Body-worn sensors have become an indispensable tool for health research in the last decades. Not only they are used in major large-scale cohort studies in public health, they are also increasingly being used in clinical trials for developing novel therapies. The accelerometer used in actigraphy devices is the most used sensor in the health research, with a history dating back to the 1950s. When the advent of technology enabled mass production of reliable, affordable, and user-friendly devices in the 90s, several manufacturers made actigraphy devices commercially available and firmly entered the scientific research community. Ambulatory Monitoring, Inc developed one of the first commercially available activity monitor devices under the name MotionloggerTM.1 Mini-Mitter Co., Inc. developed an improved version of the Motionlogger™ product in 1999 (Actiwatch®, later acquired by Respironics Inc., then Philips) which received FDA approval and gained popularity among the research community.) In 1993, Computer Science and Applications (or CSA) developed the AM7164 (a smaller, more user-friendly solution) which was then used by Freedson, et.al. in the landmark study on the use of accelerometer in modelling energy expenditure.2 Shortly thereafter in 2005, ActiGraph, LLC (formerly CSA) released its the GT1M activity monitor which leveraged a micro-electro-mechanical systems (MEMS) based accelerometer which introduced significantly lower power (longer battery life) and broaden the useability of these products for longitudinal research.
In those early days, manufacturers had to minimize the data storage and transmission due to the constraints of the onboard memory and battery life. The volume of the raw or original data collected by the accelerometer was very large, and data reduction was necessary to optimize the storage and battery life of those early devices. This data reduction was achieved by derived a summary score, so called “activity count”, for each epoch of acceleration data collected.3 The raw accelerometer data was typically collected at greater than 10Hz whereas the resultant epoch data was recorded for every 5 to 60s. The raw acceleration data was then discarded so the requirement of data storage and transmission was vastly reduced. While these technical constraints have mostly been eliminated since 2010, the use of activity counts continued. At this point, activity counts have become the common input source data to many validated algorithms for computing physical activity and sleep measures, and contributed to thousands of research articles across different age groups, clinical conditions as well as animals.2,4 Furthermore, to many non-technical researchers, activity counts are much easier to understand and analyse, than the raw acceleration data.
Activity counts, however, can represent very different metrics depending on the count algorithms and thus cannot be used interchangeably.3 For example, the values of activity counts provided by ActiGraph devices are significantly higher than the ones provided by Philips devices.5 Because the count algorithms had been held proprietary by many device manufacturers, it became very challenges to compare research findings using different actigraphy devices. This limitation has been recognized by many research groups and several open-source counts algorithms were developed to overcome this limitation. ActiGraph has also published its own count algorithm and made it open-source recently.6 But these open-source algorithms require the collection and retention of raw data, which is not provided by some device manufacturer, such as Philips, and / or might not have been retained by the researcher due to data storage requirements.
This issue has recently become acute as Philips discontinued its actigraphy product line in 2022. As it was a major actigraphy device provider, thousands of research groups are now facing the urgent needs to continue their work with alternative solutions. As the count algorithm from Philips is proprietary, it could be very difficult to compare future research findings using different devices with the historical data collected by Philips devices. We hereby provide a conversion method of activity counts between Philips and ActiGraph devices based on two datasets where data was collected simultaneously from both devices. One dataset was collected in laboratory condition and used for deriving the conversion equation.5 The second and larger dataset was collected in free-living condition and used for testing the conversation equation.7