Pathology-Informed Augmentation Improves Cross-Cohort IMU-to-vGRF Estimation Between Healthy Adults and Adults With...

Estimating vertical ground reaction forces (vGRF) from wearable inertial measurement units (IMUs) degrades when models are applied across populations with different gait characteristics. We evaluated pathology-informed augmentation to improve generalization between healthy adults and adults with lower-limb osteoarthritis. We collected instrumented-treadmill walking data from 47 adults, including 29 healthy participants and 18 with radiographic osteoarthritis, predominantly knee osteoarthritis. Force-plate vGRFs were synchronized with body-worn IMUs. We developed a Transformer-based IMU model with low-rank multi-stream fusion (LMF), which learns compact shared patterns linking accelerometer and gyroscope signals across segments. We then applied a novel ratio-based augmentation that transforms source-cohort strides into target-like strides using donor-derived gait-feature ratios for stance duration, peak vGRF magnitude, dominant acceleration frequency, and segment-level acceleration amplitude. Generalization was tested in two directions, healthy-to-osteoarthritis (H2K) and osteoarthritis-to-healthy (K2H), using participant-disjoint splits. Augmentation reduced source-target separation in a low-dimensional gait feature space by 87% in H2K and 88% in K2H. On held-out participants, mass-normalized root-mean-square error (RMSE, N/kg) decreased by 10.5% in H2K and 11.4% in K2H. Expanded training without augmentation yielded smaller changes, with 4.9% in H2K and 0.2% in K2H. Peak accuracy was high: 96.7% of matched peaks in H2K and 97.0% in K2H were within 10% of the reference peak height. An interpretable, ratio-based augmentation that targets temporal, kinetic, spectral, and segment-amplitude features improves cross-cohort generalization of IMU-to-vGRF models and may reduce reliance on large target cohort datasets.