Beyond the Laboratory: Deep Learning-Based Reconstruction of Force-Plate Measurements Using Wearable Pressure-Sensitive Insoles

Reconstructing force-plate measurements from wearable pressure insoles could support portable biomechanical assessment, but generalization to unseen wearers and fast loading events requires rigorous validation. We investigated pressure-only regression of six right-foot force and moment components and derived center-of-pressure (COP) coordinates during in-place walking and quiet standing. Twenty-one university volunteers performed both tasks using an 89-sensor insole (50 Hz) and AMTI force plates (1000 Hz). Force-plate signals were filtered, downsampled to 50 Hz, and aligned by cross-correlation. Each task used subject-wise seven-fold cross-validation. For every fold, sensor Z-score normalization was fitted using only its training participants and then applied unchanged to its held-out test participants. Mean ± SD of participant-level R2 during walking was 0.88 ± 0.071 (Fx), 0.80 ± 0.062 (Fy), 0.89 ± 0.075 (Fz), 0.95 ± 0.090 (Mx), 0.89 ± 0.078 (My), and 0.43 ± 0.390 (Mz); COPx and COPy achieved 0.99 ± 0.039 and 0.76 ± 0.055. During standing, Fz, Mx, and My each reached mean R2 = 0.99, while COPx and COPy reached 0.77 and 0.98, respectively. The contribution of this study is a pressure-only six-component reconstruction assessed across unseen individuals and complementary controlled loading tasks. To our knowledge, this is the first study to combine a high-density, pressure-only insole (no inertial sensors), estimation of all six force-plate components plus COP, and evaluation on participants unseen during training in both a dynamic and a static task.