The review followed PRISMA 2020-informed procedures, searching six databases including Web of Science, Scopus, PubMed, IEEE Xplore, ACM Digital Library, and the China National Knowledge Infrastructure for studies published between 1 January 2016 and 30 August 2026. To qualify, a study had to include participants aged 60 or older, examine relationships involving at least two of the three domains, and use machine learning as a substantive component of the empirical analysis rather than a passing mention. After removing duplicates and screening titles, abstracts, and full texts, the authors settled on 27 eligible studies, synthesizing them narratively because the heterogeneity of designs, measures, and outcomes made quantitative meta-analysis inappropriate.

The geographic concentration of the evidence is one of the review’s most eye-catching findings. Twenty-two of the 27 studies — 81.5 percent — were conducted in China, with single studies each from Iran, the United States, South Korea, Japan, and Finland. Cross-sectional designs dominated, accounting for 21 of the 27 studies, while only six used longitudinal, cohort, observational, or feasibility-based designs. Sample sizes ranged enormously, from sensor feasibility studies with fewer than ten participants to secondary analyses of national surveys involving more than 16,000 older adults, and one ecological study aggregating data from 169,996 residents across 995 neighborhoods.

To bring order to this scattered literature, the authors developed a five-task framework describing what machine learning actually does in these studies. Ten studies (37.0 percent) used machine learning for measurement and feature construction — extracting streetscape greenery, blue-space exposure, sky visibility, mobility trajectories, gait characteristics, and physiological stress signals from raw image, spatial, and sensor data. Eleven studies (40.7 percent) focused on prediction, building models to classify or forecast depression risk, loneliness, social isolation, and well-being. But the dominant function by far was interpretation: 24 studies (88.9 percent) used machine learning to characterize associations, ranking predictor importance and mapping nonlinear relationships, thresholds, and interactions that conventional linear models would miss.

The technical toolkit reveals how far the field has moved beyond simple regression. Random forest and gradient-boosting methods such as XGBoost, GBDT, and LightGBM appeared repeatedly, often paired with SHAP values, partial-dependence plots, and GeoShapley to explain which environmental and behavioral variables mattered most. In one study, fully convolutional neural networks extracted green and blue space from Tencent Street View imagery to compare visual exposure measures against satellite-derived indices in Beijing. In another, semantic segmentation of street scenes combined with XGBoost and a CNN-BiLSTM model assessed how multi-scale built environments shape older adults’ walkability, reporting R-squared values between 0.806 and 0.906. These pipelines illustrate how deep learning now routinely converts raw imagery and sensor streams into scientifically usable variables.

The nonlinear findings are among the most scientifically interesting. Studies in Zhongshan, Nanjing, and Lanzhou used gradient boosting and random forest to show that the built environment’s relationship with walking, cycling, and light physical activity is rarely linear; instead, effects plateau, threshold, and interact in ways that partial-dependence plots can expose. A bus-use study found XGBoost achieved a testing pseudo-R-squared of 0.838 compared with 0.160 for multiple linear regression, underscoring how much signal linear models leave on the table. Meanwhile, a Japanese cohort study combined propensity score matching with causal forest analysis to estimate how the effect of social participation on depressive symptoms varies across subgroups — one of only two studies in the entire review that attempted explicit causal-effect estimation.

That scarcity of causal work is a central caution of the review. Only four studies (14.8 percent) examined pathways or mechanisms, integrating machine learning with mediation and structural equation modeling to probe whether environments affect mental health through physical activity, perceived stress, or social cohesion. Only two (7.4 percent) estimated causal effects under stated assumptions. The authors are emphatic that these levels of explanation are not interchangeable: a strong nonlinear association does not establish a mechanism, and a plausible pathway does not establish a causal effect. Machine learning, however sophisticated, does not by itself transform associational evidence into causal evidence — that depends on study design and the inferential framework surrounding the model.

Risk-of-bias assessments reinforced the need for caution. Using JBI critical appraisal tools for observational studies and PROBAST+AI for prediction models, the reviewers found that concerns clustered around internal-only validation, small and unrepresentative samples, and cross-sectional designs that cannot rule out residual confounding. Several prediction studies rated high risk, including a gait-based depression recognition model that achieved an impressive AUC-ROC of 0.911 but was trained and tested on an outcome-balanced subsample of just 50 people without external validation. A multimodal stress-detection study reported ensemble bagged-tree accuracies above 98 percent, yet analyzed physiological data from only nine participants. Impressive numbers, the review suggests, can reflect leakage and small-sample optimism as much as genuine generalizable signal.

The review also highlights what is missing: integration. Nine studies examined living environment–activity behavior links, nine examined activity behavior–mental health links, and five examined environment–mental health links, but only four considered all three domains simultaneously. Because older adults experience their neighborhoods through the activities they perform in them — walking to parks, taking buses, meeting friends — the authors argue that treating these domains separately leaves the central question of healthy aging only partially answered. They call for longitudinal and multisite designs, independent external and spatial validation, and more consistent definitions of exposures and outcomes so that findings can accumulate rather than remain scattered across partially connected relationships.

For now, the practical payoff lies in evidence generation rather than prescription. Fine-grained environmental and behavioral measures derived from street imagery, GPS trajectories, and wearable sensors can help planners identify which places and population subgroups show different patterns of activity and well-being, complementing conventional environmental assessment. But the review stops short of endorsing deployment: models built in one population or city may not transfer elsewhere, and evidence directly evaluating whether machine-learning insights improve real-world planning decisions remains limited. What the synthesis offers instead is a clear map of a young, fast-moving field — one where the algorithms have outpaced the study designs needed to turn their predictions into trustworthy knowledge about how the places we age in shape the minds we age with.

Subject of Research: Machine learning applications in research on living environments, activity behaviors, and mental health among older adults

Article Title: Machine learning applications across living environments, activity behaviors, and mental health among older adults: a systematic review

Article References: Wang, Y., & Lan, J. (2026). Machine learning applications across living environments, activity behaviors, and mental health among older adults: a systematic review. Frontiers in Public Health, 14, Article 1942765. https://doi.org/10.3389/fpubh.2026.1942765

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1942765

Keywords: machine learning, older adults, living environment, activity behaviors, mental health, systematic review, depression, built environment, XGBoost, SHAP, causal inference, healthy aging