A community screening tool for neuropsychiatric symptoms in the elderly: integrating cortisol, microbiome, and social factors with machine learning

Abstract

Neuropsychiatric symptoms (NPS) are early indicators of cognitive decline due to neurodegenerative diseases, and their timely detection is of the utmost importance. We aimed to develop and validate methods for large-scale NPS screening among elderly individuals and explore underlying metabolic mechanisms. This observational, cross-section study involved 138 and 200 participants in the modeling and external validation cohorts, respectively, chosen from community healthcare centers in Chongqing, China. Data collection involved demographic questionnaires, saliva samples for oral microbiome analysis, and assays for other biomarkers (IL-6, IL-1β, TNF-α, Cath-B and cortisol). EXtreme gradient boosting(XGBoost), support vector machine(SVM), and logistic regression(LR) were developed with RFE and LASSO. The models were primarily evaluated using AUROC and F1 scores. The best model was interpreted using SHAP values, while the LR model was transformed into a nomogram. Additionally, BioCyc function pathway analysis was used to predict the functional shift of biomarkers. The genus-augmented XGBoost model achieved the highest performance, with an AUROC of 0.936 and an F1 score of 0.864, outperforming other models. The LR model was converted into a nomogram to facilitate NPS-risk assessment in community settings. The external validation confirmed the strong predictive power (AUROC = 0.986, F1 score = 0.944). Enrichment and correlation analyses revealed cortisol and microbial interactions with pathways such as the pentose phosphate pathway and enterobacterial common antigen biosynthesis. The XGBoost-augmented model and nomogram offer promising tools for community-based NPS screening, while enrichment analysis provides insights into biological mechanisms.

Introduction

Neuropsychiatric symptoms (NPS) encompass a range of symptoms and behaviors resulting from neurodegenerative diseases, including apathy, depression, anxiety, irritability/aggression, sleep disturbances, changes in appetite, motor behavior abnormalities, delusions, disinhibition, euphoria, and hallucinations [1]. NPS are prevalent in 30.8–80% of individuals across the cognitive spectrum [2]. They often emerge in the early stages of cognitive decline and are significant predictors of dementia risk, increasing the likelihood of developing cognitive impairment [3, 4]. Additionally, caring for patients with NPS is resource-intensive, with approximately 30% of dementia care resources being dedicated to NPS treatment, making the early detection of NPS in the elderly crucial [5].

The Neuropsychiatric Inventory (NPI) is the standard tool for NPS assessment, but its implementation faces challenges in community settings due to its length, variability in scoring, and limited resources [6–9]. In China, approximately 90% of older adults prefer home-based care, thereby increasing the demand for community-based services [10]. Therefore, simplified and objective screening tools suitable for community settings are urgently needed to facilitate early NPS detection.

The oral microbiome has emerged as a promising candidate for early detection [11]. The oral microbiome is essential for oral health and is associated with systemic diseases and neuropsychiatric well-being [12]. Studies have shown a strong association between oral microbiome imbalances and the onset of NPS, such as anxiety, depression, and agitation [13]. For instance, alterations in the oral microbiome—such as shifts in Neisseria, Prevotella, Fusobacterium, and Actinomycetaceae—have been associated with anxiety and depression symptoms in large population-based cohorts [14]. These mechanisms involve multiple biological pathways, including inflammatory responses, metabolic byproducts, and immune system activation [15]. First, an imbalance in the oral microbiome can trigger systemic inflammation, thus elevating pro-inflammatory cytokines such as IL-6 and TNF-α. These cytokines can circulate throughout the body and reach the central nervous system, inducing neuroinflammation, which can lead to neuronal damage and emotional dysregulation [16]. Additionally, metabolic byproducts from oral bacteria, such as short-chain fatty acids and lipopolysaccharides, can cross the blood-brain barrier and directly influence neurotransmitter production and release, thereby exacerbating NPS [17]. Given this evidence, the oral microbiome presents a unique opportunity for the noninvasive early detection of NPS and offers new intervention possibilities. Other biological factors, such as elevated levels of cytokines (IL-1β, IL-6, TNF-α), cortisol, and cathepsin B (Cath-B), are also linked to NPS progression, contributing to chronic neuroinflammation, oxidative stress, and cognitive decline [18–23]. These findings bolster the biological plausibility and translational potential of using oral microbiome and inflammatory markers for early detection of NPS in elderly populations. While mounting evidence supports the biological link between oral microbiota, inflammatory markers, and neuropsychiatric symptoms, real-world implementation remains limited. Notably, analyses of the NHANES 2011-2012 cycles based on nationally representative U.S. community samples have demonstrated that reduced oral microbiome α-diversity is significantly associated with higher risks of depression, cognitive impairment, and poorer executive function [24]. In a large-scale community cohort, salivary microbial composition, such as elevated Prevotella histicola and reduced Haemophilus sputorum, was significantly associated with symptoms of anxiety and depression [14]. These real-world studies underscore the feasibility of applying oral microbiome and inflammatory biomarkers in large-scale community screening for neuropsychiatric symptoms. The noninvasive nature and low cost of oral microbiome sampling, combined with advances in portable sequencing and biomarker detection technologies, offer promising avenues for practical application [25].

While biological markers provide crucial insights, NPS risk is inherently multifactorial and shaped by easily obtainable socio-demographic and behavioral characteristics [26, 27]. For example, alcohol consumption status is a well-established behavioral risk factor linked to accelerated NPS via neuroinflammation and oxidative stress pathways [28]. Lower educational attainment, often reflecting socioeconomic disparities and limited cognitive reserve, increases vulnerability to NPS onset [29]. Moreover, factors such as smoking habits, age, sex, marital status, BMI, and cognitive function jointly influence neuroendocrine regulation and the dynamics of social support systems [30–32]. Critically, sex differences have been shown to significantly influence the onset and progression of NPS, with women often experiencing higher levels of anxiety, depression, and agitation due to hormonal fluctuations, social roles, and caregiving burdens [33]. Essentially, these variables can be readily captured through brief questionnaires or routine health records in community settings, offering practical complements to biological assessment. Yet at the same time, reliance on any single biomarker or social determinant fails to capture the complex, interconnected biological–social system underlying NPS. This intrinsic complexity necessitates a multimodal data integration approach capable of synthesizing diverse features to improve predictive power and interpretability.

Recent advances in multimodal data integration, which combine demographic factors with emerging biomarkers like the microbiome, offer enhanced predictive accuracy [34]. Machine learning algorithms are particularly effective in processing complex multidimensional data and identifying patterns within biological systems. To date, most studies have focused on feature sets from single data types, such as clinical or genetic data. Future predictive models should integrate multimodal data—including clinical assessments, imaging, microbiome analysis, and genetic profiling—to offer a holistic view of patient health. The stress process model provides a valuable theoretical framework for further understanding the complexity of NPS. This model considers not only individual stressors but also patient background characteristics, primary and secondary stressors, and mediating factors that interact to influence health outcomes [35]. Expanding research horizons by incorporating the stress process model and integrating multiple biological and social factors can lead to a comprehensive understanding of NPS mechanisms (Fig. 1A).

Fig. 1. Conceptual framework for machine learning model development and interpretation.

A The conceptualframework of this study. B Overview of the analysis workflow, a: Data preparation, b: Classifier training, internal validation and external validation set in cross-validation, c: Models selection and d: Model interpretation based on SHAP. XGBoost eXtreme gradient boosting, SVM support vector machine, LR logistic regression, SHAP Shapley additive explanations, Avg Average.

In this study, we aimed to develop and validate a community-based screening tool for NPS in the elderly using interpretable machine learning algorithms. We also compared the performance of these models from the perspective of background characteristics (sex, age, marital status, education, body mass index [BMI], Mini-Mental State Examination [MMSE] score, smoking, drinking, diabetes, hypertension) and inflammatory biomarkers (Cath-B, cortisol, IL-6, IL-1β, TNF-α) and oral microbiome data. Additionally, we transformed the logistic regression (LR) model into a nomogram, thus enabling rapid manual calculation of the likelihood of NPS occurrence. Furthermore, we analyzed the model features to explore the potential metabolic pathways underlying their association with NPS, thus providing deeper insights into the biological relevance of biomarkers (Fig. 1B).

Methods

Study setting and subject recruitment

This study adhered to the Declaration of Helsinki guidelines and was approved by the Research Ethics Committee of the Affiliated Stomatological Hospital of Chongqing Medical University (Permit Number: 2024074). Informed consent was obtained from all participants or their legal surrogates and that patient anonymity has been preserved. The modeling cohort was recruited from three urban districts of Chongqing (Yuzhong, Jiangbei, Nan’an) between August-September 2021, while the external validation cohort was recruited from three other districts in Chongqing (Shapingba, Banan, Yubei) in July-August 2024. Participants aged 60 and above were invited through local healthcare centers, with inclusion/exclusion criteria in Supplementary Table 1. Participants were excluded based on predefined inclusion and exclusion criteria. All included participants completed the questionnaires under direct supervision, resulting in no missing data. Frequency matching by sex was conducted primarily to control for confounding, with the benefit of reducing class imbalance between the NPS and non-NPS groups. This research was registered with the Chinese Clinical Trial Registry (ChiCTR2400086760). Power analysis for the Receiver Operating Characteristic (ROC) curve was performed using PASS 2025 (NCSS, LLC, Kaysville, Utah, USA). With an assumed null Area Under the Curve (AUC) of 0.5, alternative AUC of 0.9, α = 0.05 (two-sided), and the available sample size (64 positive and 74 negative cases), the statistical power was approximately 0.99, indicating sufficient sample size [36].

Data collection

Data elements were obtained from a questionnaire survey, an enzyme-linked immunosorbent assay (ELISA), and 16S rRNA gene sequencing data analysis and used as predictors to train and build the model.

Questionnaire survey

Demographic and medical data, including age, sex, marital status, education, smoking, drinking, BMI and histories of hypertension and diabetes, were collected. The two neurologists who assessed cognitive status and NPS using the MMSE and Chinese Neuropsychiatric Inventory Questionnaire (CNPI-Q) were blinded to the purpose of the study [37, 38]. They were unaware of the study’s objectives and group allocations during the data collection process. All researchers received standardized training on NPS concepts, symptoms, and evaluation. Before administering the CNPI-Q, they completed online certification and a one-month trial phase (https://www.alz.washington.edu/npiq/Signin2.html). Daily quality control was conducted to ensure consistency and minimize bias, with random reassessments of half the patients and third evaluations by a psychiatrist for discrepancies. CNPI-Q assessors were blinded to the purpose of the study to reduce bias. Detailed information on the scale is available in Supplementary Table 2.

Saliva sample collection

Saliva samples were collected by professionally trained personnel the questionnaire, with participants providing informed consent and confirming no antibiotic use in the past six months. They were instructed to avoid food, beverages (other than water), gum, tobacco, and oral hygiene for two hours before collection. Saliva was collected using a mouthwash method, in which participants swished and gargled 10 mL of Scope mouthwash for 30 s and then spit it into a screw-cap container. Samples were collected between 9 and 11 a.m., frozen with dry ice, and transferred to the laboratory on the same day for aliquoting and storage [39]. This mouthwash-based collection method has been widely used and validated in previous studies involving oral microbiome profiling and biomarker detection, supporting its feasibility and reproducibility in population-based research [40, 41].

16S rRNA gene sequencing data analysis

DNA was extracted from various samples using the E.Z.N.A. Stool DNA Kit (D4015; Omega Bio-tek, Inc., Norcross, GA, USA), following the manufacturer’s protocol. The quality of the extracted DNA was assessed via 2% agarose gel electrophoresis, while detection was performed using an ultraviolet spectrophotometer. The 16S rRNA gene (V3-V4 region) was amplified via PCR using primers 341 F (5′-CCTACGGGNGGCWGCAG-3′) and 805 R (5′-GACTACHVGGGTATCTAATCC-3′).

ELISA analysis

ELISA kits (Gianglai, Shanghai, China) were used to measure salivary pro-inflammatory cytokines (IL-1β, TNF-α, IL-6), Cath-B, and cortisol, following the manufacturer’s instructions. The concentrations were determined by comparing the average absorbance of each sample to a standard curve.

Machine learning analytical framework

Pre-selection of microbial taxa

Empirically, microbiome that are present in only a small fraction of the entire population may have a negligible effect on the classifier performance due to their unavailability in most samples [42]. Microbiota present in less than 10% of the population were excluded from the classification analysis to avoid influencing classifier performance.

Feature selection

To address the high dimensionality of oral microbiota data at the genus and species levels, we adopted a stepwise modeling strategy. First, a reference model was built using sociodemographic characteristics (sex, age, education, BMI, MMSE score, smoking, drinking, diabetes, hypertension) and inflammatory biomarkers (Cath-B, cortisol, IL-6, IL-1β, TNF-α) with recursive feature elimination (RFE) for dimensionality reduction [43]. Next, microbiota data were modeled separately to avoid overshadowing lower-dimensional variables such as sociodemographic and inflammatory factors. Microbiota models were developed using RFE to optimize feature selection. Finally, we created a combined model integrating context characteristics, biomarkers, and microbiota using RFE and LASSO to enhance predictive performance and generalization [44].

Model development

We employed three widely used machine-learning classifiers to develop the binary classification model. SVM was selected because of its strong generalization ability, XGBoost for its efficient ensemble learning and performance, and LR as a reference due to its simplicity and widespread use in medical research [45]. Given the relatively small sample size and potential class imbalance, we adopted 10-fold cross-validation to maximize data utilization and ensure more stable performance estimates [46]. The model performance was evaluated using accuracy, balanced accuracy, precision, recall, F1 score, and AUROC averaged over 95% confidence intervals. The optimal model was chosen based on the highest AUROC. External validation using an independent cohort confirmed the reproducibility of the top models. Calibration was assessed using the Brier score and calibration curve, with lower Brier scores indicating better performance.

Interpretation of the best model

Once the optimal model was selected, we conducted SHAP values, which quantify each feature’s marginal contribution to the model. Higher SHAP values suggest a stronger influence [47]. By analyzing samples’ SHAP values, this method provides insights into feature importance and the direction of the influence of features on the model output.

Statistical analysis

Data input was completed using Epidata 3.1, with analysis conducted in Python 3.11.4 and R 4.2.1. Models were implemented in Python. Categorical variables were reported as counts/proportions, and continuous variables as medians (interquartile ranges). Chi-square and t tests were used to evaluate differences between groups.

Alpha diversity was assessed using QIIME2 and ggplot2, with significant differences detected by the Wilcoxon test. Nonparametric multidimensional scaling with the Vegan package in R 4.2.1 was used to assess microbial distribution. Mantel testing and Pearson’s correlation explored relationships between microbial composition and immune factors. Functional predictions of the microbial metabolic pathways were reconstructed using BioCyc pathway/genome databases [48], which provide organism-specific pathway annotations based on curated genomic evidence. This approach precisely maps metagenomic reads to enzymatic reactions and pathways within oral bacterial taxa, capturing niche-specific functional potentials. Statistical analysis was conducted with Statistical Analysis of Metagenomic Profiles (STAMP) [49], and a co-occurrence network of biomarkers and pathways was constructed using Spearman’s correlation and visualized in Cytoscape [50].

Results

Participant demographic and clinical characteristics

A total of 138 participants from the modeling cohort and 200 from the external validation cohort completed all questionnaires and saliva sample collection (Supplementary Figure 1). Frequency matching by sex ensured balanced comparison groups. As shown in Table 1, no significant differences were found between cohorts in terms of age, sex, marital status, education, BMI, smoking, or drinking. In both cohorts, MMSE scores were significantly lower in the NPS group (P = 0.016 for modeling, P = 0.002 for validation). Additionally, cortisol, Cath-B, TNF-α, IL-1β, and IL-6 levels were significantly higher in the NPS group than in the non-NPS group in both cohorts (all P < 0.001).

Table 1.

Comparison of demographic and clinical characteristics in the modeling cohort.

Variables	Modeling cohort				External validation cohort
	Total (n = 138)	non-NPS (n = 74)	NPS (n = 64)	P	Total (n = 200)	non-NPS (n = 103)	NPS (n = 97)	P
Age(years), Mean ± SD	74.88 ± 6.01	74.53 ± 5.83	75.30 ± 6.23	0.455a	74.40 ± 6.09	74.87 ± 6.01	73.90 ± 6.17	0.258a
Sex, n(%)				1.000b				1.000b
male	69 (50.00)	37 (50.00)	32 (50.00)		99(49.50)	51 (49.51)	48 (49.48)
female	69 (50.00)	37 (50.00)	32 (50.00)		101 (50.50)	52 (50.49)	49 (50.52)
Marital Status, n(%)				0.796b				0.887 b
With no spouse now	89 (64.49)	47 (63.51)	42 (65.62)		100 (50.00)	52 (50.49)	48 (49.48)
With spouse now	49 (35.51)	27 (36.49)	22 (34.38)		100 (50.00)	51 (49.51)	49 (50.52)
Education, n(%)				0.316b				0.111b
Junior middle school and the following	58 (42.03)	34 (45.95)	24 (37.50)		108 (54.00)	50 (48.54)	58 (59.79)
High school or above	80 (57.97)	40 (54.05)	40 (62.50)		92 (46.00)	53 (51.46)	39 (40.21)
Smoking, n(%)				0.818 b				0.559b
no	87 (63.04)	46 (62.16)	41 (64.06)		174 (87.00)	91 (88.35)	83 (85.57)
yes	51 (36.96)	28 (37.84)	23 (35.94)		26 (13.00)	12 (11.65)	14 (14.43)
Drinking, n(%)				0.350 b				0.100b
no	96 (69.57)	54 (72.97)	42 (65.62)		129 (64.50)	72 (69.90)	57 (58.76)
yes	42 (30.43)	20 (27.03)	22 (34.38)		71 (35.50)	31 (30.10)	40 (41.24)
Diabetes, n(%)				0.456 b				0.230b
no	86 (62.32)	44 (59.46)	42 (65.62)		136 (68.00)	74 (71.84)	62 (63.92)
yes	52 (37.68)	30 (40.54)	22 (34.38)		64 (32.00)	29 (28.16)	35 (36.08)
Hypertension, n(%)				0.526b				0.389b
no	73 (52.90)	41 (55.41)	32 (50.00)		97 (48.50)	53 (51.46)	44 (45.36)
yes	65 (47.10)	33 (44.59)	32 (50.00)		103 (51.50)	50 (48.54)	53 (54.64)
BMI, median (IQR)	21.64 ± 1.29	21.55 ± 1.44	21.73 ± 1.10	0.404a	23.26 ± 2.89	23.42 ± 2.69	23.08 ± 3.10	0.414a
MMSE, median (IQR)	23.44 ± 6.55	24.70 ± 5.73	21.98 ± 7.15	0.016a	27.55 ± 2.85	28.17 ± 2.13	26.91 ± 3.35	0.002a
Cath-B, median (IQR)	25.36 ± 8.62	22.66 ± 9.25	28.48 ± 6.63	<0.001a	38.05 ± 6.52	36.08 ± 6.64	40.15 ± 5.73	<0.001a
Cortisol, median (IQR)	7.59 ± 3.14	6.72 ± 3.40	8.60 ± 2.47	<0.001a	8.91 ± 2.22	8.18 ± 2.48	9.68 ± 1.57	<0.001a
IL-6, median (IQR)	36.66 ± 8.98	34.70 ± 8.90	38.93 ± 8.58	0.005 a	46.69 ± 8.07	44.40 ± 8.37	49.12 ± 7.01	<0.001a
IL-1β, median (IQR)	60.77 ± 16.31	56.23 ± 12.20	66.02 ± 18.81	<0.001 a	64.35 ± 13.84	59.73 ± 15.34	69.24 ± 9.98	<0.001a
TNF-α, median (IQR)	56.21 ± 13.52	53.22 ± 12.75	59.67 ± 13.66	0.005 a	67.56 ± 12.58	62.59 ± 13.35	72.83 ± 9.16	<0.001a

a: t test, b: Chi-square test; NPS neuropsychiatric symptoms, non-NPS elderly without neuropsychiatric symptoms, SD standard deviation, BMI body mass index; Smoking refers to having smoked continuously or cumulatively for more than six months in one’s lifetime and having smoked within the 30 days prior to the survey; Drinking alcohol = based on The Alcohol Use Disorders Identification Test (AUDIT-C). AUDIT-C score = 0 is Not drinking, AUDIT-C score > 0 is Drinking; MMSE Mini-Mental State Examination, Cath-B Cathepsin B, IL-6 interleukin-6, IL-1β interleukin-1β; TNF-α tumor necrosis factor α.

Oral microbial composition differences between the NPS and Non-NPS subjects

Saliva samples underwent 16S rRNA gene sequencing, with an average sequencing depth of roughly 2.253 million reads per sample. The rarefaction curve indicated that as sequencing depth increased, the Goods coverage index for the observed samples from both groups neared saturation and approached a value of 1 (Fig. 2A). Regarding alpha diversity, metrics such as the Chao1, Simpson, Pielou’s evenness, and Shannon indices (Fig. 2B) were lower in the non-NPS group than in the NPS group, though these differences were not statistically significant (P = 0.732, 0.143, 0.551, and 0.426, respectively). Beta diversity, visualized through principal component analysis and nonparametric multidimensional scaling based on amplicon sequence variant abundance, revealed distinct clusters between the groups. Different analytical approaches consistently showed significant microbial structural differences between the two groups (R = 0.0258, P = 0.026; Fig. 2C).

Fig. 2. The differences of salivary microbiota in elderly with NPS and healthy controls according to the 16S rRNA data.

A The rarefaction curve. B Alpha-diversity indices of microbiome in NPS saliva and non-NPS saliva. C Principal components analysis and nonparametric multi-dimensional scaling (NMDS) of salivary microbiota on the basis of unweighted_unifrac dissimilarity. D Relative abundance of the top 60 at the genus level between two groups. E the Statistical Analysis of Metagenomic Profiles (STAMP) depicts the disparities in microbial abundance between the NPS and non-NPS groups.

To minimize alignment artifacts and reduce noise, we concentrated on the prevalent bacterial taxa, filtering based on relative abundances and retaining only those found in at least 10% of the samples. This filtering yielded a comprehensive dataset comprising 29 phyla, 81 classes, 169 orders, 278 families, 537 genera, and 786 species. The overall composition of the salivary microbiota was further explored at the genus level (Fig. 2D and E). Analyses at the phylum, class, order, family, and species levels are shown in Supplementary Figure 2.

The stacked bar plot (Fig. 2D) illustrates the relative abundance of the top 60 genera across all samples, revealing considerable variability in the microbial composition between individuals, as indicated by substantial fluctuations in the relative abundances of bacterial taxa. These findings highlight individual differences in the oral microbiome among participants, even within the same group. Additionally, STAMP (Fig. 2E) revealed the disparities in microbial abundance between the NPS and non-NPS groups. Genera such as Haemophilus, Rothia, and Neisseria were notably enriched in the non-NPS group, whereas periodontal pathogens such as Tannerella, Fusobacterium, and Fretibacterium were more abundant in the NPS group. These taxa demonstrated significant differences, reinforcing their potential role in the pathogenesis or exacerbation of NPS and showing a clear microbial structural distinction between the two groups.

Oral microbiome to predict NPS in community-dwelling elderly individuals

To assess whether the oral microbiome predicts NPS, we trained and validated machine learning models for each microbiota level to avoid redundancy from overlapping taxonomic features. RFE was used to filter features (Supplementary Table 3), and class-level microbial communities were excluded due to excessive features post-reduction. Prediction accuracy improved at more detailed levels(Fig. 3A), with LR achieving the highest AUROC at the species level (0.822[95% CI: 0.795–0.848]), followed by genus (0.812[95% CI: 0.785–0.840]). SVM also performed best AUROC at the species level (0.843[95% CI: 0.817–0.870]). XGBoost showed strong AUROC performance across all levels, peaking at the family level (0.898[95% CI: 0.863–0.934]). Full performance metrics are provided in Supplementary Table 4.

Fig. 3. Comparison of approaches for the identification of NPS using oral microbiota features.

A The performance of different models for NPS identification, comparing a reference model (red) with microbiome-only models (blue). Errors bars represent mean and IQR. Horizontal dashed lines mark the mean performance of reference model. B Models of context characteristics, inflammation factors and oral microbiome data improved the identification of NPS over reference model. Area under the ROC curve (AUROC) for three algorithms using genus-level (orange) or species-level (blue) oral microbiome data together with other factors or a reference model (red). AUROC to determine the predictive ability of excellent models in the validation cohort, representing the sensitivity on the ordinate axis and specificity in abscissa. Error bars represent mean and IQR. Difference in performance between two models was tested using one-sided Wilcoxon-sighed rank test. ^***P < 0.001. Horizontal dashed lines mark the mean performance of reference model. C The remaining evaluation indicators of the models. D AUROC for external validation of the model. E Calibration curve of the models. F Positive and negative explanation of each feature impact on the NPS in the genus-augmented model by the SHAP values in the XGBoost algorithm. G Nomogram for identification NPS risk based on genus-augmented model of logistic regression.

Suboptimal performance of the reference model

To assess the value of various risk factors, we built a reference model using context characteristics and inflammatory biomarkers to compare with the prediction capacity of microbiome-based models. The reference model achieved an average AUROC of 0.679[95% CI: 0.633–0.726] (Fig. 3A), indicating the model has some discriminatory power, as AUROC values between 0.5 and 0.7 suggest that the model’s performance is only slightly better than random chance and may lack sufficient accuracy for practical application [45]. In contrast, the microbiome-based models exhibited superior predictive performance compared with that of the reference models.

Superior performance in genus-augmented models for NPS screening

To evaluate a microbiome-augmented prediction model for NPS, we combined microbiome features with the reference model and focused on the genus and species levels, which had shown superior performance. LASSO regression was applied for model simplification (Supplementary Table 3). As shown in Fig. 3B, XGBoost performed best, followed by LR, with SVM showing weaker results. The genus-augmented XGBoost model achieved the highest AUROC (0.936[95% CI: 0.903–0.969]), outperforming the species-augmented model (AUROC: 0.924[95% CI: 0.894–0.954], P < 0.001). For LR, the genus-augmented model also performed slightly better than the species-augmented model (AUROC: 0.823[95% CI: 0.801–0.844] vs. 0.832[95% CI: (0.802–0.863), P < 0.001). The genus-augmented XGBoost model also had the highest recall (0.878[95% CI: 0.779–0.978]), and its calibration curve showed strong alignment with ideal predictions, enhancing its clinical reliability. While LR performed well, its F1 score, recall, and accuracy were slightly lower than those of XGBoost but still showed better results at the genus level (Fig. 3C).

External validation confirms the robust predictive performance of augmented models

During the external validation phase, we further assessed the performance of the genus- and species-augmented models using the LR and XGBoost algorithms. Data from the external validation cohort were incorporated, and the final results are presented in Fig. 3D.

For the XGBoost algorithm, both the genus- and species-augmented models performed exceptionally well, with AUROC values of 0.986 and 0.945, and F1 score of 0.944 and 0.922, respectively, indicating high predictive accuracy for practical applications. The Brier scores for the genus- and species-augmented models were 0.0380 and 0.0424, respectively, indicating good model calibration, as demonstrated by the calibration curve shown in Fig. 3E.

For the LR algorithm, the genus- and species-augmented models achieved AUROC values of 0.746 and 0.744, respectively, with corresponding F1 scores of 0.668 and 0.661. Although these values were lower than those of the XGBoost models, they still demonstrated good predictive ability. The Brier scores for the genus- and species-augmented LR models were 0.2019 and 0.2001, respectively, indicating reasonable calibration, though not as optimal as the XGBoost models, as shown by the calibration curves in Fig. 3E.

External validation confirmed that both genus- and species-augmented models, using LR or XGBoost, generalized well and maintained strong predictive accuracy on unseen data.

Model interpretation for binary classification

Given the strong predictive performance of the genus-augmented model using XGBoost, we conducted a SHAP analysis to better understand the model and feature importance. Figure 3F presents the SHAP values for each variable in a summary plot. The y-axis lists the variables, whereas the x-axis shows the effect of each variable on species prediction using SHAP values. Dots of different colors represent the attribution of all patients to NPS: red denotes higher variable values, and blue denotes lower values. Each dot represents a single SHAP value for a specific feature and data point, highlighting its influence on the model output. The intensity of the color indicates the strength of the variable’s influence on the model’s output; darker shades correspond to stronger effects.

Higher levels of cortisol, Tannerella, Paludibacteraceae_F0058, Fretibacterium, and alcohol consumption were associated with a higher probability of NPS occurrence, as indicated by positive SHAP values. Conversely, features such as Filifactor, Lachnospiraceae_unclassified, and low education showed negative SHAP values, indicating a negative association with NPS occurrence. This suggests that higher levels of these features are associated with a reduced likelihood of NPS development.

Nomogram of the genus-augmented model

We translated the genus-augmented model into a nomogram format to enhance its visual interpretability and ease of use in clinical settings (Fig. 3G). The LR model, which performed exceptionally well, second only to the XGBoost model, along with the superior accuracy of the augmented model (compared with that of the individual models), justified this approach. Presenting the model as a nomogram will allow staff at community healthcare centers to easily assess the probability of NPS development using individual patient data, thereby facilitating informed decision-making and personalized treatment strategies. This format not only improves user comprehension, but also enables quick and accurate risk assessment, making it a valuable tool for large-scale implementation in community settings.

In the nomogram, each variable is represented by a scale along a line segment reflecting its value range, with the length of the segment indicating its contribution to the outcome. A vertical line is drawn to intersect the “Points” axis to determine the points for each variable. The “Total Points,” which comprise the sum of these individual points, are then used to ascertain the probability of NPS occurrence by drawing another vertical line that intersects the “Probability of NPS” axis. This tool allows medical teams to swiftly identify patients at high risk of NPS, making it highly suitable for widespread application in community healthcare environments.

Interactions between microbial communities and host immune factors

Previous studies have demonstrated a strong connection between inflammation, the microbiome, and NPS. This study utilized Mantel testing and Pearson correlation analysis to explore the interactions between microbial communities and host immune factors (Fig. 4A).

Fig. 4. Interactions between oral microbiota, inflammatory factors, and metabolic pathways in NPS and healthy controls.

A Interactions between microbiota genus and inflammatory factors based on mantel analysis. B Volcano plot showing significantly changed metabolites between the two groups samples. C BioCyc function pathway analysis showing metabolites-associated metabolic pathways. D A co-occurrence network constructed from relative abundances of differential oral bacteria, saliva metabolies in NPS subjects versus HCs. the spearman correlation analysis was used to explore the co-occurrence network. The relevant networks with Spearman’s correaltion (r) å 0.20 or <−0.20 were shown partly (P < 0.05).

Figure 4A illustrates the significant positive correlation between Tannerella and cortisol, suggesting that Tannerella may facilitate the regulation of cortisol levels. Similarly, Fretibacterium also exhibits a significant positive correlation with cortisol, as well as with Cath-B, IL-1β, and TNF-α, all showing statistical significance. These results imply that Fretibacterium may play an active role in inflammatory processes, potentially modulating key inflammatory mediators as part of host immune regulation. Furthermore, the significant positive correlation between Filifactor and TNF-α suggests that this microorganism may enhance TNF-α expression, indicating a possible role for Filifactor in promoting or sustaining inflammatory responses. In contrast, the significant negative correlation between Paludibacteraceae_F0058 and Cath-B suggests that an increase in Paludibacteraceae_F0058 abundance may lead to a corresponding decrease in Cath-B levels and vice versa. Additionally, Lachnospiraceae_unclassified shows a significant negative correlation with Cath-B, cortisol, and TNF-α, suggesting that Lachnospiraceae_unclassified may contribute to the suppression of these inflammatory factors and stress hormones. This negative correlation may reflect the potential role of this microorganism in host anti-inflammatory mechanisms.

These results highlight the role of specific microbial genera, such as Tannerella, Fretibacterium, and Lachnospiraceae_unclassified, which influence host immune and stress responses. Significant associations between key inflammatory and stress-related factors underscore their importance in health and disease. Additionally, the positive correlations between cortisol and certain microbiome suggest that stress may affect microbiome regulation.

Correlation between the identified combination of microbiota and function pathways

Volcano plots showed significantly different function pathways between the two sample groups (Fig. 4B). BioCyc function analysis revealed 116 altered metabolite-associated function pathways (Fig. 4C). These differential pathways were mainly involved in some biological processes: amino acid metabolism, carbohydrate degradation, organic acids and other compounds, vitamins and cofactors, fermentation pathways, benzenoids, nucleoside and derivatives, lipid metabolism, and biosynthetic pathways. This metabolic profiling provides insights into the distinct biochemical alterations associated with NPS and highlights the potential metabolic targets for therapeutic interventions.

Using Spearman’s correlation analysis, a co-occurrence network explored interactions between cortisol, key genera, and function pathway (Fig. 4D). Salivary cortisol, as the central node, showed negative correlations with the pentose phosphate pathway (PPP) and positive correlations with 2-aminophenol, 4-hydroxyphenylacetate, and D-galactarate degradation pathways. Tannerella was positively correlated with colanic acid and O-antigen biosynthesis but negatively with the PPP. Filifactor and Paludibacteraceae_F0058 were linked to fermentation pathways and pyrimidine deoxyribonucleotide biosynthesis, while Fretibacterium correlated with enterobacterial common antigen (ECA) biosynthesis and lipid metabolism.

Lachnospiraceae_unclassified showed positive correlations with the PPP and amino acid metabolism but negative correlations with ECA biosynthesis and lipid metabolism. These findings suggest complex interactions between oral microbiota and metabolic pathways in NPS.

Discussion

In this study, we demonstrated that the screening of NPS in the elderly substantially benefited from integrating multimodal data rather than relying exclusively on unimodal data. Overall, the results emphasize the high predictive value of oral microbiology, compared with those of clinical and sociodemographic features, whether used independently in the model or as part of a full-featured set. Specifically, the model performance improved significantly when factors such as drinking, education level, cortisol, and oral microbiota were combined. Our findings replicate those of previous psychiatric studies, suggesting that multimodal modeling may outperform unimodal approaches [34]. By transforming an LR model into a nomogram, we developed a simple and practical tool that can be rapidly applied to resource-limited communities and rural settings (Fig. 3G). The non-invasive collection of oral microbiome samples further enhances patient compliance, making it suitable for widespread implementation. SHAP analysis identified the importance of eight key features within the model. This study also outlined the landscapes and interaction networks of differential microbiota and other markers and pathways in the NPS and non-NPS groups (Fig. 4). These findings not only provide effective tools for the early screening of NPS, particularly in community health initiatives, but also suggest that oral microbiota disturbances could potentially contribute to NPS pathogenesis by modulating the host’s amino acid, nucleotide, cofactor, vitamin, and carbohydrate metabolism, thus providing a new avenue for understanding the basis of NPS (Fig. 4D).

The findings of our SHAP analysis are consistent with those of a previous NPS study, which showed that drinking and education level are predictive of NPS [51, 52]. Long-term alcohol consumption induces structural changes in the brain, particularly in areas such as the hippocampus and prefrontal cortex. The hippocampus, which is crucial for memory formation and spatial navigation, can suffer from neurodegeneration due to chronic alcohol exposure, leading to impairments in memory and learning. Similarly, the prefrontal cortex, which is responsible for executive functions such as decision-making, impulse control, and emotional regulation, is also highly susceptible to the damaging effects of alcohol. These changes can manifest as emotional dysregulation, impulsivity, and increased vulnerability to mood disorders, all of which are key components of NPS [53].

Education level is closely linked to cognitive reserve, which refers to the brain’s ability to compensate for damage and maintain functionality. Individuals with higher education levels tend to develop a high cognitive reserve, which provides them with a greater buffer against neurological decline. Conversely, lower educational levels have been associated with reduced cognitive reserve, making individuals more susceptible to neuropsychiatric disorders [53]. This reduced reserve can result in an increased likelihood of developing conditions such as dementia, depression, and anxiety as cognitive demands surpass the brain’s compensatory capabilities. These findings suggest that future interventions should focus on individuals with lower educational levels and alcohol-consumption behaviors to slow the progression of NPS through lifestyle modifications and early interventions.

In our study, the SHAP analysis identified cortisol as a top-ranked feature. Salivary cortisol, despite its known diurnal fluctuations and individual variability (e.g., due to exercise or stress), remains a widely used biomarker for stress [54]. Previous studies have highlighted the challenges in using cortisol due to these fluctuations, which are influenced by both the timing of sample collection and individual factors [54, 55]. In this study, we chose to collect saliva samples between 9:00 and 11:00 a.m., a period that follows the early morning cortisol peak, to minimize the influence of the cortisol awakening response while still capturing basal cortisol levels. This choice aligns with recommendations from prior research, which suggest that this time window is stable and less subject to the sharp fluctuations seen immediately after awakening [39]. Additionally, the 9:00–11:00 a.m. sampling period facilitates comparability across participants and enhances the reliability of cortisol as a stress marker.

While previous studies have extensively explored cortisol’s role in stress responses [21], our study further revealed a negative relationship between cortisol and the PPP, a neuroprotective mechanism crucial for neuronal survival, particularly in mitigating damage by reactive oxygen species [20, 56]. This aligns with the findings of Wu et al., who reported a connection between cortisol and metabolic disturbances in patients with depression, suggesting that cortisol may influence NPS by affecting neuronal metabolic pathways [20]. However, our findings extend this by highlighting that elevated cortisol levels may weaken the PPP, thus potentially leading to reduced oxidative stress resistance in neurons.

Mechanistically, we observed that the negative correlation between cortisol and the PPP may lead to reduced nicotinamide adenine dinucleotide phosphate production, thus impairing the neuronal capacity to counteract oxidative stress [57]. This result is consistent with previous reports that cortisol affects glucose, fat, and protein metabolism, thereby contributing to broader metabolic imbalances across multiple biological systems [58]. Compared with Dubar et al., who focused on the impact of cortisol on periodontal tissue metabolism and its link to increased inflammatory susceptibility, our study broadens this perspective by connecting metabolic disruptions to neurological conditions [59]. We propose that cortisol inhibition in the PPP may exacerbate NPS development, providing new evidence for its role in neurodegenerative processes.

In addition to cortisol, we found significant interactions between oral microbiota and the PPP. Specifically, Fretibacterium and Tannerella were negatively correlated with the PPP, mirroring the effects of cortisol. These findings support the hypothesis that these bacteria contribute to NPS pathogenesis via cortisol-like metabolic mechanisms. While previous studies have confirmed the pro-inflammatory role of species such as Tannerella forsythia in periodontal disease [17], our research is the first to suggest that these bacteria may directly influence neuronal metabolic pathways by affecting the PPP. This microbiota–metabolic interaction may be one of the underlying mechanisms linking chronic stress to neuroinflammation.

Conversely, Lachnospiraceae_unclassified showed a positive correlation with the PPP, suggesting that it may promote nicotinamide adenine dinucleotide phosphate production and enhance antioxidant defenses, thereby reducing oxidative stress. This finding aligns with prior research indicating that Lachnospiraceae produce short-chain fatty acids (particularly butyrate) that have anti-inflammatory properties [60, 61]. Our study expands on this by showing that these bacteria may reduce ECA production, thereby mitigating chronic low-grade inflammation and associated neuroinflammation [62]. This study provides a theoretical foundation for future strategies aimed at regulating the PPP to prevent NPS through probiotic interventions.

In summary, our study uncovered critical interactions between cortisol, oral microbiota, and metabolic pathways, such as the PPP, in driving NPS. Cortisol and specific bacteria contribute to neuroinflammation, whereas Lachnospiraceae_unclassified may counteract these effects. These findings not only enrich the existing theoretical framework but also offer new perspectives on early intervention strategies for NPS, particularly by modulating cortisol levels and maintaining microbial balance to prevent neuroinflammation and NPS.

Limitations

Our study had several limitations. First, while 16S rRNA sequencing was employed in this study to characterize the oral microbiome, we acknowledge that such techniques are not yet widely accessible or affordable in many real-world clinical or rural settings. To improve the translational potential of our findings, future work should explore the feasibility of translating key microbial signatures into low-cost, rapid detection platforms such as qPCR, loop-mediated isothermal amplification (LAMP), or microarray-based methods. Second, model validation was limited to regional data, restricting its generalizability. In the future, we will collaborate with community healthcare centers in other regions of China to expand the sample base and carry out cross-regional data collection and model validation work. Third, specific parameters were focused on, possibly missing factors like pain assessment and social support. Fourth, we did not include neuroimaging data, which is a key component in the clinical prediction of neurodegenerative disorders. In the absence of imaging, future studies should consider using imaging data or validated surrogate markers such as blood-based biomarkers or digital assessments. Moreover, salivary cortisol levels may be influenced by individual differences and external factors, this study minimized interference by standardizing the sampling time to enhance the comparability and reliability of the data. Lastly, cost data related to predicting NPS from oral microbiota—such as those for model development, implementation, and ongoing management—were not collected in the present study, and thus a formal cost-effectiveness analysis could not be conducted. We plan to incorporate it into future research as part of a more comprehensive assessment of the model’s clinical utility.

Conclusions

The XGBoost-augmented model effectively screened for NPS in the elderly. The nomogram may be a beneficial management tool with enhanced practical utility in clinical settings. This approach not only improves early screening accuracy but also offers valuable insights into the biological pathways involved, making it a useful tool in community healthcare settings. Additionally, the BioCyc function analysis provided insights into specific microbial groups and their potential roles in the biological mechanisms underlying NPS.

Author contributions

Original draft writing: P.L., Z.Y.; Methodology: P.L., Z.Y., Q.Y., X.J.; Data curation: P.L., Z.Y., Q.Y., Y.D.; Conceptualization: P.L., Z.Y.; Formal analysis: P.L., Investigation: Q.Y., B.T.; Review & editing: X.J., Y.D., Y.L., X.X., Y.C., B.Y.; Supervision: B.T., B.Y.; Conceptualization: B.Y.; Funding acquisition: B.Y.

Data availability

The raw sequencing datasets supporting this study have been deposited in the NCBI Sequence Read Archive (SRA) under BioProject PRJNA1355217. All SRA run accessions (SRR35916334–SRR35916671) have been released and are openly accessible to the public through the NCBI SRA repository. In the meantime, we are happy to provide the raw data upon any reasonable request to the corresponding author.

Competing interests

The authors declare no competing interests.

Disclosure statement

An unauthorized version of the Chinese MMSE was used by the study team without permission, however this has now been rectified with PAR. The MMSE is a copyrighted instrument and may not be used or reproduced in whole or in part, in any form or language, or by any means without written permission of PAR (www.parinc.com).

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-025-03797-3.