Accurate and real-time calculation of whole-brain strain and strain rate in head impact is much in need for medical professionals and managers in sports teams to better understand the potential damage at specific brain regions and how serious it is and to monitor traumatic brain injury (TBI) and give targeted suggestions.Conventional methods of calculating the brain strain and strain rate typically involve Finite Element Analysis (FEA), which is time-consuming, resource-consuming, and difficult to model efficiently. These factors prevent an efficient, deployable solution for players and sport-team managers for real-time brain impact analysis. Instead, a faster method that is easily accessible to players and medical professionals is required for more practical applications. Although previous studies have developed machine learning head models (MLHM) to compute brain strain in a faster manner, they are limited in strain calculation and the accuracy deteriorates on on-field data, which indicates the generalizability issue of MLHM.To improve the generalizability across types of head impacts, particularly the TBI risk estimation accuracy on small on-field impact datasets, and further optimize the MLHM to be both fast and accurate to predict strain rate from the perspective of model optimization, we developed two strategies for more powerful MLHM: the data fusion and the transfer learning. With these two strategies, we developed a series of MLHM on the impact datasets from simulation (HM, 12,780 impacts), college football (CF, 302 impacts), mixed martial arts (MMA, 457 impacts), car crashes (NHTSA, 48 impacts) and NFL impacts (NFL, 36 impacts), to be able to predict both maximum principal strain (MPS) and MPS rate (MPSR) accurately for specific types of head impacts. The reference MPS and MPSR were calculated with the validated KTH model. To develop the MLHM, we extracted 510 temporal and spectral features by taking a union of features extracted in our two previous studies, from the four kinematics: linear acceleration, angular velocity, angular acceleration, and angular jerk.We partitioned dataset HM into 70% training, 15% validation, and 15% test data and developed the basis MPS/MPSR models on the HM training data. Then, we partitioned the on-field datasets (CF/MMA/NFL/NHTSA) into training/validation/test data. Two strategies were developed to adapt the prior knowledge of brain dynamics from dataset HM to better predict the MPS/MPSR of on-field impacts: 1) data fusion: we combine dataset HM and on-field training data to train the MPS/MPSR models; 2) transfer learning: we pre-train the models on dataset HM and then fine-tune the weights on the on-field training data. The deep neural network was of a five-layer structure. The accuracy of the models was assessed on the test set with mean absolute error (MAE). We compared the accuracy of the models developed with the two proposed strategies and those models trained on the on-field training data.On dataset HM, the basis MPS/MPSR models showed accuracy with averaged MAE of 0.015 for the MPS prediction, and 2.818 for the MPSR prediction over 20 parallel experiments. On dataset CF and MMA, the transfer-learning models were generally the most accurate models (p&lt;0.01) and reached the averaged MAE smaller than 0.015 in MPS prediction and smaller than 4 in MPSR prediction. Furthermore, we have shown that the best-performing transfer-learning models were accurate across different brain regions on the datasets with slightly higher accuracy on the brainstem, cerebellum, and thalamus. On the two smaller on-field datasets: NFL and NHTSA, the transfer-learning-based models were also the most accurate models with averaged MAE smaller than 0.03 for MPS prediction and smaller than 7 for MPSR prediction. The models trained on the limited on-field impacts (18 impacts for NFL and 24 impacts for NHTSA) generally failed to predict the MPS and MPSR for these two datasets.The MLHM based on transfer learning enables fast and accurate computation of brain strain and strain rate resulting from various on-field impacts for real-time TBI risk estimation, which cannot be achieved with the FEM and previous MLHM developed by researchers.

Rapidly Estimate Brain Strain and Strain Rate on Various Types of Head Impacts With Transfer Learning and Data Fusion on Deep Neural Network

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.