Optimizing choice and timing of behavioral outcome tests following repetitive mild traumatic brain injury: A machine learning-based approach on multiple preclinical experiments

Repetitive mild traumatic brain injury (rmTBI) is a potentially debilitating condition with long-term sequelae. Animal models are used to study rmTBI in a controlled environment, but there is currently no established standard battery of behavioral tests used. Primarily, we aimed to identify the best combination and timing of behavioral tests to distinguish injured from non-injured animals in rmTBI studies, and secondarily, to determine whether combinations of independent experiments have better behavioral outcome prediction accuracy. Data of 1,203 mice from 58 rmTBI experiments, some of which has already been published, was used. In total, 11 types of behavioral tests were measured by 37 parameters at 13 timepoints during the first 6 months after injury. Univariate regression analyses were used to identify optimal combinations of behavioral tests and whether the inclusion of multiple heterogenous experiments improved accuracy. k-means clustering was used to determine whether a combination of multiple tests could discriminate mice with rmTBI from non-injured mice. We found that a combination of behavioral tests outperformed individual tests alone when discriminating animals with rmTBI from non-injured animals. The best timing for most individual behavioral tests was 3-4 months after first injury. Overall, Morris water maze (MWM; hidden and probe frequency) was the behavioral test with best capability of detecting injury effects (AUC=0.98). Combinations of open field tests and elevated plus mazes also performed well (AUC=0.92), as well as the forced swim test alone (AUC=0.90). In summary, multiple heterogeneous experiments tended to predict outcome better than individual experiments and MWM 3-4 months after injury was the optimal test, but several combinations performed well too. In order to design future preclinical rmTBI trials, we have included an interactive application available online utilizing the data from the study.