A Neural Network Algorithm-Based Prediction Model for Depression Among Elderly Individuals with Diabetes Mellitus

Abstract

Objective: To address the "syndemic" of depression among elderly diabetic patients in China by developing a non-invasive machine learning predictive model for large-scale community screening.

Methods: Data from 4,827 participants (aged 60–85) were extracted from the 2020 CHARLS database. Random Forest (RF), a tree-based machine learning technique, was used to select important variables. Then, a Backpropagation Neural Network (BPNN), which can capture complex, non-linear relationships, was built. Model performance was evaluated using Accuracy (correct prediction rate) and AUC (area under the curve, a measure of discrimination ability).

Results: The prevalence of depressive symptoms was 33.7%. RF identified age, physical pain distribution, and sleep duration as the top predictors. Although post-noon nap duration was not significant in a simple statistical analysis, it emerged as important in a neural network, which detects hidden patterns. The model achieved an accuracy of 0.67 (percentage of correct predictions) and AUC of 0.59 (ability to distinguish cases) on the test set, indicating strong negative predictive value for risk triaging.

Conclusion: Physical pain, sleep patterns, and social engagement are decisive warning signs for depression in elderly diabetics. This non-invasive model is an essential, scalable solution for early identification and intervention within primary healthcare systems.