Machine learning maps how neighborhoods, activity, and mental health
A large review of research has found that machine learning is quickly changing how scientists study the links between where older adults live, how they move around and spend time with others, and how they feel mentally. However, the evidence is still scattered across different studies. The review, published in Frontiers in Public Health by Yuxuan Wang and Jun Lan of the University of the Chinese Academy of Sciences, examined 4,337 research records and found only 27 studies that met its strict requirements. Nearly three-quarters of these studies were published in 2024 or later, showing how quickly this field has grown in recent years. Yet almost none of the studies have examined all three areas—living environment, activity, and mental health—together in one research framework.
The review followed methods based on the PRISMA 2020 guidelines. The authors searched six databases: Web of Science, Scopus, PubMed, IEEE Xplore, ACM Digital Library, and the China National Knowledge Infrastructure. They looked for studies published between 1 January 2016 and 30 August 2026. To qualify, a study had to include people aged 60 or older, examine links between at least two of the three areas, and use machine learning as an important part of its analysis rather than just mention it. After removing repeated records and checking titles, summaries, and full articles, the authors selected 27 studies. They described the results in words rather than combining them in a statistical meta-analysis because the studies used very different methods, measurements, and outcomes.
One of the clearest findings was that most of the research came from China. Twenty-two of the 27 studies, or 81.5 percent, were carried out there. Iran, the United States, South Korea, Japan, and Finland each had one study. Most studies took a snapshot of people at one point in time: 21 of the 27 used this approach. Only six followed people over time or used cohort, observational, or feasibility-based designs. Study sizes also varied greatly. Some tests of sensors included fewer than ten people, while analyses of national surveys included more than 16,000 older adults. One study examined combined environmental data from 169,996 residents across 995 neighborhoods.
To organize this scattered research, the authors created a framework with five types of tasks that machine learning performed in the studies. Ten studies (37.0 percent) used machine learning to measure things and create useful features from raw data. These included greenery along streets, exposure to water and other blue spaces, how much of the sky could be seen, people's movement patterns, the way they walked, and signs of physical stress. Eleven studies (40.7 percent) focused on prediction, using models to classify or estimate the risk of depression, loneliness, social isolation, and poor well-being. But the most common use was understanding patterns: 24 studies (88.9 percent) used machine learning to study relationships, identify the most important factors, and uncover patterns that change in complex ways, including thresholds and interactions that ordinary linear models might miss.
The methods used show how much the field has moved beyond simple regression. Researchers often used random forests and gradient-boosting methods, including XGBoost, GBDT, and LightGBM. They combined these with tools such as SHAP values, partial-dependence plots, and GeoShapley to explain which environmental and behavioral factors mattered most. In one study, fully convolutional neural networks examined Tencent Street View images to measure green and blue spaces and compare what people could see with measurements taken from satellite data in Beijing. In another, researchers used a method called semantic segmentation to identify parts of street scenes, then combined XGBoost with a CNN-BiLSTM model to study how features of the built environment at different scales affected how easy it was for older adults to walk. The reported R-squared values ranged from 0.806 to 0.906. These examples show how deep learning can turn raw images and sensor data into measurements that scientists can use in their research.
Some of the most interesting findings involved relationships that did not follow a straight line. Studies in Zhongshan, Nanjing, and Lanzhou used gradient boosting and random forests to show that features of the built environment do not always have a simple, steady relationship with walking, cycling, or light physical activity. Instead, effects may level off, appear only after a certain point, or depend on other factors. Partial-dependence plots can help reveal these patterns. One study of bus use found that XGBoost achieved a testing pseudo-R-squared of 0.838, compared with 0.160 for multiple linear regression. This suggests that the linear model missed much of the useful information in the data. A study that followed a group of people in Japan combined propensity score matching with causal forest analysis to estimate how the effect of taking part in social activities on symptoms of depression differed between groups. It was one of only two studies in the entire review that tried to estimate causal effects directly.
The lack of research into cause and effect is an important reason to be careful when interpreting the findings. Only four studies (14.8 percent) examined pathways or mechanisms. These studies combined machine learning with methods such as mediation analysis and structural equation modeling to explore whether environments might affect mental health through physical activity, perceived stress, or social connections. Only two studies (7.4 percent) estimated causal effects under clearly stated assumptions. The authors stress that these different levels of explanation are not the same. Finding a strong, complex relationship does not prove how it works, and identifying a possible pathway does not prove that one factor causes another. Even advanced machine learning cannot turn evidence of an association into proof of cause and effect on its own. That depends on how a study is designed and the methods used to draw conclusions from its results.
The review's checks for possible bias showed further reasons for caution. The researchers used JBI critical appraisal tools for observational studies and PROBAST+AI for prediction models. They found common problems, including testing models only within the original study, using small or unrepresentative groups, and relying on studies that measured people at just one point in time and could not rule out other factors that might explain the results. Several prediction studies were judged to have a high risk of bias. For example, a model that used walking patterns to recognize depression achieved an impressive AUC-ROC score of 0.911. However, it was trained and tested on a sample of just 50 people, selected to have a balanced mix of outcomes, and was not tested on an independent group. Another study that combined different types of data to detect stress reported accuracy above 98 percent using an ensemble of bagged decision trees. Yet it analyzed physical measurements from only nine participants. The review warns that impressive scores can result from data leakage or overly optimistic results from small samples, rather than from a model that will work well for new people.
The review also points out a major gap: researchers rarely study all three areas together. Nine studies looked at links between living environments and activity, another nine examined links between activity and mental health, and five studied links between environments and mental health. But only four examined all three areas at the same time. Older adults experience their neighborhoods through the things they do there, such as walking to parks, taking buses, and meeting friends. The authors argue that studying each area separately leaves important questions about healthy aging unanswered. They call for studies that follow people over time and include multiple locations, independent testing in new groups and geographic areas, and more consistent ways of defining and measuring environmental conditions and outcomes. These changes could help researchers build on one another's findings instead of producing separate results about only parts of the problem.
For now, the main benefit of this research is that it helps scientists gather better evidence, rather than giving planners firm instructions about what to do. Detailed measurements from street images, GPS movement records, and wearable sensors could help planners identify places and groups of people with different patterns of activity and well-being. These tools could add useful detail to traditional environmental assessments. However, the review does not recommend putting these models into practice without further testing. A model developed in one city or population may not work in another, and few studies have directly tested whether insights from machine learning lead to better real-world planning decisions. Instead, the review provides a clear picture of a young and rapidly growing field. Its algorithms have advanced faster than the study designs needed to turn their predictions into reliable knowledge about how the places where we grow older affect our mental health as we age.
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