Development of an artificial intelligence model to estimate psychiatrist-assessed mental health-related presenteeism

OBJECTIVES: This study aimed to contribute to the development of an AI-based system that supports worker health and productivity by enabling early detection of presenteeism. We tested whether an AI model could assess mental health-related presenteeism with accuracy comparable to that of psychiatrists and whether the frequency of application use was comparable between avatar-based and real-person interfaces. METHODS: This study aimed to design a multivariable prediction model among white-collar employees in a Japanese company. The participants comprised 117 white-collar workers who provided a total of 1,631 video responses to a standardized health-status question over a period of 10 working days. The primary outcome measure was the accuracy of the AI model in estimating workers' mental health-related presenteeism. The secondary outcome was the frequency of application use when inquiring about workers' health conditions, comparing the avatar-based interface with the real-person interface. RESULTS: The AI model achieved an overall accuracy of 72.2%, a macro F1 score of 0.64, and a weighted F1 score of 0.73 compared with psychiatrists' ratings. Agreement between the two psychiatrists was 84.5%. Participants were allocated to either an avatar-based interface or a real-person interface, with no significant differences observed between groups in baseline characteristics or frequency of application use, whereas a significant difference was observed in the frequency of missing values. CONCLUSIONS: The developed AI model demonstrated performance comparable to psychiatrists in estimating mental health-related presenteeism from video data. This approach offers a novel, objective alternative to traditional questionnaire-based methods.