Standardized synchronization and validation pipeline for physiological biomarkers across multiple devices

AI Summary

User-friendly, generalisable pipeline synchronises any two physiological devices without coding, using resampling, dynamic time warping, wavelet amplitude correction, and signal standardisation.
Validation on 31 participants' up to 48 h recordings from Empatica E4 and EmbracePlus showed near-perfect BVP and strong phasic EDA agreement.
Tonic EDA, temperature, and accelerometry had systematic amplitude biases and axis-dependent variability; relative dynamics preserved but movement signals require calibration.

AI Summary

Wearable devices enable continuous monitoring of physiological signals in real-world settings, yet a standardized approach for synchronizing signals across devices remains lacking.

We compared signal-level agreement across waveform similarity, amplitude distribution, spectral content, and extracted features between the two research-grade devices to determine their interchangeability for longitudinal and multi-site studies.

Results showed near-perfect agreement for BVP (concordance correlation coefficient (CCC) ≈ 1.0; coherence ≤ 0.98) and phasic EDA features (CCC 0.85-0.99), whereas tonic EDA, temperature, and accelerometry exhibited systematic amplitude biases (EmbracePlus lower) and axis-dependent variability (Z-axis CCC = 0.85; Y-axis = 0.19).

Basic summary

Behav Res Methods. 2026 Jul 16;58(8):239. doi: 10.3758/s13428-026-03097-8.

ABSTRACT

Wearable devices enable continuous monitoring of physiological signals in real-world settings, yet a standardized approach for synchronizing signals across devices remains lacking. We present a generalizable, user-friendly pipeline that enables synchronization of any two devices capturing the same physiological signals, without requiring coding expertise. The pipeline performs resampling, dynamic time warping alignment, amplitude correction via wavelet transforms, and signal standardization, followed by agreement analyses at both waveform and feature levels. To demonstrate its validity, we applied the pipeline to a case study comparing two research-grade devices, the Empatica E4 and the EmbracePlus, using up to 48 h of concurrent recordings from 31 participants. We compared signal-level agreement across waveform similarity, amplitude distribution, spectral content, and extracted features between the two research-grade devices to determine their interchangeability for longitudinal and multi-site studies. Specifically, we aimed to determine how well these devices agree at the signal level and to identify which physiological signals are most robust to device-specific variability. Four signals were examined (blood volume pulse, electrodermal activity, accelerometry, and temperature) using NeuroKit2 and FLIRT. We computed Pearson and concordance correlation coefficients, Bland-Altman bias and limits of agreement, root mean squared error (RMSE), KL divergence, spectral coherence, mutual information, and feature-level correlations using NeuroKit2 and FLIRT. Results showed near-perfect agreement for BVP (concordance correlation coefficient (CCC) ≈ 1.0; coherence ≤ 0.98) and phasic EDA features (CCC 0.85-0.99), whereas tonic EDA, temperature, and accelerometry exhibited systematic amplitude biases (EmbracePlus lower) and axis-dependent variability (Z-axis CCC = 0.85; Y-axis = 0.19). Relative signal dynamics were preserved across devices despite differences in absolute levels. These findings support integration of BVP, EDA, and TEMP data across E4 and EmbracePlus with proper preprocessing, while highlighting calibration needs for movement signals.

PMID:42463554 | DOI:10.3758/s13428-026-03097-8

Document this CPD