ML Models Predicting Depression Risk in Diabetic Patients: Systematic Review & Meta-Analysis
Overview
This systematic review and meta-analysis evaluated 64 machine learning models from 14 studies predicting depression risk in diabetic patients. The pooled best-model performance showed good predictive ability with an AUC of 0.822, though all studies had high risk of bias and clinical applicability concerns.
Background
Depression is a common comorbidity in patients with diabetes mellitus (DM), adversely affecting disease management and outcomes. Early identification of patients at risk for depression can facilitate timely interventions. Machine learning (ML) models have been increasingly applied to predict depression risk using clinical and demographic data. However, the performance and clinical utility of these ML models in diabetic populations require systematic evaluation to guide healthcare professionals in model selection and optimization.
Data Highlights
Model Type
Pooled AUC (95% CI)
Traditional Regression Models
0.765 (0.706-0.829)
General Machine Learning Models
0.789 (0.747-0.834)
Deep Learning Models
0.802 (0.769-0.836)
5-fold Cross-validation
0.913 (0.781-1.067)
10-fold Cross-validation
0.819 (0.781-0.858)
Random Split Validation
0.747 (0.648-0.862)
Overall Best Models
0.822 (0.789-0.858)
Key Findings
64 ML models from 14 studies were analyzed, all with high risk of bias and clinical applicability concerns.
The pooled AUC for the best-performing models was 0.822, indicating relatively good predictive performance.
Logistic regression was the most commonly used ML method (10 studies) for depression risk prediction in diabetic patients.
Subgroup pooled AUCs were 0.765 for traditional regression, 0.789 for general ML, and 0.802 for deep learning models.
Validation strategies influenced performance: 5-fold cross-validation showed highest pooled AUC (0.913), followed by 10-fold (0.819) and random split (0.747).
Common predictors included age, sex, and body mass index, which are easily accessible in clinical practice.
Clinical Implications
Clinicians should consider ML models, particularly logistic regression-based ones, as useful tools for identifying diabetic patients at risk of depression. However, given the high risk of bias and heterogeneity among studies, caution is warranted in interpreting model predictions. Incorporating routinely collected predictors such as age, sex, and BMI can facilitate practical implementation of these models in clinical settings.
Conclusion
ML models demonstrate promising accuracy for predicting depression risk in diabetic patients, but methodological limitations and heterogeneity highlight the need for further validation and optimization before widespread clinical adoption.
References
Systematic Review and Meta-Analysis of Machine Learning Models for Predicting Depression Risk in Diabetic Patients
