Foundation Model for Biosignal

FMBIO-Model

FMBIO-Model to-do: explain FMBIO-Model

Abstract

We will attempt to develop the foundation model using bio-signal (ECG, Heart Rate) [1] for applying the sleep stage classification from the personal data of edge devices [2], such as Apple Watch or Fit-bit. It is hard to get high performance by only using personally own data from edge devices and to train the model, as limitation of the amount of data for train and low hardware resources of edge devices. We expect that the foundation model generates informative representative feature from large bio-signal dataset, and it can improve the downstream task in the restricted environment that people cannot share bio-signal data to others.

Overview

Our framework has two main strategies:

• MultiModalLoss: develop the foundation model using bio-signal (ECG, Heart Rate)
• Subject-Invariant:

Environment setup

create and activate conda environment named FMBIO with python=3.8.18

conda create -n FMBIO python=3.8 -y
conda activate FMBIO
pip install -r requirements.txt

Dataset

We used two dataset MESA dataset and Apple Watch dataset. The description of dataset is below. Due to access issue, the MESA dataset MESA Dataset cannot share to others. In addition, we only share 1 subject Apple Watch dataset for demo.

Preprocessing

The pre-processing scripts are included in this repo. For preprocessing the MESA dataset and Apple Watch dataset, we cited GitHub

Foundation Model pre-training on MESA dataset

python ./src/foundation/train.py

Baseline model Training on Apple Watch dataset

Train using Jupyter Notebook basemodel/subj_baseline.ipynb

Classifier with Foundation Model to-do

Results

(1) Evaluation Metrics

Client	Model	Accuracy	F1-Score
1	Base	0.804	0.705
	FM	0.816	0.689
2	Base	0.621	0.317
	FM	0.641	0.380
3	Base	0.784	0.400
	FM	0.791	0.404
4	Base	0.736	0.595
	FM	0.648	0.506
5	Base	0.812	0.692
	FM	0.716	0.518

(2) Confusion Matrix

Reference

[1] Large-scale Training of Foundation Models for Wearable Biosignals

[2] Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device