Automated detection of the head-twitch response using wavelet scalograms and a deep convolutional neural network

A new automated method detects head-twitch responses (HTR) in mice, a behavior induced by hallucinogens, with 99.4% sensitivity. Traditional methods relying on amplitude-time features are unreliable due to variability in twitch characteristics and similarity to other behaviors like jumping. The approach transforms magnetometer recordings into time-frequency visual representations (scalograms), extracts deep features using a pretrained convolutional neural network (ResNet-50), and classifies them with a Support Vector Machine algorithm. Tested on recordings from 237 mice containing 11,312 HTR, the method correctly identified 11,244 twitches and was insensitive to other behaviors such as jumping and seizures. Deep learning on scalograms enables robust, reliable automated HTR detection.

Study at a glance

Abstract

Hallucinogens induce the head-twitch response (HTR), a rapid reciprocal head movement, in mice. Although head twitches are usually identified by direct observation, they can also be assessed using a head-mounted magnet and a magnetometer. Procedures have been developed to automate the analysis of magnetometer recordings by detecting events that match the frequency, duration, and amplitude of the HTR. However, there is considerable variability in the features of head twitches, and behaviors such as jumping have similar characteristics, reducing the reliability of these methods. We have developed an automated method that can detect head twitches unambiguously, without relying on features in the amplitude-time domain. To detect the behavior, events are transformed into a visual representation in the time-frequency domain (a scalogram), deep features are extracted using the pretrained convolutional neural network (CNN) ResNet-50, and then the images are classified using a Support Vector Machine (SVM) algorithm. These procedures were used to analyze recordings from 237 mice containing 11,312 HTR. After transformation to scalograms, the multistage CNN-SVM approach detected 11,244 (99.4%) of the HTR. The procedures were insensitive to other behaviors, including jumping and seizures. Deep learning based on scalograms can be used to automate HTR detection with robust sensitivity and reliability.