This research introduces a novel data mixture optimization strategy for Human Activity Recognition (HAR) systems, adapting techniques from Large Language Models to improve performance across diverse datasets.

• Develops HAR-DoReMi, a specialized framework that optimizes data mixture weights for pre-training HAR models
• Addresses challenges of working with continuous, multi-channel accelerometer data unlike text-based LLM inputs
• Demonstrates improved model generalization across heterogeneous IMU (Inertial Measurement Unit) datasets
• Establishes a robust foundation for cross-dataset HAR applications

Medical Impact: This advancement enables more reliable patient movement monitoring, fall detection, physical therapy tracking, and continuous health assessment across varied sensor environments and patient populations.

HAR-DoReMi: Optimizing Data Mixture for Self-Supervised Human Activity Recognition Across Heterogeneous IMU Datasets