Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces MOSAIC, a continual learning framework designed to address the challenges of modality-incremental settings in Parkinson's disease gait assessment. By tackling issues such as the Toxic Teacher phenomenon and modality-specific statistical shifts, MOSAIC employs Modality-Specific Warm-Up and a statistics-decoupled architecture to enhance learning stability and performance. Experimental results demonstrate that MOSAIC significantly improves final performance and reduces forgetting across three multimodal gait datasets.
MOSAIC not only stabilizes learning in the face of new sensor modalities but also recovers lost capacity without sacrificing legacy knowledge.
Gait-based Parkinson's disease assessment increasingly relies on heterogeneous sensors, but clinical systems rarely collect all modalities simultaneously. New sensors may arrive through device upgrades, protocol changes, or multi-center deployment, while historical patient data are often unavailable because of privacy and storage constraints. This modality-incremental setting faces three challenges: unreliable cross-modal distillation, modality-specific statistical shifts, and reduced plasticity after preservation. We propose MOSAIC, a compact continual learning framework. First, we identify the Toxic Teacher phenomenon and introduce Modality-Specific Warm-Up to stabilize newly learned modality representations before distillation. Second, we propose a statistics-decoupled MSBN architecture that isolates sensor statistics while maintaining a shared semantic backbone. Third, we design a curriculum-guided repulsive objective for Plasticity Recovery, preserving legacy knowledge while recovering modality-specific capacity. Experiments on three multimodal Parkinson's gait datasets show that MOSAIC improves final performance and mitigates forgetting. Project code is available at: https://github.com/minlinzeng/MOSAIC_Modality-Specific-Adaptation-for-Incremental-Continual-Learning-in-PD-Gait-Assessment.git