Solutions to Increase Reliability of Clinical Trial Data from Wearable Devices
​[Thursday, February 06, 2020]
Data collected from wearable devices contain several sources of “noise” that reduce the reliability of such data in support of clinical/medical claims. A report this week presents a summary of the various factors that contribute to erroneous results and reasonable tips to address them. While the use of wearable devices in clinical trials has almost doubled in the last 5 years, based on review of published reports, the associate issues with the data quality have also become more evident. The larger the data set, the greater the noise due to lack of standardization of data collection and processing. The authors of the report reviewed data related on only one health indicator, the heart rate, collected via wearable devices worn on the wrist to evaluate the contributors to erroneous conclusions, but these finding can be extrapolated to most other clinical measures made using wearable devices. The contributors to errors can be grouped into two categories: (1) technological factors such as device type, firmware version, and sampling rate, and (2) biological factors such as BMI, wrist circumference, and demographic characteristics. There is significant variability in the core technology used by various devises to measure health-related data, and these get updated every time there is software/firmware update. So, not surprisingly multiple devices in the same study will likely lead to variability in data. The technological factors can be controlled by limiting the kind of device, software, firmware version and data sampling rate. This is feasible for studies involving fewer participants. For larger studies and studies based on BYOD format, one must address the variability due to device factors in the statistical models used to analyze the results. For controlling the biological factors, it is critical to consider the data collection sampling rates as they vary the most between devices. Secondly,  data should be cleaned to address outliers and missing data points while using measures for bias reduction. The studies using wrist wearable devices should control non-wear time for device charging and non-adherence, placement of the device on the wrist, tightness of wear, dominant vs non-dominant hand use, degree to wrist movement. Other factors such as BMI, sex, skin tone (for wrist wearable devices), and physical activity of the participants have not been found to be contributors to noise but must be collected and reported. These principles can be applied to studies with all kinds of wearable devices. The report includes two checklists that should be useful for all studies planning to use using wearable devices for clinical endpoints.