MOSAIC: Modality-Specific Adaptation for Incremental Continual Learning in Parkinson's Disease Gait Assessment

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Internet of Things (IoT) & Connected Devices, AI in Healthcare · Depth: Expert, extended

Summary

MOSAIC, a novel continual learning framework, addresses critical challenges in modality-incremental settings for Parkinson's disease (PD) gait assessment using heterogeneous sensors. It tackles unreliable cross-modal distillation, modality-specific statistical shifts, and reduced plasticity after preservation. The framework introduces Modality-Specific Warm-Up to stabilize newly learned modality representations before distillation, a statistics-decoupled MSBN architecture to isolate sensor statistics while maintaining a shared semantic backbone, and a curriculum-guided repulsive objective for Plasticity Recovery. Experiments on three multimodal PD gait datasets (WearGait, FBG, FOG) demonstrate that MOSAIC consistently improves final performance and mitigates forgetting, achieving near-ideal Normalized Average Accuracy (NAA ≥ 96.79%) and significantly reducing Backward Transfer (BWT) compared to existing baselines.

Key takeaway

For Machine Learning Engineers developing clinical AIoT systems for Parkinson's disease gait assessment, integrating new sensor modalities sequentially without historical data presents significant challenges. Your current approaches likely suffer from "Toxic Teacher" phenomena and statistical overwriting, degrading model performance. You should adopt MOSAIC's framework, which stabilizes new modality representations, decouples sensor statistics, and recovers plasticity, ensuring robust model evolution and preventing catastrophic forgetting in your deployments.

Key insights

MOSAIC resolves modality-incremental learning challenges in PD gait assessment by stabilizing new sensor representations and decoupling statistics.

Principles

• Stabilize new modality representations before distillation.
• Isolate modality-specific statistics with a shared backbone.
• Recover plasticity using a repulsive objective.

Method

MOSAIC employs a three-stage pipeline: Modality-Specific Warm-Up, Statistics-Decoupled MSBN Architecture, and Plasticity Recovery via a curriculum-guided repulsive objective, all integrated into a unified model.

In practice

• Pre-train new encoders with cross-entropy loss.
• Use Modality-Specific Batch Normalization (MSBN).
• Apply repulsive loss to expand latent space.

Topics

• Parkinson's Disease Assessment
• Continual Learning
• Multimodal Learning
• Gait Analysis
• Modality-Specific Batch Normalization
• Knowledge Distillation

Code references

• minlinzeng/MOSAIC_Modality-Specific-Adaptation-for-Incremental-Continual-Learning-in-PD-Gait-Assessment

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• MOSAIC: Modality-Specific Adaptation for Incremental Continual Learning in Parkinson's Disease Gait Assessment — MOSAIC enables robust continual learning for multimodal Parkinson's gait assessment by addressing modality-specific challenges.
• Treatment-Conditioned Diffusion for Forecasting Neurodegenerative Disease Progression — A treatment-conditioned diffusion framework accurately forecasts neurodegenerative disease progression by integrating imaging and pharmacological data, preserving anatomical detail.
• Continual Learning for fMRI-Based Brain Disorder Diagnosis via Functional Connectivity Matrices Generative Replay — A continual learning framework uses generative replay and knowledge distillation to diagnose brain disorders from sequential fMRI data.
• Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation — Adapting tabular foundation models with a survival-aware head significantly improves clinical time-to-event prediction.
• Learning Emergent Modular Representations in Multi-modality Medical Vision Foundation Models — DEX uses modular networks with experts and a director to manage modality-specific and shared representations in multi-modality medical FMs.
• UniMotion: A Unified Framework for Motion-Text-Vision Understanding and Generation — UniMotion unifies continuous motion, text, and vision understanding/generation in a single architecture.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.