Prediction of late-onset depression in the elderly Korean population using machine learning algorithms

Abstract

Late-onset depression (LOD) refers to depression that newly appears in elderly individuals without prior depression episodes. Predicting future depression is crucial for mitigating the risk of major depression in prospective patients. This study aims to develop machine learning models to predict future depression. Using public data from the nationwide panel survey 'Korean Longitudinal Study of Aging,' we employed latent growth modeling and growth mixture modeling to identify four latent classes of depression trajectories in the elderly Korean population. Based on the results of binary logistic regression, we selected 12 variables capable of distinguishing the LOD population from the reference population and tested 12 machine learning (ML) algorithms. While most ML algorithms showed acceptable predictive capability, Random Forest Classifier and Gradient Boosting Classifier demonstrated superior performance. Consequently, we successfully established new ML-based LOD prediction programs. These programs could be further developed into self-checking online tools, expected to serve as decision support systems for primary medical care and health screening services.

Keywords: Depression trajectories; Late-onset depression; Longitudinal study of aging; Machine learning algorithms; Predictive performance.

Fig. 1

Schematic diagram of study population and system framework. These study panels include a variety of data on demographics, family composition, health information, medical history, and socio-economic status. Depressive symptoms were evaluated based on CES-D10 scores. Four latent classes of depression trajectories were identified. The class of preexisting depression was excluded to focus on predicting new-onset depression. Two classes of no symptoms were used as the reference group. Twelve ML algorithms were trained and tested using 12 variables.

Fig. 2

Prediction of the depression trajectories in latent classes. Statistically analysis of the growth factors in the quadratic growth model classifies depression trajectories into 4 latent classes.

Fig. 3

ROC curves of 12 machine learning models for predicting LOD. Different colors represent the different machine learning classifiers used in this study.

Fig. 4

Confusion matrices of representative ML models in predicting RF (a) and GBC (b) models exhibited superior performances based on their AUC scores. In the confusion matrices, columns represent the actual values (identified latent classes) while rows represent the predicted values calculated by the corresponding MLs.