MAPPING AFFECTIVE PROFILES IN DEPRESSION, BURNOUT, NORMAL SADNESS, AND EUTHYMIC STATE: A SELF-REPORT SCREENING TOOL DEVELOPED THROUGH A MACHINE LEARNING APPROACH.

Background: Modern post-industrial society is facing a complex of challenges, such as including epidemiological threats, high demands from employers, aggressive forms of corporations' management, stress at the work place, as well as geopolitical and economic instability worldwide. These factors bring a significant impact on mental health of the general population, contributing to an increased prevalence of mental disorders, particularly, affective states. The aim of this study was to develop a sensitive screening tool based on a self-questionnaire approach for accurate differentiation of affective spectrum state, from preclinical / at-risk to severe clinical conditions. To achieve this goal, we focused on identifying key affective symptoms' domains and application of machine learning (ML) methods to perform a comprehensive data analysis on classifying the respondents into preclinical and clinical subgroups.

Subjects and methods: The study consisted of two stages. At the first stage, we developed and conducted an online survey among the experimental population consisting of university staff and students. This survey version included 19 questions. The study was interrupted to make adjustments. At the second stage, the survey was finalized based on data analysis (descriptive and inferential) and classification tasks. The revised survey was redistributed with additional criteria for inclusion and exclusion of the respondents applied to the study design. The final version contained 34 questions, excluding unreliable questions characterized by p > .05. 381 individuals (269 employees and 112 students) were interviewed, of whom 99 showed signs of depression, normal sadness or emotional burnout. We conducted correlation, descriptive, and inferential analyses and classification of respondents using ML-based methods.

Results: The results confirmed the presence of significant differences (p < .001) between the groups with euthymia, normal sadness, emotional burnout and depression. However, there were no statistically significant differences for respondents with a pre-known emotional state and for respondents whose condition has been classified using machine learning technologies. The final distribution by category was as follows: euthymia - 38.8%, normal sadness - 27.3%, emotional burnout - 25.2%, depression - 8.7%. Our developed self-report tool has demonstrated statistical benefit, but requires further clinical research to clarify sensitive symptoms' domains for updating its items content.

Conclusions: ML-based analysis of the self-report screening tool-related data demonstrated its sensitivity to classify affective states spectrum onto the separate states of depression, emotional burnout, normal sadness and euthymia (i.e. affective or emotional profiles of the respondents) with 100% accuracy at the final iteration. The problem of assessing mental health lies in the difficulty of obtaining fast, accurate, and emotionally neutral determination of the affective state in individual respondents and across populations. Development of a sensitive self-questionnaire / screening benefits from the the integration of clinical assessments along with the modern ML-based algorithms, as well as targeting the approach that helps to reduce costs and increase the diagnostic accuracy of existing psychometric tools.