When Mixture-of-Experts Fails: An Empirical Study of Expert Routing Limitations in Long-Tailed Gait Regression

This study presents a methodological evaluation of the conditions required for successful Mixture-of-Experts (MoE) deployment in continuous biomechanical regression, rather than proposing a new predictive architecture. Using plantar-pressure-based stride length estimation as a case study, we examined whether MoE routing provides stable performance gains over a single convolutional neural network (CNN) under a long-tailed target distribution. Plantar pressure data from nine force-sensitive resistor sensors were collected from 22 male participants, including five older adults and 17 young adults, and represented as 36-channel time-series inputs comprising raw pressure, first-order differences, second-order differences, and cumulative integrals. Under leave-one-subject-out evaluation with five repeated runs, a CNN baseline was compared with four MoE variants using statistical gating, BiLSTM encoder gating, fully connected autoencoder gating, and zone-specific expert architectures. The CNN achieved the most stable overall performance (RMSE: 7.54 ± 0.14 cm). MoE-AEG significantly improved short-stride prediction but significantly deteriorated long-stride prediction. In MoE-FCAE, expert-only analyses confirmed zone-specific specialization, whereas gating weights remained predominantly allocated to the middle-stride expert across all zones. These findings show that expert specialization alone is insufficient for improved MoE regression; input separability, expert specialization, and gating alignment should be verified separately, together with subject-wise robustness under subject-independent evaluation.