Swaminathan Krithika, Choe Dabin K, Kim Daekyum, Barde Flore, Baker Teresa C, Wendel Nicholas C, Chin Andrew, Bergamo Gregoire, Siviy Christopher J, Lee Christina, Awad Louis N, Ellis Terry D, Walsh Conor J
IEEE Trans Neural Syst Rehabil Eng. 2025;33:2273-2285. doi: 10.1109/TNSRE.2025.3577961.
Gait rehabilitation is critical for regaining locomotor independence after neuromotor injuries like stroke. Rehabilitation literature indicates the need for such therapy to continue beyond the clinic in order to maintain motor function and support recovery. However, implementing community-based rehabilitation requires the ability to monitor gait in the real-world with clinically relevant accuracies. Despite advances in machine learning, achieving this performance with single sensing modalities has been challenging using wearable sensors like inertial measurement units (IMUs) and pressure insoles. Here, we investigate the benefits of multi-modal sensing by integrating IMU and insole data to develop individualized machine learning models in people post-stroke that estimate propulsion, a key biomechanical variable. We show that in the lab, IMU + Insole models improve performance relative to IMU only and Insole only models, with an average root-mean-squared-error (RMSE) of 0.80 %bodyweight (%BW) across the stance phase. We obtain RMSEs of 0.71%BW for peak paretic propulsion and 0.19%BW s for paretic propulsion impulse, which are within corresponding clinical thresholds. We then explore the application of this algorithm to track propulsion changes in the real-world for two participants during variable-speed walking and two participants during active gait interventions, either functional electrical stimulation or exosuit-applied resistance. For these participants, we observe similar changes in measured propulsion in the lab and estimated propulsion out of the lab across speeds and interventions. Overall, this work aims to address the challenges in applying machine learning methods for individuals post-stroke and presents an investigation into the feasibility of developing estimation methods for real-world propulsion estimation during gait rehabilitation.
步态康复对于中风等神经运动损伤后恢复运动独立性至关重要。康复文献表明，此类治疗需要在临床环境之外持续进行，以维持运动功能并支持恢复。然而，实施基于社区的康复需要具备在现实世界中以临床相关精度监测步态的能力。尽管机器学习取得了进展，但使用惯性测量单元（IMU）和压力鞋垫等可穿戴传感器通过单一传感模式实现这种性能一直具有挑战性。在此，我们通过整合IMU和鞋垫数据来研究多模式传感的益处，以开发针对中风后患者的个性化机器学习模型，该模型可估计推进力，这是一个关键的生物力学变量。我们表明，在实验室中，IMU + 鞋垫模型相对于仅使用IMU和仅使用鞋垫的模型性能有所提高，在整个站立阶段的平均均方根误差（RMSE）为0.80%体重（%BW）。我们获得的患侧峰值推进力的RMSE为0.71%BW，患侧推进力冲量的RMSE为0.19%BW·s，均在相应的临床阈值范围内。然后，我们探索了该算法在现实世界中的应用，以跟踪两名参与者在变速行走过程中以及另外两名参与者在主动步态干预（功能性电刺激或外骨骼施加阻力）过程中的推进力变化。对于这些参与者，我们观察到在实验室中测量的推进力和实验室外估计的推进力在不同速度和干预下有相似的变化。总体而言，这项工作旨在解决将机器学习方法应用于中风后个体的挑战，并对开发步态康复期间现实世界推进力估计方法的可行性进行研究。
