Deep Neural Networks and Transfer Learning on a Multivariate Physiological Signal Dataset

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Physiological Signal Processing

Background:

• Deep Neural Networks (DNNs) and Transfer Learning (TL) show promise in medical applications but are underutilized for multivariate physiological datasets.
• Current methods often focus on single signal types, limiting the transferability of trained models to diverse physiological data.

Purpose of the Study:

• To develop and release a comprehensive dataset of multiple physiological signals for DNN and TL research.
• To demonstrate the efficacy of DNNs and TL in classifying diverse physiological signals and acquisition devices.
• To facilitate broader adoption of DNN and TL for multivariate physiological data analysis.

Main Methods:

• Compiled a dataset of 813 samples comprising six physiological signal types (ECG, EDA, EMG, PPG, Respiration, Acceleration) from 232 subjects.
• Utilized a DNN to classify physiological signal types.
• Employed TL by using DNN-extracted features to train a Support Vector Machine for classifying data acquisition devices.

Main Results:

• The DNN successfully classified the types of physiological signals within the dataset.
• Transfer learning enabled the classification of data acquisition devices using features extracted by the DNN trained on signal types.
• The dataset, code, and DNN parameters are publicly available.

Conclusions:

• This work establishes a foundation for applying DNNs and TL to complex, multivariate physiological datasets.
• The findings highlight the potential of TL to leverage DNNs across different physiological signal analysis tasks.
• Publicly releasing the resources encourages further research and development in this domain.