Background: Depression in older adults is often underdiagnosed due to atypical symptom presentation and generational stigma, leading to delayed intervention. Early identification of individuals at risk of developing elevated depressive symptoms is therefore critical, but traditional approaches show limited predictive accuracy. To date, no study has applied machine learning models to predict clinically significant depressive symptoms at 2-year follow-up in older adults in the UK using data from the English Longitudinal Study of Ageing (ELSA). Moreover, the impact of encoding strategies for categorical healthcare variables has not been examined.

Objective: This study aimed to develop and evaluate machine learning (ML) models to predict the clinically significant depressive symptoms at 2-year follow-up in older adults using ELSA data. We further compared ordinal and one-hot encoding strategies across different ML architectures and identified key predictors of depressive symptoms at follow-up.

Methods: Data were drawn from four consecutive waves of ELSA, including participants aged ≥50 years without significant depressive symptoms at the baseline wave (Waves 6-9). Clinically significant depressive symptoms were defined as CES-D-8 ≥4 at the subsequent wave (Waves 7-10). Over 120 features spanning sociodemographic, psychological, and health-related domains were analysed. Eight ML models were applied including tree-based ensembles, deep learning architectures for tabular data, distance-based, probabilistic, and linear methods. Model performance was assessed using area under the receiver operating characteristic curve (AUROC), and F1-score. Model interpretability was examined using SHapley Additive exPlanations (SHAP). Sensitivity analyses assessed the robustness of results across alternative CES-D-8 thresholds (≥3, ≥4, ≥5) and encoding strategies.

Results: Across waves, the best-performing models achieved mean AUROC scores of 0.72–0.73, with a peak of 0.75 in the highest-performing wave. Ordinal encoding consistently outperformed one-hot encoding across all ML models, yielding improvements in AUROC and F1-score, with the greatest increase in tree-based methods. SHAP consistently identified loneliness, sleep disturbances, and low social engagement as strong predictors of elevated depressive symptoms at follow-up. Sensitivity analyses across CES-D-8 thresholds demonstrated robust feature importance, with AUROC ranging from 0.67 to 0.82. Traditional machine learning models (Random Forest, XGBoost, Support Vector Machines) generally achieved higher performance than the deep learning models for this task.

Conclusions: Our findings demonstrate the feasibility of predicting clinically significant depressive symptoms at 2-year follow-up in UK older adults with moderate accuracy. Ordinal encoding demonstrates superior performance for healthcare datasets with inherently ordered categorical features. The identification of consistent risk factors highlights opportunities for developing targeted clinical screening tools and preventive interventions. This study provides new evidence on depressive symptoms prediction in the UK context, leveraging longitudinal data from ELSA, and contributes to advancing digital mental health research for aging populations.