Adolescence is a critical period of life when people explore, experiment, and are exposed to particular risk factors such as substance use, violence, bullying, and academic problems. It is when the course is set for health in adulthood [1]. The current debate about an adolescent and young adult mental health crisis in Western countries [2,3] is therefore all the more important, especially because it is substantiated by a number of recent findings: a study comparing 2 UK birth cohorts found that while no difference was found during childhood, there were markedly more emotional problems during adolescence in individuals born 2000‐2002 compared to individuals born 1991‐1992 [4]. Consistently, a recent comparison of 2 methodologically standardized and comparable Dutch population surveys showed almost a doubling of affective disorders between the years around 2008 and around 2021, with the increase being particularly pronounced among young adults [5]. In the United States, after a years-long decline, rising depression among adolescents has been reported since about 2010 [6], accompanied by an increase in suicide rates in this age group [7]. While prevalence rates for externalizing mental health problems among children and adolescents in Germany were declining before the COVID-19 pandemic, emotional problems also increased here between 2003‐2006 and 2014‐2017, especially among girls [8]. These alarming results are in line with findings from other countries, although there is heterogeneity and effect sizes appear slightly smaller in meta-analytic summaries [9]. While reduced stigma and an increased readiness to report mental health problems might partly contribute to these trends, it is remarkable that rising prevalence rates are also reported in studies that used the same standardized and well-validated clinical interview over different time periods [5]. Moreover, investigations relating time trends of self-reported mental disorders to time trends of stigmatization of mental disorders in the same regions did not find evidence that reduced stigma explains the increase in self-reported mental health problems [10].

Summarizing the evidence suggests that since about 2010, there seems to be a marked cross-national increase in mental health problems of adolescents, particularly concerning internalizing or emotional problems and affective disorders. The causes for this increase are insufficiently understood [2]. Proposed risk factors cover rising screen time and social media consumption [11], worsened peer relationships with loneliness and social isolation [12], and physical inactivity [13]. However, the predictive strength and the causal interplay of these factors are still unclear and debated.

To address this research gap, we used multivariable machine learning models to assess the development of mental health from early to late adolescence in members of longitudinal cohort studies. This allows us to model nonlinear relationships between numerous predictors in early adolescence and the future development of mental health, as well as to test the quality of the resulting predictions by means of cross-validation. We investigated 2 large independent population-based cohort studies of adolescents, the UK Millennium Cohort (MC) study [14] and the German Health Interview and Examination Survey for Children and Adolescents (KiGGS) [15]. This approach allows a robust comparison of results between 2 independent cohorts from different countries.

The analyzed cohort studies were conducted in countries affected by the mental health crisis, and the investigated time period covers exactly the “critical period” of the crisis development from 2010 onwards. Thus, the developed models can also be used to add further evidence to the debate on driving factors of this deterioration. More specifically, it can be examined which simulated changes in risk factors that have been implicated in the ongoing mental health crisis actually led to worse predicted mental health. However, assessing beyond pure predictive relationships the causal impact of these risk factors on the development of mental health involves stronger assumptions and challenges interpretation [16]. In this work, we draw on our multivariable models in conjunction with recent research on the causal interpretation of predictive models [17] and on the appropriate selection of covariates [18]. Based on this approach, we assess the independent effect of proposed driving factors for the adolescent mental health crisis (increased screen time, less physical activity, and worsened peer relationships).

We aimed to answer 3 related research questions. (1) How well can mental health in late adolescence be assessed based on assessments in early adolescence? (2) Which assessments in early adolescence are important for these assessments? (3) What are the independent impacts of physical activity, screen time, and peer relationships on unfavorable mental health trajectories of adolescents in the years after 2010?

In this study, we used interpretable machine learning pipelines on 2 large European cohort studies for 2 related goals: First, we tested how well and based on which assessments the development of mental health during adolescence can be assessed. Second, we assessed the causal effects that high screen time, low physical activity, and worsened peer relationships had on an unfavorable development of mental health. Because the analyzed time period lies in the critical years after 2010, we thereby added evidence to the discussion on reasons for the mental health crisis unfolding during that time period.

Using cross-validation, we showed that multivariable models had medium to high predictive accuracy with Pearson correlations between predicted and actual mental health in late adolescence between 0.38 and 0.61. Although direct comparisons are complicated by differences in the study design and outcome measures, this predictive accuracy is in the upper range of the few studies published to date on the prediction of mental health problems in adolescents using comprehensive multivariable models [38,39]. Predictive performance was improved with nonlinear GBMs, which are known to perform well for medium-sized tabular datasets [31]. However, these figures also make it clear that the calculations of the models should be interpreted as assessments and not as absolutely accurate predictions. The relative superiority of the nonlinear model compared with the linear model suggests that at least some relationships between the investigated predictors and the development of mental health are not optimally modeled in a linear and additive fashion. This is also reflected, for example, in the disproportionately strong increase in predicted mental health problems with very severe peer problems.

We used univariable and multivariable feature importance analyses to identify predictive factors. Here, we were able to replicate and extend previously described risk and resilience factors. As expected and consistently demonstrated in previous studies [40,41], mental health problems in early adolescence were the strongest predictors of mental health problems in late adolescence. This result captures a certain stability of mental health problems during adolescence, but the contribution of other factors to the predictions shows that the development of adolescent mental health depends on multiple determinants that go beyond a mere “autoregressive” continuation of baseline mental health problems. Among family factors, mental health problems of the parents were strongly related to mental health problems in their adolescent children. For preventive strategies, it is important to understand the mechanisms of this intergenerational transmission of risk and to develop targeted interventions based on this understanding [42]. Indicators of physical health problems (chronic health conditions, self-reported pain, and impairments in general health) were also associated with future mental health problems. This finding substantiates current evidence that physical health problems in early life predict a worsening of mental health in adolescence and that children with chronic conditions might hence be particularly in need of support [23]. Low socioeconomic status (ie, income and education) of the parents predicted an unfavorable development of their children’s mental health during adolescence, which corroborates previous findings with different cohorts [43,44]. In both cohorts, we moreover found that female sex predicted worse mental health in late adolescence also when statistically correcting for baseline mental health symptoms in young adolescence. This finding is in accordance with a well-established gender gap in depressive symptoms, which starts to manifest during puberty, increases until adulthood, and has multifaceted reasons [28].

We leveraged recent advances in covariate selection and causal interpretation of predictive models [17,18] to approach the causal impacts of factors implicated in the deterioration of adolescent mental health in the past decade (less physical activity, more screen time, and worse peer relationships). This analysis revealed improved predicted mental health in late adolescence when more physical activity was assumed in early adolescence. However, the direct effect was small, particularly in the MC. We found that, except small effects of very much screen time in the KiGGS cohort, reducing screen time alone did not directly entail meaningful improvements in the predicted development of mental health in both cohorts. Conversely, good peer relationships predicted better mental health trajectories in both cohorts even when all other factors were held constant.

What do these results suggest in the context of current debates on the causes of the youth mental health crisis?

First, according to these analyses, adolescents would have had a more favorable mental health trajectory during the period of the observed deterioration in adolescent mental health if they had better peer relationships at ages 12-14 years, even if a wide range of other determinants of mental health had remained unchanged. It is noteworthy in this context that an overall deterioration in peer relationships of adolescents was observed across countries during the same period as the deterioration of mental health [45] and that proposed influencing factors such as increased screen time were shown to potentially have a detrimental impact on peer relationships [46]. It thus seems plausible that increased peer problems, themselves the result of a complex change in behavioral patterns of adolescents, represent an important direct cause of the deterioration of mental health. Given the growing evidence for the effectiveness of general prevention programs (such as antibullying school interventions) and individual therapeutic strategies (such as social skills training or self-esteem reflections) to improve peer relationships [47], it appears that a combination of these interventions would be a promising public health approach to the adolescent mental health crisis.

Second, adolescents would not have altered mental health trajectories if they had less screen time at ages 12-14 years when the other determinants of mental health had remained unchanged. This finding ties well with a recent panel network analysis showing that when multiple determinants of mental health were taken into account, social media use only had a very small influence on the development of mental health [48]. There is an intense debate around relationships between screen time, particularly social media use, and mental health in adolescents and young adults. The debate is fueled by a remarkable coincidence between the mass spread of social media and the deterioration of adolescent mental health around 2010 [49]. However, longitudinal studies could not prove a clear association between social media use and adolescent mental health [50,51]. It seems plausible that heavy social media use overall is part of a complex change in adolescent behaviors that might, potentially mediated through increased peer problems, contribute to mental health problems [11]. As illustrated in Figure 3, the absence of a direct effect in our study would be compatible with a model in which screen time exerts a negative effect on mental health via other downstream factors (eg, factors impacted by screen time and themselves impacting mental health, such as time spent with friends, body dissatisfaction, and sleep problems). Such modifiable factors on the causal pathway between exposures such as screen time and mental health outcomes could be targets for intervention strategies. Characterizing them more precisely would be an important undertaking for future research, which could use causal inference in combination with higher frequency surveys [18].

Third, adolescents would have a slightly improved mental health trajectory if they had been more physically active at ages 12-14 years, even when the other determinants of mental health had remained unchanged, especially those with very low levels of physical activity. In line with this finding, exercise interventions were shown to have small beneficial effects on adolescent mental health [52], and a recent review of reviews concluded that there is only “partial evidence” for a causal impact of physical activity on adolescent mental health [53]. Interestingly, the authors of this review discuss the possibility that the association between physical activity and adolescent mental health “may be linear, curvilinear, or contain a threshold, after which no further gains in mental health are made” [53], which would be consistent with our findings that increases in physical activity only have a beneficial direct effect in the case of low baseline levels. However, the small effect strengths, together with meta-studies that used aggregate data to demonstrate that the level of physical activity in adolescents has not relevantly decreased in Western countries between 2001 and 2016 [54], render it unlikely that the observed substantial worsening of adolescent mental in the past decade is due to decreased physical activity.

Some limitations of our study are relevant when interpreting the study results. First, causal inference from observational data is challenging, and results can depend on assumptions that are hard to test empirically. If the research questions cannot be investigated in a randomized controlled manner and one has to resort to observational data (as in our case), the best possible selection of covariates must be undertaken to approach causal interpretations. We applied principles developed for this purpose in our study [18]. Nevertheless, it should be kept in mind that the evidence for causal relationships in observational studies is generally lower compared to randomized controlled trials [55]. The results should therefore be interpreted in conjunction with other research work (eg, cohort studies, clinical observations, and analysis of aggregated time series) in order to gain a comprehensive picture of the factors contributing to the deterioration of young people’s mental health. Moreover, the inability to reliably distinguish confounders from mediators only allows for the calculation of direct effects, and it is possible that the factors under investigation have additional effects through the modification of mediator variables. To characterize such mediating relationships more precisely, future research could explore multiple-wave longitudinal studies in conjunction with causal inference [18]. Second, although the mutual validation of results in 2 independent cohorts is a strength, both cohorts differed in some design features, which reduces the comparability. In particular, the MC is age-homogenous, whilst in the age-heterogeneous KiGGS cohort, participants with an age similar to the MC participants were selected, leading to a smaller case number. Due to a longer interval between baseline and follow-up examination, participants in the KiGGS cohort were younger at baseline. This could contribute to observed discrepancies, for example, to the result that the direct effect of physical activity was more pronounced in the KiGGS cohort. In addition, predictors were only partially overlapping between the 2 studies. Despite these differences, it is noteworthy that key predictors of future mental health, such as the parents’ mental health, the participants’ baseline mental health, sex, peer problems, and physical activity, were found in both cohorts and there was a qualitative agreement on the estimated direct causal effects. The robustness of these results would be further strengthened by an analysis of additional cohorts from other countries. Third, not all proposed driving factors of the observed adolescent mental health deterioration could be investigated with our study. For instance, factors affecting society as a whole, such as growing social inequality and pressure to perform in education, have also been suggested [56,57] but could not be analyzed with our research design, which leveraged interindividual differences in risk factors to assess their impact on mental health development. Relatedly, the influence of new stressors such as the COVID pandemic, the war in Ukraine, and the increasingly noticeable effects of climate change could not be examined due to the time period under study. However, once appropriate data are available, our ML pipeline could be readily adapted in future work to predict the risk for mental health problems in the face of societal crises and identify its determining factors.

In summary, we showed that the development of adolescent mental health can be assessed by a combination of variables from early adolescence. Our analyses suggest a strong influence of peer relationships, and a deterioration of these through different reasons may contribute as a causal factor to the observed adolescent mental health crisis. Interventions to improve social relationships with peers could be explored as countermeasures against further deterioration of adolescent mental health.