JMIR Publications

ABSTRACT

Background:

Work hours are typically investigated by an employee’s time spent at the worksite. However, difficulties to identify break times at the worksite and remote work outside of the worksite would mix up the work hour estimations. Machine learning have the potential to differentiate between human-smartphone interactions at work and off work.

Objective:

This study aimed to develop the probability in work mode based on human-smartphone interaction patterns with corresponding Global Positioning System (GPS) location data.

Methods:

The participants’ screen events, including the timestamps of notifications, screen-on/off, and types of apps used, as well as the GPS locations at work and off work were passively and continuously recorded by “Staff Hours”, an app developed by our team. We used extreme gradient-boosted trees to transform human-smartphone interaction patterns into a probability, and the one-dimension convolutional neural networks yielded the successive probabilities given the previous probability in a sequence. We then used the probability in work mode to identify the period of office-working, off-working, break at the worksite and remote work.

Results:

Among the 121 participants with 5503 person-days, the average prediction performance of the developed machine learning model was the area under the receiver operating characteristic curve of 0.915±0.064. The average work hours of 11.2±2.8 hour/day defined by the probability in work mode (higher than 0.5) were significantly longer than the GPS-defined counterparts of 10.2±2.3 hour/day (P<.001). This difference resulted from the higher remote work time of 111.6±106.4 minutes than the break time of 54.7±74.5 minutes.

Conclusions:

The probability in work mode, which is based on human-smartphone interaction patterns, and generated via machine learning models, could improve the precision and accuracy of work hour investigation.