Neuropsychiatric symptoms (NPS), including anxiety, depression, sleep disturbances, agitation, apathy, and psychotic manifestations, are highly prevalent in older adults. They frequently emerge during the early stages of cognitive decline and are associated with an increased risk of dementia, accelerated functional deterioration, and a substantial burden on the caregivers and healthcare system. Although the Neuropsychiatric Inventory is the standard assessment tool, its length and resource needs limit feasibility in community and primary care settings, specifically in regions where home-based elderly care predominates. Emerging evidence suggests that the oral microbiome, a non-invasive and low-cost biological source, plays a significant role in neuropsychiatric health through inflammatory, immune, and metabolic pathways. Integrating microbiome data with sociodemographic and biological markers using machine learning may provide a practical solution for community-based NPS screening.
The aim of this study was to develop and validate an interpretable, community-based screening model for NPS in elderly individuals by integrating oral microbiome profiles, inflammatory biomarkers, and sociodemographic and clinical characteristics. It also compared the predictive performance of unimodal models with multimodal machine learning methods and translated the optimal model into a nomogram suitable for real-world clinical application. This study explored potential biological mechanisms linking oral microbiota, stress-related hormones, inflammatory markers, and neuropsychiatric symptoms.
This study was conducted in accordance with the Declaration of Helsinki and involved a modelling cohort recruited from three urban districts of Chongqing in 2021 and an external validation cohort recruited from different districts in 2024. Community-dwelling adults aged 60 years and older were enrolled by local healthcare centers using predefined inclusion and exclusion criteria with sex-based frequency matching to decrease confounding. All participants completed standardised questionnaires assessing demographic factors, lifestyle behaviors, and medical history, and cognitive and neuropsychiatric evaluation using the Mini-Mental State Examination (MMSE) and the Chinese Neuropsychiatric Inventory Questionnaire.
Saliva samples were collected under standardized conditions and analyzed for pro-inflammatory cytokines (IL-1β, IL-6, TNF-α), cortisol, and cathepsin B using enzyme-linked immunosorbent assays (ELISA). Oral microbiome composition was characterized by 16S rRNA gene sequencing of the V3-V4 region. Machine learning models, including logistic regression, support vector machine, and XGBoost, were trained by recursive feature elimination, LASSO regularization, and 10-fold cross-validation. Model performance was evaluated by using AUROC, accuracy, recall, F1 score, and calibration metrics, with external validation conducted to evaluate generalizability. Model interpretability was examined using SHAP analysis, and functional pathway predictions were performed by BioCyc-based metabolic reconstruction.
A total of 138 participants in the modelling cohort and 200 in the external validation cohort completed all assessments. Individuals with NPS showed significantly lower MMSE scores and higher levels of cortisol and inflammatory biomarkers in both cohorts. Although the alpha diversity of the oral microbiome did not differ significantly between groups, beta diversity analyses revealed clear structural differences in the microbial communities. Health-associated genera like Haemophilus, Rothia, and Neisseria were enriched in the non-NPS participants, whereas periodontal and inflammation-related genera, including Tannerella and Fretibacterium, were more abundant in the NPS group.
Microbiome-only models showed stronger predictive performance than models based on sociodemographic and inflammatory variables. Integrating microbiome data with contextual characteristics substantially improved prediction accuracy. The genus augmented XGBoost model got the highest performance with excellent discrimination and calibration in external and internal validation. SHAP analysis detected cortisol, alcohol consumption, education level, and specific microbial general as the most influential predictors. Functional analyses showed changes in the metabolic pathways related to amino acid, lipid, carbohydrate, and nucleotide metabolism, with notable involvement of the pentose phosphate pathway.
This study demonstrates that multimodal integration of oral microbiome profiles, inflammatory biomarkers, and sociodemographic factors significantly enhances early screening of neuropsychiatric symptoms in older adults. The genus-augmented XGBoost model showed robust performance and strong generalizability, while the nomogram derived from the logistic regression model offers a practical and interpretable tool for community healthcare settings. These findings provide mechanistic insights into how oral microbiota and stress-related metabolic pathways contribute to NPS pathogenesis. This non-invasive and scalable approach holds significant promise to improve early detection, risk stratification, and preventive interventions for NPS in aging populations, specifically in resource-limited community environments.
Reference: Liu P, Yang Z, Yin Q, et al. A community screening tool for neuropsychiatric symptoms in the elderly: integrating cortisol, microbiome, and social factors with machine learning. Transl Psychiatry. 2026;16:26. doi:10.1038/s41398-025-03797-3
