They tested its generalizability in an independent cohort of 6,424 participants from the Health and Retirement Study. The study assessed whether predictions made with repeated observations from one national population would remain accurate in a second aging cohort. The outcome was the presence of clinically significant depressive symptoms at the final survey wave, rather than a diagnosis of depression at the moment of prediction.

The input comprised 28 predictors covering sociodemographic characteristics, health, childhood experiences and intergenerational factors. The researchers evaluated discrimination, calibration, Brier score, threshold-dependent metrics and Decision Curve Analysis. For comparison, the investigators evaluated the sequence model against three conventional non-sequential methods under the same framework: L2-regularized logistic regression, random forest and XGBoost.

The attention-based LSTM outperformed all three in both the internal test set and the external validation cohort. Its performance was comparable to that of a plain LSTM without the attention component, showing that the addition of attention was not associated with a distinct improvement in the reported discrimination results. The model’s attention weights provided a window into how it used the longitudinal record.

The attention-weight plots indicated that more recent survey waves received greater relative weight than earlier waves. Shapley additive explanations were used to identify the factors most influential to individual predictions. Childhood relationships with parents emerged as the strongest predictor, with mother-child relationships particularly prominent.

Cognitive ability, self-rated memory and employment status followed. Calibration plots showed generally acceptable agreement between predicted and observed outcomes, particularly for the LSTM-based models. Some conventional models showed modest calibration deviations in the external validation cohort.

The researchers also used Decision Curve Analysis to assess the models’ practical decision-analytic performance, although the supplied evidence does not report individual threshold results or establish that the model should be used in clinical screening. Its applicability to younger populations, selected clinical groups or demographic settings not represented by the two cohorts remains uncertain. Overall, the strongest finding was that the attention-based LSTM combined good discrimination with external validation performance, while early parent-child relationships were among its most important predictive markers.

Subject of Research: Longitudinal machine-learning prediction of clinically significant depressive symptoms in middle-aged and older adults

Article Title: Development and validation of an LSTM model for predicting depressive symptoms in middle-aged and older adults: a prospective multi-cohort transcultural study

Article References: Shi, X., Li, Y., Wu, L., & Zhu, L. (2026). Development and validation of an LSTM model for predicting depressive symptoms in middle-aged and older adults: a prospective multi-cohort transcultural study. BMC Psychology. https://doi.org/10.1186/s40359-026-05546-7

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05546-7

Keywords: depressive symptoms, long short-term memory, attention mechanism, machine learning, longitudinal data, external validation, aging populations, parent-child relationships, risk prediction, SHAP, Development, validation