AI facial expression valence: reproducibility in healthy Japanese adults for digital environments

Introduction: Artificial intelligence (AI)-based facial expression analysis can estimate emotional valence rapidly and noninvasively without the need for human training. However, reproducibility data remain limited, particularly in East Asian populations, where facial expressions may be more subtle. Establishing reproducibility in this context is essential for evaluating whether AI-derived valence measures may be useful in digital environments and mental health-related monitoring. Therefore, this study had two aims: (1) to evaluate state-specific reproducibility for positive, negative, and neutral facial expressions, and (2) to examine short-term trait-like stability by integrating valence estimates across all emotional conditions. Materials and methods: Healthy young adults in Japan (N = 40) completed two facial expression recording sessions separated by 30 min. Positive, neutral, and negative facial expression conditions were recorded using an AI-based system. Test–retest reproducibility was assessed using intraclass correlation coefficients (ICCs) and Bland–Altman analyses. To examine short-term trait-like stability across conditions, all condition-specific valence values were additionally integrated into a single dataset for each participant. Results: The ICC values were 0.82, 0.61, and 0.05 for the positive, negative, and neutral conditions, respectively. Bland–Altman analyses indicated no notable systematic bias across sessions. When data from all conditions were combined, the integrated dataset yielded a high ICC of 0.94, indicating short-term trait-like stability. Conclusions: AI-based emotional valence values showed good to moderate reproducibility for positive and negative expressions, whereas the neutral condition showed low ICC values, likely because of restricted score variability. Integrating valence data across conditions yielded a high ICC, indicating short-term trait-like stability at the individual level. These findings support the potential of AI-based valence as a low-burden measure for future research on emotional assessment in digital environments. However, further validation under ecologically valid real-world conditions is needed.