AI Spots Which Older Diabetics Are Most at Risk of Depression

Depression and diabetes are two of the most burdensome chronic conditions of aging, and when they arrive together, the consequences are far worse than either alone. Older adults with diabetes mellitus who also develop depressive symptoms face poorer blood sugar control, higher rates of complications, greater treatment nonadherence, and increased healthcare costs. Yet depression in this population often goes undetected, buried under the everyday fatigue and physical complaints that accompany both aging and chronic disease. A new study published in Current Psychology by Yuran Qiu of Renmin University of China takes a distinctly modern approach to this old problem: training machine learning algorithms to flag which older Chinese adults with diabetes are most likely to be experiencing high-risk depressive symptoms.

The research drew on data from 825 older adults with diabetes mellitus, using a publicly available national database as its foundation. Depressive symptoms were measured with the 10-item Center for Epidemiological Studies Depression Scale, or CESD-10, a widely validated screening instrument designed specifically for use in older populations. Rather than relying on a single statistical technique, the study compared five different machine learning algorithms, each competing to see which could most accurately classify participants as high or low risk based on their demographic, health, and lifestyle characteristics.

The technical evaluation was rigorous. The models were judged on six complementary metrics: accuracy, sensitivity, specificity, F1 score, positive predictive value, and the Brier score, which quantifies how well-calibrated a model’s probability estimates are. This matters because a classifier can look impressive on accuracy alone while quietly failing at the task that matters most in clinical screening—correctly catching the people who truly are at risk. Sensitivity and the F1 score, which balances precision with recall, are therefore the numbers clinicians watch most closely.

Among the five full-feature algorithms, XGBoost emerged as the clear winner in terms of balanced overall performance. XGBoost, short for extreme gradient boosting, is an ensemble method that builds a sequence of decision trees, with each new tree trained to correct the errors of the previous ones. It has become a workhorse of applied machine learning because it handles heterogeneous data types, captures nonlinear interactions between variables, and resists overfitting through regularization. In this study, XGBoost achieved an accuracy of 0.843, a sensitivity of 0.775, an F1 score of 0.802, and a Brier score of 0.141—figures that indicate a model capable of correctly identifying roughly three out of four high-risk individuals while keeping false alarms manageable.

But a high-performing black box is of limited use in medicine if nobody can see inside it. To open up the model, the study employed SHapley Additive exPlanations, or SHAP values, a technique borrowed from cooperative game theory that assigns each input feature a precise contribution to every individual prediction. SHAP analysis has become the standard tool for interpretable machine learning because it satisfies mathematical fairness properties: the contributions of all features sum exactly to the model’s output, and each feature’s contribution reflects its marginal effect across all possible combinations of the other variables.

The most striking finding to emerge from the SHAP analysis was that the single most influential feature driving XGBoost’s predictions was whether participants reported feeling energetic. This is more than a statistical curiosity. Energy and vitality sit at the intersection of physical and mental health in later life—fatigue is a recognized early indicator of aging processes, a common somatic symptom of depression, and a frequent complaint in diabetes itself. Its centrality in the model suggests that a simple question about energy levels, easily asked in any clinic, carries outsized diagnostic information about depression risk in older diabetic patients.

The study’s population focus is significant in its own right. China is aging rapidly, and diabetes prevalence has climbed steeply over recent decades, making the country home to one of the largest populations of older adults living with diabetes in the world. Meta-analytic evidence indicates that depressive symptoms are common among older adults in mainland China, and research consistently shows that comorbid depression in diabetes is associated with worse self-management, more complications, and higher mortality. Screening tools that can be deployed cheaply and at scale are therefore a pressing public health need, particularly in primary care settings where time and specialist mental health resources are scarce.

It is worth noting that the study is cross-sectional, meaning it identifies associations at a single point in time rather than proving that the measured features cause or predict future depression. The author also acknowledges a broader debate in the clinical prediction literature: systematic reviews have found that machine learning does not always outperform traditional logistic regression for clinical prediction tasks. What this study adds is a demonstration that, for this specific and previously understudied population—older adults with diabetes—ensemble methods can achieve strong, interpretable classification performance, and that the approach is feasible with data already collected in large national surveys.

The practical implications are considerable. A trained XGBoost model with SHAP-based explanations could, in principle, be embedded into routine diabetes care, allowing clinicians to enter a handful of easily obtainable variables and receive an immediate risk estimate along with the factors driving it. Such a tool would not replace clinical judgment or formal psychiatric assessment, but it could serve as a triage layer, directing limited mental health resources toward the patients most likely to need them. The finding that perceived energy level dominates the model also offers a low-cost screening cue that community health workers could act on even without computational support.

Looking ahead, the author suggests that machine learning models of this kind hold real promise for the early identification of CESD-10-defined high-risk depressive symptoms and could assist screening efforts among older adults with diabetes. Future work building on this foundation—particularly longitudinal designs that track individuals over time—will help determine whether these algorithms can forecast the onset of depression before symptoms fully emerge. For now, the study offers a concrete example of how artificial intelligence can be aimed at one of the quiet epidemics of aging: the hidden depression of millions of older people managing a chronic metabolic disease, waiting to be noticed.

Subject of Research: Machine learning identification of high-risk depressive symptoms among older adults with diabetes mellitus in China

Article Title: Identifying high-risk depressive symptoms among older adults with diabetes mellitus in China: a machine learning approach

Article References: Qiu, Y. (2026). Identifying high-risk depressive symptoms among older adults with diabetes mellitus in China: a machine learning approach. Current Psychology, 45(17), Article 1485. https://doi.org/10.1007/s12144-026-10040-9

Image Credits: AI Generated

DOI: 10.1007/s12144-026-10040-9

Keywords: depression, diabetes mellitus, machine learning, XGBoost, older adults, CESD-10, SHAP values, China, screening, mental health, geriatrics, predictive modeling