This study employed a non-randomized controlled trial involving four primary schools, with one intervention and one control school located in Limburg, Belgium, and the other intervention and control school located in Limburg, the Netherlands. All participating schools were located in urban areas according to the urban–rural classification system (38), which defines urban areas as those with a higher population density and built-up environment. Estimated schoolyard area per child and percentage schoolyard greening within this study per school are provided in Supplementary Table 1. We carried out interventions involving creating designated play areas aimed at increasing biodiversity through the incorporation of height variation and structural diversity, using native plants, and providing varied vegetation structure. This initiative not only aimed at increasing vegetation but also aimed at supporting greater species diversity, contributing to a richer ecological setting. The green schoolyards were designed to create interactive spaces where children could engage with nature, encouraging both free play and structured outdoor activities. The control schools maintained their usual paved playgrounds with minimal vegetation during the study period. Randomization was not possible due to the necessity for complete voluntary cooperation and participation of the intervention schools in greening the schoolyard.

Data were collected at five time points: the baseline assessment took place in November 2021, and subsequent assessments in April 2022 (except for the Dutch schools), November 2022, April 2023, and November 2023 (except for the Dutch schools) (Supplementary Table 2). The school management was personally informed about the importance, purpose, design, and possible risks of the study. After permission from the management, the children and parents were informed by a class visit for the children and a short info flyer with an attached comprehensive information letter for the parents. Before the baseline examination, written consent was obtained from the parents, and the children gave their oral permission to participate. The information letter and related materials were provided in Dutch. Copies of these documents are available from the authors upon request.

Data collection during all assessment moments included several self-reported questionnaires to assess socio-emotional wellbeing, and cognitive performance assessment in combination with eye-tracking to assess attentional bias. An overview of all outcome measures collected at each assessment is given in Supplementary Table 2. A general questionnaire regarding information about parental education, parity, ethnicity, parental occupation, income, household smoking behavior, illness, socio-economic demographics, the birth date of the child, the sex of the child, and the use of medication was filled out by a parent at home before the baseline examination. These variables were selected a priori on the basis of background knowledge about the relationship of the variable to the outcomes. Completion of privacy-sensitive questions, such as parental body mass index (BMI), socioeconomic status (SES), and disease status, was optional.

Cognitive functioning was assessed using a battery of five consecutive cognitive tests on a computer, guided by a trained staff member using the mental information processing and neuropsychological diagnostic system (MINDS) (39). Cognitive performance assessments were conducted individually in a separate room during regular class hours. At least one researcher was present to supervise them and address any questions. The task order was fixed for all participants during all assessments and took approximately 30 min. The cognitive battery consisted of five tasks. The continuous performance test (CPT) was administered to evaluate attention and concentration (40–42). In this task, a series of stimuli were sequentially displayed on a screen, and participants were required to respond (using a space bar, mouse click, touchscreen, or external button box) only when a specific target combination of two stimuli appeared (e.g., the letter X following the letter A). The frequency and type of errors made during the task provided insights into potential impulsivity or attention deficits. The average reaction time of correct responses (ms) was used as the outcome variable. The second assessment, the symbol digit modalities test (SDMT), measured information processing speed (42). Participants used a coding key of nine abstract symbols, each paired with a number, to quickly identify and match numbers corresponding to the presented symbols. The average time per response (seconds) is used as the outcome measure. The third test, the SPANNE test, assessed short-term memory (42). During this task, participants were instructed to recall and reproduce sequences of numbers both in the forward and backward directions. In the analyses, the maximum number of digits recalled was used as the outcome measure. The fourth test, the signal detection test (SDT), evaluated the visual information processing speed. Participants were tasked with identifying whether a deviant symbol was present within a series of visual stimuli as quickly as possible (42, 43). In the analyses, the average reaction time (ms) for correct responses is presented. Lastly, the fifth test, the STROOP task, assessed selective attention (42, 44, 45). Participants were presented with a color word displayed in an incongruent color (e.g., the word “red” printed in blue) and instructed to identify the color of the text rather than the word itself. For the last test, the average reaction time after log transformation was used as the analysis outcome. To assess mental effort, the Tobii Pro-Nano eye-tracking device (Tobii, Stockholm, Sweden) was used to determine the pupil dilation of the participants during all tests (46). The eye-tracking device was placed on the bottom of the screen of the tablets on which the tests were performed. Prior to beginning the eye-tracking session, a calibration process was conducted as recommended by the software. Using the Tobii Pro lab software (Version 24.21), the pupil diameter was calculated during each task. Pupil dilation can be used as a biomarker of effort in cognitive tasks (46).

Attentional bias was assessed using a Tobii Pro eye-tracking device while participants viewed emotional faces, following procedures similar to those described by Goeleven et al. (47). Attentional bias assessment was conducted individually in a separate room during regular class hours directly after the battery of cognitive tasks. Areas of interest (AOIs) were defined within the Tobii Pro lab software for faces depicting happy, anxious, and angry expressions (Supplementary material 1). Eye movement data (first fixation duration in milliseconds) were then used to calculate the ratio of attention directed toward the happy AOI relative to the anxious and angry AOIs. Specifically, a higher ratio for happy versus anxious/angry AOIs suggests a stronger initial preference or bias toward positive stimuli (i.e., a positive attentional bias) (48, 49). Conversely, a lower ratio may indicate that attention is preferentially captured by negative expressions, consistent with a negative attentional bias. This ratio-based approach for measuring attentional bias is grounded in findings that individuals with specific emotional or clinical traits (e.g., anxiety, depression) often show systematic differences in how they orient and maintain their gaze on emotional stimuli (49, 50). By comparing the duration of the first fixation on happy AOIs with those on anxious and angry AOIs, we gain insight into whether participants’ attention is initially guided by positive or negative emotional cues.

Visual preference was assessed using the same emotional faces as during the attentional bias test. Eye movement data (first fixation duration in milliseconds) were used to calculate the proportion of first fixation duration specifically allocated to the eye region compared with the forehead and mouth. The eye region is often considered the primary region for emotion recognition and can be particularly important in detecting social cues (51). A visual example illustrating how the AOIs were defined in Tobii Pro software can be found in Supplementary material 1.

During each timepoint, the child’s behavior was evaluated using the Strengths and Difficulties Questionnaire (SDQ) filled out by a parent (44, 52, 53). Questionnaires were sent to the parents via mail to be completed at home. This publicly available, validated tool measures the emotional and behavioral wellbeing of children and can be found online (54). The SDQ consists of five subscales: emotional problems, conduct problems, hyperactivity/inattention, peer relationship problems, and prosocial behavior. Responses are rated on a three-point scale (0, 1, 2), and a total difficulty score is calculated by summing the scores from the first four subscales (emotional symptoms, conduct problems, hyperactivity/inattention, and peer relationship problems). A higher total difficulty score reflects more risk of abnormal behavior, whereas for prosocial behavior a lower score indicates more risk of experiencing abnormal behavior.

Data were analyzed using RStudio software, version 4.4.0 (RStudio Inc., Boston, USA). Continuous data are presented as means and standard deviations (SD), and categorical data are presented as numbers and frequencies (%). Due to the explorative nature of this study, no sample size calculations were performed. Data normality was checked using the Shapiro–Wilk test. Baseline differences in demographic characteristics between the control and intervention groups were assessed using independent samples t-test for continuous variables and Chi-Squared tests for categorical variables. These preliminary tests were used solely to check for potential systematic bias at baseline and were not part of the main outcome analyses. All four research questions were analyzed using linear mixed-effects regression models, accounting for fixed and random effects. Fixed effects included group (intervention vs. control) and examination number (baseline, after 6 months, after 12 months, after 18 months, and after 24 months), while random intercepts accounted for individual variability among participants, a priori chosen covariates included age (years), sex, country in which the school is located, education level of the household (low, middle, or high when parents indicated their highest education level as ‘no diploma or primary school’, ‘secondary school’, or ‘college of university’, respectively), the capacity to manage household incomes (very difficult, difficult, average, rather easy, or very easy), and the season at the moment of the examination (spring and summer were coded as warm season, and winter and fall were coded as cold season). An interaction term for the examination (consecutive moment of examination) and group (control group vs. intervention group) was included to assess differences over time per treatment group. School identity was not included as a separate random factor, as the combination of country and treatment group already accounts for this structure. This method provides an indication of the effect of the intervention when the intervention is progressing. The assumption of linearity was assessed using residual plots. Akaike information criterion (AIC) and Bayesian information criterion (BIC) were used for model selection after the addition of a quadratic term to the model when patterns indicative of non-linearity were observed. All pre-specified covariates and interaction terms were retained in the final models regardless of AIC values. Selection bias was assessed for all outcomes by checking for baseline differences between the intervention and control schools using an independent samples t-test (Supplementary Table 3). Beta estimates, 95% confidence intervals, and p-values were provided for key outcomes. Beta estimates are abbreviated as beta (β) and can be interpreted as a change in outcome having the green school intervention compared to the control schools over time. Statistical significance was set at p ≤ 0.05. Estimated means and standard deviations for each examination in comparison to the baseline measurement are provided in Supplementary Table 4.

Out of 482 students, a total of 169 children aged 7–12 participated in the study of which 86 in the intervention group and 83 in the control group (Supplementary Figure 1). The participation rate was 44.1% and 28.9% for the intervention and control schools, respectively. Demographic characteristics did not differ significantly between the intervention and control schools (Table 1). The average age of the participating children was 10 years. The majority of the children were girls (55% vs. 45%). The mean BMI z-score was 0.40. Approximately one-third of mothers (36.1%) had attained a medium level of education, defined as completing secondary school. A substantial proportion of fathers (39.1%) are highly educated; holding a college or university degree. Parents most commonly reported that their household income allows them to get by without significant difficulty but also without ease (28.4%).

The intervention group showed more increased selective attention and larger pupil diameter in multiple cognitive tests during greening as compared to the control group. An inverse significant association was found for the interaction effect between intervention and control group throughout greening on selective attention during the stroop test (β = −0.01, 95% CI −0.03 to −0.002, p = 0.02) after adjustment for age, sex, country in which the school is located, education level of the household, the capacity to manage household incomes, and season. No significant interaction effects were found for the outcomes of the continuous performance test, signal detection modalities test, spanne test, and signal detection test during greening. An interaction effect between intervention and control group throughout greening was observed for pupil diameter during the signal detection modalities test (β = 0.09, 95% CI −0.01 to 0.19, p = 0.08), spanne test (β = 0.11, 95% CI 0.02–0.20, p = 0.01), signal detection test (β = 0.11, 95% CI 0.01–0.21, p = 0.03), and stroop test (β = 0.08, 95% CI −0.02 to 0.17, p = 0.1) independent of age, sex, country in which the school is located, education level of the household, the capacity to manage household incomes, and season. An overview of estimates is provided in Table 2.

The intervention group showed improved social behavior throughout greening compared to the control group, based on the mixed-effect models. A positive significant association was found for the interaction effect between intervention and control group throughout greening on prosocial behavior over time (β = 0.21, 95% CI 0.01–0.40, p = 0.03, Figure 1), after adjustment for the following covariates: age, sex, country in which the school is located, education level of the household, the capacity to manage household incomes, and season. No significant intervention effects over time were found for emotional problems, conduct problems, hyperactivity/inattention, peer relationship problems, and the total difficulty score. Over the course of the examinations, the intervention group showed a trend of an increase in the proportion of first fixation duration to the eyes as compared to the forehead and mouth (β = 0.03, 95% CI −0.001 to 0.05, p = 0.06). No effect of greening was shown on the proportion of the first fixation duration to the emotion happy as compared to the neutral face and angry and fearful emotions (β = −0.004, 95% CI −0.02 to 0.02, p = 0.67).