Stress prediction for employees based on an individually optimized model using attendance records and wearable devices

Abstract

Since the number of cases of depression has been increasing, care in presymptomatic stages based on stress prediction is significant. Although several indices such as sleep and heart rate are known to be relevant to stress, they are affected by many factors like work styles, and it is difficult to apply an identical stress prediction model to all cases. We propose an individually optimized stress prediction algorithm that utilizes data from people who have similar work styles to the prediction object. The proposed algorithm uses data from wearable devices, activity questionnaires, and employee attendance records as inputs and predicts the stress condition of the next week. The proposed algorithm was evaluated on the data collected from 192 employees in SHIONOGI group. It achieved 68% of precision, 59% of recall, and 0.85 of ROC-AUC, which were superior to those of the model trained on the whole of the training dataset.