Entraînement d’un réseau de neurones BiLSTM pour estimer la cinématique des trois axes de rotation du genou à partir de données issues de centrales inertielles : étude préliminaire

Abstract

Knee motion analysis is essential for understanding gait dynamics and optimizing patient diagnosis and rehabilitation. Currently, kinematic curves are obtained using precise but costly and restrictive optoelectronic systems. Inertial measurement units (IMUs) have emerged as a promising alternative, enabling real-time acquisition in uncontrolled environments. However, ensuring a stable anatomical reference frame remains a major challenge in achieving clinically accurate measurements. Machine learning, particularly neural networks, provides an effective solution for leveraging IMU data without a reference frame, thereby reducing estimation errors. Nonetheless, their reliability remains suboptimal, especially in anatomical planes where motion amplitudes are low. Moreover, while increasing the number of IMUs improves accuracy, it also complicates the experimental protocol. This study proposes an algorithm based on a bidirectional long short-term memory (BiLSTM) network designed to accurately estimate knee kinematic curves in all three anatomical planes using only two IMUs. The objective is to balance accuracy and practicality while aiming for generalization to a more heterogeneous population. A preliminary analysis demonstrated the consistency of the obtained biomechanical factors, regardless of the choice of gait cycles, thereby eliminating the need for additional preprocessing. The evaluation of the BiLSTM model revealed that separating the output layers by axis limits error propagation, thus improving measurement accuracy for tibial internal/external rotation and flexion/extension axes. Compared to existing approaches, the proposed method achieves high performance while reducing the number of required IMUs. However, an overfitting phenomenon was observed despite a data partitioning strategy that preserved the distribution of pathological levels. This observation raises concerns about the reproducibility of results reported in the literature and highlights the need to refine both data partitioning strategies and model complexity. The proposed approach proves particularly promising, with several optimizations still possible to enhance its robustness and generalization capability.

Date	5 May 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Nicola Hagemeister (Supervisor) & Neila Mezghani (Co-supervisor)