基于深度神经网络和逐层相关性传播技术探究“高-低”里程跑者步态模式差异

目的 通过深度神经网络(deep neural network, DNN)分类模型揭示高里程跑者( high-mileage runner, HMR) 和低里程跑者 ( low-mileage runner, LMR) 跑 步 步 态 模 式 差 异, 并 探 讨 逐 层 相 关 性 传 播 ( layer-wise relevance propagation, LRP)技术解释 DNN 分类器模型的决策有效性。 方法 通过 DNN 对 HMR 和 LMR 总计 1 200 组跑步 步态特征数据进行训练分类识别,采用 LRP 计算相关变量在不同步态阶段的相关性得分( relevance score, RS),提 取高相关变量对步态模式差异进行解释性分析。 结果 DNN 对 HMR 和 LMR 的跑步步态模式特征分类精度达到 91. 25% 。 LRP 计算结果显示支撑前期(1% ~ 47% )各变量的成功分类贡献率高于支撑后期(48% ~ 100% )。 踝关节 相关轨迹变量 RS 的贡献率总和达到 43. 10% ,膝、髋关节贡献率分别为 37. 07% 、19. 83% 。 结论 膝、踝关节相关 生物力学参数对识别 HMR 和 LMR 步态特征的贡献程度最高。 跑步支撑早期可能包含更多步态模式信息,能够提 升步态模式识别的有效性和敏感性。 LRP 实现了对模型预测结果的可行性解释,从而为分析步态模式提供了更有 趣的见解和更有效的信息。

Exploration of Gait Pattern Differences Between High-Mileage andLow-Mileage Runners Based on Deep Neural Network and Layer-Wise Relevance Propagation

Objective To reveal the gait pattern differences between higher-mileage runners ( HMR) and low-mileage runners ( LMR) by using the deep neural network ( DNN) classification model, and investigate the interpretability analysis of successfully recognized gait patterns by layer-wise relevance propagation ( LRP) technique. Methods Through DNN, 1 200 groups of gait feature data from HMR and LMR were trained and classified. Then, the LRP was used to calculate the relevance score ( RS) of relevant variables at each time point, and the high relevance variables were extracted to analyze the interpretability of gait pattern differences. Results The DNN model achieved 91. 25% accuracy in gait feature classification between HMR and LMR. The contribution of variables during 1% -47% stance phase was higher than the contribution of variables during the 48% -100% stance phase to the successful classification. The sum contribution rate of the ankle joint related trajectory variable RS reached 43. 10% , and that of the knee joint and hip joint was 37. 07% and 19. 83% , respectively. Conclusions The ankle and knee provide considerable information can help recognize gait features between HMR and LMR. The early stages of the stance are very important in the term of gait pattern recognition because it may contain more effective information about gait patterns. LRP completes a feasible interpretation of the predicted result of the model, thus providing more interesting insights and more effective information for analyzing gait patterns.