This study used anonymized data from the fifth wave of the CHARLS. The project was approved by the Peking University Biomedical Ethics Committee in June 2008 (approval IRB00001052–11015). All personal identifiers were removed before our analysis. Patients and the public were not involved in the design, conduct, reporting, or dissemination of this research. All participants provided written informed consent, and the study followed the ethical principles set out in the Declaration of Helsinki. No financial or material compensation was offered to study participants.

We carried out a retrospective study using data from the fifth wave of the CHARLS collected between July 2020 and August 2020. Participants had to be aged at least 45 years and have both hypertension and diabetes. We excluded anyone who was missing key details such as CESD-10 score, diagnosis of hypertension or diabetes, or age. After these exclusions, 1504 people were left for the analysis. Depression status was based on the CESD-10, a brief questionnaire that mixes positively and negatively worded statements. Each item is scored from 0 to 3, yielding total scores from 0 to 30; higher scores mean more severe depressive symptoms [12,13]. In Chinese older adults, a score above 10 is generally taken to indicate possible depression, and this cutoff has been shown to work well in previous studies and in latent profile analyses [13,14]. Using this cutoff, of the 1504 participants, we labeled 635 (42.2%) as the depression group and 869 (57.8%) as the nondepression group.

Guided by earlier studies and clinical practice, we chose 42 factors that might relate to depression risk [15-17]. These factors cover basic personal details, overall health, daily habits, blood test results, and how well people manage daily tasks. The full list is as follows: sex, age, marital status, smoking status, drinking status, cancer, lung disease, heart disease, stroke, mental illness, arthritis, high blood fat, liver disease, kidney disease, stomach disease, asthma, memory problems, hip fracture, number of painful body sites, exercise level, sleep hours, social activity engagement, whether the patient was retired, medical insurance, pension insurance, years of schooling, household size, number of living children, household income, recent physician visits, memory test score, activity of daily living (ADL) score, instrumental ADL (IADL) score, self-rated health, and satisfaction with life. Regarding “social activity engagement,” it was categorized into the following eight types: (1) interacted with friends (visiting friends or socializing with friends); (2) played mahjong, played chess, played cards, or went to community clubs; (3) provided help to family members, friends, or neighbors who do not live with the participant and who did not pay them for the help; (4) went to a sports, social, or other type of club; (5) took part in community-related organization activities; (6) did voluntary or charity work; (7) attended an educational or training course; and (8) other.

Participants completed the questionnaire based on their actual situation: participation in 1 type of the above activities was recorded as “1,” participation in 2 types was recorded as “2,” and a higher count indicated a higher level of social activity engagement of the participant. For “life satisfaction,” the CHARLS survey asked respondents about their overall satisfaction with their lives, with 5 response options: “extremely satisfied,” “very satisfied,” “relatively satisfied,” “not very satisfied,” and “not satisfied at all.” For the convenience of statistical analysis, we assigned numerical values to these levels: “1” represented “not satisfied at all,” “5” represented “extremely satisfied,” and so on, with each level corresponding to a specific number.

We used least absolute shrinkage and selection operator (LASSO) regression to pick out the most useful risk factors while keeping the list short. Ten-fold cross-validation helped us find the best penalty value (λ). We then split the data into a training set (1052 people) and a testing set (452 people) in a 7:3 ratio. The training set was used to find the key factors linked to depression. These factors were entered into a standard logistic regression, and we drew a simple chart (nomogram) to show how each one affected the risk of depression. To keep the model easy to use, we removed predictors with little impact, reran the logistic regression with the remaining ones, and built the final nomogram.

To check how well the model performed, we used receiver operating characteristic curves and the area under these curves (AUC) to measure overall ability to separate people with and without depression, calibration plots to check whether the predicted chances matched what really happened, and decision curve analysis to weigh the benefits of true positive predictions against the harms of false ones, giving a clear picture of both accuracy and clinical value.

To make the model easier to understand, we used Shapley additive explanations (SHAP) values to show how much each factor mattered. We drew 4 simple charts (an overall importance plot, a swarm plot, a waterfall plot, and a force plot) that displayed how every factor pushed the risk of depression up or down. These pictures help readers see which factors carry the most weight and how they work together in the model.

All statistical analyses and figure generation were performed using R (R Foundation for Statistical Computing). When less than 20% of the data were missing, we used the mice package to run 5 rounds of imputation and kept the set that best matched the overall pattern. We report counts and percentages for categorical variables and means with SDs for continuous variables. We compared groups using 2-tailed t tests, chi-square tests, or nonparametric tests as needed; the latter were chosen when the data did not follow the rules for parametric tests. A P value <.05 was considered statistically significant.

The study flowchart is presented in Figure 1.

We included 1504 adults with an average age of 65.99 (SD 8.28) years and 64.93 (SD 8.72) years for those with and without depression, respectively; 678 (45.1%) were men, and 826 (54.9%) were women. Of these 1504 participants, 635 (42.2%) were classed as having possible depression, and 869 (57.8%) were classed as nondepressed. Table 1 shows that the 2 groups differed on many key points. Those with possible depression were more often women, widowed or separated, and less active and reported more illnesses, such as lung disease, heart disease, stroke, arthritis, mental illness, liver disease, kidney disease, stomach disease, asthma, hip fracture, and memory problems. They also had more pain sites; poorer sleep; lower income; fewer years of schooling; smaller households; more physician visits; fewer social activities; and lower scores on memory, daily activities, self-rated health, and life satisfaction (P<.05 in all cases).

As the regularization strength (λ) increases, the model error initially decreases and then increases. We chose the simplest model within 1 SE of the minimum error. This left 16 variables whose coefficients stayed above 0 (Figure 2A). Each line in the plot shows one variable’s coefficient; as λ grows (moving right), many lines drop to 0 and are removed. In the end, 16 factors remained: life satisfaction score, sleep hours, self-rated health score, IADL score, memory score, pain sites, ADL score, retirement status, arthritis, sex, memory problems, social activity engagement type, recent physician visits, stroke, years of schooling, and heart disease, listed from most to least important (Figure 2C). Figure 2D shows the direction of each factor: red bars mean higher risk, and blue bars mean lower risk.

We took the 16 variables and ran them together in a new regression model and then drew a simple point chart (nomogram). The chart showed which factors mattered the most so that we could drop the weaker ones. After this step, 9 factors remained: retirement status, self-rated health score, life satisfaction score, number of pain sites, sleep hours, social activity engagement type, memory score, ADL score, and memory-related disease. We put these 9 factors into a final regression and built the last nomogram (Figure 3). On the chart, each line shows how many points a person obtains for that factor; the bottom scale turns the total points into the chance of depression. Higher total points mean higher risk.

We tested how well the model separated people with and without depression using the AUC. In the training data (Figure 4A), the AUC was 0.819, and in the testing data (Figure 4D), it was 0.825, showing good discrimination. The predicted probabilities matched the real outcomes closely in both sets; the average error was only 0.006 in training (Figure 4B) and 0.018 in testing (Figure 4E). Decision curve analysis (Figures 4C and 4F) also showed clear clinical benefit: the model yielded much larger net benefits than the “treat all” or “treat none” approaches. Overall, the nomogram was accurate and clinically useful.

We used SHAP values to look at the 9 main variables. Global and swarm plots (Figure 5A-B) show how each variable affects the entire sample. ADL score, memory problems, and number of pain sites pushed the risk upward, whereas retirement status, social activity engagement type, sleep hours, life satisfaction score, self-rated health score, and memory score pushed it downward. To see how these factors worked for individual people, we picked 2 participants at random and used waterfall and force plots (Figure 5C-D) to show each factor’s exact contribution.

This study establishes a clinically operable prediction model for depression risk among middle-aged and older patients with hypertension and diabetes, leveraging multidimensional predictors to achieve robust discrimination (AUC=0.819‐0.825) and calibration (mean absolute error≤0.018). Nine core variables emerged as significant predictors in the final nomogram: ADL score, memory disorders, number of pain sites, sleep duration, life satisfaction score, self-rated health score, social activity engagement, retirement status, and memory score. The centrality of functional impairment as a mechanistic pathway is underscored by the ADL score’s dominant predictive role [18,19], where individuals with depression exhibited markedly elevated scores (mean 1.20, SD 1.60 vs mean 0.34, SD 0.85; P<.001). Each unit increase in ADL score corresponded to a 3-fold increase in depression risk, particularly at critical thresholds beyond ADL score ≥2. Retirement status provided substantial protection against depression [19,20], with prevalence rates nearly halved among retirees (103/635, 16.2%) compared to nonretirees (269/869, 31.0%; P<.001) mediated by significantly greater social participation (mean activity score 0.94, SD 1.14 vs mean 0.65, SD 0.90; P<.001). SHAP value distributions (Figure 5A-B) corroborated a triadic pathogenesis framework. First, functional autonomy erosion (ADL and IADL decline interacting with multisite pain) imposes physical constraints that amplify helplessness [21,22]. Second, affective-cognitive burden manifests through memory disorders (there were memory disorders in 97/635, 15.3% vs 47/869, 5.4% of participants with and without depression, respectively; P<.001), and reduced life satisfaction (mean score 2.87, SD 0.87 vs mean score 3.41, SD 0.66 among participants with and without depression; P<.001) depletes emotional reserves [23]. Third, socioeconomic disengagement was evidenced by significantly lower household incomes among participants with depression (mean ¥32,470.65 [US $4720.71], SD ¥88,119.77 [US $12,811.20] vs mean ¥54,055.20 [US $7858.73], SD ¥79,512.32 [US $10,687.50], respectively; P<.001), constraining access to health-promoting resources [24,25]. The nonlinear inflection point observed in the ADL-depression relationship indicates that moderate functional decline triggers disproportionately severe psychological sequelae, offering an identifiable intercept for targeted interventions [26].

These findings advance beyond prior depression prediction paradigms in several critical dimensions. Traditional screening instruments reliant on mood-based symptomatology such as the Patient Health Questionnaire series demonstrate limited applicability in cardiometabolic populations due to diagnostic circularity: they cannot resolve whether fatigue or anhedonia stems from depression vs underlying chronic disease pathology [27,28]. Our model effectively dissociates these confounds by operationalizing ADLs as an independent construct validated through SHAP’s feature dissociation capability. For instance, while arthritis prevalence was markedly higher among participants with depression (389/635, 61.3% vs 324/869, 37.3%; P<.001), LASSO selected number of pain sites rather than arthritis diagnosis as the principal pain predictor due to its ability to quantify functional impact severity [29]. When benchmarked against models with similar AUC values to those of the model evaluated in this study, such as the depressive symptom predictor by Zheng et al [30,31], our algorithm demonstrates superior translational efficiency by using information already captured in routine chronic disease management ADL documentation during nursing assessments, medication review for memory issues, and social history documenting retirement status. This obviates the need for additional screening instruments, reducing assessment time to under 3 minutes. Moreover, we identify novel modifiable pathways absent in prior frameworks. The protective effect of retirement when linked to sustained social activity engagement reveals life transitions as windows for resilience-building interventions [32]. Enhancing life satisfaction emerged as another actionable leverage point given its strong negative association with depression (β=−0.63 via LASSO), suggesting that psychosocial enrichment may disrupt pathogenic cascades more effectively than purely biomedical approaches [33].

Methodologically, the integration of LASSO regression with SHAP value interpretation constitutes a primary strength. LASSO refined the predictor pool from 42 candidates to 9 nonredundant variables, efficiently eliminating noncontributory markers such as cancer or lung disease. SHAP analysis then transformed statistical coefficients—for example, the negative β value for retirement (−0.89)—into clinically interpretable narratives. At an individual level (Figure 5D), retirees maintaining active social roles exhibited an 18-fold depression risk reduction compared to their socially isolated counterparts. The CHARLS cohort’s population representativeness further enhances generalizability to China’s aging cardiometabolic patient demographic. Four principal limitations warrant acknowledgment. Due to the cross-sectional nature of wave 5 CHARLS data, causal precedence between functional decline and depression could not be definitively established. However, the biological gradient observed in ADL risk proportionality (higher scores predicting depression probability stepwise) lends credibility to functional deterioration as a driver rather than a secondary consequence [34,35]. Second, depression ascertainment using a CESD-10 score >10 may overlook subsyndromal presentations, potentially underestimating risk in highly functional individuals [36]. Third, unmeasured variables influencing both pathways, particularly medication adherence patterns or pharmacodynamic interactions between psychotropics and antihypertensives, were unaccounted for. Finally, sociocultural proxies unique to Chinese aging populations’ children’s financial support demonstrating borderline significance (P=.02-.03) were excluded from the final model despite their potential clinical relevance [37]. Future iterations should address this through culturally adapted predictor sets.

This study was conducted based on statistical data from the CHARLS database. Therefore, the results of this study have limitations in terms of applicability to young and middle-aged populations, as well as populations from different regions and ethnic groups. In future research, we may consider combining data from public databases in the Americas, Europe, and other regions to conduct joint modeling, thereby improving the generalization performance of the model. These findings suggest priority areas for future research and clinical application. Longitudinal analyses using multiwave CHARLS data should test the proposed retirement–social engagement–protection pathway, examining whether weekly social activity sustains resilience to depressive symptoms. A 3-tier risk stratification protocol based on nomogram scores may guide care: individuals scoring ≤150 can continue annual screening; those scoring 150 to 250 should receive functional preservation strategies, including task-oriented exercise and cognitive behavioral approaches to improve ADLs, sleep, and social engagement; and those scoring ≥250 would require multidisciplinary management due to high comorbidity. Embedding this algorithm into electronic health records within China’s National Essential Public Health Service Package could enable automated alerts when thresholds are met. Pilot programs should focus on communities where the baseline ADL score is ≥1.5; in this study, 51.2% (325/635) of participants with depression were in this category. Randomized trials are needed to evaluate mediator-targeted interventions, including combined ADL and pain management, structured postretirement social role adaptation, and hemoglobin A_1c–stratified memory screening in diabetes. Targeting domain-specific vulnerabilities rather than symptoms alone may reduce depression incidence while optimizing primary care resources.

We developed and validated a depression risk model for middle-aged and older adults with hypertension and diabetes. The nomogram, integrating 9 routinely collected predictors, including ADL score, memory impairment, and pain burden, showed strong discrimination and calibration. Depression risk rose 3-fold for each unit increase in ADL score, whereas retirement was protective. Functional impairment, emotional cognitive burden, and socioeconomic disadvantage formed the dominant risk pathway. LASSO-SHAP integration streamlined predictor selection, enabling screening in under 3 minutes. Prospective trials should test interventions targeting postretirement social engagement and functional rehabilitation, alongside implementing a 3-tier risk stratification protocol in primary care. Targeting these domain-specific risks could reduce depression incidence.