Development and Internal Validation of Machine Learning-Based Risk Prediction Models for Depression in Older Adults Undergoing Maintenance Hemodialysis

Abstract

Background

The high prevalence of depressive symptoms in older maintenance hemodialysis (MHD) patients necessitates early screening. This study aimed to identify core predictors of depressive symptoms in older MHD patients and develop internally validated risk prediction models.

Methods

We recruited 226 eligible older MHD patients from hemodialysis units of the Affiliated Hospital of Zunyi Medical University and its branches. Data collection included general characteristics, the MHD Self-Management Scale, Dialysis Symptom Index (DSI), Short Physical Performance Battery, Mini Nutritional Assessment-Short Form (MNA-SF), 15-item Geriatric Depression Scale (GDS-15), and self-reported sensory function. Key variables were identified using univariate logistic regression, Least Absolute Shrinkage and Selection Operator regression, and backward stepwise multivariable logistic regression. Five machine learning algorithms were trained using 10-fold repeated cross-validation and internally validated via a stratified 7:3 train–test split. Model performance was evaluated using area under the receiver operating characteristic curve (AUC) with 95% bootstrap CIs, calibration metrics (Brier score, calibration intercept, and slope), F1-scores, and decision curve analysis. Shapley Additive Explanations and partial dependence plots were generated for the Naive Bayes model.

Results

Among 226 patients, the prevalence of GDS-15-defined depressive symptoms was 54.4%. All models demonstrated comparable discriminative performance (test bootstrap AUC range: 0.851–0.865), with DeLong’s test confirming no statistically significant pairwise differences (all P > 0.15). Naive Bayes achieved the optimal overall performance with the highest discrimination (AUC = 0.865, 95% CI: 0.772–0.944) and best calibration (Brier = 0.151; calibration slope = 0.904). Top predictors were vision impairment, hearing loss, DSI, MNA-SF, and self-management.

Conclusion

We identified five key predictors and developed internally validated machine learning models for depression risk screening in older MHD patients. The models require external validation before clinical application.

Introduction

Chronic Kidney Disease (CKD) is a prevalent condition involving progressive structural kidney damage and functional decline,^1,2 irreversibly advancing to end-stage renal disease (ESRD).³ Hemodialysis (HD)—the predominant form of renal replacement therapy, accounting for approximately 69% of all modalities and 89% of all dialysis procedures^4,5—remains the primary life-sustaining treatment for most ESRD patients because of the high cost, limited organ availability, and substantial risks associated with kidney transplantation. Maintenance Hemodialysis (MHD) refers to regular HD treatment administered for more than three months as part of a prescribed regimen.⁶ Globally, an estimated 2.6 million patients received renal replacement therapy in 2010, a figure projected to reach 5.44 million by 2030, with Asia anticipated as the fastest-growing region.⁷ Concurrently, driven by population aging and advances in HD technology, older patients aged ≥ 60 years now constitute nearly 70% of the total dialysis population,⁸ establishing them as the primary MHD cohort.

The physiological complications of MHD (e.g, electrolyte disturbances, anemia), treatment burden (dependence on thrice-weekly dialysis sessions), and impaired social functioning collectively contribute to a high vulnerability to mental health disorders in this patient population.⁹ Depression, one of the most prevalent psychological complications among MHD patients,^10,11 is characterized by a higher prevalence and more severe consequences within the older subgroup. While evidence regarding depression in MHD patients remains limited, existing studies indicate that advanced age constitutes a significant risk factor. The prevalence of depression among older MHD patients (91.84%) is markedly higher than rates observed in the general older population (6–52%)^12,13 and non-older MHD patients (77.59%).^14,15 Beyond impairing quality of life, depression is significantly associated with reduced dialysis adherence, elevated suicide risk, and increased all-cause mortality, establishing it as a critical risk factor for survival outcomes in older MHD patients.^16–18

Despite confirmed detrimental effects of depression in older MHD patients, important limitations persist in current research. On one hand, existing studies investigating influencing factors of depression in older MHD patients largely overlook the more crucial exploration of predictive modeling for early depressive states.¹⁵ On the other hand, existing research developing depression prediction models for MHD populations has primarily focused on the overall all-age MHD population, with no dedicated model development targeting the older adult subgroup.¹⁹ Moreover, these studies typically rely on single, conventional modeling approaches.²⁰ Implementing early screening and timely psychological interventions prior to the onset of depression in older MHD patients can effectively prevent depressive episodes and enhance psychological well-being.²¹ Concurrently, this approach improves adherence to MHD therapy while reducing depression-related adverse events, including suicide and mortality,^22,23 thereby improving patients’ quality of life.¹⁸

Traditional modeling approaches exhibit limited capacity to capture nonlinear relationships, coupled with inadequate predictive accuracy and model interpretability. These limitations impede the precise determination of individual feature contributions to prediction outcomes, thereby restricting their clinical utility. In contrast, machine learning algorithms effectively address these limitations, enabling significant performance leaps.²⁴ Firstly, ensemble algorithms automatically capture nonlinear interaction effects through feature subspace sampling, overcoming the limitation of traditional methods that rely on a priori parameter specification. Secondly, regularization techniques, such as the Least Absolute Shrinkage and Selection Operator (LASSO), synergistically optimize model complexity with embedded feature selection, significantly enhancing generalization capability for high-dimensional data.²⁴

The Shapley Additive Explanations (SHAP) swarm plot serves as one of the most widely utilized global interpretability visualization tools for predictive models, demonstrating both the directionality and magnitude distribution of each feature’s impact on model predictions across all samples.²⁵ Partial Dependence Plots (PDPs) represent a visualization technique for interpreting machine learning models, illustrating the marginal effect of one or multiple features on the predicted outcome.²⁶ Integrating machine learning prediction models with these two interpretability techniques enhances model transparency, thus overcoming the inherent “black box” limitation of traditional models.²⁷

In light of this, our study employs a multidimensional factor integration approach, constructing a depression risk prediction model for older MHD patients through LASSO regression combined with five machine learning algorithms. Furthermore, we incorporate SHAP and PDP methodologies to analyze feature importance and risk-response patterns. This work aims to establish a theoretical foundation for early identification and targeted interventions for depression in this population, thereby improving quality of life and reducing adverse outcome risks among older MHD patients.

Materials and Methods

Participants

Patients meeting the inclusion criteria were enrolled from hemodialysis units at Zunyi Medical University Affiliated Hospital and its branch hospitals between July and November 2024. The inclusion criteria were as follows: (1) age ≥ 60 years; (2) diagnosed with ESRD and receiving MHD for >3 months; (3) functional independence in activities of daily living (including ambulation with assistive devices); (4) voluntarily participated after providing informed consent. The exclusion criteria were as follows: (1) individuals with communication barriers (e.g, aphasia or severe dementia) precluding valid questionnaire completion; (2) persons with congenital or acquired physical disabilities; (3) patients diagnosed with life-limiting comorbidities (e.g, active malignancies).

Outcome Definition and Candidate Predictive Variables

Outcome Definition

The outcome variable in this study was assessed using the 15-item Geriatric Depression Scale (GDS-15) to measure psychological status. The GDS-15, developed by Yesavage et al²⁸ in 1986 as a short form of the Geriatric Depression Scale (GDS), is specifically validated for geriatric populations. This instrument comprises 15 dichotomous items, with a total score ranging from 0 to 15. Higher scores indicate greater symptom severity, with a score of ≥5 identifying individuals at risk for depression. For conciseness, the term “depression” is used throughout to represent this threshold-defined symptomatic state. It should be noted that the GDS-15 is a screening instrument; a score ≥ 5 identifies individuals at elevated risk warranting further clinical evaluation, and does not constitute a clinical diagnosis of depressive disorder.

Candidate Predictive Variables

A total of 21 candidate predictor variables were selected based on literature review and clinical expertise, categorized as follows:

• 1)Sociodemographic factors: age, Gender, educational level (Primary school and below/Middle school/High school or technical high school/Junior college and above), health insurance type (Resident/Employee), drinking and smoking.
• 2)Dialysis-related variables: dialysis frequency per week (within recent 2 weeks) (Once or twice times/Three times), vascular access (Autogenous arteriovenous fistula/Central venous catheters) and primary disease (Hypertension/Diabetes/Glomerulonephritis/Others).
• 3)Physiological indicators: serum potassium (K), serum calcium (Ca), serum phosphorus (P), serum creatinine (Cr), serum albumin (Alb) and hemoglobin (Hb).
• 4)Symptom burden: Assessed using the Dialysis Symptom Index (DSI).²⁹ This scale consists of 30 items, including 25 physical symptoms and 5 psychological symptoms. Participants were asked to report the presence of dialysis-related symptoms they had experienced in the past 7 days. Distress levels were rated on a 5-point Likert scale: None = 0, Mild = 1, Moderate = 2, Severe = 3, and Extreme = 4. The total score ranges from 0 to 120, with higher scores indicating greater symptom distress. Of the 30 DSI items, 25 assess physical symptoms and 5 assess psychological symptoms (e.g, anxiety, sadness, irritability); the potential overlap between these psychological symptom items and the GDS-15 depression outcome is addressed in the Limitations.
• 5)Self-management level was evaluated using the validated Chinese version of the Self-Management Scale for MHD Patients. This 20-item instrument encompasses four domains: therapeutic partnership, emotional regulation, problem-solving, and self-care, with total scores ranging from 20 to 80. Higher scores indicate better self-management proficiency. Scores were categorized into three tiers: 20–40 (inadequate), 41–60 (moderate), and 61–80 (competent) self-management level.³⁰ In the present analysis, the total score was treated as a continuous variable.
• 6)Physical functioning: Assessed using the Short Physical Performance Battery (SPPB). Developed by Guralnik et al³¹ this scale includes three dimensions: “balance ability”, “4-meter gait speed” and “chair stand test” with a total score of 12. Higher scores indicate stronger physical performance.
• 7)Nutritional status: Assessed using the Mini-Nutritional Assessment Short-Form (MNA-SF). This scale, developed by Rubenstein et al, shows a very high correlation with the Mini-Nutritional Assessment (MNA) (r = 0.945),³² supporting its validity for comprehensive nutritional assessment and its particular suitability for use in older populations.³³ It consists of 6 items, with a total score of 14, and higher scores indicate better nutritional status.
• 8)Sensory function (vision and hearing) was assessed through self-reported measures using two standardized questions: “Has your eyesight declined to the extent that it affects your daily activities?” “Has your hearing declined to the extent that it affects your daily activities?” Participants were instructed to respond based on their corrected functional status (i.e, with habitual use of eyeglasses or hearing aids if applicable). Affirmative responses to either question defined sensory impairment in the respective domain.

Statistical Analysis

All statistical analyses were performed using SPSS version 29.0 (IBM Corp., Armonk, NY) and R version 4.4.3 (R Foundation for Statistical Computing, Vienna, Austria). The main R packages used were glmnet (LASSO regression), caret (model training and tuning), pROC (receiver operating characteristic [ROC] analysis and DeLong’s test), ggplot2 (visualization), randomForest, e1071 (Naive Bayes), nnet (multilayer perceptron [MLP]), dplyr and tidyr (data manipulation), rms (calibration assessment), rmda (decision curve analysis), fastshap (SHAP computation), pdp (partial dependence plots), and boot (bootstrap confidence intervals).

Continuous variables were assessed for normality using the Shapiro–Wilk test. Normally distributed variables are presented as mean ± standard deviation (SD) and compared between groups with independent samples t-tests. Non-normally distributed (skewed) variables are summarized as median (first quartile [Q1], third quartile [Q3]) and analyzed using the Mann–Whitney U-test. Categorical variables are expressed as frequencies and percentages (n [%]), with group comparisons performed by Chi-square tests or Fisher’s exact tests as appropriate. P<0.05 was taken as the difference was statistically significant. Spearman’s rank correlation analysis was performed using the full dataset (n = 226) to evaluate relationships among the five final predictors and depression. Correlation coefficients (r) and corresponding P values were calculated. The magnitude of correlation was interpreted using the following criteria: |r| < 0.3 indicated a weak correlation, 0.3 ≤ |r| < 0.6 indicated a moderate correlation, and |r| ≥ 0.6 indicated a strong correlation. The correlation matrix was visualized as a heatmap using the R package corrplot.

Predictor Screening Methods

Feature selection, the process of identifying significant predictors for model construction, is a critical pre-modeling step. LASSO employs L1-penalized regularization by introducing a hyperparameter lambda (λ).³⁴ This technique imposes a penalty on the sum of absolute regression coefficients during model estimation, which sparsifies the coefficient vector by forcing non-informative coefficients to exactly zero. Consequently, LASSO retains only variables with genuine predictive power (non-zero coefficients), thereby enhancing model generalizability, mitigating overfitting, and improving predictive accuracy.³⁵

To mitigate data leakage risks, the dataset was preemptively partitioned into training and test set at a 7:3 ratio via stratified random sampling by depression outcome.³⁶ Feature selection was strictly confined to the training set prior to model development. An initial screening of the 21 candidate predictors was performed using univariate logistic regression, retaining variables demonstrating statistical significance. Subsequently, LASSO regression was applied to these selected predictors. Model optimization employed 10-fold cross-validation by partitioning the dataset into 10 mutually exclusive folds. Through an iterative rotation process, each fold served sequentially as the validation set while the remaining nine folds constituted the training set. For each candidate λ value, this procedure generated 10 replicated model fits with corresponding validation errors. The mean squared error (MSE) averaged across all iterations quantified model performance, effectively minimizing contingency from single train-test splits while maximizing data utilization efficiency.”

The optimal regularization parameter λ was selected via the one-standard-error rule, identifying the largest λ value (denoted λ.1se) that produced a mean squared error within one standard error of the minimal MSE (MSE_min + SE_min). This λ.1se value imposes the strongest regularization penalty, forcibly shrinking coefficients of weakly associated predictors to exact zero while maximally reducing model complexity. Consequently, only the most robust core predictors with non-zero coefficients were retained. These LASSO-selected variables were subsequently entered into a multivariable logistic regression model to identify statistically significant predictors for the final predictive framework.

Model Development and Hyperparameter Tuning

All five models were developed using the five predictors selected by LASSO and multivariate logistic regression as independent variables. The dataset was partitioned into training (70%) and test (30%) sets using stratified random sampling by depression outcome to preserve class proportions. All models were trained using 10-fold cross-validation repeated five times with the area under the receiver operating characteristic curve (AUC) as the optimization metric. Continuous predictors were centered and scaled to unit variance (z-score normalization) for scale-sensitive algorithms (KNN and MLP);³⁷ all preprocessing parameters were derived exclusively from the training data and then applied to the test data to prevent information leakage. No missing data were present in the dataset. The random seed was set to 1,122 to ensure full reproducibility of all stochastic procedures. A post-hoc sample size assessment using the pmsampsize framework³⁸ was performed. Under a conservative pre-specification of anticipated performance (Cox-Snell R² = 0.10), the most stringent criterion suggested a larger sample than the present study provided; however, a less stringent criterion that still penalizes overfitting required only 125 patients with 68 events—a threshold the present sample (n = 226, 123 events) comfortably exceeded. With five final predictors and 123 events (events-per-predictor ≈ 24.6), the model structure also exceeded the commonly recommended minimum of 10–20 events per variable. The implications of sample size are addressed in the Limitations. Hyperparameters were tuned via grid search within the training set:

Logistic Regression: No hyperparameter tuning was required (fixed model with binomial family and logit link).³⁹

K-Nearest Neighbors (KNN): The number of neighbors k was searched over the set {3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25}, with the optimal value selected as k = 25.^40,41

Naive Bayes: Three hyperparameters were jointly tuned: Laplace smoothing (fL ∈ {0, 0.5, 1}), kernel density estimation (usekernel ∈ {TRUE, FALSE}), and bandwidth adjustment (adjust ∈ {0.5, 1, 1.5}). The optimal configuration was fL = 0, usekernel = TRUE, adjust = 1.5.^42,43

MLP: A single-hidden-layer feedforward architecture was used, with the number of hidden neurons (size ∈ {2, 3, 5, 7, 10}) and weight decay regularization (decay ∈ {0.001, 0.01, 0.1, 1}) tuned jointly. Training proceeded for a maximum of 1,000 iterations (maxit = 1,000). The optimal configuration was size = 10, decay = 1.⁴⁴

Random Forest: The number of randomly sampled candidate variables at each split (mtry ∈ {1, 2, 3, 4, 5}) was tuned, with the number of trees fixed at ntree = 1,000 to ensure stable convergence. The minimum node size (nodesize) was set to 5. The optimal mtry was 1.^45,46

All models were implemented using the caret package (version 6.0–94) in R. The rationale for algorithm selection and key methodological choices is detailed in Table S1. Model development and reporting followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis and Artificial Intelligence (TRIPOD+AI) guidelines.⁴⁷

Model Performance Evaluation

Model performance was assessed from three complementary perspectives: discrimination, calibration, and clinical utility.

Discrimination was evaluated using the AUC derived from the ROC curve. Bootstrap 95% confidence intervals (CIs) for AUC were computed using 2,000 resamples with replacement. Additional classification metrics—including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score—were computed at the default probability threshold of 0.5, with 95% bootstrap CIs likewise obtained from 2,000 resamples. Pairwise comparisons of test-set AUCs across the five models were performed using DeLong’s test, with a two-sided significance level of P < 0.05.

Calibration was assessed through calibration plots with loess smoothing, which graphically compare predicted probabilities against observed outcome proportions. Quantitative calibration metrics included the calibration-in-the-large (intercept; ideal = 0, with negative values indicating systematic overestimation and positive values indicating underestimation), calibration slope (ideal = 1, with values > 1 indicating overfitting and < 1 indicating underfitting), and the Brier score (range: 0–1; lower values indicate better overall predictive accuracy). These metrics were computed using the rms package.

Clinical utility was evaluated using decision curve analysis (DCA) via the rmda package, which quantifies the net benefit of model-guided intervention across a range of threshold probabilities. The net benefit curves of the five models were compared against two reference strategies: intervening in all patients (“treat-all”) and intervening in none (“treat-none”).

Model Interpretability

Model interpretability was addressed through two complementary approaches using the Naive Bayes model, which demonstrated the optimal combination of discrimination and calibration among the five algorithms.

Shapley Additive Explanations (SHAP) values were computed using KernelSHAP via the fastshap R package with 100 Monte Carlo simulations and 50 background samples randomly drawn from the training set. SHAP values were calculated for all test-set observations and visualized using a beeswarm plot, which displays the direction and magnitude of each feature’s contribution to individual predictions. Feature importance was quantified as the mean absolute SHAP value and presented as a bar chart.^48,49

Partial dependence plots (PDPs) were generated using the pdp R package to visualize the marginal effect of each continuous predictor on the predicted probability of depression, averaged over the distribution of all other predictors. Bivariate PDPs were additionally constructed to examine potential synergistic interaction patterns among key predictors.

Results

Characteristics of the Participants

This study enrolled a total of 226 older patients with MHD, with a prevalence of GDS-15-defined depressive symptoms of 54.4%. The mean age was (68.76 ± 7.61) years, and the median age of the depression group was significantly greater than that of the non-depression group (70 vs 66 years, P = 0.014). There were 127 males (56.19%) and 99 females (43.81%). Age, Ca, Alb, DSI, self-management level, SPPB, MNA-SF, Gender, educational level, medical insurance, vision impairment and hearing loss demonstrated statistically significant differences between the groups (P< 0.05), while other variables did not show significant differences (Table 1).

Table 1.

Comparison of Characteristics Between Depression Group and Non-Depression Group in Older Patients with MHD

Variables	Total (n = 226)	Non-Depression (n = 103)	Depression (n = 123)	Statistic	P
K (mmol/L)	4.00 ± 0.73	4.02 ± 0.69	3.99 ± 0.77	t=0.33	0.743
Age	67.50 (61.00, 75.00)	66.00 (61.00, 72.00)	70.00 (62.00, 76.00)	Z=−2.45	0.014
Ca (mmol/L)	2.17 (2.02, 2.33)	2.24 (2.10, 2.37)	2.11 (2.00, 2.30)	Z=−3.09	0.002
P (mmol/L)	1.50 (1.07, 1.93)	1.56 (1.13, 1.96)	1.46 (1.03, 1.89)	Z=−1.25	0.211
Cr (μmol/L)	500.50 (320.75, 814.25)	496.00 (299.50, 799.50)	521.00 (341.50, 815.50)	Z=−0.78	0.438
Alb (g/L)	37.60 (35.12, 39.60)	38.10 (36.60, 40.50)	37.00 (34.03, 39.25)	Z=−3.51	<0.001
Hb (g/L)	119.00 (103.25, 131.00)	119.00 (106.00, 130.50)	117.00 (101.00, 131.00)	Z=−0.60	0.549
DSI	11.00 (7.00, 15.00)	9.00 (5.00, 12.00)	13.00 (9.00, 19.50)	Z=−5.43	<0.001
Self-management level	50.00 (45.00, 56.00)	53.00 (47.00, 58.00)	48.00 (43.00, 53.00)	Z=−3.75	<0.001
SPPB	4.50 (3.00, 7.00)	6.00 (4.00, 7.50)	3.00 (2.00, 5.00)	Z=−5.75	<0.001
MNA-SF	12.00 (10.00, 13.00)	13.00 (11.00, 14.00)	11.00 (8.00, 12.00)	Z=−6.34	<0.001
Gender				χ2=4.78	0.029
Male	127 (56.19)	66 (64.08)	61 (49.59)
Female	99 (43.81)	37 (35.92)	62 (50.41)
Educational level				χ2=19.39	<0.001
Primary school and below	69 (30.53)	33 (32.04)	36 (29.27)
Middle school	36 (15.93)	20 (19.42)	16 (13.01)
High school/technical high school	26 (11.50)	20 (19.42)	6 (4.88)
Junior college and above	95 (42.04)	30 (29.13)	65 (52.85)
Medical insurance				χ2=10.71	0.001
Resident medical insurance	119 (52.65)	42 (40.78)	77 (62.60)
Employee medical insurance	107 (47.35)	61 (59.22)	46 (37.40)
Drinking				χ2=0.04	0.843
Non-Drinker	205 (90.71)	93 (90.29)	112 (91.06)
Drinker	21 (9.29)	10 (9.71)	11 (8.94)
Smoking				χ2=0.88	0.349
Non-Smoker	192 (84.96)	85 (82.52)	107 (86.99)
Smoker	34 (15.04)	18 (17.48)	16 (13.01)
Dialysis frequency				χ2=0.02	0.899
Once or twice times a week	78 (34.51)	36 (34.95)	42 (34.15)
Three times a week	148 (65.49)	67 (65.05)	81 (65.85)
Vascular access				χ2=1.25	0.263
Autogenous arteriovenous fistula	198 (87.61)	93 (90.29)	105 (85.37)
Central venous catheters	28 (12.39)	10 (9.71)	18 (14.63)
Primary disease				χ2=2.20	0.531
Hypertension	82 (36.28)	34 (33.01)	48 (39.02)
Diabetes	50 (22.12)	27 (26.21)	23 (18.70)
Glomerulonephritis	67 (29.65)	29 (28.16)	38 (30.89)
Others	27 (11.95)	13 (12.62)	14 (11.38)
Vision impairment				χ2=32.70	<0.001
No	141 (62.39)	85 (82.52)	56 (45.53)
Yes	85 (37.61)	18 (17.48)	67 (54.47)
Hearing Loss				χ2=23.20	<0.001
No	149 (65.93)	85 (82.52)	64 (52.03)
Yes	77 (34.07)	18 (17.48)	59 (47.97)

Notes: P < 0.05: statistically significant. Non-normal continuous variables are described by the median and interquartile range, M (Q₁, Q₃): Age, Ca, P, Cr, Alb, Hb, DSI, Self-management level, SPPB, MNA-SF.

Abbreviations: SD, Standard deviation; M, Median; Q₁, First quartile; Q₃, Third quartile; K, serum potassium; Ca, serum calcium; P, serum phosphorus; Cr, serum creatinine; Alb, serum albumin; Hb, Hemoglobin; DSI, Dialysis Symptom Index; SPPB, Short Physical Performance Battery; MNA-SF, Mini Nutritional Assessment Short-Form.

Variable Selection

The results of variable selection using univariate logistic regression analysis, the LASSO algorithm, and multivariate logistic regression are as follows: univariate logistic regression of the 21 candidate predictors initially retained 8 statistically significant variables (DSI, Self-management Level, SPPB, MNA-SF, Vision impairment, Hearing loss, Alb, and Gender). Subsequent LASSO regression with 10-fold cross-validation (λ.1se = 0.056) excluded Alb, SPPB, and Gender via coefficient shrinkage, yielding five final predictors: DSI, Self-management Level, MNA-SF, Vision impairment, and Hearing loss. These five variables were then entered into a backward stepwise multivariable logistic regression model, which confirmed all five as statistically significant and provided adjusted odds ratios. The forest plot of these results is shown in Figure 1(A).

Figure 1.

Screening results of predictor variables. (A) Multivariate logistic regression forest plot of five predictors and depression. (B) Multicollinearity diagnostics for the final predictors.

Abbreviations: DSI, Dialysis Symptom Index; MNA-SF, Mini Nutritional Assessment Short-Form; VIF, Variance Inflation Factors.

To assess multicollinearity among predictors, variance inflation factors (VIF) were calculated. As depicted in Figure 1(B), all independent variables exhibited VIF values <5, indicating absence of severe multicollinearity, thereby ensuring the reliability of regression estimates.

Figure 2 showed the Spearman correlation heatmap among the five final predictors and depression. Overall, weak to moderate correlations were observed between most predictor pairs (maximum absolute correlation coefficient = 0.352, between DSI and MNA-SF), and all absolute coefficients were below 0.4. Specifically, DSI was moderately negatively correlated with MNA-SF (r = −0.352, P < 0.001) and weakly positively correlated with vision impairment (r = 0.241, P < 0.001). MNA-SF showed weak negative correlations with vision impairment (r = −0.307, P < 0.001) and hearing loss (r = −0.241, P < 0.001). Vision impairment and hearing loss exhibited a weak positive correlation (r = 0.309, P < 0.001). Self-management level was weakly correlated with the remaining four predictors, none reaching statistical significance (all |r| < 0.1, P > 0.05). These results further support the absence of problematic multicollinearity (consistent with VIF < 5 in Figure 1B) and justify the simultaneous inclusion of all five predictors in the multivariate model.

Figure 2.

Spearman correlation heatmap among the five final predictors and depression.

Notes: The color gradient represents the correlation coefficient (red for positive correlation, blue for negative correlation). Numerical values indicate the Spearman’s rho, with asterisks denoting statistical significance (***P < 0.001).

Abbreviations: DSI, Dialysis Symptom Index; MNA-SF, Mini Nutritional Assessment Short-Form.

Model Construction and Validation

Performance metrics for the five machine learning-based depression risk prediction models are presented in Figure 3, Table 2, and Figure S1 (calibration plots). All five models demonstrated acceptable discriminative performance on the test set (bootstrap-derived AUC range: 0.851–0.865), with substantially overlapping 95% bootstrap confidence intervals (Table 2). The Naive Bayes model achieved the highest test-set AUC (0.865, 95% CI: 0.772–0.944), while the Random Forest model yielded the highest F1-score (0.783, 95% CI: 0.667–0.880).

Figure 3.

ROC curves for five machine learning prediction models.

Notes: Pink curves represent the training set; green curves represent the test set. The semi-transparent shaded bands around each ROC curve represent the 95% bootstrap confidence intervals (2,000 resamples). The grey dashed diagonal line represents the performance of a random classifier (AUC = 0.5).

Abbreviations: ROC, Receiver Operating Characteristic; KNN, K-Nearest Neighbors; MLP, Multilayer Perceptron.

Table 2.

Performance of Five Machine Learning Models on the Test Set

Model	AUC (95% CI)	F1 (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	Brier	Calib Intercept	Calib Slope
LR	0.851 (0.752, 0.931)	0.754 (0.627, 0.857)	0.812 (0.667, 0.938)	0.694 (0.531, 0.842)	0.703 (0.558, 0.846)	0.806 (0.647, 0.933)	0.163	−0.396	1.031
KNN	0.857 (0.753, 0.934)	0.765 (0.636, 0.866)	0.812 (0.667, 0.938)	0.722 (0.575, 0.864)	0.722 (0.564, 0.857)	0.812 (0.667, 0.938)	0.159	−0.295	1.624
Naive Bayes	0.865 (0.772, 0.944)	0.733 (0.593, 0.846)	0.688 (0.522, 0.846)	0.833 (0.700, 0.944)	0.786 (0.625, 0.926)	0.750 (0.618, 0.881)	0.151	0.162	0.904
MLP	0.858 (0.763, 0.935)	0.754 (0.625, 0.857)	0.812 (0.667, 0.935)	0.694 (0.543, 0.839)	0.703 (0.545, 0.848)	0.806 (0.667, 0.933)	0.164	−0.470	1.455
RF	0.863 (0.764, 0.944)	0.783 (0.667, 0.880)	0.844 (0.710, 0.966)	0.722 (0.568, 0.871)	0.730 (0.581, 0.871)	0.839 (0.700, 0.966)	0.155	−0.362	1.246

Notes: AUC, F1, sensitivity, specificity, PPV, and NPV are reported with 95% bootstrap confidence intervals (2,000 resamples). Calibration intercept: ideal = 0 (negative values indicate overestimation; positive values indicate underestimation). Calibration slope: ideal = 1 (values > 1 indicate overfitting). Brier score: range 0–1, lower values indicate better overall predictive accuracy.

Abbreviations: LR, Logistic Regression; KNN, K-Nearest Neighbors; MLP, Multilayer Perceptron; RF, Random Forest; AUC, Area Under the Receiver Operating Characteristic Curve; PPV, Positive Predictive Value; NPV, Negative Predictive Value.

DeLong’s test revealed no statistically significant differences between any pair of models in the test set (all P > 0.15; Table S2), indicating that the five algorithms performed comparably within the bounds of sampling uncertainty. Therefore, no single model can be definitively declared superior to the others based on the current data. The Random Forest model exhibited a noticeable discrepancy between training (bootstrap AUC = 0.938, 95% CI: 0.900–0.969) and test (bootstrap AUC = 0.863, 95% CI: 0.764, 0.944) performance (ΔAUC = 0.075), suggesting moderate overfitting. In contrast, Logistic Regression (training AUC = 0.847 vs test AUC = 0.851, ΔAUC = −0.004), KNN (ΔAUC = −0.032), and MLP (ΔAUC = −0.013) maintained strong consistency across datasets, with test AUC values slightly exceeding training estimates, likely reflecting sampling variability in the small test set.

Calibration assessment revealed that all models achieved reasonably low Brier scores on the test set (range: 0.151–0.164; Table 2). The Naive Bayes model demonstrated the best overall calibration (Brier score = 0.151, calibration intercept = 0.162, calibration slope = 0.904), with its calibration slope approaching the ideal value of 1.0. The remaining models exhibited moderate calibration slopes exceeding 1.0 (range: 1.031–1.624), indicating a tendency toward overconfident predictions in extreme probability regions—a pattern commonly observed with distance-based and ensemble algorithms in small samples.

In decision curve analysis, all five models yielded positive net benefit within the clinically relevant threshold probability range of 0.4–0.6, consistently outperforming both the “treat-all” and “treat-none” strategies (Figure 4). These results indicate that the models provide meaningful clinical value for identifying high-risk patients warranting depression screening.

Figure 4.

Decision curves for five machine learning models. (A) Training Set. (B) Test Set.

Notes: The x-axis represents the threshold probability at which a clinician would consider intervening for depression; the y-axis represents the net benefit. The gray “All” line assumes all patients receive intervention; the black “None” line assumes no patients receive intervention. A model is clinically useful when its curve lies above both reference lines within the threshold range of interest (0.4–0.6). (A) Training set. (B) Test set.

Abbreviations: DCA, Decision Curve Analysis;LR, Logistic Regression; KNN, K-Nearest Neighbors; MLP, Multilayer Perceptron.

Collectively, these findings indicate that the five machine learning approaches offer broadly comparable discriminative performance, with modest differences falling within the range of sampling variability. However, the models diverged substantially in calibration: Naive Bayes was the only model whose calibration slope approached the ideal value of 1.0 (0.904), whereas the remaining models exhibited slopes ranging from 1.031 to 1.624, indicating a systematic tendency toward overconfident predictions. Given that well-calibrated probability estimates are essential for meaningful risk stratification, the Naive Bayes model was selected for subsequent interpretability analysis using SHAP and partial dependence plots. This model combined the highest discrimination (test AUC = 0.865) with the best calibration (Brier score = 0.151; calibration intercept = 0.162; calibration slope = 0.904).

Feature Importance

SHAP values were computed for the Naive Bayes model to quantify the contribution of each predictor to individual depression risk predictions (Figure 5). In the SHAP beeswarm plot (Figure 5A), each point represents a single test-set observation, with the x-axis indicating the SHAP value—positive values (right of zero) increase the predicted probability of depression, while negative values (left of zero) decrease it. Continuous variables (DSI, MNA-SF, and self-management) are color-coded by their original feature values using a continuous viridis scale, whereas binary variables (vision impairment and hearing loss) are displayed as two distinct colors representing absence and presence. For vision impairment and hearing loss, the presence of impairment (yellow points) was predominantly associated with positive SHAP values, indicating increased depression risk, while the absence of impairment (dark purple points) was predominantly associated with negative SHAP values. Among continuous predictors, lower MNA-SF scores and lower self-management levels were associated with higher depression risk (positive SHAP values), whereas higher values of these features shifted SHAP values toward the negative direction. Conversely, higher DSI scores were associated with positive SHAP values, indicating elevated depression risk.

Figure 5.

SHAP Interpretation of the Naive Bayes Model. (A) Beeswarm plot of SHAP values across the five predictors on the test set. (B) Feature importance ranked by mean absolute SHAP value in descending order.

Notes: Each point in (A) represents one test-set observation; positive SHAP values (right of zero) increase the predicted probability of depression, negative values (left of zero) decrease it. Continuous variables are color-coded by feature value using a continuous viridis scale; binary variables are displayed as purple (absent) and yellow (present). SHAP values were computed via KernelSHAP with 100 Monte Carlo simulations and 50 background samples randomly drawn from the training set. All SHAP values were calculated on the test set.

Abbreviations: SHAP, Shapley Additive Explanations; DSI, Dialysis Symptom Index; MNA-SF, Mini Nutritional Assessment Short-Form.

Feature importance, quantified as the mean absolute SHAP value (Figure 5B), ranked the five predictors in descending order as follows: vision impairment, hearing loss, DSI, MNA-SF, and self-management. Vision impairment was the dominant predictor (mean |SHAP| > all other features), consistent with its identification as the most important variable in the multivariate logistic regression. Hearing loss ranked second, followed by DSI, MNA-SF, and self-management, reflecting a hierarchy in which sensory impairments exerted the strongest influence on model predictions, followed by dialysis-related symptom burden and nutritional status. SHAP values averaged across all five algorithms yielded a broadly consistent feature importance ranking (Figure S2). The per-algorithm importance ranking for each variable further confirmed that the predictor hierarchy was qualitatively stable across model types (Figure S3).

Partial Dependence Analysis

PDPs were generated from the Naive Bayes model to visualize the marginal effect of each continuous predictor on the predicted probability of depression, averaged over the distribution of all other predictors (Figure 6A–C). Each PDP is accompanied by a histogram indicating the empirical distribution of the feature in the training data.

Figure 6.

Visualizing Depression Risk Factors via Partial Dependence Plots. (A) Partial Dependence of Depression Probability on MNA-SF. (B) Partial Dependence of Depression Probability on DSI. (C) Partial Dependence of Depression Probability on Self-management level. Bivariate partial dependence interactions: (D) MNA-SF × DSI, (E) MNA-SF × Self-management, and (F) DSI × Self-management.

Abbreviations: MNA-SF, Mini Nutritional Assessment Short-Form; DSI, Dialysis Symptom Index.

The PDP for MNA-SF (Figure 6A) revealed a precipitous decline in predicted depression probability as nutritional scores increased from 2 to 12. At the 5th percentile of MNA-SF (score = 7), the predicted probability of depression was 0.937, which fell sharply to 0.618 at the 25th percentile (score = 10) and further to 0.464 at the median (score = 12). Beyond a score of 13, the probability stabilized near the minimum value of 0.418. The steepest gradient occurred between scores of 7 and 10, where each one-point increase in MNA-SF was associated with a substantial reduction in predicted depression risk. This threshold range (7–12) closely brackets the clinically established malnutrition cutoff (≤ 11), reinforcing that patients classified as malnourished or at risk of malnutrition by the MNA-SF fall within the highest-risk zone for depression.

The PDP for DSI (Figure 6B) demonstrated a continuous increase in predicted depression probability across the full range of symptom burden. The predicted probability rose from 0.438 at the 5th percentile (DSI = 3) to 0.442 at the 25th percentile (DSI = 7), 0.492 at the median (DSI = 11), 0.596 at the 75th percentile (DSI = 15), and 0.902 at the 95th percentile (DSI = 23). The rate of increase was modest in the lower range of DSI and accelerated noticeably between scores of 11 and 23, where the probability nearly doubled. At the maximum observed DSI value of 30, the predicted probability approached 1.0 (0.997). This monotonic upward trajectory indicates that each increment in dialysis-related symptom burden is associated with an increase in depression risk, with the effect becoming disproportionately pronounced among the most symptomatic patients.

The PDP for self-management level (Figure 6C) showed a more modest and gradual decline in predicted depression probability compared with DSI and MNA-SF. The predicted probability was 0.769 at the minimum score (20), declining to 0.644 at the 5th percentile (score = 33), 0.544 at the 25th percentile (score = 45), 0.516 at the median (score = 50), and 0.447 at the 95th percentile (score = 65). The overall range of predicted probabilities spanned approximately 0.34 (from 0.77 to 0.43), which was narrower than the ranges observed for DSI (0.56) and MNA-SF (0.58). This attenuated gradient is consistent with its lower SHAP importance ranking relative to the other predictors and suggests that self-management exerts a modest modulating effect on depression risk rather than serving as a dominant determinant.

Bivariate PDPs further visualized synergistic interaction patterns. The MNA-SF × DSI interaction (Figure 6D) produced the widest range of predicted depression probabilities among all pairwise combinations: the lowest predicted probability (0.292) occurred at MNA-SF = 14 and DSI = 5, whereas the highest (1.000) occurred at MNA-SF = 2 and DSI = 30, spanning nearly the full 0–1 probability scale. The MNA-SF × self-management interaction (Figure 6E) yielded a broadly similar range (0.315 to 1.000), while the DSI × self-management interaction (Figure 6F) exhibited the narrowest range (0.333 to 1.000). Across all three interaction plots, the combination of low nutritional status (MNA-SF < 8) and either high symptom burden (DSI > 20) or low self-management (score < 35) consistently mapped to predicted probabilities exceeding 0.90. These risk compounding patterns highlight the potential value of integrated interventions that simultaneously target nutritional support, symptom management, and self-management enhancement.

Discussion

To our knowledge, this is the first depression risk model for older MHD patients developed via comparative machine learning approaches rather than traditional methods. The results demonstrate that these machine learning algorithms can accurately identify individuals at risk of depression within this population. In this study, we analyzed the risk factors for depression in older MHD patients. Univariate logistic regression and LASSO regression with 10-fold cross-validation were used to select 5 core predictive features from the initial 21 candidate features, with backward stepwise multivariable logistic regression confirming their statistical significance. The top three predictors (vision impairment, hearing loss, and DSI) may facilitate efficient screening of depressive states in this population after further validation, potentially supporting early intervention. Furthermore, we integrated SHAP interpretability framework with PDPs to achieve dual-perspective analysis of the model’s internal mechanisms. The predictor hierarchy identified in this study may inform the prioritization of intervention targets for older adults at risk of depression, pending prospective validation. The internally validated prediction model provides a foundation for developing depression risk assessment tools for older MHD populations, though external validation is required before clinical deployment. We anticipate these findings will offer actionable insights for predicting psychological states in this vulnerable cohort.

Our findings demonstrate that, in a limited-sample cross-sectional study (n = 226), Naive Bayes achieved the most favorable overall performance profile among five candidate algorithms. Beyond its highest test-set AUC (0.865, 95% CI: 0.772–0.944), Naive Bayes exhibited substantially superior calibration (Brier = 0.151; calibration slope = 0.904), outperforming the remaining models whose slopes deviated further from unity (range: 1.031–1.624). This calibration advantage is clinically consequential: well-calibrated models produce probability estimates that can be directly interpreted as risk levels, enabling clinicians to set meaningful decision thresholds for intervention. The biological plausibility of Naive Bayes’ probabilistic framework, which treats predictor–outcome relationships as conditional probabilities, may further contribute to its robust performance given the inherently probabilistic nature of depression risk in this population. DeLong’s test confirmed no statistically significant differences between any pair of models (all P > 0.15; Table S2), indicating that the observed variations in AUC likely reflect sampling variability rather than genuine algorithmic superiority. This pattern is consistent with the broader machine learning literature, which often finds that simpler models perform comparably to more complex ensemble methods in small-to-moderate sample settings with a limited set of well-selected predictors. Notably, the Random Forest model exhibited a train–test AUC discrepancy (ΔAUC = 0.075), consistent with modest overfitting, which further supports the observation that complex ensemble models do not necessarily confer advantages in this context. The Naive Bayes model additionally demonstrated the best calibration (Brier = 0.151; calibration slope = 0.904), a finding that warrants attention as calibration is critical for clinical decision-making. For subsequent interpretability analysis, we selected the Naive Bayes model, which provided the optimal combination of discrimination (test AUC = 0.865), calibration (Brier = 0.151; calibration slope = 0.904), and probabilistic interpretability. Its calibration slope of 0.904, the only model approaching the ideal value of 1.0, indicates that its predicted probabilities are well-aligned with observed event rates, a prerequisite for reliable risk stratification in clinical settings. Nevertheless, we emphasize that the choice of preferred model cannot be firmly established based on the current data alone; external validation in independent cohorts is required to determine which algorithm generalizes most robustly.

The results of this study indicated that the prevalence of GDS-15-defined depressive symptoms among older patients receiving MHD was 54.4%. A recent global meta-analysis by Adejumo et al reported a pooled depression prevalence of 26.5% in the general CKD population, rising to 29.9% among hemodialysis patients.¹¹ The substantially higher prevalence observed in our cohort likely reflects the combined effects of advanced age, the use of a screening instrument (GDS-15) rather than diagnostic criteria, and regional or cultural factors. A previous study involving 107 MHD patients provided a stratified analysis by age,¹⁵ and its findings showed that the prevalence of depression was 91.84% among 49 older patients, which was significantly higher than the 77.59% prevalence observed in non- older patients. The depression prevalence reported in that study was higher than that in our survey. This discrepancy may be attributed to two main factors: first, the sample size included in that study was smaller; second, the Beck Depression Inventory (BDI) used in it is more suitable for general inpatients rather than the older population.⁵⁰ There is limited literature on depression in the older with MHD, but most studies confirm that age is a risk factor for depression in MHD patients.¹⁸ Meanwhile, a large body of research has identified female gender as a risk factor for depression in patients receiving MHD.²⁰ Although “gender” was not identified as a predictive factor in the present study, statistical results showed that the prevalence of depression among older female MHD patients (62.63%) was significantly higher than that among older male MHD patients (48.03%), which further corroborates this perspective. Overall, the prevalence of depression among older patients receiving MHD remains at a relatively high level. As previously noted, MHD patients themselves, due to the chronic nature of ESRD and dependence on dialysis, are prone to developing a range of psychological issues. Many older patients require accompaniment by family members to travel to hospitals multiple times a week for HD treatments lasting several hours each time. This prolonged process not only causes fatigue in patients but also imposes a similar burden on accompanying family members.⁵¹ Such persistent caregiving burden may induce “self-stigma” in older patients, manifested as intense feelings of guilt about becoming a burden to their families, which ultimately exacerbates the psychological distress of the older MHD population.¹⁵ Greater attention should be paid to the decline in mental health among MHD patients, particularly the older subgroup. We advocate for the integration of depression risk prediction models into the management system for older MHD patients to enable early screening.

In our study results, vision impairment was the most important predictor of depression in older patients with MHD, even more important than DSI and self-management level, which are two unique indicators of MHD. Older MHD patients may develop vision impairment due to diabetic nephropathy or uremic optic neuropathy, which in turn leads to loss of independence and limitations in activities of daily living, further increasing feelings of frustration in this population.⁵² Studies have shown that the probability of depression in older individuals with vision impairment is 43% higher than that in those with normal vision.⁵³ Furthermore, older individuals who reported self-reported vision impairment at baseline were more likely to present with persistent depressive symptoms during the 1-year follow-up.⁵⁴ These findings are consistent with the results of the present study, further validating the reliability of the proposed model. Studies have demonstrated a significant association between self-reported vision impairment and severe depressive symptoms, which is independent of objectively measured visual acuity.⁵⁵ Frank et al⁵⁶ argued that self-reported vision impairment may be a better measure of changes in mental health symptoms, describing it as a unique and valuable concept. This study confirms that this view is also applicable in the older MHD population, suggesting that subjective perception in the older may be more predictive of mental health outcomes than biological impairment itself. Nevertheless, several methodological considerations warrant caution in interpreting the magnitude of this association. Because both vision impairment and depressive symptoms were assessed via self-report, shared method variance may partially inflate the observed relationship, and the ranking of vision impairment above multi-item validated instruments (DSI and self-management) could partly reflect measurement properties rather than a genuinely larger effect size. The fact that vision impairment is a single-item measure, which tends to produce more variable responses than multi-item scales, further complicates direct comparisons of predictor importance. Future studies incorporating objective vision assessments (e.g, visual acuity testing) alongside self-reported measures would help disentangle the relative contributions of subjective perception and biological impairment to depression risk in this population, as advocated by Frank et al.⁵⁶

Hearing loss was the second most important predictive factor in this study. Hearing loss may induce negative emotions in older MHD patients through communication barriers and social withdrawal.^57,58 Currently, although there is evidence suggesting that self-reported hearing loss in older adults is associated with an increased risk of depression,⁵⁹ conclusions remain controversial across different studies. Studies have indicated an association between self-reported vision impairment and depression, whereas no such association has been found with self-reported hearing loss.⁵⁷ Of note, this study found that the ranking of hearing loss in terms of importance in SHAP analysis(2nd place) was lower than that of vision impairment, and this finding provides a valuable reference for the aforementioned controversy. Furthermore, hearing loss in older MHD patients remains non-negligible. Renal impairment in MHD patients can also affect hearing function. Studies have shown that the ear and kidney begin to develop during the same stage of embryonic development, sharing numerous similarities in anatomical structure, physiological characteristics, and even genomic profiles.⁶⁰ There are parallels between the mechanisms of the inner ear and renal impairment,⁶¹ and most nephrotoxic drugs also exhibit ototoxicity. Consequently, Yuan et al⁵⁸ have hypothesized the potential existence of the “ear-kidney homology” phenomenon. As with vision impairment, the reliance on a single-item self-reported measure of hearing loss should be considered when interpreting its rank among predictors.

A higher DSI score indicates a greater number of dialysis-related complications, which was associated with an increased risk of depression. Dialysis-induced refractory symptoms (e.g, constipation, dizziness, difficulty falling asleep) not only directly cause physical distress, but also lead to multidimensional loss of patients’ physical endurance, social role performance, and life satisfaction due to persistent symptoms, thereby undermining psychological well-being.^62,63 Meanwhile, the weekly dependence on hemodialysis compels patients to endure unavoidable physical discomfort, which predisposes them to negative emotions and ultimately leads to the development of depression.^20,51 The PDP for DSI confirmed a monotonic increase in predicted depression probability across the full range of symptom burden, from 0.438 at the 5th percentile (DSI = 3) to 0.902 at the 95th percentile (DSI = 23) and approaching 1.0 at the maximum observed DSI score of 30. The rate of increase accelerated notably between DSI scores of 11 and 23, suggesting that the effect of symptom burden on depression risk becomes disproportionately pronounced among the most symptomatic patients.

MNA-SF was the fourth most important predictive factor in the SHAP analysis. A higher MNA-SF score was associated with a relatively lower risk of depression in older MHD patients. For patients requiring MHD treatment, the protective effect of nutritional status against depression exhibits unique characteristics. On one hand, hemodialysis itself leads to substantial loss of neuroprotective substances such as amino acids and water-soluble vitamins, which in turn causes nerve damage and impairs emotional regulation.^63–65 On the other hand, MHD patients are required to adhere to strict dietary restrictions (including phosphorus, potassium, and sodium limitations) and fluid intake control. Such restrictions easily result in insufficient protein intake, which not only reduces MNA-SF scores but also leads to a monotonous and distressing dietary experience, thereby inducing the development of depressive feelings.^63,66 Furthermore, in the present study, the MNA-SF was treated as a continuous variable, with no specific score set as the cutoff for malnutrition. As clearly illustrated in the PDP, a significant change in the probability of depression among patients was observed when the MNA-SF score ranged between 10 and 12. This finding not only suggests that timely intervention for patients with an MNA-SF score of 10–12 may enhance the risk-benefit ratio but also indicates that this score range (10–12) is highly consistent with the clinically commonly used malnutrition threshold (less than or equal to 11).⁶⁷ Such an observation supports the rationality of the existing cutoff value, demonstrating that a score of ≤11 falls exactly within the critical transition region where the depression risk increases at an accelerated rate. However, it should be emphasized that this finding only indicates the possibility of changes in risk patterns based on MNA-SF scores and cannot replace validation studies of identification criteria.

Beyond the behavioral and symptom-based pathways discussed above, chronic inflammation and oxidative stress represent important biological mechanisms linking MHD to depression. Han et al recently reported that the systemic immune-inflammation index (SII), a composite marker integrating peripheral blood cell counts, was independently associated with a 4.7-fold increased risk of depression in MHD patients.⁶⁸ Although direct evidence for oxidative stress pathways in MHD-related depression remains limited, disturbances in pro/antioxidant balance and trace element homeostasis have been implicated in mood disorders more broadly. Rog et al demonstrated that a machine learning model incorporating oxidative stress markers (4-hydroxynonenal and glutathione peroxidase) distinguished patients with bipolar disorder from healthy controls with an AUC of 90.5%, highlighting the potential diagnostic value of redox-related biomarkers.⁶⁹ While the present study did not include laboratory-based inflammatory or oxidative stress markers among the final predictors—LASSO preferentially retained variables with the strongest and most consistent signal in this sample—future studies incorporating these biomarkers alongside the current clinical predictors may further improve model performance and shed light on the biological underpinnings of depression risk in older MHD patients.

In older MHD patients, a high self-management level buffers the risk of depression through three dimensions: enhancing treatment control (to reduce complications), optimizing symptom coping, and delaying role deprivation.^70,71 Essentially, it is the behavioral manifestation of executive willingness, whereby patients still retain autonomy over their lives despite disease-related constraints, delaying the erosion of “social identity” by “patient identity”.⁷² Existing studies have utilized a “Transitional Care Unit”⁷³ to ensure that patients newly initiating hemodialysis receive appropriate preparation. This approach enables patients to engage more actively in their own care, make informed decisions regarding the choice between in-center hemodialysis and home hemodialysis, and simultaneously improve their self-management level as MHD patients.

Our study also holds novel clinical implications. The findings enhance the understanding of interrelationships among the five identified predictive factors and depression: increased dialysis symptom burden, malnutrition, and poor self-management ability are each associated with a higher risk of depression. This suggests that clinical interventions should adopt a comprehensive strategy—simultaneously targeting symptom management, nutritional support, and self-management enhancement—rather than addressing these factors in isolation. Our results therefore provide evidence-based support for stratified interventions over the traditional “one-size-fits-all” model.⁶⁴ Following external validation, the proposed model may serve as a convenient screening tool to assist healthcare providers in identifying older MHD patients at elevated risk of depression during routine care. A recent cross-sectional study by Kubanek et al demonstrated that routine depression screening in hemodialysis units, followed by psychiatric consultation for screen-positive patients, is both feasible and effective for improving diagnostic accuracy,⁷⁴ lending further support to the integration of risk-prediction models as a first-step triage tool. In resource-constrained settings, such an approach could enable early warning and facilitate precise prevention and control of depression in this vulnerable population.⁵³

Our study had several strengths. First, to our knowledge, this is the first depression risk prediction model for older MHD patients. Second, breaking the limitation of traditional single models, our study employed five machine learning algorithms in parallel and further validated the performance advantage of Naive Bayes, particularly its calibration, in small samples (n=226). Furthermore, this study analyzed the model from two perspectives—quantification of individual contribution via SHAP and trend visualization through PDP—endowing the model with clinical interpretability and mechanistic discoverability. These two approaches work synergistically to overcome the limitations of “black-box” models, advancing the prediction model from a statistical tool to a decision-making engine and enhancing the efficacy of clinical depression prevention and control.

This study has limitations. First, the model only underwent internal validation, with no independent external cohort; its generalizability thus requires verification in more diverse settings. Second, the sample size (n = 226, 123 events) was adequate under criterion 2 of the pmsampsize framework but fell below criterion 1, which targets minimal shrinkage; this may modestly affect the stability of performance estimates at extreme predicted probabilities. Nevertheless, the five-predictor model (events-per-predictor ≈ 24.6) and the use of bootstrap resampling (2,000 iterations) for confidence interval estimation partially mitigate the imprecision associated with the limited sample. Third, this study adopted a cross-sectional design, precluding causal inference and leaving open the possibility of reverse causation between depression and self-reported predictors. Fourth, both the outcome (GDS-15) and key predictors (vision impairment, hearing loss) relied on self-report, which may introduce shared method variance; the GDS-15 is a screening instrument rather than a diagnostic tool; and sensory function was assessed using single-item questions rather than standardized instruments. Fifth, the DSI includes five psychological symptom items (of 30 total) that address constructs related to mood; however, this overlap is partial given that GDS-15 captures a broader depressive syndrome, and DSI’s predictive contribution was predominantly driven by its 25 physical symptom items. The fact that DSI was retained by LASSO—which aggressively shrinks redundant predictors—further indicates that its physical symptom content carried independent predictive value beyond any shared variance with the outcome. In summary, future studies should verify the current model in independent cohorts with larger sample sizes across multiple centers, incorporate objective assessments of sensory function, employ prospective designs, and explore additional potentially meaningful predictors to further optimize model performance.

Conclusion

The results of this study showed that the prevalence of GDS-15-defined depressive symptoms among older MHD patients was 54.4%. Using five machine learning methods with internal validation, we developed depression risk prediction models for this population, addressing an important research gap. DeLong’s test revealed no statistically significant differences among the five algorithms (all P > 0.15), suggesting that simpler models may perform comparably to more complex alternatives. For interpretability analysis, Naive Bayes achieved the optimal overall performance (test AUC = 0.865, 95% CI: 0.772–0.944; Brier = 0.151; calibration slope = 0.904). Five predictive variables were identified, with SHAP-based importance ranked as: vision impairment, hearing loss, DSI, MNA-SF, and self-management. However, given the single-center design, modest sample size, lack of external validation, and reliance on self-reported measures, these models should be considered preliminary and are not yet ready for clinical implementation. Future multicenter prospective studies with external validation, objective sensory assessments, and refined predictor sets are required before these models can be deployed as clinical screening tools.

Funding Statement

This study was supported by the China National Key R&D Program (grant number 2020YFC2008500); the Medical Research Union Fund for High-quality health development of Guizhou Province (grant number 2024GZYXKYJJXM0163).

Abbreviations

Alb, Serum Albumin; AUC, Area Under the Receiver Operating Characteristic Curve; Ca, Serum Calcium; CKD, Chronic Kidney Disease; Cr, Serum Creatinine; DCA, Decision Curve Analysis; DSI, Dialysis Symptom Index; ESRD, End-Stage Renal Disease; GDS-15, 15-item Geriatric Depression Scale; Hb, Hemoglobin; HD, Hemodialysis; K, Serum Potassium; KNN, K-Nearest Neighbors; LASSO, Least Absolute Shrinkage and Selection Operator; LR, Logistic Regression; MHD, Maintenance Hemodialysis; MLP, Multilayer Perceptron; MNA-SF, Mini Nutritional Assessment Short-Form; PDP, Partial Dependence Plot; ROC, Receiver Operating Characteristic; SHAP, Shapley Additive Explanations; SPPB, Short Physical Performance Battery; VIF, Variance Inflation Factor.

Ethics Approval and Consent to Participate

This study was reviewed by the Ethical Review Committee of the Affiliated Hospital of Zunyi Medical University (KLLY-2024-045). The study was conducted in accordance with the principles of the Declaration of Helsinki. All the participants provided signed informed consent at the time of participation.

Disclosure

The authors declare no competing interests in this work.