AI-driven speech analysis for screening common psychiatric symptoms and investigating their impact on occupational performance in non-clinical hospital staff

Background and Aims

Psychiatric disorders severely impair occupational functioning, while the lack of self-awareness among employees regarding their own mental health status exacerbates both clinical and economic burdens.[1] Traditional assessments depend on subjective self-reports and are often complicated by the fear of stigmatization, necessitating more objective and reliable monitoring tools.[2] In this context, the human voice has proven its validity as an objective, low-cost, and non-invasive biomarker for monitoring mental health through its intrinsic acoustic and linguistic features.[3] This study aims to develop an automated screening model that detects common psychiatric symptoms among employees using short voice recordings and artificial intelligence(AI). Furthermore, the study examines the relationship between these objective data and employees’ work performance. By doing so, it facilitates the early detection of psychological risks that cause labor loss and leads to the development of preventive mental health policies that enhance employee well-being and institutional efficiency.

Methods

The study included 206 active hospital employees (security, administrative, technical, and cleaning staff). Clinical diagnoses were established through face-to-face expert psychiatric evaluations using the SCID-5 CV; additionally, a sociodemographic form, PHQ-9, GAD-7, and Chalder Fatigue Scale(CFS) were utilized. Occupational functioning was calculated by averaging supervisor scores covering performance, shift compliance, and workplace relationships. For voice analysis, participants provided approximately two-minute recordings consisting of text-reading and free-speaking tasks in quiet environments. Multimodal AI analyzed prosodic and linguistic features to examine correlations between extracted vocal biomarkers and mental health parameters. Statistical analyses were conducted using IBM SPSS Statistics version 26.0.

Results

The mean age of the participants was 38.13±9.32 years; 113(54.9%) were male and 93(45.1%) were female. In the SCID-based interviews, 86 individuals(41.7%) received a clinical diagnosis; however, approximately 60% of these individuals(n=51) were unaware of their current condition and did not report any diagnosis. Among the participants(n=206), comorbid diagnoses were identified in 12.1%(n=25) for two conditions and in 0.5%(n=1) for three conditions. When comparing scale scores between the group diagnosed via SCID-5 and the healthy control group, significant differences were found in the mean scores for the PHQ-9, GAD-7, and CFS. (respectively, p<0.001, p<0.001, p=0.014).

In multi-modal evaluations based on voice analysis data, a 90% accuracy rate was achieved in distinguishing between patients and healthy individuals. Positive, statistically significant correlations were identified between the AI score and the severity of PHQ-9, GAD-7, and CFS(respectively, r=0.505, p<0.001; r=0.456, p<0.001; r=0.348, p<0.001).

In the regression analysis, it was found that the scores calculated using our AI algorithm had a direct and significant effect on work performance (R=0.174, F=6.333, p=0.013). When the SCID-5 variable was added to the developed hierarchical regression model, no statistically significant increase in the model’s explanatory power was observed. (ΔR²=0.011, p=0.126).

Conclusions

This study confirms multimodal voice analysis as an objective method for predicting psychiatric diagnoses and monitoring occupational performance. The model is crucial for early diagnosis and monitoring, particularly for individuals with low self-awareness. Despite sample size and single-center limitations, longitudinal studies across different sectors are recommended. In summary, this approach provides a robust foundation for developing objective clinical assessment instruments to monitor mental health and occupational functioning.

References

1. Rehm, J., Shield, K.D., 2019. Global burden of disease and the impact of mental and addictive disorders. Current Psychiatry Reports 21 (2), 1-7
2. Low, D.M., Bentley, K.H., Ghosh, S.S., 2020. Automated assessment of psychiatric disorders using speech: a systematic review. Laryngoscope Investigative Otolaryngology 5 (1), 96-116
3. Fagherazzi, G., Fischer, A., Ismael, M., Despotovic, V., 2021. Voice for health: The use of vocal biomarkers from research to clinical practice. Digital Biomarkers 5 (1), 78-88