Abstract Psychological, behavioral, and physical factors jointly contribute to heterogeneity in health-related outcomes, yet existing instruments often assess these domains separately or require lengthy questionnaires. In this study, we aimed to identify multidimensional psychosomatic profiles and develop a brief machine learning–based tool for their classification. Data from 3,207 employees (aged 18–65 years) were analyzed using responses from the NEO Five-Factor Inventory (NEO-FFI), coping questionnaires, the 12-item General Health Questionnaire (GHQ-12), body mass index (BMI), and physical activity measures. To enable integrated analysis, questionnaire responses were transformed into text-based representations and encoded using DistilBERT embeddings, followed by dimensionality reduction using principal component analysis. Spectral clustering identified latent psychosomatic profiles, resulting in three distinct subgroups characterized by different psychological, behavioral, and physical patterns. A supervised machine learning model was then trained to classify profile membership, and SHAP-based feature selection identified a minimal subset of informative items. This process yielded an eight-item instrument, the Psychosomatic Susceptibility Profile (PSSP), consisting of selected personality and coping items. An XGBoost classifier with five-fold cross-validation demonstrated good internal performance (F1 = 0.80; recall = 0.81). Overall, this study presents a data-driven framework for identifying multidimensional psychosomatic profiles and deriving a brief, interpretable classification tool for efficient profile assignment in research and applied settings.
Roohafza et al. (Tue,) studied this question.