The current study is based on a cross-sectional design using baseline information from the Walnuts Smart Snack (WSS) trial [24]. The purpose of this intervention was to determine if dietary supplementation with 30 g of raw walnut kernels per day for six months resulted in beneficial improvements in cognitive and socioemotional development when compared to a control group of healthy adolescents [25]. The study targeted adolescents aged 11–16 years attending regular schools in Barcelona. Although the initial protocol aimed to recruit 12–15-year-olds, the inclusion criteria were expanded to accommodate slightly younger and older students willing to participate, as facilitated by the schools. Exclusion criteria comprise individuals who regularly consume omega-3 PUFA supplements, eat walnuts daily, or have allergies to walnuts or gluten. Additionally, participants were excluded if they reported lactose intolerance or allergies to cereals, dried fruits, peanuts, soy, sesame, or sulfites, due to potential traces of these substances in walnut packages as a result of walnut industry practices. Participants with neuropsychological disorders were not excluded. In our study, there are only 3 cases of dyslexia and 1 case of attention deficit without hyperactivity. The rest of the cohort population in this analysis does not report neuropsychological or mental health disorders.

Over a year (2015–2016), we recruited 771 participants from 11 high schools in Barcelona which were evenly distributed geographically. We invited both public and private schools to participate in the project, and our goal was to include at least one high school per municipal district. Participants completed several neuropsychological tests and provided information on their lifestyle and dietary preferences prior to randomization. All information about the clinical trial procedure is described in the WSS protocol [24]. Participants who had complete baseline data on the food frequency questionnaire and the neuropsychological tests were eligible for the present study (n = 643). The trial study received permission from Parc Salut Mar’s Clinical Research Ethics Committee (approval number: 2015/6026/I).

A fieldwork technician provided questionnaires to the adolescent participants and other questionnaires were given to the parents to complete at home and return to us through the school instructors. Additionally, the technician asked about dietary habits using a food-frequency questionnaire (FFQ) of 60 food items adapted to the adolescent age from a validated questionnaire for the Spanish population [26, 27]. The adaptation consisted on increasing the portion sizes for some items (from small to medium size for meat, fish, vegetables, rice, pasta, and ultra-processed foods) and adding some extra items (e.g. Plant milk and soy yogurt). The technicians requested to the adolescents to report the average frequency of consumption for the specified standard units or portion size for each food item of the FFQ. The questionnaire included 9 frequency categories for each item ranging from “never or < 1 time/month’’ to ‘‘≥6 times/day” (Supplementary Table 1). All the selected frequency categories for each food item were converted to a daily intake.

Several primary endpoints concerning the neuropsychological function of adolescents were assessed at baseline (pre-intervention). The administration of all neuropsychological tests was carried out at the school by one trained psychologist and two fieldwork technicians. Similar procedure was applied for anthropometric measurements. Standard techniques were used to measure height, weight, and waist circumference (SECA 214 stadiometer for height, SECA 770 weighing scales for weight, and SECA 201 tape for waist measurement).

Based on WSS outcome data, three primary outcomes were selected for the current study. The first primary outcome was the self-reported Strengths and Difficulties Questionnaire (SDQ), a child mental health assessment tool with five hypothesized subscales. The scale was originally developed for the measurement of five aspects related to mental health screening namely four “difficulty” domains: hyperactivity/inattention, emotional symptoms, conduct problems, and peer problems. The SDQ additionally gathers data on prosocial behavior as a strength domain [28]. The SDQ externalizing score, which ranges from 0 to 20, is calculated by adding the conduct problems and hyperactivity scales. Emotional symptoms and peer problems scores are added to create the SDQ internalizing score, which has also a range of 0 to 20. In this case, we looked at the two SDQ profiles, where a higher score indicates abnormal behaviors.

The second measurement used was the Attention Network Test (ANT), a computer-based neuropsychological test to determine the attention function performance and the integrity of the three attentional networks [29, 30]: Alerting, the capacity to achieve and sustain maximum vigilance and performance while performing tasks; executive attention, which entails recognizing and resolving shifting attention to sensory inputs; and orienting, which includes shifting attention to sensory stimuli. Participants had to quickly determine the direction of the middle arrow when five arrows appeared on the screen. The impulsivity index was selected for the current research, which is calculated by deducing the reaction times in incorrect responses from the reaction times in correct responses, measured in milliseconds (ms). Lower scores indicate better attention performance (less impulsivity). The ANT impulsivity index is useful for identifying mental health changes, behavioral regulation, and risk-taking behavior in the adolescent population.

The third measurement analyzed, to assess emotional face recognition, was the Emotion Recognition Task (ERT). The neuropsychological test is a computer-generated paradigm for assessing the recognition of six fundamental facial emotional expressions: anger, contempt, fear, pleasure, sadness, and surprise. One at a time, the screen shows computer-generated images that have been warped from actual people’s facial characteristics to represent various emotions [31]. Each face is shown for 200 ms and then immediately covered up to prevent residual processing of the image. Finally, the participant must choose one of six possible emotions based on the expression on the face. A total of 60 images were used for emotion recognition. The outcome measures for ERT cover the correct total responses of facial emotions; higher scores indicate better emotional recognition.

The food frequency questionnaire (FFQ) was used to identify dietary patterns performing PCA. All food items listed in the FFQ, except for water, were used to derive the PCA in daily servings. The original food items included from the Spanish questionnaire are detailed in Supplementary Table 1. PCA is a reduction of dimensionality technique useful for analyzing complex and vast data. The main idea is to reduce the number of variables and detect “intrinsic patterns” in the data based on linear combinations of questionnaire food, illustrating the combinations of food that are usually eaten together in an individual’s diet of group study [32]. Since this FFQ version has only 60 items, we preferred to proceed with this agnostic approach (PCA) without any prior item manipulation, such as regrouping the items to common nutrient domains.

The adequacy of the data for factor analysis was evaluated in advance of PCA (Supplementary Table 2). The Kaiser-Meyer-Olkin test and Bartlett’s test of sphericity were conducted before PCA analysis to observe the relationship between variables. PC factors were retained considering factors with eigenvalues greater than 1.8 (Supplementary Table 3). Then, the varimax rotation was performed on PCA to simplify the factor structure and to enhance their interpretability. The rotated component fed the factor loadings for each food question contained in the factor. To interpret the results, variables with loadings greater than |0.2| were considered to contribute significantly to the pattern. Finally, factor standardized scores were also saved, as continuous and tertiles, for each PC for posterior regression analysis.

PCA standardized scores are a standardized system for representing an individual’s position within a component. As the score increases, it indicates that the individual is more represented in the foods with higher positive loadings of the PC. Conversely, as the score decreases, it means that the individual is less represented by the positive loadings of the PC. In the case of a component with negative loadings (i.e., PC1 and PC5), as the score increases (positive values), it indicates that the individual is more represented by eating less food with negative values.

After the creation of the PCA derived dietary patterns, multivariate linear regression models were performed to study the association between PCA standardized scores (continuous and tertiles) and neuropsychological outcomes (SDQ externalizing score, SDQ internalizing, ANT impulsivity index, ERT score), and all final models were adjusted for confounding variables.

First, we performed the multivariate linear regressions between PCA standardized scores and neuropsychological outcomes to study their linear association. Then, the PCA standardized scores were divided into three tertiles to avoid potential problems with non-linear relationships between the exposure and the outcome. By converting the tertiles into continuous variables (coded as 1, 2, and 3), we calculated the p-for-trend to evaluate the dose-response effect of the exposure variable. The confounding variables included in the regression models were as follows: age, gender, body mass index (BMI), physical activity and maternal education. Age was studied as a continuous variable (12–16 years old), while gender was categorical (female / male). The BMI, based on the World Health Organization (WHO) referent, was computed weight (kg)/length (m) [2] and then converted as z-scores (BMI z-scores). Physical activity was evaluated as high-intensity activity practice in four categories: once, twice, three times and more than three times a week. Finally, maternal education (presence or absence of university studies) was expressed as dichotomous.

For all regression models, the p-value of < 0.05 was considered statistically significant. Specifically, in the analysis of the ordinal exposure variable (categories), we considered the test as p for trend, with p-values < 0.05 indicating statistical significance. We did not conduct an in-depth analysis when significance was found only in the categorical exposure variable without the presence of a significant p for trend. To reduce the Type I error rates when performing multiple regression models, we conducted the False Discovery Rate test. We considered a threshold value of 0.05 to calculate the q-values. Only p-values below the q-values were considered as new thresholds for statistical significance. All analyses were conducted using RStudio version 4.2.3.