Activity-function transitions and interpretable machine learning for predicting incident depressive symptoms among ACE-exposed middle-aged and older adults: a multi-cohort study.
Lin Z, Shangguan Y, et al. • Frontiers in public health • 2026
An interpretable machine-learning framework integrating physical performance, physical activity, and health-related factors achieved moderate and externally validated prediction of incident depressive symptoms among ACE-exposed adults, with random forest demonstrating the best performance (AUC 0.701 in ELSA; 0.693 in HRS).
Key Findings
Results
Random forest was the best-performing machine learning algorithm for predicting incident depressive symptoms among ACE-exposed middle-aged and older adults.
AUC of 0.701 in the development cohort (ELSA) and 0.693 in the external validation cohort (HRS)
Multiple machine learning algorithms were compared before selecting random forest as the top performer
Performance was characterized as 'moderate' by the authors
Authors noted 'limited case-detection ability,' cautioning against clinical application without further validation
Methods
LASSO regression identified stable predictors of incident depressive symptoms primarily involving physical function, physical activity, chronic conditions, and sociodemographic characteristics.
LASSO (Least Absolute Shrinkage and Selection Operator) regression was used for predictor selection prior to model comparison
Key predictor domains included physical function, physical activity, chronic conditions, and sociodemographic characteristics
The study used ELSA for model development and HRS for external validation to assess generalizability
Participants with baseline depressive symptoms were excluded from both cohorts
Results
SHAP analysis identified walking time, grip strength, falls, arthritis, and physical activity as the most important contributors to predicting incident depressive symptoms.
SHAP (SHapley Additive exPlanations) analysis was used to provide interpretability to the machine learning model
The top five contributors were walking time, grip strength, falls, arthritis, and physical activity
These features reflect the physical performance and health-related domains identified by LASSO
The use of SHAP analysis was described as part of an 'interpretable machine-learning framework'
Results
Participants with unfavorable activity-function status at both pre-outcome assessments showed higher odds of incident depressive symptoms in both cohorts.
Pre-outcome activity-function transition profiles were examined in relation to observed incident depressive symptoms
The association between persistently unfavorable activity-function status and depressive symptoms was replicated in both ELSA and HRS cohorts
This finding provided 'additional information on subsequent depressive-symptom risk' beyond the main predictive model
Transition profiles captured changes across at least two pre-outcome assessment time points
Background
Adverse childhood experiences (ACEs) are associated with increased risk of depressive symptoms in later life, motivating the development of a predictive model specifically for ACE-exposed middle-aged and older adults.
The study population was restricted to adults with a history of ACEs
Data were drawn from two large longitudinal cohorts: the English Longitudinal Study of Aging (ELSA) and the Health and Retirement Study (HRS)
The study focused on incident (new-onset) depressive symptoms, excluding those with baseline depressive symptoms
The multi-cohort design spanning the UK (ELSA) and USA (HRS) was intended to support external generalizability
Conclusions
The authors characterized the machine learning framework as exploratory, noting that moderate discrimination and limited case-detection ability preclude clinical application without further validation and refinement.
AUC values of 0.701 (ELSA) and 0.693 (HRS) were described as reflecting 'moderate' performance
The authors explicitly stated the framework 'should be regarded as exploratory and requires further validation and refinement before any clinical application'
The framework was described as 'interpretable' due to the inclusion of SHAP analysis
External validation in HRS provided evidence of cross-national generalizability despite the moderate performance level
What This Means
This research suggests that machine learning models can be used to predict which adults who experienced adverse childhood experiences (ACEs) — such as abuse, neglect, or household dysfunction — are more likely to develop depression in middle age or later life. The researchers built a predictive model using data from a large UK study (the English Longitudinal Study of Aging) and tested it on a large US study (the Health and Retirement Study). Among several algorithms tested, the 'random forest' model performed best, correctly distinguishing between those who did and did not develop depressive symptoms about 70% of the time in both datasets — a result described as moderate rather than strong.
The most important factors the model used to make predictions were physical in nature: how long people could walk, how strong their grip was, whether they had fallen recently, whether they had arthritis, and how physically active they were. People who had poor physical function and low physical activity at multiple time points before the study outcome were especially likely to develop depressive symptoms. This highlights a connection between declining physical ability and mental health risk among people who experienced childhood adversity.
This research suggests that monitoring physical activity and physical function in ACE-exposed older adults could help identify those at higher risk for depression. However, the authors themselves caution that the model's accuracy is only moderate and should not yet be used in clinical settings to make decisions about individual patients. Further research and refinement are needed before such tools could be responsibly applied in healthcare.
Check Your Own Numbers
Upload your bloodwork. We'll cross-reference your results against this study and 4,700 others.
Lin Z, Shangguan Y, Sim Y, Chen D, Wang X. (2026). Activity-function transitions and interpretable machine learning for predicting incident depressive symptoms among ACE-exposed middle-aged and older adults: a multi-cohort study.. Frontiers in public health. https://doi.org/10.3389/fpubh.2026.1924836