Research Objective:
The goal is to develop a publicly available foundation model for accelerometry that can generalize across various downstream health-related tasks for the research community to use.
Importance:
Researchers will be able to leverage the pre-trained model, fine-tuning it for specialized tasks such as identifying early signs of neurodegenerative diseases, predicting recovery progress in rehabilitation patients, or detecting abnormal movement patterns. The availability of this foundation model will thus accelerate innovation in health and motion analytics, while also ensuring consistency and reliability in data analysis across studies.
Data Required:
Need to Have:
Raw accelerometry x,y,z data for self-supervised training (ID: 90001)
Methods Used:
Self-supervised learning (SSL) with Relative Contrastive Learning will be the core methodology, which learns semantic similarities between motion segments.
Data augmentation will be applied to maintain rotation invariance and preserve accelerometer-specific signal integrity.
Evaluation of model performance will be done on tasks such as activity recognition, gait metric regression, and energy expenditure estimation
Expected outcomes:
A general-purpose, robust motion foundation model that outperforms existing approaches in tasks such as human activity recognition, gait analysis, and other health metrics without fine tuning, exceeding the performance of fully-supervised models.
Scientific Rationale
This study is set to deliver a transformative tool for the research community, one that will not only speed up the process of scientific inquiry but also deepen our understanding of human motion and its implications for health. By providing a robust, generalizable model, we aim to foster innovation and reproducibility in health research, impacting various domains from preventative medicine to personalized health strategies.