Elevated depressive symptoms and depressive disorders are prevalent, debilitating, and costly, particularly among older adults. Globally, 350 million people suffer from depression annually (Smith, 2014). Depression is expected to be the leading cause of global disease burden by 2030 (Malhi and Mann, 2018), and recent reports using age-standardized rates and estimated annual percentage change from the Global Burden of Disease (GBD) 2019 study have shown that incidence of depression has increased by at least 0.6 % globally since 1990 (Liu et al., 2024), particularly in high socio-demographic index regions (Yang et al., 2023). Among older adults, the GBD 2019 study reported that the incidence and disability-adjusted life years of depression were highest among 60–64 year-olds (Liu et al., 2024). Likewise, GBD analysis in 2021 reported that the age-standardized incidence rate, age-standardized prevalence rate, and age-standardized disability-adjusted life years rate of late-life depression, defined as depression among adults ≥65 years, increased between 1990 and 2021 (Sun et al., 2025). This is especially concerning given that depression is both a risk factor for and comorbidity of cancer and cardiovascular, metabolic, inflammatory, and neurological diseases (Gold et al., n.d.), and has an estimated economic burden of over $210.5 billion in the United States alone (Greenberg et al., 2021). Moreover, recent global estimates showed that only 9.1 % of people with Major Depressive Disorder were receiving minimally adequate treatment with pharmacotherapy and/or psychotherapy (Santomauro et al., 2024), highlighting the continued need for identification of potentially accessible and low-cost lifestyle behaviours that could facilitate prevention and treatment of depressive symptoms and disorders.

Our recent meta-analytic evidence showed that across 111 prospective cohort studies involving >3 million adults, physical activity exposure reduced the odds of incident cases of either depression or an attenuated elevation in subclinical depressive symptoms by 21 % after full adjustment (Dishman et al., 2021). We more recently addressed calls from the World Health Organization Guidelines Development Group for a better understanding of physical activity dose-response derived from adequately powered large-scale studies (DiPietro et al., 2020), showing 7 % lower rates of depressive symptoms for physical activity doses of 1- < 600 metabolic equivalent of task (MET) minutes per week (MET.min.week−1), 7 % for 600- < 1200 MET.min.week−1, 16 % for 1200- < 2400 MET.min.week−1 and 23 % for ≥2400 MET.min.week−1, respectively, among 4016 older adults from The Irish Longitudinal Study on Ageing (TILDA) (Laird et al., 2023). Similarly, odds of Major Depression were 44 %, 41 %, and 49 % lower for physical activity doses of 600- < 1200 MET.min.week−1, 1200- < 2400 MET.min.week−1, and ≥ 2400 MET.min.week−1, respectively (Laird et al., 2023). However, the primary novel findings were that a physical activity dose of only 400- < 600 MET.min.week−1 was associated with 16 % lower rates of depressive symptoms and 43 % lower odds of Major Depression compared to no activity, suggesting that even minimal doses of physical activity—the equivalent to 20 min/day for five days/week of moderate-intensity activity (e.g., brisk walking)—may reduce the risk of depressive symptoms and Major Depression over time among older adults (Laird et al., 2023).

These findings add to a growing body of literature, from our group and others, supporting a nuanced dose-response relationship between physical activity and mental health outcomes. Although there is evidence of differences in biopsychosocial factors between sedentary and active samples (Condello et al., 2017; Gerber et al., 2025), the parametric, frequentist approaches used in those studies rely on restrictive assumptions (e.g., linearity, homoscedasticity, normal residuals) that often fail to accommodate the non-linear, hierarchical, and multicollinear patterns that are typical of behavioural research. While these models are well suited for hypothesis testing and inference, and often provide interpretable and mechanistic insight, they can be limited when it comes to generalizing patterns beyond the samples in which they are observed, particularly when relationships are complex and high-dimensional. Machine learning (ML) methods complement these approaches by emphasizing prediction, using discrete error metrics (e.g., classification accuracy), cross-validation, and flexible algorithms to learn complex, multivariate patterns. In the long term, such models could help inform a precision medicine approach for activity prescription to guide physical activity practice and policy. Importantly, certain ML architectures emphasize interpretability, offering insight into how individual features contribute to a multivariate phenotype. At the same time, the flexibility that makes ML powerful also introduces the risk of overfitting, a challenge often described as the ‘curse of dimensionality.’ This tension underscores the importance of principled feature selection. Thus, herein, we used a two-step approach in which large-scale univariate analyses were used to identify biopsychosocial variables that were consistently different between groups (meeting/exceeding the optimal dose; not meeting the optimal dose) among 6747 older adults from Waves 1–5 of TILDA as candidate features. These features subsequently served as inputs (predictors) to a machine learning (random forest) classifier based on physical activity dose reported at Wave 1 (outcome).