Convolutional neural networks detected changes in facial expressions of mice that had conditioned fear stress

Animal models are essential for studying aversive states such as fear and pain. Facial expressions may provide non-invasive readouts of aversive states in animals. This study investigated whether changes in facial expressions, which are potentially consistent between humans and mice, can serve as objective indicators of fear responses and distinguish fear from pain. We analyzed changes in the facial expressions of mice associated with conditioned fear stress (CFS) using convolutional neural networks (CNNs). Photographs of CFS and control mice were analyzed using four advanced CNN models: VGG16, ResNet50, DenseNet121, and InceptionV3. The CNNs identified CFS mice from facial images under contrasts: control/non-freezing vs CFS/freezing and control/non-freezing vs CFS/non-freezing, with consistently high performance (Control/non-freezing vs CFS/freezing: sensitivity 0.942, specificity 0.929, accuracy 0.935, precision 0.929, AUC 0.966; Control/non-freezing vs CFS/non-freezing: sensitivity 0.912, specificity 0.900, accuracy 0.906, precision 0.902, AUC 0.950). The ability to detect CFS without freezing decreased as stress intensity weakened, from an AUC of 0.950 to 0.701, suggesting that CNNs can detect facial changes depending on the degree of stress exposure. Facial changes were particularly pronounced in freezing mice, further supporting their association with CFS-related emotional responses. During testing, mice were returned to the conditioning chamber without shock; therefore, this facial expression could reflect fear response rather than pain response. These findings demonstrate the potential of CNNs to serve as non-invasive tools for detecting stress-induced affective changes in mice.