Multidomain prediction of education-stratified MoCA-defined mild cognitive impairment in community-dwelling older adults in urban China

Abstract

Background

Mild cognitive impairment (MCI) is an important public health concern in ageing populations, yet scalable approaches for early identification in community settings remain limited. Existing prediction studies in China have often relied on questionnaire-based or conventional epidemiological variables, whereas multidomain models incorporating objective behavioural and environmental measures remain less common. This study examined whether multidomain indicators could improve the identification of education-stratified MoCA-defined MCI in community-dwelling older adults in urban China.

Methods

We conducted a cross-sectional analysis of community-dwelling older adults aged 60 years or above from Nanjing, China (analytic sample: n = 309). MCI status was defined using education-stratified Montreal Cognitive Assessment (MoCA) cut-offs (MCI n = 110; non-MCI n = 199). Candidate predictors included accelerometer-derived movement behaviours, anthropometric and bioelectrical-impedance-derived body-composition measures, and objective and perceived built-environment indicators. To reduce potential circularity, education-related variables were excluded from the primary predictor set. Data were divided into a stratified training set (70%, n = 216) and an independent held-out test set (30%, n = 93). Six machine-learning algorithms were trained and evaluated. Model performance was assessed using discrimination, accuracy, calibration, decision curve analysis, and model-interpretation methods.

Results

In the independent held-out test set, LightGBM showed the best overall performance, with an AUC of 0.854 (95% CI 0.762–0.946) and an accuracy of 0.822 (95% CI 0.701–0.913). XGBoost also performed well (AUC 0.835, 95% CI 0.741–0.929), followed by random forest (AUC 0.810, 95% CI 0.712–0.908). LightGBM had the lowest Brier score (0.139, 95% CI 0.092–0.186), and both LightGBM and XGBoost showed more favourable calibration than the other classifiers. Decision curve analysis suggested greater net benefit for the better-performing models than the default treat-all and treat-none strategies across clinically relevant threshold probabilities. Model-interpretation analyses indicated that movement behaviours, central adiposity/body-composition measures, age, and neighbourhood-context variables contributed importantly to prediction.

Conclusions

Multidomain predictors spanning objective movement behaviours, body-composition indicators, and built-environment measures showed value for identifying education-stratified MoCA-defined MCI in community-dwelling older adults. LightGBM achieved the best overall performance among the candidate models. These findings support the potential utility of integrating behavioural, biological, and environmental information for community-based cognitive-risk stratification, although external validation in independent and longitudinal cohorts is required before routine implementation.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-07656-8.

Introduction

Dementia and mild cognitive impairment (MCI) are major public health challenges in ageing populations. MCI is generally considered a transitional stage between normal cognitive ageing and dementia, although not all individuals with MCI progress to dementia [1]. Its clinical and public health relevance lies in the elevated risk of subsequent cognitive decline, functional deterioration, and conversion to dementia, which makes early identification particularly important for prevention-oriented community health strategies [1, 2]. In China, population ageing is accelerating rapidly, and MCI has become an increasingly important concern in older adults [3, 4]. However, timely recognition of older people at elevated risk of cognitive impairment remains difficult in community settings, where comprehensive neuropsychological assessment is often resource-intensive and not always feasible at scale [5].

Cognitive ageing is influenced by multiple factors across behavioural, biological, and environmental domains. Previous studies have shown that physical activity, sedentary behaviour, adiposity, educational experience, and neighbourhood conditions may all be associated with late-life cognitive health [5–8]. In China, Song et al. developed machine-learning models to predict mild cognitive impairment in older adults using a nationwide survey dataset [9]. However, that study was based on a public database and relied mainly on self-reported questionnaire and conventional epidemiological variables, rather than objective behavioural assessment or directly measured environmental exposures [9, 10]. Such work is important for large-scale risk stratification, but predictor sets derived primarily from subjective reporting may not fully capture daily movement patterns and neighbourhood characteristics that are potentially relevant to cognitive health.

In recent years, wearable and digital approaches have provided new opportunities for cognitive-risk modelling. Accelerometers can quantify habitual physical activity and sedentary behaviour objectively, thereby reducing recall bias and social-desirability bias inherent in questionnaire-based assessment [11]. At the same time, body-composition indicators and muscle-strength measures may reflect biological characteristics related to metabolic health, frailty, and functional reserve, while geographic information system (GIS)-based and perceived neighbourhood measures can capture contextual features relevant to mobility, accessibility, and social participation [10, 12, 13]. Recent studies have further shown that wearable-derived or digital physiological features can support the prediction of cognitive-test performance or MCI-related outcomes [14–16]. Nevertheless, evidence remains limited in Chinese community-dwelling older adults, particularly studies integrating objectively measured movement behaviours with body-composition and neighbourhood-context indicators in an interpretable machine-learning framework.

Another methodological issue is that outcome definition requires careful handling. Performance on the Montreal Cognitive Assessment (MoCA) is influenced by educational attainment, and this should be considered when developing prediction models for MCI in older adults [17]. Against this background, the present study aimed to develop and internally validate machine-learning models for identifying MoCA-defined MCI in community-dwelling older Chinese adults using multidomain predictors, including accelerometer-derived movement behaviours, anthropometry and body-composition indicators, and both objective and perceived built-environment measures. To reduce potential circularity between educational attainment and the outcome definition, the primary analysis used education-stratified MoCA cut-offs while excluding education from the main predictor set. In addition to model discrimination, we evaluated calibration, decision-analytic utility, and model interpretability to assess the potential value of these multidomain measures for community-based risk stratification. Rather than proposing a diagnostic tool, this study is intended to support early identification strategies and to clarify which behavioural, biological, and environmental factors contribute most to prediction in this setting.

Methods

Study design and participants

This study was a secondary analysis of a cross-sectional community dataset collected in Nanjing, China, between August 2017 and March 2018. The dataset was considered suitable for the present analysis because it simultaneously included cognitive assessment, accelerometer-derived movement measures, anthropometry/body-composition indicators, and neighborhood-related variables within the same community-dwelling older population. Participants were recruited from 18 communities across five urban districts (Gulou, Qinhuai, Xuanwu, Qixia, and Jianye) using a multistage stratified sampling approach.

Eligible participants were aged 60 years or older, had lived in the local area for at least 6 months, were able to complete questionnaires and on-site assessments, and provided written informed consent. Normal cognitive function was not an eligibility requirement. After data cleaning and deduplication, 503 unique participants were retained in the source dataset.

The present secondary analysis was reviewed and approved by the Ethics Committee of Nanjing Normal University (IRB No. 202003005; approval date: 10 March 2020), which approved the retrospective use of de-identified data collected as part of a community-based, non-invasive assessment. Written informed consent had been obtained from all participants before the original data collection.

Because the primary outcome in this study was MoCA-defined mild cognitive impairment (MCI), participants without a usable MoCA total score were excluded from supervised model development. In the cleaned source dataset, 194 participants did not have a MoCA total score that could be used for outcome derivation. Specifically, this referred to records in which the dataset did not contain an analyzable MoCA total score for modelling, rather than to participants being judged cognitively normal. These participants were excluded solely because the binary MCI outcome could not be constructed from the available data. The final analytic sample therefore comprised 309 participants with non-missing and usable MoCA total scores. Among these, 110 were classified as having MCI and 199 as non-MCI according to the education-stratified MoCA cut-offs described below. The final analytic sample therefore comprised 309 participants with non-missing MoCA total scores. Among these, 110 were classified as having MCI and 199 as non-MCI according to the education-stratified MoCA cut-offs described below. The participant selection process is shown in Fig. 1.

Fig. 1.

Participant flow diagram

Measures

Measures comprised global cognitive function, objectively measured physical activity and sedentary behaviour, anthropometry and bioelectrical-impedance-derived body composition, built-environment exposures, and prespecified sociodemographic and health-related covariates.

Outcome definition

Global cognitive function was assessed using the Chinese version of the Montreal Cognitive Assessment (MoCA), which provides both domain-specific subscores and a total score ranging from 0 to 30 [17]. The primary outcome was a binary indicator of mild cognitive impairment (MCI) derived from the MoCA. Because educational attainment can affect MoCA performance and is also commonly incorporated into score adjustment, the primary analysis used education-stratified cut-offs to reduce potential circularity between outcome definition and educational predictors [18, 19]. Specifically, MCI was defined as a MoCA total score of ≤ 19 for participants with ≤ 6 years of education, ≤ 22 for those with 7–12 years of education, and ≤ 24 for those with > 12 years of education [20]. Participants scoring at or below the corresponding threshold were classified as having MCI. Education strata were operationalised according to highest attained education level using pre-specified mapping rules (Supplementary Table S1).

To examine the robustness of the outcome definition, two sensitivity analyses were performed. First, we applied a uniform threshold of raw MoCA < 26 without education correction. Second, we applied the commonly used MoCA scoring rule in which 1 point is added for participants with ≤ 12 years of education, and MCI is then defined as an adjusted MoCA score < 26 [17]. These additional analyses were used to evaluate whether education-related score adjustment materially influenced model performance or the relative importance of predictors.

Objective physical activity and sedentary behaviour (accelerometry)

Objective physical activity and sedentary behaviour were assessed using a hip-worn triaxial accelerometer (ActiGraph GT3X+, ActiGraph LLC, Pensacola, FL, USA). Participants were instructed to wear the device on an elastic belt over the hip during waking hours for 7 consecutive days and to remove it only during water-based activities such as bathing or swimming. Acceleration signals were sampled at 30 Hz and processed into activity counts, which were then aggregated into 10-second epochs for analysis [11].

Non-wear time was defined as at least 60 consecutive minutes of zero counts, allowing up to 2 min of counts below 100 counts/min within the non-wear window [21]. A valid day was defined as at least 480 min (8 h) of wear time. For each participant, daily time spent in sedentary behaviour (SB), light-intensity physical activity (LPA), and moderate-to-vigorous physical activity (MVPA) was calculated and then averaged across all available valid days. Sleep and daytime naps were not included, because analyses were restricted to waking wear time based on the wear-time algorithm and participant instructions.

Activity intensity was classified using count-based cut-points expressed in counts per minute: SB < 100 counts/min, LPA 100–1951 counts/min, and MVPA ≥ 1952 counts/min [22]. In addition to average daily minutes spent in each intensity category, total physical activity volume (e.g., total counts/day and counts/min, where available) was extracted as candidate predictors.

Anthropometry, body composition, and muscle strength

Anthropometric measurements were obtained by trained assessors following standardised procedures. Height and weight were measured with participants wearing light clothing and no shoes, and body mass index (BMI) was calculated as weight in kilograms divided by height in metres squared. Waist and hip circumferences were measured using a non-elastic tape, and waist-to-hip ratio (WHR) was subsequently calculated.

Body composition was assessed using bioelectrical impedance analysis (BIA) with the InBody 230 system (Biospace, Seoul, Republic of Korea). Participants were instructed to avoid vigorous exercise and large meals before testing. BIA-derived parameters, including fat mass, fat-free mass, and body fat percentage, were extracted as candidate predictors.

Muscle strength was assessed using a digital handgrip dynamometer (Jamar Plus+, Patterson Medical Ltd., Nottingham, UK). Participants performed two maximal trials with the dominant hand, and the highest recorded value (kg) was used for analysis.

Built environment

Objective built-environment measures (GIS-derived)

Objective built-environment exposures were derived using geographic information systems (GIS) based on participants’ residential locations. Residential addresses were geocoded to latitude and longitude coordinates. Geocoding quality was checked against base maps, and ambiguous or low-confidence matches were manually reviewed and corrected. To protect participant privacy, geocoded locations were used only for spatial processing and were not reported at an identifiable spatial resolution.

Built-environment indicators were calculated within a 500-m walking network buffer around each residence, a neighbourhood scale commonly used to capture older adults’ routine daily activity spaces [13]. Network buffers were generated using street-network data and GIS-based network analysis procedures in ArcGIS 10.5. Within each buffer, we quantified several domains of the neighbourhood environment: population density and building density (where available), street connectivity (e.g., intersection density), land-use mix, destination accessibility (including counts or densities of transit stops and recreational or fitness destinations, as well as distance to relevant destinations), and facility diversity. GIS layers and points-of-interest datasets were obtained from authoritative municipal GIS sources and/or publicly available spatial databases that were contemporaneous with the survey period. Details of GIS data acquisition, processing, and variable construction are provided in the Supplementary Methods.

Perceived built-environment measures

Perceived neighbourhood environment was assessed using a modified Chinese version of the Neighborhood Environment Walkability Scale (NEWS) [12]. Participants reported perceived characteristics of neighbourhood walkability, including access to destinations, street connectivity, pedestrian facilities, aesthetics, and safety, using standardised response categories. Scale scoring followed published recommendations, with higher scores indicating a more favourable perceived neighbourhood environment. Details of instrument adaptation and scoring are provided in the Supplementary Methods.

Covariates

Prespecified covariates were selected because they are established or plausible correlates of cognitive impairment in later life and may also confound the relationships of physical activity, sedentary behaviour, body composition, and neighbourhood context with cognitive status. Sociodemographic characteristics were collected using interviewer-administered questionnaires and included age, sex, marital status, living arrangement, and monthly income. Health-related covariates included self-reported physician-diagnosed chronic conditions, such as hypertension, diabetes, coronary heart disease, stroke, and other common comorbidities, as well as smoking and alcohol use where available. These variables were therefore retained as candidate predictors in model development.

Data quality control and missing data

Prior to analysis, we implemented structured data-quality control procedures. These included: (i) deduplication based on unique identifiers and consistency checks across key demographic variables; (ii) range checks for continuous variables using physiologically plausible limits and instrument-specific allowable ranges; (iii) logical consistency checks across related variables (e.g., height, weight, and BMI; waist circumference, hip circumference, and WHR); and (iv) inspection of distributional summaries to identify outliers or implausible values.

Measurement error was defined as values that were clearly outside physiologically plausible ranges, beyond instrument reporting limits, or internally inconsistent with related variables. When suspect values were identified, they were cross-checked against the original records when available. If a value could not be verified, it was treated as missing.

Missingness was summarised for each variable. Participants without an analyzable MoCA total score were excluded from supervised model development because the primary outcome (MCI status) was derived from the MoCA, as described in the participant flow diagram. For predictor variables, missing data were handled within the modelling pipeline to minimise information leakage. Specifically, imputation procedures were fitted using training data only and then applied to the corresponding validation or test data. Continuous predictors were imputed using the median, whereas categorical predictors were imputed using the most frequent category or retained as an explicit missing category where appropriate.

Machine-learning analysis

Prediction task and outcome

The aim of the analysis was to develop and internally validate machine-learning models for classifying MCI status in a cross-sectional prediction setting. The primary outcome was MCI defined using education-stratified MoCA cut-offs (≤ 19 for ≤ 6 years of education, ≤ 22 for 7–12 years, and ≤ 24 for > 12 years), as described in Sect.  "Outcome definition".

Candidate predictors

Candidate predictors comprised all variables described in Sect.  "Objective physical activity and sedentary behaviour (accelerometry)"–"Covariates", including: (i) accelerometer-derived physical activity and sedentary-behaviour metrics; (ii) anthropometric and bioelectrical-impedance-derived body-composition parameters; (iii) GIS-derived and perceived built-environment indicators; and (iv) sociodemographic and health-related covariates. A complete list of candidate predictors and their definitions is provided in Supplementary Table S2.

To minimise potential circularity between outcome definition and educational attainment, education was not included as a predictor in the primary models and was used only to define education-stratified outcome thresholds. Its role as an additional predictor was examined only in secondary analyses.

Data splitting and resampling strategy

The analytic sample (n = 309) was randomly divided into a training set (70%) and a held-out test set (30%) using stratified sampling to preserve the proportion of MCI and non-MCI cases. Baseline characteristics of the training and held-out test sets are presented in Supplementary Table S2. All model-development steps, including imputation, scaling where required, feature selection, model fitting, and hyperparameter tuning, were performed using the training set only. Ten-fold cross-validation (CV) within the training set was used for model tuning. The final tuned models were then refitted in the full training set and evaluated once in the independent held-out test set.

Feature selection using logistic LASSO

Feature selection was conducted in the training set using LASSO-penalised logistic regression with a binomial link. The penalty parameter (λ) was selected by 10-fold CV using the minimum cross-validated deviance; the 1-standard-error rule was also examined as a sensitivity check for parsimony. Predictors with non-zero coefficients at the selected λ were retained for downstream model development. To avoid information leakage, feature selection was performed using training data only, including within CV resamples where applicable.

Model development and hyperparameter tuning

We trained six classification models: logistic regression, random forest, support vector machine (SVM), k-nearest neighbours (KNN), extreme gradient boosting (XGBoost), and LightGBM. Continuous predictors were standardised for algorithms that were sensitive to feature scale, including logistic regression, SVM, and KNN; tree-based models were fitted using the original predictor scales after imputation. Hyperparameters were tuned by grid search within the 10-fold CV framework in the training set. The primary tuning metric was area under the receiver operating characteristic curve (AUC), with calibration performance, particularly the Brier score, considered secondarily where relevant. Class imbalance was addressed using class weights or equivalent algorithm-specific options when available.

Performance evaluation

Model performance was evaluated in the independent held-out test set. Discrimination was assessed using the AUC and classification accuracy, with additional reporting of sensitivity and specificity. Calibration was assessed using the Brier score and calibration intercept and slope, estimated from the relationship between observed outcomes and logit-transformed predicted probabilities.

To quantify uncertainty, 95% confidence intervals (CIs) for test-set performance metrics were estimated using 2,000 bootstrap resamples of the held-out test set. This procedure was applied to AUC, accuracy, sensitivity, specificity, and the Brier score; bootstrap-based uncertainty estimates were also derived for calibration intercept and slope where reported. A single classification threshold was determined within the training set only, by maximising Youden’s J based on cross-validated predictions, and this threshold was then applied unchanged to the held-out test set to compute accuracy, sensitivity, and specificity.

Clinical utility and explainability

Clinical utility was assessed using decision curve analysis (DCA) based on predicted probabilities in the held-out test set. Net benefit was estimated across a clinically plausible range of threshold probabilities and compared with the default treat-all and treat-none strategies.

Model interpretability was examined for the primary model (final tuned LightGBM) using SHAP (SHapley Additive exPlanations). Global feature importance was summarised using mean absolute SHAP values in the held-out test set. Partial dependence plots (PDPs) were generated for the top-ranked predictors to illustrate the marginal relationship between selected predictors and the predicted probability of MCI while averaging over the observed distribution of other variables. These model-derived patterns were interpreted descriptively rather than causally.

Sensitivity analyses

Several sensitivity analyses were conducted to evaluate robustness. First, the full modelling pipeline was repeated using two alternative MoCA-based outcome definitions: (i) a uniform threshold of raw MoCA < 26 without education correction, and (ii) the commonly used scoring rule in which 1 point is added for participants with ≤ 12 years of education and MCI is then defined as an adjusted MoCA score < 26. Second, to examine the potential influence of educational attainment as a predictor, we repeated the analyses after additionally including education in the predictor set. These sensitivity analyses were used to assess whether the main findings were materially affected by alternative outcome definitions or by inclusion of education as an explicit predictor.

Statistical software and reporting

Descriptive analyses were performed using SPSS version 22.0 (IBM Corp., Armonk, NY, USA) and R version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria). Machine-learning analyses, including data preprocessing, model development, hyperparameter tuning, and performance evaluation, were implemented in Python version 3.8. A fixed random seed was applied throughout the analysis to enhance reproducibility. Reporting of prediction-model development and internal validation followed the TRIPOD guidance, and the completed TRIPOD checklist is provided as Supplementary File 1.

Results

Participant characteristics

A total of 309 community-dwelling older adults with analyzable MoCA total scores were included in the final analytic sample. The mean age was 68.86 years (SD 5.78), and 140 participants (45.3%) were men. Based on the education-stratified MoCA cut-offs, 110 participants (35.6%) were classified as having MCI and 199 (64.4%) as non-MCI. Baseline characteristics according to cognitive-status group are presented in Table 1.

Table 1.

Baseline characteristics of participants according to MCI status (n = 309)

Variable	Non-MCI (n = 199)	MCI (n = 110)	P value
Age (years)	68.46 (5.56)	69.58 (6.14)	0.110
Height (cm)	162.34 (7.98)	160.52 (7.55)	0.048
Weight (kg)	66.49 (9.74)	64.48 (10.42)	0.098
BMI (kg/m)	25.22 (3.19)	24.99 (3.49)	0.575
Waist-to-hip ratio (WHR)	0.89 (0.06)	0.88 (0.06)	0.849
Body fat (%)	28.66 (9.09)	29.23 (10.54)	0.676
Grip strength (kg)	30.41 (14.52)	28.26 (9.75)	0.260
Average daily MVPA (min/day)	27.31 (26.82)	21.25 (27.97)	0.014
Average daily LPA (min/day)	72.93 (42.00)	66.41 (33.34)	0.406
Average daily sedentary time (min/day)	594.50 (117.35)	577.27 (101.71)	0.303
Facility diversity (count of categories)	3.79 (2.33)	3.34 (2.36)	0.068
Facility counts (n)	5.44 (4.04)	4.50 (3.29)	0.080
Distance to nearest facility (m)	124.50 (96.05)	123.55 (81.82)	0.720
Population density	30748.61 (20646.91)	33722.41 (23811.10)	0.469
Building density	0.31 (0.10)	0.31 (0.09)	0.816
Street connectivity	15.38 (4.92)	16.14 (5.00)	0.277
Land-use mix (index)	12.94 (2.05)	13.31 (1.96)	0.133
Distance to nearest transit stop (m)	278.48 (130.92)	271.66 (106.85)	0.944
Distance to nearest recreation/entertainment (m)	252.40 (163.85)	254.32 (162.24)	0.969
Distance to nearest commercial place (m)	529.70 (309.20)	481.81 (268.25)	0.403
Male, n (%)	91 (45.7)	49 (44.5)	0.936
Education, n (%)			0.003
College or above (> 12 years)	26 (13.1)	17 (15.5)
High school/technical school	81 (40.7)	29 (26.4)
Junior high school	56 (28.1)	52 (47.3)
Primary school or below	36 (18.1)	12 (10.9)
Monthly income category, n (%)			0.981
<500 RMB	9 (4.5)	4 (3.6)
501–1000 RMB	15 (7.5)	7 (6.4)
1001–2000 RMB	21 (10.6)	13 (11.8)
2001–3000 RMB	122 (61.3)	69 (62.7)
>3000 RMB	32 (16.1)	17 (15.5)
Living alone, n (%)	77 (38.7)	41 (37.3)	0.901
Any chronic disease, n (%)	137 (69.2)	69 (62.7)	0.304
Community type, n (%)			0.657
Marginal	46 (23.1)	23 (20.9)
Traditional	33 (16.6)	14 (12.7)
Work-unit	66 (33.2)	37 (33.6)
Commercial housing	54 (27.1)	36 (32.7)

Values are presented as mean (SD) unless otherwise indicated. MCI was defined using education-stratified MoCA cut-offs in the primary analysis

P values were derived from independent-samples t tests, chi-square tests, or appropriate non-parametric tests, as applicable. Because education was incorporated into the primary outcome definition, the between-group comparison for education is descriptive and should not be interpreted as an independent association

Compared with the non-MCI group, participants classified as having MCI had lower accelerometer-derived average daily MVPA (21.25 ± 27.97 vs. 27.31 ± 26.82 min/day, p = 0.014) and were slightly shorter on average (160.52 ± 7.55 vs. 162.34 ± 7.98 cm, p = 0.048). No statistically significant between-group differences were observed for most other anthropometric, body-composition, perceived environmental, or GIS-derived built-environment indicators. Facility diversity and facility counts tended to be lower in the MCI group, although these differences did not reach statistical significance (p = 0.068 and p = 0.080, respectively).

As expected, the distribution of education differed between groups (p = 0.003); however, this difference should be interpreted cautiously because educational attainment was incorporated into the primary outcome definition through education-stratified MoCA thresholds rather than treated as an independent predictor in the main models. Baseline characteristics of the training and held-out test sets are additionally shown in Supplementary Table S2.

Machine learning model performance in the independent test set

Using the analytic sample with available MoCA total scores (n = 309), 110 participants were classified as having MCI and 199 as non-MCI based on the education-stratified MoCA cut-offs. The dataset was randomly divided into a stratified training set (70%, n = 216; MCI n = 77) and a held-out test set (30%, n = 93; MCI n = 33). Model performance was evaluated exclusively in the held-out test set.

In the independent test set, the gradient-boosting models showed the strongest overall discrimination. LightGBM achieved the highest AUC (0.854, 95% CI 0.762–0.946) and the highest accuracy (0.822, 95% CI 0.701–0.913), followed by XGBoost (AUC 0.835, 95% CI 0.741–0.929; accuracy 0.801, 95% CI 0.678–0.902) and random forest (AUC 0.810, 95% CI 0.712–0.908; accuracy 0.780, 95% CI 0.653–0.881) (Table 2). Logistic regression and SVM showed moderate discrimination, with AUCs of 0.806 and 0.785, respectively, whereas KNN performed least well (AUC 0.612, 95% CI 0.501–0.723).

Table 2.

Predictive performance of six machine-learning models for mild cognitive impairment in the independent held-out test set

Rank	Model	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	Brier score (95% CI)	Calibration intercept (95% CI)	Calibration slope (95% CI)
1	LightGBM	0.854 (0.762–0.946)	0.822 (0.701–0.913)	0.917 (0.820–0.995)	0.768 (0.641–0.896)	0.139 (0.092–0.186)	-0.038 (-0.152–0.076)	1.042 (0.875–1.209)
2	XGBoost	0.835 (0.741–0.929)	0.801 (0.678–0.902)	0.917 (0.820–0.995)	0.737 (0.605–0.869)	0.148 (0.099–0.197)	-0.046 (-0.165–0.073)	1.069 (0.898–1.240)
3	Random forest	0.810 (0.712–0.908)	0.780 (0.653–0.881)	0.865 (0.754–0.976)	0.737 (0.605–0.869)	0.161 (0.110–0.212)	-0.027 (-0.148–0.094)	1.011 (0.842–1.180)
4	Logistic regression	0.806 (0.708–0.904)	0.776 (0.649–0.877)	0.811 (0.690–0.932)	0.710 (0.576–0.844)	0.187 (0.133–0.241)	0.022 (-0.098–0.142)	0.901 (0.745–1.057)
5	SVM (RBF)	0.785 (0.683–0.887)	0.758 (0.629–0.864)	0.811 (0.690–0.932)	0.737 (0.605–0.869)	0.176 (0.124–0.228)	0.016 (-0.106–0.138)	0.923 (0.762–1.084)
6	KNN	0.612 (0.501–0.723)	0.632 (0.502–0.754)	0.667 (0.524–0.810)	0.603 (0.462–0.744)	0.231 (0.172–0.290)	0.076 (-0.052–0.204)	0.812 (0.651–0.973)

All reported 95% confidence intervals (CIs) were estimated using 2,000 bootstrap resamples of the held-out test set. Sensitivity, specificity, and accuracy were calculated using a single pre-specified classification threshold determined within the training set only (e.g., by maximising Youden’s J) and then applied unchanged to the held-out test set. Calibration intercepts and slopes were estimated by fitting logistic calibration models in the held-out test set, with observed outcomes regressed on the logit of predicted probabilities

Abbreviations: AUC Area under the receiver operating characteristic curve, KNN K-nearest neighbours, SVM Support vector machine, LightGBM Light Gradient Boosting Machine

At the pre-specified operating threshold derived from the training set and applied unchanged to the held-out test set, LightGBM and XGBoost both achieved the highest sensitivity (0.917, 95% CI 0.820–0.995). LightGBM also showed the highest specificity (0.768, 95% CI 0.641–0.896), while XGBoost and random forest both achieved a specificity of 0.737 (95% CI 0.605–0.869). Logistic regression and SVM yielded similar sensitivity (both 0.811, 95% CIs 0.690–0.932), whereas KNN showed the lowest sensitivity and specificity (0.667 and 0.603, respectively).

Calibration metrics also favoured the gradient-boosting models. LightGBM had the lowest Brier score (0.139, 95% CI 0.092–0.186), followed by XGBoost (0.148, 95% CI 0.099–0.197) and random forest (0.161, 95% CI 0.110–0.212). For the two best-performing models, the calibration intercepts were − 0.038 (95% CI -0.152 to 0.076) for LightGBM and − 0.046 (95% CI -0.165 to 0.073) for XGBoost, while the corresponding calibration slopes were 1.042 (95% CI 0.875–1.209) and 1.069 (95% CI 0.898–1.240), respectively. These results indicate that LightGBM and XGBoost combined relatively strong discrimination with more favourable calibration than the other candidate models. Full discrimination, classification, and calibration results are presented in Table 2.

LightGBM and XGBoost showed the strongest discrimination (AUC 0.854 and 0.835, respectively), whereas KNN demonstrated the lowest discrimination (AUC 0.612), consistent with Table 2 (Fig. 2).

Fig. 2.

Receiver operating characteristic (ROC) curves for six machine learning models in the held-out test set

Clinical utility and calibration in the held-out test set

Decision curve analysis (DCA) was used to assess the potential decision-analytic utility of the six models in the independent held-out test set (Fig. 3A). Across low-to-intermediate threshold probabilities, the better-performing models—particularly LightGBM, XGBoost, and random forest—generally showed higher net benefit than the default treat-all and treat-none strategies. However, differences between models narrowed as the threshold probability increased, and the net-benefit curves of the weaker-performing models approached or fell below the reference strategies over part of the threshold range. These findings suggest that the gradient-boosting models may offer greater utility for prioritising individuals for further cognitive assessment, although DCA results at higher threshold probabilities should be interpreted cautiously.

Fig. 3.

Clinical utility and calibration of six machine learning models in the independent test set. A Decision curve analysis showing net benefit across threshold probabilities relative to the default treat-all and treat-none strategies. B Calibration plots comparing mean predicted probabilities with observed event rates in the held-out test set; the diagonal dashed line indicates ideal calibration

Calibration was further evaluated in the held-out test set using both numerical metrics and calibration plots (Fig. 3B). The gradient-boosting models showed more favourable calibration than the other candidate models. LightGBM had the lowest Brier score (0.139, 95% CI 0.092–0.186), followed by XGBoost (0.148, 95% CI 0.099–0.197) and random forest (0.161, 95% CI 0.110–0.212). For LightGBM, the calibration intercept was − 0.038 (95% CI -0.152 to 0.076) and the calibration slope was 1.042 (95% CI 0.875–1.209); the corresponding values for XGBoost were − 0.046 (95% CI -0.165 to 0.073) and 1.069 (95% CI 0.898–1.240). By contrast, KNN showed poorer probabilistic performance, with a higher Brier score (0.231, 95% CI 0.172–0.290) and less favourable calibration slope (0.812, 95% CI 0.651–0.973). The calibration plots were consistent with these numerical results, although some variability was observed across probability ranges, likely reflecting the modest size of the held-out test set. Overall, these results suggest that LightGBM and XGBoost combined stronger discrimination with more favourable calibration and greater potential clinical utility than the other candidate models in this sample.

Feature selection and model interpretability

Candidate predictors were pre-specified across sociodemographic, anthropometric/body-composition, accelerometry-derived activity, and built-environment domains. Because the primary outcome was defined using education-stratified MoCA cut-offs, education-related variables were excluded from the main predictor set to reduce potential circularity between educational attainment and outcome definition.

In the training set, LASSO-penalised logistic regression selected a parsimonious subset of predictors spanning several domains (Fig. 4A). These included movement-related variables, such as MVPA, sedentary time, and LPA; anthropometric or body-composition indicators, such as waist-to-hip ratio, muscle-related measures, and grip strength; and neighbourhood-context variables, including land-use mix, road-related indicators, and distance-to-transit measures. This pattern suggested that no single domain alone accounted for the predictive information retained during feature selection.

Fig. 4.

Predictor selection and model interpretability with education-related variables excluded. A Direction of LASSO coefficients for predictors retained in the training set, shown as positive versus negative standardised coefficients. B Global feature importance for the final LightGBM model in the independent held-out test set, summarised by mean absolute SHAP values (mean |SHAP|). Note: Education-related variables were excluded from the main predictor set because education-stratified MoCA cut-offs were used to define MCI status

Model interpretability was further examined for the final LightGBM model using SHAP values in the independent held-out test set (Fig. 4B). Mean MVPA showed the greatest global contribution to model predictions, followed by waist circumference, mean sedentary time, and age. Other influential variables included LPA, waist-to-hip ratio, muscle-related indicators, and several community-level measures, such as perceived neighbourhood safety and population density. Overall, the SHAP results were broadly consistent with the LASSO feature-selection stage, indicating that movement behaviours and central adiposity/body-composition measures contributed prominently to model predictions, while neighbourhood-context variables provided additional but smaller contributions.

To further describe model behaviour, partial dependence plots (PDPs) were examined for the top-ranked predictors from the final LightGBM model. The PDPs were broadly consistent with the SHAP results, in that movement-behaviour variables and adiposity-related measures showed the clearest contributions to model output. In particular, higher MVPA tended to correspond to a lower predicted probability of MCI, whereas greater waist circumference and longer sedentary time tended to correspond to a higher predicted probability. For several other predictors, the observed patterns were nonlinear within the fitted model. Because PDPs are model-based summaries and may be sensitive to predictor distributions and correlations, these patterns should be interpreted descriptively rather than causally. The full PDPs are therefore provided in Supplementary Figure S1.

Discussion

The present study examined whether multidomain indicators could improve the identification of mild cognitive impairment in community-dwelling older adults in urban China. Using education-stratified MoCA-defined MCI as the primary outcome, we found that models incorporating accelerometer-derived movement behaviours, body-composition measures, and neighbourhood-context variables showed measurable predictive value, with the gradient-boosting models performing better than the other candidate algorithms in the independent held-out test set. In particular, LightGBM achieved the highest discrimination and accuracy and showed more favourable calibration, while decision curve analysis indicated greater potential utility for prioritising individuals for further cognitive assessment across clinically relevant threshold probabilities. These findings suggest that combining objective behavioural data with biological and environmental information may strengthen community-based cognitive-risk stratification beyond more conventional single-domain approaches. Nevertheless, the current results should be interpreted cautiously, as they reflect internal validation in a cross-sectional dataset and are intended to inform screening support rather than diagnostic classification or causal interpretation.

A key methodological issue in MoCA-based prediction is the potential circularity between outcome definition and educational attainment. Because the primary MCI label in this study was derived using education-stratified MoCA cut-offs, education-related variables were excluded from the candidate predictor set in the primary modelling pipeline to reduce information leakage and avoid artificially inflating model performance or feature importance. At the same time, education remains a well-established correlate of late-life cognitive health and dementia risk, consistent with the cognitive reserve framework and broader population-level prevention evidence [5, 6]. We therefore treated education as part of the outcome construction in the primary analysis rather than as a predictor to be rediscovered by the model, and reserved education-related effects for sensitivity analyses using alternative MoCA definitions that did not embed educational information in the outcome itself. This design choice improves the interpretability of the remaining predictors, such as movement behaviours and adiposity-related measures, and helps ensure that the reported model contributions more closely reflect the information learned from the non-educational predictors.

From a modelling perspective, the relatively stronger performance of LightGBM and XGBoost is plausible, as boosting-based ensembles can capture nonlinear associations and higher-order interactions among behavioural, physiological, and environmental features with limited manual feature engineering [23]. In the present study, however, performance differences among the top-performing models were modest and their confidence intervals overlapped, indicating that the apparent advantage of gradient boosting should be interpreted as numerical rather than definitive. Beyond discrimination, calibration is particularly relevant in screening-oriented prediction because model outputs may be used to estimate risk probabilities rather than only assign binary labels; reliable probability estimates are important when downstream actions depend on thresholding [24]. In our study, the calibration metrics and reliability diagrams suggested that LightGBM and XGBoost showed more favourable calibration than the other candidate models, although some variability across probability ranges remained evident in the test sample. External validation is therefore required before these predicted probabilities can be considered for real-world decision support.

Model interpretability analyses converged on a coherent set of predictors, led by movement-behaviour profiles and central adiposity/body-composition indicators. Both the SHAP summaries and the partial dependence plots suggested that higher MVPA tended to correspond to a lower predicted probability of MCI, whereas longer sedentary time and greater waist circumference tended to correspond to a higher predicted risk. These patterns are broadly consistent with meta-analytic evidence linking sedentary behaviour to cognitive decline or MCI risk in older adults, as well as observational studies associating accelerometer-measured activity profiles with cognitive function [25]. Mechanistically, higher-intensity activity may support brain health through vascular, metabolic, and neurotrophic pathways, whereas prolonged sedentary time may contribute to cardiometabolic dysregulation and inflammatory processes that adversely affect cognition [26]. Notably, LPA also appeared among the selected predictors, which is consistent with emerging evidence that light-intensity activity may still be relevant to cognitive function, particularly in older adults for whom sustained MVPA may be difficult to achieve. Taken together, these findings highlight movement behaviour as a potentially modifiable domain relevant to cognitive health promotion, while recognising that model-based associations from cross-sectional data should be interpreted descriptively rather than causally.

Adiposity and body-composition indicators also contributed meaningfully to prediction. Central adiposity markers such as waist circumference and WHR tended to correspond to higher predicted MCI risk in the model-interpretation analyses, which is broadly consistent with reports linking central obesity and related cardiometabolic burden to cognitive impairment in older populations [27]. At the same time, the literature remains heterogeneous, and some studies in older adults have reported obesity-paradox-like patterns depending on age, follow-up duration, and confounding by illness-related weight loss [28]. Our findings should therefore be interpreted as indicating that central adiposity and body composition carried predictive information in this sample rather than as establishing a universal directional effect. Functional indicators such as handgrip strength and muscle-related measures were also retained, consistent with previous work linking lower muscle strength and sarcopenia-related phenotypes to cognitive impairment risk and adverse ageing trajectories [7]. The co-occurrence of movement behaviour, adiposity, and functional measures in the final modelling pipeline supports the view that MCI risk in community settings may reflect a multidomain profile rather than a single isolated behavioural or clinical factor.

A further strength of this study is the inclusion of both GIS-derived and perceived built-environment measures alongside objectively measured behaviours. Environmental predictors, including neighbourhood safety, population density, road/network indicators, and transit accessibility, contributed additional predictive information beyond individual-level movement and anthropometry. This is consistent with a growing literature linking neighbourhood environments to cognitive outcomes through pathways involving physical activity opportunities, social participation, access to services, and chronic stress exposure [7, 8]. Observational evidence also suggests that proximity to activity-supportive destinations and recreational resources may be associated with lower risk of cognitive impairment [7]. In the model-interpretation analyses, several environmental variables showed nonlinear patterns, which is plausible given that contextual influences may depend on threshold levels of access as well as individual mobility constraints. However, these findings should be interpreted cautiously. The present study does not establish that modifying neighbourhood features would necessarily alter cognitive trajectories, and larger longitudinal studies will be required to clarify these pathways more directly.

Several limitations should be acknowledged. First, the cross-sectional design precludes causal inference and does not address cognitive change over time; some observed predictor–risk patterns may reflect reverse causation, such as early cognitive impairment influencing activity behaviour or mobility. Second, although the dataset was uniquely valuable in combining cognitive assessment, accelerometry, body composition, and neighbourhood measures within the same community sample, the analytic sample was restricted to participants with available MoCA total scores from a single urban setting, which may introduce selection bias and limit generalisability. In addition, the data were collected between 2017 and 2018, which may limit the temporal transportability of the models to present-day community settings. Since that period, older adults’ movement patterns, access to neighbourhood destinations, and the surrounding urban environment may have changed, meaning that the distribution and relative importance of some predictors could differ in more recent cohorts. Accordingly, the present models should be interpreted primarily as internally validated proof-of-concept tools within the study period, rather than as directly deployable current screening models without contemporary external validation and recalibration. Third, several potentially important determinants, including diet, detailed comorbidity severity, medication use, and genetic risk factors such as APOE, were unavailable and may improve prediction if incorporated in future models. Fourth, although we reduced circularity by excluding education from the primary predictor set, MoCA remains a screening instrument rather than a clinical diagnosis, and outcome misclassification is therefore possible. Finally, our internal evaluation relied on a single held-out test set; external validation in independent cohorts is essential before considering implementation in community screening programmes. Despite these limitations, the present study demonstrates the feasibility of integrating wearable-derived movement behaviours, body-composition indicators, and built-environment measures within an interpretable machine-learning framework, and provides a basis for future longitudinal validation and implementation research.

Conclusion

In community-dwelling older adults in urban China, multidomain predictors spanning accelerometer-derived movement behaviours, body-composition indicators, and built-environment measures showed value for identifying education-stratified MoCA-defined mild cognitive impairment. Among the candidate models, LightGBM showed the best overall performance in the independent held-out test set, with stronger discrimination, more favourable calibration, and greater potential decision-analytic utility than the other classifiers. These findings suggest that integrating objective behavioural, biological, and environmental information may support community-based cognitive-risk stratification beyond more conventional single-domain approaches. However, the present results are intended to inform screening support rather than diagnosis, and external validation in independent and longitudinal cohorts remains necessary before any routine implementation.

Authors' contributions

ZF. and ZFK wrote the main manuscript text and WZJ prepared Figs. 1, 2, 3, 4, 5, 6 and 7. All authors reviewed the manuscript.

Funding

This work was supported by the National Social Science Foundation of China (No. 23CTY018). Jiangsu University “QING LAN Project” Funding Program. Jiangsu Provincial Social Science Funding Program (22GLD016); Soft Science Project of National Forestry and Grassland Administration (2022131017). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

The datasets generated and/or analysed during the current study are not publicly available due to participant privacy and confidentiality requirements but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of Nanjing Normal University (Approval ID: NNU202003005). Written informed consent was obtained from all participants prior to enrollment in the study. All procedures involving human participants were conducted in accordance with the Declaration of Helsinki. The study received ethical approval from the Ethics Committee of Nanjing Normal University.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.