Identifying Youth Somatic Symptom Risk Through Interpretable Machine Learning and Behavioral Phenotyping: A Temporal Validation Study

Abstract BACKGROUND & AIMS: Psychological characteristics are closely associated with somatic symptoms. However, commonly used psychological assessments often include items related to somatic symptoms. As a result, using such assessments for somatic symptom risk prediction may introduce information leakage and inflate predictive performance. In addition, existing studies have often examined individual lifestyle factors independently, overlooking the potential contribution of combined lifestyle patterns to somatic symptom risk. This study aims to develop an interpretable machine learning approach for youth somatic symptom risk prediction and evaluate its generalizability through temporal validation. METHODS: Using data from the National Health and Nutrition Examination Survey (NHANES), we conducted a two-stage analysis of participants aged 12-24 years. In stage one, we used 1,599 participants from the 2017-2018 cycle for model development. Somatic symptom risk was defined using somatic-related Patient Health Questionnaire (PHQ) items, and psychological indicators were reconstructed after removing overlapping somatic items. Lifestyle characteristics were clustered using K-means clustering to derive behavioral phenotypes. In stage two, an ensemble multilayer perceptron (MLP) model was trained using demographic characteristics, refined psychological indicators, and behavioral phenotype features, and was temporally validated using 1,986 participants from the 2021-2023 cycle. SHAP analysis examined feature contributions to model predictions. RESULTS: K-means clustering identified three behavioral phenotypes. The ensemble MLP model achieved 93% overall accuracy in the independent 2021-2023 validation dataset and correctly identified 84% of individuals with elevated somatic symptom risk. SHAP analysis showed that behavioral phenotype features were important for model interpretation and helped characterize risk differences across lifestyle profiles. CONCLUSION: Behavioral profiles characterized by unhealthy lifestyle patterns, including sleep insufficiency and excessive screen exposure, were identified as key drivers of elevated somatic symptom risk among youth. The interpretable machine learning model demonstrated strong performance in an independent temporal validation dataset, although further validation is needed before broader application in screening settings.