Video-based detection of pentylenetetrazol induced severe convulsions in cynomolgus monkeys

Abstract

Central nervous system (CNS) toxicity is one of the important toxicities in nonclinical studies for drug development. Abnormal behaviors often serve as indicators of CNS abnormalities. However, the exact timing of such symptoms' occurrence is usually uncertain, and accurate detection of abnormal behaviors is difficult since real-time observation or long-time observation of video data is required. To address this, there is a need for automated continuous monitoring methods that can assist observers and enhance detection sensitivity. In this study, we attempted to develop an advanced model for detecting drug-induced convulsions using video data of cynomolgus monkeys (n = 7) after administration of pentylenetetrazol, a GABA receptor antagonist, at a dose of 70 mg/kg. To detect convulsions, we tested various algorithms including classical image processing techniques, supervised learning models that use frequency information derived from wavelet transformation of each body part based on pose estimation, and unsupervised learning models that identify abnormalities through training only on videos of animals exhibiting ordinary behavior. Our results showed that the most feasible approach was to track regularly sampled points on the animal's body using optical flow. This method calculates the apparent motion of pixels between consecutive frames by analyzing the intensity patterns. This enabled the detection of convulsions based on characteristic frequencies in the tracking point trajectories, observed via Fourier transformation. Despite the limitation of available data, our model captured several convulsion patterns and demonstrated encouraging scores (accuracy: 77.8 %, false positive rate: 1.7 per hour). In conclusion, we have established a prototype of the convulsion detection model although fine-tuning is still necessary. The use of optical flow, which is also employed in human seizure detection models, as a benchmark for automatically detecting convulsions in monkeys holds potential in combination with other methods and analysis, with the possibility of further improvement with larger datasets.