Machine learning-based risk predictive models for depression in patients with diabetes: a systematic review and meta-analysis.
Currently, numerous studies have employed machine learning (ML) methods to develop predictive models for depression risk in patients with diabetes mellitus (DM); however, the findings remain inconsistent. Therefore, this study aims to clarify the current state of research and emerging trends in this field by systematically evaluating the performance, strengths, and limitations of existing prediction models.
This systematic review evaluates the performance and clinical applicability of ML-based depression risk prediction models for patients with DM, providing reliable evidence to assist healthcare professionals in selecting and optimizing more appropriate prediction models.
We conducted a systematic search of clinical studies employing ML approaches to predict depression risk in patients with DM across the PubMed, Embase, Cochrane Library, and Web of Science databases, from their inception to January 2026. The primary performance metric for the models was the area under the receiver operating characteristic curve (AUC) along with its 95% confidence interval (95% CI). Two independent researchers screened the literature, extracted data, and used PROBAST-AI to assess the risk of bias and clinical applicability of the included studies. Pooled AUC was estimated using the Der Simonian and Laird random-effects model.
A total of 14 studies comprising 64 distinct ML models were included. All included studies were assessed as high risk of bias and high clinical applicability. A pooled analysis of the best-performing ML prediction models reported in each study showed a pooled AUC of 0.822 (95% CI, 0.789-0.858), indicating relatively good overall predictive performance. However, there was substantial heterogeneity among the studies (I² = 97.4%; P < 0.001). Subgroup analysis based on ML model types revealed the following pooled AUC values: 0.765 (95% CI 0.706-0.829) for traditional regression models, 0.789 (95% CI 0.747-0.834) for general machine learning models, and 0.802 (95% CI 0.769-0.836) for deep learning models. Notably, logistic regression (LR) (n = 10) was the most frequently employed ML method for developing depression risk prediction models in patients with DM. To evaluate model generalizability and avoid overfitting, the included studies adopted three validation strategies: 5-fold cross-validation yielded a pooled AUC of 0.913 (95% CI 0.781-1.067), 10-fold cross-validation yielded 0.819 (95% CI 0.781-0.858), and random split validation yielded 0.747 (95% CI 0.648-0.862). The most commonly used predictors in the included models were age, sex, and body mass index (BMI), which are readily available in clinical settings and strongly associated with depression risk.
ML-based depression risk prediction models for patients with DM demonstrate overall satisfactory predictive performance. However, most existing studies had relatively small sample sizes and lacked external validation. Future research should prioritize refining study design and optimizing clinical data processing to improve the generalizability and stability of these models in clinical practice.
https://www.crd.york.ac.uk/PROSPERO/view/CRD420251243343, identifier CRD420251243343.
This systematic review evaluates the performance and clinical applicability of ML-based depression risk prediction models for patients with DM, providing reliable evidence to assist healthcare professionals in selecting and optimizing more appropriate prediction models.
We conducted a systematic search of clinical studies employing ML approaches to predict depression risk in patients with DM across the PubMed, Embase, Cochrane Library, and Web of Science databases, from their inception to January 2026. The primary performance metric for the models was the area under the receiver operating characteristic curve (AUC) along with its 95% confidence interval (95% CI). Two independent researchers screened the literature, extracted data, and used PROBAST-AI to assess the risk of bias and clinical applicability of the included studies. Pooled AUC was estimated using the Der Simonian and Laird random-effects model.
A total of 14 studies comprising 64 distinct ML models were included. All included studies were assessed as high risk of bias and high clinical applicability. A pooled analysis of the best-performing ML prediction models reported in each study showed a pooled AUC of 0.822 (95% CI, 0.789-0.858), indicating relatively good overall predictive performance. However, there was substantial heterogeneity among the studies (I² = 97.4%; P < 0.001). Subgroup analysis based on ML model types revealed the following pooled AUC values: 0.765 (95% CI 0.706-0.829) for traditional regression models, 0.789 (95% CI 0.747-0.834) for general machine learning models, and 0.802 (95% CI 0.769-0.836) for deep learning models. Notably, logistic regression (LR) (n = 10) was the most frequently employed ML method for developing depression risk prediction models in patients with DM. To evaluate model generalizability and avoid overfitting, the included studies adopted three validation strategies: 5-fold cross-validation yielded a pooled AUC of 0.913 (95% CI 0.781-1.067), 10-fold cross-validation yielded 0.819 (95% CI 0.781-0.858), and random split validation yielded 0.747 (95% CI 0.648-0.862). The most commonly used predictors in the included models were age, sex, and body mass index (BMI), which are readily available in clinical settings and strongly associated with depression risk.
ML-based depression risk prediction models for patients with DM demonstrate overall satisfactory predictive performance. However, most existing studies had relatively small sample sizes and lacked external validation. Future research should prioritize refining study design and optimizing clinical data processing to improve the generalizability and stability of these models in clinical practice.
https://www.crd.york.ac.uk/PROSPERO/view/CRD420251243343, identifier CRD420251243343.