A computer-based search was conducted across four English databases (Web of Science, PubMed, Embase, and Cochrane Library) and four Chinese databases (CNKI, Wanfang, VIP, and SinoMed) to identify studies related to depression risk prediction models in CHD patients. The search period spanned from database inception to September 29, 2024. A combination of subject terms, free terms, and Boolean operators was used for both Chinese and English searches. Chinese search terms included coronary heart disease, acute coronary syndrome, myocardial infarction, post-PCI, depression, depressive state, prediction, predictive factors, influencing factors, risk assessment, model, and tool. English search terms included coronary disease, acute coronary syndrome, myocardial infarction, post-PCI, depression, depressive disorder, prediction, predictors, influencing factors, risk assessment, model, tool, and score. For each database, a tailored search strategy was developed based on its unique features. Additionally, references in included studies were reviewed to identify supplementary relevant literature.

Two researchers independently screened the literature and extracted data based on the inclusion and exclusion criteria, followed by cross-checking the data results. Any disagreements were resolved through discussion or decided by a third researcher. The extracted data included the following details: first author, year, country, study type, study population, depression diagnostic criteria, sample size, modeling method, area under the receiver operating characteristic curve (AUC), and model presentation format.

The Prediction model Risk Of Bias ASsessment Tool (PROBAST) (10) was used to assess the risk of bias and applicability of the included models. The risk of bias evaluation covers four domains: participants, predictors, outcomes, and analysis, with a total of 20 specific questions. Based on the “shortcoming theory,” each domain's results were synthesized as follows: if all items are marked as “probably yes” or “yes,” the domain is rated as “low risk”; if any item is rated as “no” or “probably no,” the domain is deemed “high risk”; if an item lacks sufficient information, the domain is rated as “unclear.” For the overall risk of bias, a “low risk” is assigned only when all four domains are rated as “low risk”; if any domain is rated as “high risk,” the overall risk of bias is rated as “high risk”; if any domain is rated as “unclear,” the overall risk of bias is categorized as “unclear.” The applicability assessment includes three domains: study population, predictors, and outcomes, using the same evaluation approach as the risk of bias assessment.

Stata 17.0 software was used for quantitative analysis of the AUCs of the included models. Cochrane's Q test was applied to assess heterogeneity among the models, with I² used to quantify the degree of heterogeneity. If P > 0.05 and I² ≤ 50%, it indicates no significant heterogeneity among studies, and a fixed-effects model is applied. Conversely, if P ≤ 0.05 or I² > 50%, it suggests substantial heterogeneity. In such cases, subgroup analysis is conducted to explore the sources of heterogeneity, and sensitivity analysis is performed by sequentially excluding individual studies. If heterogeneity persists, a random-effects model is used for the analysis. To detect potential publication bias, Egger's test was performed, with P > 0.05 suggesting a low likelihood of publication bias.

The initial search yielded 2,421 relevant articles. After removing 792 duplicates, a further 1,590 articles were excluded based on title and abstract screening for topic relevance. Full texts of 39 articles were reviewed, and ultimately, 8 articles were included. The screening process is shown in Figure 1.

This study included a total of 8 studies (8, 9, 11–16), comprising 3 prospective cohort studies (8, 12, 14), 4 retrospective studies (9, 13, 15, 16) (including 2 retrospective cohort studies (13, 16) and 2 retrospective case-control studies (9, 15), with one being a nested case-control study (15), and 1 cross-sectional study (11). Data for 2 studies (9, 11) were sourced from the U.S. Centers for Disease Control and Prevention's National Health and Nutrition Examination Survey (NHANES), while data for the remaining 6 studies (8, 12–16) were drawn from clinical databases. A total of 8,035 CHD patients were included in the analysis. Details are provided in Table 1.

The 8 studies constructed a total of 13 prediction models. Hou et al. developed 5 models and ultimately selected an optimal model for nomogram construction, while Wang et al. developed 2 models; each of the remaining studies constructed 1 model. In terms of variable selection, Hou et al.'s optimal model used the best subset selection method, Li Cexing et al. applied Lasso regression, and the other 6 studies utilized univariate analysis to select variables. Regarding modeling methods, Wang et al. employed Lasso regression in their second model, while all other models used logistic regression. In terms of model performance, 8 models reported the area under the curve (AUC), ranging from 0.772 to 0.961; 1 model reported the C-index. Calibration was assessed in 7 studies through calibration plots, 1 study reported the Brier score, and 3 studies conducted the Hosmer-Lemeshow test (Figure 2). Details are provided in Table 2.

For model validation, only 2 studies performed external validation, while 5 conducted internal validation. Regarding model presentation, 7 studies used nomograms, and 1 used regression equations. The number of candidate predictors ranged from 9 to 27, with the final number of included predictors between 3 and 10 (distribution of the top five predictive factors, Figure 3). Additional details are available in Table 3.

Among the 8 included studies with a total of 9 models, significant heterogeneity was observed (I² = 90.2%, P < 0.001), prompting the use of a random-effects model for analysis. The pooled AUC estimate was 0.84 (95% CI: 0.80–0.88), indicating good overall predictive performance (Figure 4a). The substantial heterogeneity among studies may be attributed to differences in study design, patient characteristics, and outcome evaluation metrics. Subgroup analysis based on different outcome measurement scales showed that heterogeneity was I² = 74.4% for the PHQ-9 group and I² = 95.1% for the HAMD group, with no significant reduction in heterogeneity, suggesting that measurement tools were not the primary source of heterogeneity (Figure 4b). Sensitivity analysis, conducted by sequentially excluding individual studies, revealed no significant changes in heterogeneity, further supporting the robustness of the findings. Egger's test yielded a P-value of 0.767 (P > 0.05), indicating no significant publication bias.

The commonly used metrics for evaluating model performance include AUC and calibration. In this study, all 9 included models demonstrated AUCs above 0.7, with 6 models reaching or exceeding 0.8, indicating good discriminative ability. Additionally, the calibration curves for 8 models closely aligned with the diagonal line, suggesting strong agreement between predicted probabilities and actual occurrence rates. Four models also underwent Hosmer-Lemeshow (H-L) testing, with P-values ≥0.05, further supporting good calibration. Overall, these models effectively identify high-risk patients for depression, showing favorable performance; however, the high risk of bias persists.

First, there is bias in the data sources. (1) Most data are from single-center studies with insufficient sample sizes, and five studies are retrospective analyses, introducing recall bias to some extent and affecting the model's quality. (2) Depression diagnosis is highly subjective and requires professional assessment; however, five studies did not mention the training of evaluators, which may lead to outcome bias. Future studies should prioritize multicenter, large-sample, high-quality prospective studies to minimize recall bias. Additionally, outcome evaluators should be uniformly trained independent third parties to enhance the consistency and accuracy of depression diagnoses.

Second, biases also exist in model construction. (1) In terms of variable selection, six studies relied on univariate analysis to identify predictors, potentially omitting important factors and leading to model overfitting, which weakens predictive power (17). (2) Six studies did not report missing values, and one study directly excluded missing data. This approach may introduce bias in the associations between predictors and outcomes, and even in the absence of bias, missing data can reduce precision, widening confidence intervals. (3) Six studies used traditional logistic regression to build models, limiting the ability to capture complex relationships among variables, which affects model accuracy and stability (18). Therefore, future research should combine domain knowledge and clinical experience and cautiously select variables using methods like LASSO regression and stepwise regression (19). When handling missing data, methods such as multiple imputation and single imputation should be used to mitigate the adverse effects of missing data on statistical analysis and model stability (20). Additionally, incorporating machine learning and deep learning approaches can enhance the accuracy and adaptability of predictive models (21).

Finally, there are some limitations in model validation. The predictive performance of a model can be affected by variations in populations and regions, underscoring the need for thorough validation during model development. Internal validation assesses model reproducibility and prevents overfitting, while external validation evaluates transferability and generalizability, regarded as the “gold standard” of validation (22, 23). In this study, three models underwent external validation but lacked internal validation, potentially impacting model performance and reliability; five models performed internal validation without external validation, with study populations mainly composed of Chinese individuals, limiting the model's generalizability and applicability. Future research should emphasize external validation, particularly across diverse regions, ethnicities, cultural backgrounds, and lifestyle factors, to enhance model generalizability. Additionally, variations in coronary heart disease types (e.g., stable angina, acute coronary syndrome, post-PCI), as well as different disease stages, should be considered. Treatment modalities and levels of social support may also influence predictive performance. Taking these factors into account comprehensively will contribute to improving model reliability and applicability.

The nine models in this study included between 3 and 10 predictors, primarily categorized into four groups: demographic factors (e.g., age, gender), psychological factors (e.g., PHQ score, personality), clinical factors (e.g., hypertension, number of stents), and lifestyle factors (e.g., smoking, alcohol consumption). Despite variations in predictor selection due to study types and included variables, some commonalities were identified. Predictors frequently appearing across eight studies included age, education level, gender, and cardiac function classification. Wang's study (9) found a negative correlation between age and depression, a finding supported by Murphy et al. (24). Conversely, Zhu Hupei (13) and Chen Hongyu (15) indicated that age over 60 is an independent risk factor for post-PCI depression. This may be because younger patients feel disoriented by sudden illness, while older patients, experiencing functional decline, may perceive themselves as a burden. Research (12) suggests that patients with higher education levels have stronger comprehension and application abilities when receiving health education, resulting in better prognosis and lower risk of depression. Multiple studies (25–27) found that females are more prone to depression, likely due to hormonal fluctuations during menstruation, menopause, and perinatal periods, which contribute to emotional instability. Additionally, women often bear more responsibilities and pressures in social and family roles, increasing depression risk. Higher cardiac function classifications are associated with more pronounced symptoms of dyspnea and chest tightness, as well as greater limitations in daily life and physical activity. Such physiological discomfort amplifies psychological stress, eroding confidence in life, and leading to negative emotions or even self-harm and suicidal behaviors (28). Thus, early screening should focus on these common factors to promptly identify high-risk individuals. However, this study also found that many current predictors are challenging to directly intervene upon, limiting nursing interventions. Future research should consider including modifiable factors, such as sleep quality, psychological state, and medication adherence, to enable more targeted nursing interventions and enhance clinical outcomes.

(1) Seventy-five percent (6/8) of the studies included were based on data from China, which may introduce regional bias and limit the applicability of the findings to Western populations. (2) Seventy-five percent (6/8) of the studies did not perform external validation, restricting the generalizability of the models. (3) The included studies were limited to those published in Chinese and English, potentially introducing language bias and failing to capture findings from studies published in other languages.