Abstract
INTRODUCTION: Insomnia in shift workers has distinctive features due to circadian rhythm disruption caused by reversed or unstable sleep-wake cycle work schedules. While previous studies have primarily focused on a limited number of predictors for insomnia severity in shift workers, there is a need to further explore key predictors, and develop a data-driven prediction model for insomnia in shift workers. This study aims to identify potential predictors of insomnia severity in shift workers using a machine learning (ML) approach and evaluate the accuracy of the resulting prediction model. METHODS: We assessed the predictors of insomnia severity in large samples of individuals (4,572 shift workers and 2,093 non-shift workers). The general linear model with the least absolute shrinkage and selection operator (LASSO) was used to determine an ML-based prediction model. Additional analyses were conducted to assess the interaction effects depending on the shift work schedule. RESULTS: The ML algorithms identified 41 key predictors from 281 variables: 1 demographic, 7 physical health, 13 job characteristics, and 20 mental health factors. Compared to the non-shift workers, the shift workers showed a stronger association between insomnia severity and five predicting variables: passiveness at work, authoritarian work atmosphere, easiness to wake up, family and interpersonal stress, and medication. The prediction model demonstrated good performance with high accuracy and specificity overall despite a limited F1 score (classification effectiveness) and recall (sensitivity). Specifically, a prediction model for shift workers showed better balance in F1 scores and recall compared to that for non-shift workers. DISCUSSION: This ML algorithm provides an effective method for identifying key factors that predict insomnia severity in shift workers. Our findings align with the traditional insomnia model while also reflecting the distinctive features of shift work such as workplace conditions. Although the potential for immediate clinical application is limited, this study can serve as guidance for future research in improving a prediction model for shift workers. Constructing comprehensive ML-based prediction models that include our key predictors could be a crucial approach for clinical purposes.