Cross-Center Machine learning classification of gait in patients with ankle or subtalar Arthrodesis: A comparison of optical and IMU-Based systems.

Optical motion capture (OMC) is the clinical gold standard for instrumented gait analysis. Inertial measurement units (IMU) offer a faster and more scalable alternative. A key prerequisite for its clinical application is a comparable potential for pathology detection. This prospective, two-center study compared the classification performance of IMU based 2-Segment Foot Model kinematics data with OMC based Oxford Foot Model data. Included were healthy controls (IMU: n = 30; OMC: n = 20), patients after tibiotalar (IMU: n = 15; OMC: n = 19), and subtalar arthrodesis (IMU: n = 15; OMC: n = 21), assessed via both IMU and OMC. Kinematic waveforms were converted into scalar features using functional principal component analysis (FPCA), which preserves the temporal structure of the full gait cycle while yielding scalar scores suitable for machine learning classification Three different machine learning classification approaches were used. All approaches achieved high discriminative performance for both modalities, with consistently higher performance for OMC. Mean accuracies for the different modalities and ML models ranged between 0.67 and 0.81 for the IMU data and were consistently greater for OMC data (0.83 - 0.91). Mean AUC values remained high (IMU 0.920-0.926; OMC 0.944-0.965). Feature patterns indicated that group separation was primarily driven by FPCA-derived features of hindfoot/tibia kinematics, while differentiation between tibiotalar and subtalar arthrodesis required FPCA-derived features spanning hindfoot/tibia and forefoot/tibia across multiple planes. IMU derived 2-Segment Foot Model kinematics enable clinically meaningful classification of arthrodesis related gait patterns and separation from healthy controls, supporting clinical applicability. However, OMC using a multi segment foot model remained superior with respect to maximal classification accuracy.