This study was conducted at the Faculty of Medicine, National Autonomous University of Mexico (UNAM) in Mexico City, which is the largest medical school in the country. The sample consisted of second-year medical students enrolled in the 2010 program, (pre-clinical stage) of the medical curriculum, which emphasizes foundational sciences like Physiology. Participants were selected using a non-probabilistic convenience sampling method during scheduled class times, with no incentives for participation to ensure consistency, data collection took place in 2023. We invited all second-year medical students in the program to participate via announcements and information sessions. A total of 305 students consented and completed the assessments.

This study employed an observational, cross-sectional, correlational design. Data were collected at a single point in time to analyze the relationships between academic trajectory, intelligence, electronic media use, and attention. No experimental manipulation was conducted and the study focused on identifying associations rather than causation.

The “socio-academic” category comprised variables related to students’ academic background and status (e.g., age, years of study, number of failed courses), while “early grades” included prior academic performance indicators (such as elementary, junior high, and high school grade averages). The “study-recreation” category captured how students allocate time outside of mandatory academics (e.g., time spent in non-mandatory study or leisure activities). These variables were studied and analyzed were categorized into the following dimensions (see Figure 1):

All participants completed an online sociodemographic survey and provided information regarding their academic trajectory (an official document from the UNAM’s School Administration System) up to the time of the study. Intelligence was assessed using the standardized intelligence test Shipley-2 ^® (13). Electronic media use was measured using the Youth Screen Time Survey (14).

These particular variables were selected based on the aims of the study and prior research. The Youth Screen Time Survey targets the most common digital activities among students—watching video content, social communication (messaging/social media), and gaming—providing a broad measure of digital media use. We chose the Physiology course’s standardized test scores as the indicator of academic performance because this core course demands integration of knowledge and cognitive skills, making it a representative benchmark for academic achievement in the second year. We narrowed the scope to these constructs (digital use, intelligence, attention, and performance) to maintain a manageable survey length and because they align with our conceptual framework, whereas other potential factors (e.g., podcasts, personality traits) were beyond the scope of this study.

The instruments were administered to groups of 15 students each, ensuring adequate control of the process and direct supervision by the evaluators. Special care was taken to prepare the classroom environment, ensuring that it was free from visual and auditory distractions. These sessions were conducted in a controlled classroom setting by the research team, who provided standardized instructions. All participants gave informed consent, and it was emphasized that participation was voluntary and unrelated to course grades or academic standing (to avoid any coercion). An evaluator was present throughout to answer questions and ensure students were comfortable and not disturbed, adhering to ethical guidelines for testing.

Academic trajectory was assessed using standardized test scores for the physiological course. The physiology course grade was selected because of its evaluation of cognitive dimensions, such as integration and conceptual analysis. The department provides results from three standardized tests conducted every 4 months, with Cronbach’s alpha values of 0.857, 0.898, and 0.907. The average of these three test scores was calculated and used as a variable for standardized test performance (Str).

Cognitive intelligence was evaluated using the Shipley-2 ^® test, a psychometric instrument designed to measure various aspects of cognitive intelligence including Gf, Gc, general intelligence, and potential indicators of cognitive impairment. The Shipley-2 ^® test comprises two main sections: (I) Vocabulary scale (Gc assessment), in which participants selected synonyms. This section evaluates acquired knowledge such as vocabulary and verbal comprehension, reflecting prior experiences and learning. Gc is closely associated with skills accumulated over time. (II) Abstraction scale (Gf assessment), in which participants completed pattern-related problems. This section measures logical reasoning and problem-solving abilities, which are essential aspects of fluid intelligence that are independent of prior learning and are related to abstract reasoning. The test was administered and analyzed according to the instructions published in the original version, without modifications.

The screen time was assessed using the method developed by Sauce et al. (14). The participants were asked to rate their daily use of various applications on weekdays and weekends, including activities such as watching programs and videos, playing video games, sending messages, using social media, and video chatting. The rating scale ranged from “none” to “>4 h,” and means were calculated for both temporal contexts. Based on previous research, screen time was categorized into three main groups: (i) watching content (programs and videos); (ii) socializing (messages, social media, and video calls); and (iii) playing video games.

The databases were constructed using Microsoft Excel (Microsoft Corporation, Redmond, Washington, United States), and analyses were performed using IBM SPSS Statistics for Windows, Version 25.0 (2017; IBM Corp., Armonk, New York, United States), GraphPad Prism version 9 (Dotmatics, Boston, Massachusetts, United States), and Excel. The Kolmogorov–Smirnov test was used to evaluate normality. Descriptive statistics for the study population included measures of central tendency, dispersion, and frequency. Scatter plots and box-and-whisker plots were generated for data visualization. The hypothesis testing included Pearson’s correlation and linear regression. Additionally, network analysis was conducted in RStudio, Version 2025.05.0 + 496 (Posit Software, PBC), using the igraph package to identify clusters and analyze connectivity patterns among variables, thereby elucidating the underlying complexity of the data. Conditional Uniform Graph Tests (CUG Tests) were conducted to determine the statistical significance of the selected Graph Level Indexes (GLIs) when compared to randomly generated graphs with the same number of edges. CUG tests were performed with 100 Monte Carlo replications and p-values were reported. Using the resulting network topology, we generated a mediation model and verified the influence and interactions of the screen time and intelligence variables. The mediation model was implemented in Jamovi 2.7.6 (Sydney, Australia).

The physiological network construction was based on a previous study (15–17). First, all data were screened for extreme values and normality using the Shapiro–Wilk test. Missing values were imputed using a non-parametric approach with random forests implemented in R, using the missRanger package. No variables had more than 10% of missing values. Data sets were not normally distributed, and ordinal variables were present; consequently, Spearman’s rank correlation was selected as a measure for correlation. To select statistically significant correlations, a threshold of p < 0.001 was chosen to account for multiple comparisons and reduce the likelihood of false-positive edges in the network. The Spearman rank correlation coefficient “ρ” (rho) was squared to obtain positive values (15).

An adjacency matrix was constructed using the determination coefficient between each pair of variables (Figure 2), resulting in a weighted and undirected network (Figure 1). Notably, although the squared Spearman’s ρ (coefficient of determination) was employed to derive the adjacency matrix weights—thereby ensuring that all edge magnitudes remained positive—the original sign of each correlation was preserved for interpretative purposes. Thus, the network visualizations (the heatmap in Figure 2 and the graph in Figure 1) denote whether each association was positive or negative, even though the corresponding edge weight is derived from ρ². The matrices were rearranged in first principal component order (Figure 2). The resulting graph was plotted using a graphopt layout implemented in the igraph package (Figure 1).

We evaluated the graph topology using network density, transitivity, and average path length. Node centrality was evaluated using the sna package (18) to calculate degree, strength, eigencentrality, and betweenness. Eigencentrality corresponds to the value of each eigenvector inside the graph adjacency matrix and is interpreted as a measure of influence within the network. Betweenness was selected as a measure of intermediation within the network. Eigencentrality and betweenness show different roles that a node may play within the network (19).

To analyze the microstructure of our graph, we ranked each node according to its degree and strength. Nodes within a network can also be ranked according to several centrality definitions (19). Radial and medial centralities were assesses though eigencentrality and betweennes respectively, both values were obtained using the sna package (18).

The vicinity of the correlated variables was delimited via cluster analysis using the Louvain algorithm. Briefly, node groups were determined using the modularity optimization algorithm as calculated in the igraph package. The resulting clusters were delimited using shadows of different colors (Figure 1).

Regarding our main objective, examining whether prolonged exposure to electronic media negatively influences cognitive abilities, academic trajectories, and social determinants in medical students, we examined only the main connected component of our network. The time spent in Non-Mandatory Activities was not significantly correlated with any other variable in our study, and this single node was removed from the network (data not shown). This indicates that all other variables are systemically related to each other.

Overall, the density of the network was 0.21 (p < 0.0001), indicating that only 21% of all possible links were in this network. The network transitivity was 0.5208 (p < 0.0001), indicating that 52% of all possible connected triangles among any three individual nodes in the network were present. Average path length was 2.29, which means it takes on average 2.29 “steps” to connect any two nodes from the network. This short path length reflects the relatively small size of the network and an efficient pattern of connections. Despite this, clustering is also apparent in the network. Overall, our network has subgroups of nodes that are substantially connected among themselves, yet this high degree of connection is not maintained throughout the network, hinting at a small-world network akin to biological systems.

Our network comprised 34 variables (nodes) connected by 122 statistically significant correlations (edges) (Figure 1). Our network comprised six main clusters (Figure 1). We observed three distinct groups of variables in our heatmap (Figure 2), we can observe 3 distinct groups of variables. These relate to (1) time spent on digital media, videogames, and social media use (upper left corner); (2) different intelligences and their relationship to age, number of failed exams, and years of study (upper right corner); and (3) the relationship between different intelligences and academic performance (bottom right corner).

When computing the degree of each node, the number of connections for each variable, number of failed exams (FC, FC_Num), time spent using electronics (SCT), grades (Str), and age (Age) were the most connected inside the network. The connectedness of these same nodes was confirmed using a basic centrality measure, eigencentrality. Variables with the highest centrality (i.e., the most interconnected and influential nodes in the network) were the number of failed courses (FC_Num), total screen time (SCT), and age. This shows that these nodes are key variables in our model and have the greatest influence inside the network (these nodes had the greatest number and strength of connections to others), suggesting that they are central to academic success.

On the other hand, the highest betweenness scores were allocated to: Average grade during highschool (H_Gr), number of failed exams (FC_Num), TV time (Tv), texting time (Text), and total social media time (Soc). It is noteworthy that FC_Num ranks among the highest in both eigencentrality and betweenness, indicating that failed courses have a profound impact on academic success, both as a factor by itself and as a mediating factor between different types of intelligence and digital media use. TV time seems to bridge videogame use and overall digital media consumption, while Text and Soc are directly connected to average grades (Str).

The shortest path connecting the intelligence and electronic media clusters passes through texting. Texting had the fifth-highest betweenness among the nodes (51, p < 0.38). This indicates that any path connecting the digital behavior cluster and socio-academic factors most likely crossed the text node. Moreover, only texting was significantly associated with a type of intelligence (Gc) but was negatively correlated (ρ = −0.20, p < 0.01). Therefore, texting is one of the most common digital behaviors and is intricately related to socio-academic factors; yet it seems to hinder or discourage the use of consolidated intelligence and memory.

Our network highlights the time-dependent relationship between grades throughout academic development. Elementary school grades (Ele_Gr) are related to junior high grades (Jr_Gr), but not to high school (H_Gr) or university grades (Ah, Str). Moreover, junior high grades are related to high school grades but not to others. This suggests that previous term grades can be a predictor of academic performance, but only in a short timeframe. This highlights the importance of good grades, even early in school.

Electronic media, regardless of the form, was highly connected among themselves, but only total screen time, texting, and Tik Tok time seemed to have a relationship with the number of failed courses (FC, FC_Num).

Academic performance in university (Str, Ah) shows a relationship with Tik Tok use and Texting, yet further research is needed to clarify its relationship and if it is directly responsible for hindering academic development effect.

Louvain’s algorithm exhibited a high modularity, with a value of 0.56. Six clusters were identified in the network (Figure 1). The first cluster (red) included grades, crystallized intelligence (Gc), transportation, age, and parents’ degree of study. This cluster has most of the nodes with high eigencentrality in the network, including grades (Ah), number of failed exams (FC, FC_Num) and average grade (Str). A key node with both high eigencentrality and highest betweenness is the number of failed exams (FC_Num).

The second cluster (pink) includes the time spent on all forms of electronic media (texting, social media, videos, and television). No nodes in this cluster have high eigencentrality, yet betweenness is high for time spent on TV, texting, and Tik Tok.

The third cluster (purple) includes the gender and time spent on videogames. It is apparent that there is a broad relationship between gender and playing video games. None of these clusters had nodes with high centrality or betweenness.

The fourth cluster (navy blue) relates mandatory study time (Ms) with time spent in recreational activities. Interestingly, Mandatory study time was not related to time spent on electronic media or academic success. None of these clusters had nodes with high centrality or betweenness.

The fifth cluster (green) included cognitive impairment (CI-CB), fluid intelligence (Gf), and intelligence (G), excluding crystallized intelligence (Gc). It is noteworthy that intelligence (G) was closely related to age, years dedicated to study, and high school grades, but not to other metrics of academic achievement or electronic media use.

Finally, the sixth cluster (sky blue) solely relates elementary school grades to junior high grades. It seems apparent that there is a sequential relationship for academic performance, spanning from elementary school to university. None of these clusters had nodes with high centrality or betweenness.

To deepen our understanding of the effect of intelligence and Total Screen Time on Academic performance, we conducted a mediation analysis with only five variables: Intelligence (G), Crystallized Intelligence (Gc), Total Screen Time (SCT), Texting (Text), and Standardized Physiology test results (Str). In Figure 3, solid arrows indicate direct pathways (regression coefficients) between variables, whereas dotted arrows represent covariance relationships between the mediator variables. All Beta coefficients (β) are presented in Table 1. We observed that higher crystallized intelligence (Gc) was associated with better academic performance (β = 0.38), whereas greater total screen time (SCT) and more texting were associated with lower performance (β = −0.10 and β = −0.14, respectively).

First, the cross-sectional and observational design of the study precludes establishing any causal relationships or temporal order between screen time, cognitive function, and academic performance. This means that even though we tested a mediation model, we cannot confirm temporal precedence of the mediators; the associations should be interpreted as correlational. Second, the sample was limited to second-year medical students from a single institution (UNAM), which may affect the generalizability of the findings to other populations or settings. Importantly, our participants were in a pre-clinical year; their study habits and cognitive demands differ from those of clinical-year students or other disciplines, so digital media impacts might not be identical in those groups. Third, academic performance in this study was represented by a single course’s standardized exam score (Physiology). Fourth, the study relied on self-reported screen time, and accurately quantifying digital media use is challenging. Self-report measures may introduce recall bias or underestimation/overestimation of actual use. Fifth, in our network analysis we included some composite variables (e.g., total screen time, overall intelligence score) together with their constituent variables (specific app use measures, Gf and Gc). This part–whole inclusion can inflate certain network metrics, such as eigencentrality, because an aggregate node will naturally correlate with its components. We opted to retain these composites for completeness, but this redundancy should be kept in mind when interpreting the network structure.

This is a narrow measure of achievement; students’ abilities can vary across different subjects, and a single-course grade may not reflect their overall academic performance. Using a composite metric or multiple course outcomes in future studies would provide a more comprehensive view. Future research studies should employ longitudinal designs to track how digital media use influences cognitive and academic outcomes over time. Following students across multiple years and incorporating performance in multiple courses would help establish temporal sequences and determine long-term effects. Expanding the sample to other cohorts or institutions (including clinical-year students and different educational settings) will improve generalizability. Additionally, using objective measures of digital activity (such as device usage logs) alongside self-reports can provide more accurate assessments of screen time. Applying similar network-analytic approaches in these studies could further illuminate complex mediating pathways between digital habits and academic success, and ultimately inform interventions. This line of research can guide the development of strategies to help students balance digital engagement with learning—for instance, educational programs or tools to mitigate the distracting aspects of technology while leveraging its benefits for study.