This study utilized a convenience sampling approach to recruit participants from two phases: (a) 500 older adults were enrolled from four nursing homes in Hangzhou, Zhejiang Province, between December 2024 and March 2025; (b) An additional 112 older adults were recruited from another nursing home in the same region during March to April 2025. Participants were included if the older adults were age 60 years or older, had a minimum documented residence of 3 months within the nursing home facility, maintained preserved cognitive and communication function sufficient for study procedures, and provided documented informed consent. Individuals were excluded if they had: a formal diagnosis of Alzheimer’s disease or any other form of dementia; a history of intellectual disabilities or a history of psychiatric disorders; significant communication impairments that would compromise data collection reliability; or concurrent active participation in other interventional clinical trials.

All questionnaires were collected by a single researcher during consistent time periods to ensure assessment reliability. This study adhered to the principles outlined in the Declaration of Helsinki. All participants provided written informed consent. The study protocol was approved by the Ethics Committee of Zhejiang Chinese Medical University (No. 20241129–3).

The minimum sample size for the model was calculated based on the Events per Variable rule (14). This rule requires a sample size of at least 10 times the number of independent variables. With an estimated 10 variables to be included in this study and a reported prevalence of CF among older adults in Chinese nursing homes of approximately 24% (6), and accounting for 20% potential attrition, the required sample size was calculated as follows: 10 variables * 10 * 24% * (1 + 20%) ≈ 500 participants (15). The final sample size satisfied the empirical rule of having at least 10 events per candidate predictor variable (16).

Diagnosis of CF required meeting all of the following criteria based on established assessment standards and mild cognitive impairment guidelines (1): (1) Fried Phenotype score 3–5 (17); (2) education-adjusted Mini-Mental State Examination (MMSE) scores (18–20 for illiterate, 21–24 for primary education, 25–27 for secondary education or higher) (18); and (3) dementia was excluded using the Clinical Dementia Rating (CDR) scale, with a diagnosis of CF requiring a CDR score of 0.5 (19).

Based on a review of domestic and international literature and expert consensus (20–22), we included 19 predictors spanning four key domains: sociodemographic factors, health status indicators, physical function measures, and lifestyle factors.

Sociodemographic factors encompassed the following variables: age, gender (male/female), marital status (widowed, spouse alive, or never married), education level (illiterate, primary education, secondary education, or college/university and above), and post-retirement occupation (farmer, laborer, intellectual, or other). Health status indicators included: self-rated health (very poor, poor, fair, good, or very good), chronic pain (yes/no), history of falls (yes/no), depression (yes/no), and nutritional status (normal, at risk of malnutrition, or malnourished). Physical function measures comprised: ADL (normal, declined, or severely impaired), grip strength (reduced/normal), gait speed (slow/normal), body mass index – BMI (<19 kg/m², 19-21 kg/m², 21-23 kg/m², or ≥23 kg/m²), and sleep duration (<6 h, 6–9 h, or >9 h). Lifestyle factors were defined as: exercise frequency (0, 1–2, or ≥3 times/week), intellectual activities frequency (0, 1–2, or ≥3 times/week), smoking history (never smoker, former smoker, or current smoker), and drinking history (never drinker, former drinker, or current drinker).

Intellectual activities encompassed cognitively stimulating pursuits such as internet use, newspaper reading, calligraphy, painting, musical instrument playing, chess, and mahjong (20). Chronic pain is characterized by persistent nociception beyond the expected duration of tissue healing, typically manifested as pain lasting at least 3 months (23). BMI is calculated by dividing an individual’s weight in kilograms by the square of their height in meters (kg/m²). Grip strength was measured twice for each hand using a digital dynamometer, with the highest value from the four measurements used for analysis. Gait speed was assessed by measuring the time taken to walk 6 meters at a habitual pace. Nutritional status was assessed using the Mini-Nutritional Assessment Short-Form (24). Depressive symptoms were assessed using the 5-item Geriatric Depression Scale (25). Functional status was assessed using the ADL scale (26).

During data preprocessing, label encoding was applied to all categorical variables. To ensure consistent representation of CF across datasets, the modeling cohort (n = 500) was randomly split into training (70%, n = 350) and testing sets (30%, n = 150) using stratified sampling based on CF status. This resulted in similar minority proportions: (91/350) in the training set and (41/150) in the test set. The training set was used for model development, while the test set was used for both hyperparameter tuning and performance evaluation. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) algorithm, with the identified predictors subsequently incorporated into the predictive model. Feature selection reduces dimensionality to enhance model generalizability while mitigating overfitting risks.

K-nearest neighbors, a non-parametric learning model, excels at capturing local patterns and is well-suited for modeling non-linear relationships in smaller sample sizes. SVM addresses complex non-linear classification problems through kernel functions. LR serves as an interpretable linear baseline for validating linear relationships. RF and XGBoost, as ensemble tree models, effectively handle high-dimensional feature interactions and collinearity, with XGBoost offering further optimized computational efficiency. Moreover, given the limited sample size of the dataset, tree-based models (RF, XGBoost) demonstrate superior resistance to overfitting compared to deep learning models. Simultaneously, as the dataset contains categorical features, tree models inherently support the processing of discrete values. Therefore, this study employs five ML algorithms—KNN, SVM, LR, RF, and XGBoost—to construct the prediction models. To mitigate overfitting, we optimized the hyperparameters using 10-fold cross-validation on the training set and evaluated the model’s generalizability with an independent test set. Additionally, 112 older adults from other nursing homes were enrolled as an external validation cohort. Finally, SHAP analysis was applied to enhance the interpretability of the optimal model.

The ROC curve assessments were performed for all five models. To evaluate classification performance, accuracy, precision, recall, Brier score, and the F1 score were calculated for all models across training and validation datasets. Additionally, calibration curves were generated to assess the agreement between the predicted probabilities of a model and the actual observed probabilities. Finally, we used decision curve analysis to assess the clinical applicability of the models.

This study used R 4.4.3 and Python 3.13 for statistical analysis and predictive modeling. A comprehensive two-way analysis was performed for all variable types using a generalized linear model: continuous variables were analyzed with two-way analysis of variance; binary categorical variables with binary logistic regression; ordinal categorical variables with ordinal regression; and nominal polytomous variables with multinomial logistic regression. The significance level for all hypothesis tests was set at α = 0.05. A p-value < 0.05 was considered statistically significant.

This study enrolled 500 older adults from nursing homes, among whom 132 (26.4%) developed CF. To validate the prediction model, we recruited 112 older adults from another nursing home. CF was identified in 30 participants (26.8%). Table 1 presents the baseline characteristics of participants with and without CF. The results showed that ADL, intellectual activities frequency, age, depression, exercise frequency, education level, nutritional status, marital status, self-rated health, history of falls, BMI, post-retirement occupation, drinking history, gait speed, and chronic pain were statistically significant (p < 0.05).

Most disparities between CF and non-CF groups were consistent across modeling and validation sets. Notable interactions (p < 0.01) occurred in education level, self-rated health, sleep duration, and gait speed, suggesting cohort-specific effects in these domains.

This study included a total of 19 predictors. We used LASSO regression 10-fold cross-validation to select significant predictors of CF of the older adult in nursing homes (Figure 1). Ten feature variables were selected for the model, including ADL, frequency of intellectual activities, age, depression status, physical activity frequency, education level, nutritional status, marital status, self-rated health status, and history of fall.

We constructed five different ML models, including SVM, RF, KNN, LR, and XGBoost, and evaluated their performance to predict CF. Table 2 details the performance metrics of all five models. KNN exhibited substantial degradation from train to test in accuracy (0.916 → 0.885), precision (0.905 → 0.817), F1-score (0.907 → 0.881), and AUC (0.976 → 0.945), despite recall improvement (0.909 → 0.955) (Figures 2A,B). XGBoost showed similar declines in accuracy (0.912 → 0.895), precision (0.952 → 0.895), F1-score (0.896 → 0.880), and AUC (0.970 → 0.964), though recall increased (0.847 → 0.865). Among test set performances, XGBoost achieved the highest AUC (0.964) and best Brier score (0.075), while KNN had the highest recall (0.955) but lowest precision (0.817). There may be overfitting in KNN and XGBoost. RF demonstrated balanced metrics (accuracy: 0.865, precision: 0.852, recall: 0.843, F1: 0.847, AUC: 0.951, Brier: 0.102). RF maintained minimal AUC reduction (0.953 → 0.951) with consistent performance across all metrics. LR showed exceptional stability with near-identical accuracy (0.863 → 0.860) and F1-score (0.841 → 0.841), plus modest changes in other metrics (precision: 0.876 → 0.851, recall: 0.809 → 0.831, AUC: 0.940 → 0.932, Brier: 0.096 → 0.107). SVM exhibited controlled declines in accuracy (0.876 → 0.860), precision (0.917 → 0.859), and F1-score (0.851 → 0.839), with AUC (0.938 → 0.932) and recall (0.794 → 0.820) remaining comparable. Models without significant overfitting included RF, LR, and SVM. The results of the confusion matrix for the test set are shown in Figure 3A.

Model calibration was assessed using calibration curves (Figures 2C,D) and the Brier score (Table 2). Both RF and XGBoost demonstrate strong alignment with the ideal calibration line across most probability ranges, consistent with their relatively low test-set Brier scores (RF: 0.102; XGBoost: 0.075). SVM maintains close proximity to the ideal curve in low-to-mid probabilities but exhibits slight deviations in high-probability regions, aligning with its moderate Brier score (0.105). KNN shows significant miscalibration in low-probability zones despite achieving a competitive Brier score (0.092), suggesting that this global metric may partially mask localized inaccuracies. LR displays systematic deviations in low-to-medium probabilities and the highest Brier score (0.107), indicating pronounced calibration challenges.

The decision curve analysis for both training and testing datasets revealed distinct performance characteristics among the evaluated ML models (Figures 2E,F). The XGBoost model demonstrated superior net benefit across most threshold probabilities in both datasets, although its performance was slightly diminished in the testing set, suggesting potential overfitting. Conversely, the KNN model exhibited robust performance on the training set but a significant decline on the testing set, indicative of overfitting. The LR and SVM models showed moderate and consistent net benefits across both datasets. Notably, the RF model, which underperformed on the training set, displayed improved performance on the testing set.

In the external validation dataset (Table 3), the XGBoost model achieved the highest AUC (0.785), but exhibited substantial declines in accuracy (0.462), precision (0.454), F1 score (0.624), and probability calibration (Brier: 0.346). In contrast, SVM maintained the highest stability across metrics (accuracy: 0.735; precision: 0.731; F1: 0.685) with the best calibration performance (Brier: 0.214). The remaining models (LR, RF, KNN) demonstrated comparatively weaker overall performance and suboptimal calibration (Brier: 0.226–0.261). Collectively, while XGBoost showed suboptimal performance in external validation, SVM displayed balanced metric outcomes and superior probability calibration. The results of the confusion matrix for the external validation set are shown in Figure 3B.

Decision curve analysis was further performed to evaluate clinical utility (Figure 2G). While XGBoost achieved the highest AUC (0.785), its precision collapse (0.454) translates to a negative net benefit in decision curve analysis beyond 30% risk thresholds. This means deploying XGBoost unmodified would cause net clinical harm – misallocating resources to false positives. Conversely, SVM’s balanced precision (0.731) and recall (0.644) sustain positive net benefit across screening-relevant thresholds (10–60%), affirming its role as the safest implementation choice.

The SHAP analysis of the SVM model on both the test set and the external validation set revealed consistent directional influences of key features on the predicted risk of CF (Figure 4A). Features such as ADL, intellectual activities, age, exercise, depression, and nutrition demonstrated a stable impact direction across both datasets. The external validation set SHAP plots (Figure 4B) had minor variations in order, such as depression and exercise, suggesting a consistent directional effect across datasets, but with slight variations in relative significance. Feature importance ranking of the test set (Figure 4C) confirmed ADL and intellectual activities as the most critical predictors within the model’s logic, followed by age, depression, exercise, education, nutrition, marital status, self-rated health, and history of falls. Critically, this consistency in SHAP values indicates that the model relies on a similar set of predictors and interprets their impact directionally in the same way when making predictions on both cohorts. It does not, however, imply that the underlying distributions of these features are similar between cohorts (Table 1).

This study has several limitations. First, the cross-sectional design precludes causal inferences between CF and its associated factors. Secondly, due to the significant differences in the key predictive factors of cognitive decline among different cohorts, the prediction performance of the model on the external validation set has significantly declined. Third, not all potential predictors of CF, such as biochemical markers, were included in the analysis. Fourth, label encoding may artificially establish an ordinal relationship on nominal categorical variables. Future research will develop models incorporating more comprehensive predictive features derived from longitudinal cohorts. Additionally, one-hot encoding should be applied to purely nominal features. To enhance generalizability, adaptive techniques—such as model recalibration and domain adaptation—must be prioritized for population alignment, alongside implementing a rigorous multi-site validation framework to quantify real-world performance.