Data quality evaluation in wearable monitoring

Figures

• 

  Figure 1. Data quality assessment pipeline. The amount of data referenced in each step decreases from left to right, that is, the data completeness score is relative to the whole recording period, whereas the individual signal quality scores are relative to the amount of data that was estimated as on-body.

• 

  Figure 2. Visualization of typical artifacts in E4 data, with on-body and signal quality scores marked. Plots from top to bottom: BVP detail, ACC (x, y, z), BVP, EDA, TEMP. Off-body periods are highlighted in gray. Periods of low EDA, BVP, and TEMP signal quality scores are highlighted in red (the lighter the color the better the quality, no color = 100% quality). (a) Reference recording during rest, blood pulse wave without artifacts, no motion, baseline EDA and TEMP. (b) Wristband not on the person, BVP shows a noise-like pattern, no motion is recorded, and EDA and TEMP drop to 0 or room temperature, respectively. Note, that TEMP changes might show a delay. (c) Device body is not on the wrist, BVP cannot be recorded well as the sensor loses skin contact. (d) Wristband is not on the wrist, the EDA drops to 0 as the integrated sensor electrodes lose contact. (e) External light shining on the device and into the sensor can greatly affect the photosensor’s (BVP) recordings. (f) Random movements, (g) wrist rotation, and (h) wrist translation (e.g., moving up and down) disrupt BVP, ACC, and EDA signals, causing rhythmic changes and peaks to show across modalities.

• 

  Figure 3. Example of ~ 24 h of long-term ambulatory data from the Empatica E4 device (healthy test subject). Plots from top to bottom: ACC (x, y, z), BVP, EDA, TEMP. Off-body periods are highlighted in gray. Periods of low EDA and BVP signal quality scores are highlighted in red (the lighter the color the better the quality, no color = 100% quality). The recording is divided into three parts by the two off-body periods. Left: first day afternoon and evening; Middle: night; Right: morning and rest of the second day.

• 

  Figure 4. Swarm plots of data quality score distributions for the inpatient cohorts of (a) BCH (N = 832), (b) KCL (N = 29), (c) MCR (N = 19), and (d) UKF (N = 169).

• 

  Figure 5. Swarm plot of the data quality score distributions for the outpatient cohorts (N = 39).

• 

  Figure 6. Comparison of EDA and BVP signal quality scores when grouping data by day- and nighttime.