We conducted a cross-sectional secondary analysis using data from the Gan Dau Healthy Longevity Plan in Taiwan to examine the associations of depressive symptoms with nutritional risk and lower-extremity physical performance among community-dwelling middle- to older-aged adults. A total of 1010 participants aged 50 years or older were recruited through convenience sampling from 1 hospital and 3 community centers in northern Taiwan between January 2024 and September 2024. Participants were eligible if they were able to communicate in Mandarin or Taiwanese and were able to provide informed consent.

All assessments were completed by trained personnel after written informed consent was obtained. Standardized assessment procedures were used to reduce measurement bias, and all analytic data were fully deidentified to protect participant confidentiality. Sample size was determined by the number of eligible participants in the parent study, and no a priori power calculation was performed. Because participants were recruited through convenience sampling from 1 hospital and 3 community centers, selection bias may have occurred, and the enrolled participants may have been healthier or more motivated than nonparticipants.

Depressive symptoms were assessed using the 15-item Geriatric Depression Scale–Short Form (GDS-15). Nutritional status was assessed using the Mini Nutritional Assessment–Short Form (MNA-SF). Lower-extremity physical performance was assessed using the Short Physical Performance Battery (SPPB), gait speed, and the timed up-and-go (TUG) test. In this analysis, depressive symptom severity was treated as the main explanatory variable, whereas nutritional risk and lower-extremity physical performance were treated as study outcomes. This study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines.

This study was approved by the Institutional Review Board of Taipei Veterans General Hospital, Taiwan (number 2023-10-001B#1). Written informed consent was obtained from all participants or legally authorized representatives, when necessary, in accordance with the Declaration of Helsinki.

Participant privacy was protected by collecting data anonymously and analyzing the data in aggregate form. No personally identifiable information was reported. The data were stored securely and were accessible only to the research team. Participants did not receive any compensation for participating in this study.

Covariates included demographics (age, sex, marital status, education, living arrangement, economic status), health behaviors (regular exercise, leisure activity, fall history), anthropometrics (BMI, abdominal and calf circumferences), and multimorbidity. Multimorbidity was measured using the Cumulative Illness Rating Scale for Geriatrics (CIRS-G), developed by Linn et al [40] and adapted for older adult populations by Miller et al [41] with a 14-system scoring manual [42]. The CIRS-G total score reflects overall comorbidity burden, and the Severity Index (CIRS-G SI) represents the average severity across organ systems [42,43].

Categorical variables were summarized as n (%), and continuous variables were summarized as median and IQR. Group differences by depression status (GDS-15 score <5 vs ≥5) were examined using the Mann-Whitney U test for continuous variables and the χ² test or Fisher exact test for categorical variables, as appropriate, with a 2-sided α of .05. Nutritional status was dichotomized as well-nourished (MNA-SF≥12) versus at risk or malnourished (MNA-SF≤11). Impairment thresholds for lower-extremity performance followed established criteria: SPPB≤9, gait speed≤1.0 m/s (6-m walk), and TUG≥12 seconds [38,39]. These binary outcomes were used in all analytic models, including CHAID, logistic regression, and least absolute shrinkage and selection operator (LASSO) logistic regression. Because missingness was <5% for all variables, complete-case analysis was applied; given the low proportion of missing data, the potential impact on the results was considered limited.

The primary exploratory analytic approach used CHAID to identify data-driven thresholds and nonlinear interaction patterns related to impaired nutritional status and lower-extremity physical performance [44,45]. Outcomes were analyzed as binary variables based on the aforementioned clinical cutoffs. Candidate predictors included GDS-15 score, sociodemographic characteristics, lifestyle behaviors, anthropometric measures, and multimorbidity indicators. Predictors were retained in their original scales, and CHAID generated multiway splits using Bonferroni-adjusted χ² tests (α=.05). Tree growth was constrained to a maximum depth of 3 levels, with minimum parent and child node sizes of 100 and 50, respectively. For each terminal node, we summarized sample size, class distribution, and misclassification risk. Model evaluation was based on an internal 80% training set and a 20% testing set [46]. Classification accuracy was calculated in both datasets, and the applicability of the training set tree structure to the test set was examined descriptively. Because no cross-validation, bootstrapping, nor external validation was performed, the reported performance metrics should be interpreted cautiously as preliminary, sample-specific estimates rather than robust indicators of model performance or generalizability.

For comparison with CHAID, we fitted multivariable logistic regression models as conventional regression approaches and LASSO logistic regression models as a penalized alternative to reduce overfitting and improve variable selection. Logistic and LASSO models were used to classify outcome status within the study sample rather than to support causal inference; therefore, the results were interpreted as reflecting association patterns and classification performance rather than causal effects. The same set of candidate predictors used in the CHAID analysis was entered into the logistic and LASSO models to facilitate comparability across analytic approaches. For LASSO logistic regression, the penalty parameter was selected using internal cross-validated deviance within the training set. Classification performance was evaluated separately for each dichotomized outcome—nutritional risk (MNA-SF), impaired SPPB, slow gait speed, and impaired TUG performance—using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F₁ score in the test sets. All analyses were conducted using SPSS version 31 (IBM Corp) and R version 4.4.3. No formal subgroup, interaction, nor sensitivity analyses were conducted.

A total of 1010 community-dwelling adults were included in the study (Table 1). The median age was 68 (IQR 62.7‐73.0) years, and 757 of 1010 (75%) participants were female. Overall, 143 of 1010 (14.2%) participants screened positive for depressive symptoms, defined as a GDS-15 score ≥5. Because of variable-specific missing data, denominators differed slightly across some analyses.

Depression status was significantly associated with several sociodemographic characteristics. Educational attainment differed significantly between participants with and without depressive symptoms (χ²₂=9.42, P=.009). Specifically, a higher proportion of participants with depressive symptoms had a junior high school education or below (60/143, 42% vs 253/865, 29.2%), whereas a lower proportion had a college education or above (45/143, 31.5% vs 347/865, 40.1%). Marital status, living arrangement, religious belief, and economic status were also significantly associated with depression status. Compared with participants without depressive symptoms, those with depressive symptoms were more likely to be unmarried (61/143, 42.7% vs 251/865, 29%; χ²₁=10.68, P=.001), to live alone or in an institution (35/143, 24.5% vs 92/865, 10.6%; χ²₂=21.36, P<.001), to report having a religious belief (123/143, 86% vs 670/865, 77.5%; χ²₁=5.36, P=.02), and to report financial stress (87/143, 60.8% vs 393/865, 45.4%; χ²₁=11.68, P<.001). Employment status did not differ significantly between groups (χ²₁=3.51, P=.06).

Regarding lifestyle factors, participants with depressive symptoms were less likely to report regular exercise (92/143, 64.3% vs 741/865, 85.7%; χ²₁=38.91, P<.001) and leisure activities (77/143, 53.8% vs 557/865, 64.4%; χ²₁=5.85, P=.02) and were more likely to report a history of falls (57/143, 39.9% vs 216/865, 25%; χ²₁=14.14, P<.001).

For anthropometric measures, participants with depressive symptoms had a lower median BMI than those without depressive symptoms (23.2, IQR 20.9‐25.6 vs 24.1, IQR 21.9‐26.6 kg/m²; U=54,700.0, P=.03) and a smaller median calf circumference (34.0, IQR 32.0‐36.2 vs 35.0, IQR 33.0‐37.4 cm; U=50,160.5, P<.001), whereas abdominal circumference did not differ significantly between groups (90.0, IQR 83.0‐97.1 vs 90.0, IQR 84.0‐96.9 cm; U=62,351.5, P=.88). Clinical burden also differed significantly by depression status. Participants with depressive symptoms had higher median CIRS-G scores than those without depressive symptoms (9, IQR 6‐12 vs 5, IQR 2‐7; U=149,099.0, P<.001) and higher median CIRS-G SI scores (1.4, IQR 1.1‐1.6 vs 1.2, IQR 1.0‐1.4; U=181,284.0, P<.001).

As shown in Table 2, the median MNA-SF score was 13 (IQR 12‐14), and 159 of 1010 (15.8%) participants were classified as at risk of malnutrition or malnourished. Participants with depressive symptoms had poorer nutritional status than those without depressive symptoms, as indicated by lower median MNA-SF scores (12, IQR 11‐14 vs 14, IQR 12‐14; U=40,275.0, P<.001) and a higher proportion classified as at risk of malnutrition or malnourished (53/143, 37.1% vs 105/867, 12.1%; χ²₁=57.67, P<.001).

For lower-extremity physical performance, the median SPPB score was 11 (IQR 10‐12), and 172 of 1010 (17%) participants met the criteria for impairment. Compared with participants without depressive symptoms, those with depressive symptoms had lower median SPPB scores (10, IQR 9‐12 vs 12, IQR 10‐12; U=39,722.0, P<.001) and were more likely to be classified as impaired (53/143, 37.1% vs 118/867, 13.6%; χ²₁=50.48, P<.001). The median gait speed was 1.1 m/s, and 343 of 1010 (34%) participants were classified as impaired. Participants with depressive symptoms had slower median gait speed than those without depressive symptoms (0.9, IQR 0.8‐1.2 vs 1.1, IQR 0.9‐1.3 m/s; U=41,132.0, P<.001) and were more likely to have impaired gait speed (73/143, 51% vs 270/865, 31.2%; χ²₁=51.41, P<.001). The median TUG time was 8.2 seconds, and 80 of 1010 (7.9%) participants met the impairment threshold. Participants with depressive symptoms had longer median TUG times than those without depressive symptoms (9.3, IQR 7.8‐11.4 vs 8.1, IQR 7.0‐9.5 seconds; U=78,609.0, P<.001) and were more likely to be classified as impaired (25/143, 17.5% vs 55/864, 6.4%; χ²₁=22.02, P<.001). Overall, depressive symptoms were consistently associated with poorer nutritional status and lower-extremity physical performance.

The CHAID decision tree model for nutritional status based on the MNA-SF is presented in Figure 1. The MNA-SF ranges from 0 to 14, with higher scores indicating better nutritional status and scores ≤11 indicating risk of malnutrition or malnutrition. Nutritional status was modeled as a binary outcome classified as well-nourished versus at risk of malnutrition or malnourished.

Depressive symptom severity, measured using the GDS-15, was identified as the primary splitting variable in the decision tree (P<.001). The GDS-15 ranges from 0 to 15, with higher scores indicating greater depressive symptom severity and scores ≥5 indicating clinically relevant depressive symptoms. The model divided participants into 3 depressive symptom groups (≤1, 2‐3, and ≥4), corresponding to different levels of nutritional vulnerability within the study sample.

Within each depressive symptom stratum, BMI served as the main secondary splitting variable, indicating that anthropometric status further differentiated nutritional status patterns beyond depressive symptom severity alone. Additional hierarchical refinement was provided by calf circumference and age, which identified subgroups with distinct nutritional vulnerability profiles. Smaller calf circumference, a proxy indicator of reduced muscle mass, and older age were associated with a higher prevalence of nutritional risk within specific depressive symptoms and BMI strata.

Participants with GDS-15 scores ≤1 were predominantly well-nourished, with 91.3% (366/401) classified as having normal nutritional status. However, nutritional risk was higher among individuals with lower BMI and smaller calf circumference, with 27.8% (22/79) classified as at risk of malnutrition or malnourished within this otherwise low depressive symptom group. Nutritional vulnerability was substantially greater among participants with higher depressive symptom severity. In particular, among participants with GDS-15 scores ≥4 and BMI≤23.0 kg/m², 54.1% (45/85) were classified as at risk of malnutrition or malnourished, representing the subgroup with the highest prevalence of nutritional risk identified by the model.

Each terminal node of the decision tree represents a distinct subgroup defined by combinations of depressive symptoms, BMI, calf circumference, and age and displays the corresponding number and proportion of participants classified according to nutritional status. This hierarchical structure illustrates how depressive symptoms together with anthropometric and demographic factors differentiate subgroups with varying prevalences of nutritional risk among community-dwelling adults aged ≥50 years. In addition, the CHAID model had acceptable within-sample classification performance, with accuracies of 84.4% in the training dataset and 85.6% in the test dataset.

Across all models, the GDS-15 score consistently emerged as the primary splitting variable. Secondary predictors, including BMI, calf circumference, age, comorbidity severity, and education level, provided additional hierarchical refinement and helped differentiate subgroups with varying prevalences of nutritional risk and lower-extremity functional impairment. Overall, these findings suggest that CHAID may serve as a useful exploratory approach for characterizing within-sample patterns, although the performance estimates should be interpreted as preliminary and sample-specific.

Table 3 summarizes the classification performance of CHAID, logistic regression, and LASSO logistic regression models for nutritional status (MNA-SF) and lower-extremity physical performance outcomes, including SPPB, gait speed, and TUG.

For nutritional status assessed using MNA-SF, the CHAID model had acceptable discrimination, with an AUC of 0.80 (95% CI 0.72‐0.88), characterized by high specificity but lower sensitivity. Logistic regression analysis had higher overall discrimination, achieving an AUC of 0.85 (95% CI 0.79‐0.92), with increased sensitivity but reduced specificity compared with CHAID. The LASSO logistic regression model had the highest classification performance, with an AUC of 0.96 (95% CI 0.92‐0.99), together with the highest overall accuracy and F₁ score.

For lower-extremity physical performance outcomes, model performance varied by outcome measure. For SPPB impairment, all models had acceptable discrimination, with AUC values of 0.79 for the CHAID model, 0.87 for the logistic regression model, and 0.93 for the LASSO logistic regression model. The CHAID model had high specificity but low sensitivity, indicating a lower ability to identify impaired cases. Logistic regression and LASSO logistic regression models had more balanced sensitivity and specificity profiles, with LASSO yielding the highest overall discrimination and classification accuracy. For gait speed impairment, the CHAID model had moderate discrimination, with an AUC of 0.77 (95% CI 0.70‐0.83), characterized by lower sensitivity and higher specificity. The logistic regression and LASSO logistic regression models had higher discrimination, with AUC values of 0.92 (95% CI 0.87‐0.96) and 0.95 (95% CI 0.92‐0.98), respectively. The LASSO model also achieved the highest overall classification accuracy and F₁ score among the 3 approaches. For TUG impairment, model performance was less consistent. The CHAID model had fair discrimination, with an AUC of 0.74 (95% CI 0.71‐0.78), characterized by relatively high sensitivity but lower specificity and overall accuracy. The logistic regression model achieved an AUC of 1.00; however, the accompanying classification metrics suggested possible overfitting, as indicated by low sensitivity and overall accuracy despite perfect specificity. The LASSO logistic regression model had poor discrimination, with an AUC of 0.50, indicating limited usefulness for characterizing this outcome in the current sample.

Overall, the LASSO logistic regression model had the strongest and most consistent classification performance for MNA-SF, SPPB, and gait speed outcomes in this sample. In contrast, the CHAID model generally had lower sensitivity but offered transparent hierarchical subgroup characterization. Model performance for TUG was less consistent across approaches, suggesting that this outcome may be more difficult to characterize using the current set of predictors.

In this cross-sectional secondary analysis, depressive symptoms were associated with poorer nutritional status and lower-extremity physical performance among middle- to older-aged adults. Although both the tree-based exploratory approach and regression models identified broadly similar association patterns, the CHAID model offered a more transparent representation of subgroup structure within the study sample. Using data from a large community-based sample, this study examined the cross-sectional associations of depressive symptom severity with nutritional status and lower-extremity physical performance among community-dwelling adults aged 50 years and older. The findings suggest that greater depressive symptom severity was associated with greater nutritional vulnerability and poorer mobility-related performance. These results are consistent with prior studies linking depressive symptoms to frailty-related physiological and functional decline [10,13,17,19].

Using CHAID decision tree analysis, this study identified a hierarchical subgroup structure in which depressive symptom severity was consistently selected as the primary splitting variable across 2 outcome domains: nutritional status and lower-extremity physical performance. Additional predictors, including age, BMI, calf circumference, education level, and comorbidity severity, further refined subgroup patterns across these outcomes. Rather than serving as a definitive risk stratification tool, CHAID in this study provided rule-based subgroup characterization, with potential advantages for visualizing and communicating clinically relevant patterns within the sample. Together, these findings suggest that depressive symptoms may represent an important clinical marker within the multidimensional pathways linking psychological, nutritional, and functional health in aging populations. Importantly, our findings support the view that health status may be conceptualized as a continuous spectrum, such that older adults classified as low risk or as having only mild impairment may still exhibit substantial internal heterogeneity [46,47] and variation in scores below the conventional cutoff may still carry clinical warning value and help distinguish subgroups with different levels of nutritional vulnerability and functional impairment, which is particularly meaningful for prevention-oriented assessment.

In the comparative model analyses, the LASSO logistic regression model had the highest classification performance and the most balanced sensitivity and specificity profiles for the MNA-SF, SPPB, and gait speed outcomes, suggesting that regularized regression analysis may be useful for classification in datasets with multiple correlated clinical variables. In contrast, although the CHAID decision tree model generally had lower sensitivity, it demonstrated acceptable within-sample classification performance and offered a transparent, rule-based approach for subgroup characterization. Rather than serving as a definitive screening or stratification tool, CHAID may provide complementary value by making subgroup patterns, threshold splits, and variable interactions easier to visualize and communicate in clinical contexts. Taken together, these findings suggest complementary roles for the tree-based exploratory approach and regression-based methods: The LASSO regression model had higher classification performance, whereas the CHAID model offered transparent, rule-based subgroup characterization that may facilitate visualization and communication of clinically relevant patterns within the study sample.

These findings extend prior research on multifactorial geriatric assessment by showing that routinely collected clinical and functional measures can be examined together using a transparent, interpretable analytic approach to characterize patterns of nutritional vulnerability and mobility impairment. Consistent with previous studies emphasizing the multidimensional nature of frailty and functional decline [1,36,48,49], the CHAID decision tree model identified a hierarchical subgroup structure that differentiated participants with varying levels of nutritional and physical vulnerability. More broadly, these findings illustrate how interpretable analytic approaches may complement conventional geriatric assessment by supporting transparent pattern recognition and subgroup characterization.

Overall, 14.2% of participants in our study screened positive for depressive symptoms, which was lower than the 21.3% prevalence reported in a previous study [5]. This difference may reflect variation in sample characteristics, recruitment settings, and population composition across studies. Compared with those without depressive symptoms, participants with depressive symptoms had poorer nutritional status and lower-extremity physical performance, including higher malnutrition risk, lower SPPB scores, and slower gait speed. These associations are consistent with established physiological and behavioral mechanisms linking depression to frailty progression, as depression has been associated with reduced motivation, appetite, and physical activity, which may contribute to reduced dietary intake, sarcopenia, and functional decline [8,10-14]. Taken together, these findings support the relevance of depressive symptoms as a clinical marker of multidimensional vulnerability in aging populations. The decision tree analysis further extends previous work by illustrating how depressive symptoms interact with demographic, anthropometric, and clinical factors to differentiate subgroups with varying levels of nutritional and functional vulnerability. Nutritional and functional patterns varied according to depressive symptom severity in combination with body composition, age, education level, and comorbidity burden. These hierarchical subgroup patterns align with prior evidence linking depression, sarcopenia, and functional decline while also offering a transparent way to visualize and describe heterogeneity within the sample.

At the health systems level, these findings are consistent with international recommendations advocating integrated care models that address mental health, nutrition, and physical function as interconnected determinants of healthy aging [48,49]. Incorporating depression screening into routine geriatric assessment may help identify older adults with coexisting nutritional and functional vulnerabilities. Interpretable analytic approaches such as CHAID may support this process by providing transparent, rule-based subgroup characterization. However, external validation and prospective evaluation are still needed before such approaches can be considered for broader clinical use. In addition, this study focused on a selected tree-based exploratory approach and did not compare the full range of interpretable machine learning methods.

This study has several important strengths. First, it applied both a tree-based exploratory approach and regression-based approaches to examine depression-related patterns using routinely collected geriatric assessment measures. This comparative framework enabled assessment of both classification performance and subgroup interpretability. Second, the use of well-validated clinical instruments, including the GDS-15, MNA-SF, SPPB, gait speed, and TUG, enhances the clinical relevance of the findings. Third, the relatively large community-based sample supported model estimation and subgroup characterization across nutritional and lower-extremity functional outcomes. Finally, the tree-based exploratory approach enabled transparent, rule-based identification of hierarchical subgroup patterns, which may facilitate visualization and communication of clinically relevant heterogeneity within the sample.

Several limitations should also be considered. First, the study focused on a selected tree-based exploratory approach and did not compare the full range of interpretable machine learning methods. Second, model evaluation relied only on an internal training-testing split, without cross-validation, bootstrapping, or external validation. Accordingly, the reported performance metrics should be interpreted as exploratory and sample-specific and should not be taken as evidence of generalizability beyond the current sample. Future studies should incorporate more rigorous internal validation procedures and external validation in independent populations. Third, the cross-sectional study design precludes causal inference and does not establish temporal relationships among depressive symptoms, nutritional status, and physical performance. Fourth, participants were recruited through convenience sampling from a single geographic region, which may limit generalizability to other populations or health care settings. Fifth, depressive symptoms were assessed using a screening instrument rather than a diagnostic interview and therefore may not fully capture clinical depression severity. Finally, model performance varied across outcomes, particularly for TUG impairment, suggesting that additional predictors or alternative analytic approaches may be needed to better characterize this outcome.

In this community-based sample of adults aged 50 years and older, depressive symptom severity was associated with nutritional vulnerability and poorer lower-extremity physical performance. Using a CHAID decision tree approach, this study identified hierarchical subgroup patterns showing that depressive symptoms, including subthreshold levels, together with age, BMI, calf circumference, education level, and comorbidity burden, were related to variation in nutritional and functional vulnerability. Although LASSO logistic regression analysis had higher overall classification performance, the CHAID model offered transparent, rule-based subgroup characterization that may facilitate visualization and communication of clinically relevant patterns within the study sample. These findings suggest that combining depression screening with nutritional and functional assessment may help identify older adults with coexisting vulnerabilities and may support prevention-oriented geriatric assessment. They also support the view that health status may be better understood as a continuum, such that even depressive symptom scores below the conventional cutoff may still carry clinical relevance.