), offer insights into disease progression, social behavior, and emotional development.
• Current methods for USV analysis rely on expert classification or basic acoustic metrics, limiting standardization and scalability.
• USVs are crucial for mouse communication but occur at frequencies beyond human hearing.

Purpose of the Study:

• To replicate expert-defined mouse ultrasonic vocalization (USV) call types using computational acoustic analysis.
• To develop a supervised machine learning model for accurate and automated classification of USVs.
• To identify a parsimonious set of acoustic measures predictive of USV categories.

Main Methods:

• Utilized acoustic analysis to characterize mouse ultrasonic vocalizations (USVs).
• Employed four feature selection algorithms to identify optimal acoustic measures.
• Trained and evaluated Support Vector Machines (SVM) and Random Forests (RF) models using 10-fold cross-validation with 100 repetitions.

Main Results:

• A subset of 8 acoustic measures, when input into a Random Forest model, achieved 85% correct out-of-sample classification of USVs.
• The machine learning approach successfully replicated expert-defined USV call types.
• Demonstrated the efficacy of acoustic measures and machine learning in automating USV classification.

Conclusions:

• Acoustic measures combined with machine learning provide a robust and automated method for classifying mouse ultrasonic vocalizations (USVs).
• This approach enables labs to standardize USV analysis, compare data across studies, and gain deeper insights into vocal-behavioral patterns.
• The developed model accurately replicates expert classifications, facilitating research in disease modeling and animal behavior.