Biosignal Compression Toolbox

Compression pipelines for wearable biosignals, achieving up to 130x reduction with under 2% information loss.

Longitudinal wearable monitoring produces enormous volumes of data, and storage cost becomes a real constraint on the scale of digital biomarker research. But biosignals cannot be compressed indiscriminately — the features that matter clinically must survive the round trip.

This project built and evaluated compression pipelines combining algorithmic and encoding-based methods, and characterized the trade-off between storage footprint and recoverability for each signal type.

Methods evaluated

  • Algorithmic: singular value decomposition, discrete cosine transform, biorthogonal discrete wavelet transform, autoencoders.
  • Encoding: run-length encoding, Huffman encoding.
  • Signals: ECG, PPG, accelerometry, electrodermal activity, and skin temperature.

Different signals favored different pipelines — direct compression with Huffman encoding for ECG and PPG, SVD with Huffman encoding for EDA and accelerometry, and the biorthogonal wavelet transform for skin temperature — reaching up to 130× compression with under 2% information loss.

The resulting toolbox is released as open source through the Digital Biomarker Discovery Pipeline (Bent et al., 2021).

References

2021

  1. Biosignal Compression Toolbox for Digital Biomarker Discovery
    Brinnae Bent, Baiying Lu, Juseong Kim, and 1 more author
    Sensors, Sep 2021