Deep Transferable Intelligence for Spatial Variability Characterization and Data-Efficient Learning in Biomechanical Measurement
Deep Transferable Intelligence for Spatial Variability Characterization and Data-Efficient Learning in Biomechanical Measurement
复制标题
DOI：
10.1109/tim.2023.3265753
复制
发表时间：
2023-01-01
影响因子：
5.6
通讯作者：
Zhang, Qingxue
中科院分区：
文献类型：
--
作者：
Gangadharan, Kiirthanaa;Zhang, Qingxue
生物力学测量在康复、辅助生活和生活方式管理应用方面具有很好的应用价值。然而，对生物力学动力学的空间可变性的理解仍然有限，这对于优化运动传感器配置至关重要。此外，训练身体活动检测器通常需要大量的数据和时间。针对这两个挑战，在本研究中，我们提出了一种新的深度迁移智能框架，该框架利用深度学习来表征不同身体位置上不同运动传感器的空间变异性，并进一步利用主体间迁移学习来最大限度地提高数据效率，以挑战稀缺数据学习。更具体地说，为了表征空间变异性，我们提出了深度卷积神经网络（cnn）来研究不同传感器位置和通道对身体活动测量的能力。表征决定了最优传感器配置和最优通道配置。此外，我们提出了一种迁移学习方法来挖掘学科间的相似性，然后在学科之间共享学习到的知识，从而在可穿戴的稀缺数据学习场景中最小化训练工作量，最大化数据效率。我们的评估实验从7个选项中确定了最佳传感器位置为大腿，从42个选项中确定了最佳传感器和通道配置为大腿-加速度计-轴-Y。我们的实验进一步证明，在最优传感器和通道配置的迁移学习下，仅10%的目标对象数据用于模型微调就可以产生高达91.6%的物理活动检测（PAD）精度，与不进行迁移学习的直接学习相比，性能提高了9%。因此，深度可转移学习框架将极大地促进生物医学测量中最优传感器和通道配置的空间变异性表征以及有效的稀缺数据学习。
Biomechanical measurement is of promising value for rehabilitation, assisted living, and lifestyle management applications. Nevertheless, the understanding is still limited on the spatial variability of biomechanical dynamics that is essential for optimal motion sensor configuration. Besides, training physical activity detectors is usually data-heavy and time-consuming. Targeting these two challenges, in this study, we propose a novel deep transfer intelligence framework, which leverages deep learning to characterize the spatial variability of different motion sensors on diverse body locations, and further leverages intersubject transfer learning to maximize data efficiency in challenging scarce data learning. More specifically, to characterize the spatial variability, we propose deep convolutional neural networks (CNNs) to investigate the capabilities of both different sensor locations and channels on physical activity measurement. The characterization determines both optimal sensor configuration and optimal channel configuration. Further, we propose a transfer learning approach to mine intersubject similarity and then share learned knowledge among subjects, thereby minimizing the training effort and maximizing the data efficiency in the wearable scarce data learning scenario. Our evaluation experiments have determined the optimal sensor location from seven options as thigh, and the optimal sensor and channel configuration from 42 options as thigh-accelerometer-axis -Y. Our experiments have further demonstrated that, with transfer learning under the optimal sensor and channel configuration, only 10% of data from the target subject for model fine-tuning can yield a physical activity detection (PAD) accuracy of up to 91.6%, with a performance boosting of 9% compared with direct learning without transfer learning. Therefore, the deep transferable learning framework will greatly advance spatial variability characterization for optimal sensor and channel configuration and efficient scarce data learning in biomedical measurement.