Mobarak Rami, Mengarelli Alessandro, Khushaba Rami N, Al-Timemy Ali H, Verdini Federica, Fioretti Sandro, Burattini Laura, Tigrini Andrea
IEEE Trans Neural Syst Rehabil Eng. 2025;33:3476-3487. doi: 10.1109/TNSRE.2025.3604618.
Myoelectric pattern recognition systems serve as a promising predictive control approach for the lower limbs prostheses and exoskeletons. However, their actual deployment is challenged by the signal stochastic nature that could contaminate the decision stream with physiologically implausible transitions, posing safety and metabolic cost concerns on the potential user. Therefore, this study proposes a novel Physics-Informed Bayesian Fusion (PI-BF) post-processor that embeds biomechanical sequentiality constraints into the posterior probabilistic output of the classifiers to suppress unstable transitions and promote natural gait progression. Time-Domain (TD) and Time-Dependent Power Spectrum Descriptors (TD-PSD) features were extracted from the lower limbs muscles surface electromyography (sEMG) signals and classified using Support vector machines (SVM), Artificial neural networks (ANN), K-Nearest Neighbour (KNN), and a CNN-LSTM hybrid deep learning model to predict five phases of gait cycle. The output of these classifiers was followed by the proposed PI-BF postprocessor and it was compared against Bayesian Fusion (BF) Majority voting (MV) as well as the performance without post-processing (WPP) using different numbers of votes from the previous windows. Results shows that PI-BF can increase the classification accuracy by up to 5.5% reaching up to 85% in SIAT-LLMD dataset (40 subjects) using SVM with 3 previous decision windows. It also reduced Transition Detection Difference (TDD) to 0.1 ± 59.8 ms and improved output stability by 5%, as measured by the Instability (INS) index. The proposedPI-BF exhibited consistent improvements in real-time gait phase recognition experiments, achieving classification accuracies of around 90%. These results demonstrate that PI-BF offers a practical, low-complexity solution for enhancing the safety, reliability, and real-time performance of myoelectric control in assistive lower-limb devices.
肌电模式识别系统是下肢假肢和外骨骼一种很有前景的预测控制方法。然而，其实际应用面临信号随机性的挑战，这种随机性可能会使决策流受到生理上不合理的转换影响，给潜在用户带来安全和代谢成本方面的担忧。因此，本研究提出了一种新颖的物理信息贝叶斯融合（PI-BF）后处理器，该后处理器将生物力学顺序约束嵌入到分类器的后验概率输出中，以抑制不稳定的转换并促进自然步态进展。从下肢肌肉表面肌电图（sEMG）信号中提取时域（TD）和时变功率谱描述符（TD-PSD）特征，并使用支持向量机（SVM）、人工神经网络（ANN）、K近邻（KNN）以及CNN-LSTM混合深度学习模型进行分类，以预测步态周期的五个阶段。这些分类器的输出之后是所提出的PI-BF后处理器，并将其与贝叶斯融合（BF）多数投票（MV）以及使用来自前一个窗口的不同投票数的无后处理（WPP）性能进行比较。结果表明，PI-BF可以将分类准确率提高多达5.5%，在SIAT-LLMD数据集（40名受试者）中，使用带有3个前一个决策窗口的SVM时，准确率可达85%。通过不稳定（INS）指数衡量，它还将转换检测差异（TDD）降低到0.1±59.8毫秒，并将输出稳定性提高了5%。所提出的PI-BF在实时步态相位识别实验中表现出持续的改进，实现了约90%的分类准确率。这些结果表明，PI-BF为提高辅助下肢设备中肌电控制的安全性、可靠性和实时性能提供了一种实用的、低复杂度的解决方案。
