Abstracts

Rationale: Wearable devices are increasingly used in seizure monitoring. Recording data with high quality is crucial for a reliable analysis. However, a systematic approach regarding signal and data quality assessment from non-EEG wearables is lacking. Here, we present an approach to assessing data quality from wearable recordings across four international epilepsy centers in the inpatient and outpatient settings.

Methods: We analyzed data recorded by the Empatica E4 wrist-worn device, including accelerometry, electrodermal activity (EDA), blood volume pulse (BVP), and skin temperature (TEMP) in seven cohorts at four epilepsy centers, and we analyzed signal quality in the raw data from all cohorts. We assessed data completeness, gauging the functionality of the device and compliance of study participants. Furthermore, we estimated the amount of time the wearable was worn on the body by fusing different modalities into one “on-body” score. Lastly, we assessed the EDA, BVP, and TEMP modalities for their signal quality using established related quality measures. We evaluated EDA and TEMP raw data by a combination of simple thresholding and rate of amplitude change assessment. Spectral entropy estimated noise and artifacts in BVP signals.

Results: We included 37166 hours of wearable data from 640 patients recorded in the inpatient setting and 90788 hours from 41 patients recorded in the outpatient setting (see Figure for outpatient result summary). Artifacts were caused by arm movements and other external sources, most prominently in outpatient recordings. We produced and analyzed an overview of common artifacts for future reference. While data loss was generally higher in the outpatient cohorts ( < 72%), and when using the data streaming mode ( < 49%) for data transmission instead of onboard recording ( < 9%), the on-body scores were consistently high across cohorts (> 80%) suggesting overall high compliance. Of note, the individual signal quality scores for the autonomous modalities in the ambulatory cohorts were comparable, and in some cases better, than those in the inpatient cohorts. The BVP modality had the overall lowest quality (46-60%), due to its high susceptibility to motion artifacts, while the TEMP signals had the overall best quality (79-98%), although no statement can be made on its accuracy concerning the true skin temperature. Figure: Illustration of outpatient data quality scores for data completeness, time on-body, and EDA, BVP, and TEMP signal quality.

Conclusions: Signal recording quality affects all autonomic modalities, especially blood volume pulse recordings captured by photoplethysmography. Artifact recognition and data quality ratings may provide additional value and improve precision, may serve as a standard metric in experimental studies, and may foster further design improvements for future wearable device studies in epilepsy research.

Funding This study was supported by The Epilepsy Foundation of America’s My Seizure Gauge, Epilepsy Research Fund, Deutsche Forschungsgemeinschaft, and Innovative Medicines Initiative 2’s RADAR-CNS.