Detection of abnormal behavior and convulsion in cynomolgus monkeys using video-based deep learning

Drug-induced behavior changes in nonclinical safety studies could correlate to serious side effects in humans and hence must be detected prior to clinical trials. We investigated a video analysis using deep learning to assess locomotor activity (LA) and convulsions in conscious non-human primates (NHPs). The training data was created by annotating the waist landmarks for pose estimation based on the video data from each cage. The time series of 2D coordinates was calculated with the deep learning-based pose estimation model. Study 1. LA evaluation: Apomorphine hydrochloride (APO, 0 or 1 mg/kg, sc), non-selective dopamine agonist, was dosed (n = 4). Animal behaviors were recorded for 4 h after dosing. A momentum was calculated from the difference between the frames of 2D coordinates, and a trajectory was created from transition of 2D coordinates for each 1-h epoch. Study 2. Convulsion detection: Pentylenetetrazole (PTZ, 60 to 70 mg/kg, sc), GABAA receptor antagonist as a convulsant agent, was dosed (n = 4). Animal behaviors were recorded until the occurrence of convulsion under light-on or -off (with infrared radiation) condition. A convulsion was determined when the number of frames was greater than the threshold for ten seconds using 2D coordinates and vibration intensity calculated with wavelet transformation. In Study 1, behavioral changes characteristic of APO (circling, jumping, cage licking, etc.) were observed differently between animals. The LA increased in two NHPs, showing different trajectories correlated with circling or jumping in the cage. In Study 2, PTZ induced convulsions within about 1-h postdose in all NHPs and the algorithm could detect the convulsions in all animals under both light-on and -off conditions while this algorithm did not falsely detect any convulsions in movie files in the animals in study 1. Our deep learning model using video data is considered useful to assess LA and convulsion in NHPs.