Application of video surveillance in preclinical safety studies in canines: Understanding the interobserver reliability and validity to recognize clinical behavior

by Eline Eberhardt, Fetene Tekle, Greet Teuns, Jill Witters, Bianca Feyen, Sarah De Landtsheer, Ivan Kopljar

Preclinical in vivo studies are critical to identify potential adverse effects of drugs under development. However, a significant number of drug candidates are terminated during human clinical trials due to unexpected adverse events which were not predicted or detected in preclinical studies. Video surveillance can be a valuable tool to reduce the risk of missing and/or misclassifying adverse clinical observations (COs) in animals. To explore the applicability of detecting COs on video, the agreement between observers (reliability) and agreement between observers and experts (construct validity) of 13 important COs was evaluated. The reliability was investigated by evaluating the interobserver agreement between 23 observers with different experience levels and primary roles on defined COs and normal behavior, recorded on video during preclinical studies in canines. The validity was investigated by comparing the observers’ assessments to the ground truth confirmed by three experts. This investigation showed a substantial reliability and validity of the observers’ assessments without significant differences between experience levels or primary roles. Normal behavior was challenging to recognize (56% correct), while half of the COs appeared straightforward to identify with a validity of ≥ 90%: salivation, aggressiveness, circling, vomiting, head shaking and convulsions. Other COs were more challenging to detect with lowest scores for limb stiff/hypertonia, tremors and excitation. Regardless of experience-level, observers missed very few COs. This investigation showed the complexity when multiple COs occurred simultaneously, as well as the limitations of differentiating between visually similar COs (tremors vs. twitches; limping vs. limb stiff) on video without the possibility of in-person observation. Given the substantial overall reliability and validity, it is concluded that clinical canine behavior can be accurately detected on video by trained observers. This permits more objective and quantifiable monitoring of animal behavior and application of computer vision for future automatic monitoring of canine studies.