Introduction

Many countries, including Switzerland, are facing a shortage of medical doctors1,2,3 amid increasing rates of stress-related mental health problems in healthcare personnel4,5. Difficulties emerge already during medical education, with recent international umbrella reviews showing high prevalence rates of psychological and behavioural problems in medical students6,7. Assessing meta-analyses published until 2022 or early 2023, these reviews report pooled prevalence rates of stress between 29.6 and 49.9%7, depression between 11.0 and 37.9%, anxiety between 7.04 and 33.8%, burnout between 13.1 and 45.9%, and suicidal ideation between 2.1 and 18.7%. Erschens et al.8 found that medical students scored significantly higher on scales of emotional exhaustion and depersonalisation than other students, highlighting their distinct burden.

According to a meta-analysis9, on average 9.1% of medical students dropped out before completing their degree. More recently, a nationwide study of medical students in Switzerland reported that 34% were questioning their career choice after their first practical experiences10. In a sample of medical postgraduates in China11, 58% expressed dropout intentions—markedly more than postgraduates from other fields. This trend is echoed by Sinval et al.12 who showed stronger dropout intention in medical students with higher burnout symptoms.

Burnout in medical students and residents has also been discussed as an impeding factor to professional development, related to mistakes in patient care13 and meta-analytic evidence shows higher rates of medical errors in physicians who screened positive for depressive symptoms14. Corroborating these findings, a meta-analysis of longitudinal studies revealed that the relationship between doctors’ work demands and the quality of their clinical care is mediated via emotional exhaustion15. These findings underscore the importance of improving well-being and reducing burnout in medical students—not only for their own benefit, but also to ensure their retention in the field and to ultimately offer the best possible care for patients.

Researchers have increasingly focused on protective factors that help individuals overcome stressful situations and maintain or quickly recover mental health, i.e. achieve a resilient outcome16. Resilience research requires longitudinal study designs, where relevant health outcomes are assessed in relation to periods of stressor exposure. Multiple interrelated aspects of the stressor and the individual’s response to it, comprising structural, social, and personal factors, influence eventual outcomes. In medical students, studies have explored mainly stress, mental health problems, and dropout (intention) as outcomes, with many only focusing on burnout.

Concerning structural aspects, stress has been linked mainly to academic demands but also financial struggles17. Students may be exposed to traumatic events during their internships, with a study in the US18 reporting that 56.4% of interns had experienced work-related trauma and 19% met criteria for post-traumatic stress disorder. A comprehensive meta-analysis by ref. 19, including 48 studies with over 36,000 trainee physicians, identified work-related factors as the strongest predictors of stress and burnout. High demands, concerns about patient care, and poor work environments showed the highest odds ratios. Considering steep learning curves20 and prevalent impostor syndrome21,22 tied to the many transition phases in medical education (i.e. proceeding from school to independent study to practical work), readiness to enter the workforce has been highlighted as another important factor linked to stress and mental health23. This may be related with research showing a discrepancy between students’ expectations of medicine school and clinical work in particular and their actual experiences24,25. Students who have a more accurate idea of what awaits them may have an easier transition into practice, suffering less stress in the process.

Findings from resilience research also underscore the relevance of personal factors such as self-efficacy26, optimism27, and positive appraisal style28 in overcoming stressful situations. Some personal factors have been investigated, providing meta-analytic evidence that higher self-efficacy is linked to lower burnout and stress in medical students19. Self-reported stress reactivity has been linked to depression29 and recovery to resilience28. However, self-reported stress reactivity has not been investigated as a predictor of mental health outcomes in a medical student population.

Regarding social factors, Kadhum et al.17 found interpersonal aspects to be important determinants of medical student stress. A psychologically safe team environment, where students feel able to take interpersonal risks and contribute openly without the fear of punishment or judgement30, has been shown to be important for educational outcomes31. Psychological safety has also been highlighted in qualitative reports from the clinic32 and is especially pertinent in medical education, where learning curves are steep while practices such as bullying and mistreatment are prevalent33,34. When stressful experiences are a part of the job, social support from the clinical team and peers is vital to maintaining good mental health35,36,37. Medical interns who have to rotate through different placements must regularly integrate into new teams, making it harder for them to receive such support.

While most findings are based on cross-sectional research and retrospective self-reports (though see the Intern Health Study, https://www.internhealthstudy.org/ and the ETMED-L study38), ecological momentary assessments (EMAs) are particularly useful for obtaining process-oriented information, as these data do not suffer from retrospective bias and show excellent internal and external validity39,40,41. For instance, the inclusion of daily mood ratings substantially enhanced the predictive accuracy for depression and suicidal ideation, compared to relying solely on baseline scores42.

Although previous literature on medical students includes many large meta-analyses, the majority of studies was conducted in US samples, questioning their generalisability to medical systems in other countries. Moreover, there is a lack of longitudinal studies investigating both mental health and career outcomes in the same sample, or skills improvement at all. Given the predictive value of momentary assessments for students’ mental health outcomes, a mixed-methods approach, combining traditional self-report measures with EMA, may provide important insights to inform prevention and intervention efforts.

In this study, we collected data from Swiss medical students just before the start of their internship year (T0), as well as three (T1), six (T2), and 12 months (T3) following internship start. To provide a comprehensive overview, we report on several outcomes that represent major goals of the internship year and which are pivotal to the success and sustainability of the healthcare system. These are skills improvement as a measure of learning, well-being and burnout as measures of maintenance of good mental health, and career motivation as a measure of potential attrition. We investigated a diverse set of empirically informed predictors comprising key personal, social, and structural factors. We combined these with self-reported expectations and EMA-derived variables to examine both putative risk and resilience factors, with the aim of providing actionable insights. Furthermore, we explored students’ open reports using a data-driven qualitative analysis technique. In sum, we (1) describe reported gains in skills, changes in well-being, burnout (personal accomplishment, emotional exhaustion, and depersonalisation), and career motivation, (2) examine whether students’ mood is lower while on-duty and the effect of social company using EMA, (3) investigate whether students’ internship experiences, expectations, and predisposing factors are associated with skills, well-being, burnout, and career motivation, and (4) explore open reports in which students reflect on what helped in reducing internship-related stress.

The results highlight three recurring risk and resilience factors, namely, work dissatisfaction, social and team factors, and degree of practical involvement. Motivation to continue studying medicine is related to lower burnout and greater skills improvement. We find that negative affect is higher while on duty, which is mediated by feeling less comfortable in present company. Students suggest organisational improvements and a healthy work-life balance to reduce internship stress.

Methods

This study forms part of the HMZ STRESS project, which investigates predictors and mechanisms of stress resilience in medical students starting their first medical internships as a real-world model of stress and resilience. The analyses for the current project were preregistered (http://osf.io/pwr2a) and deviations from this preregistration are outlined in Supplementary Methods 1.

Participants

Medical students were recruited from the University of Zurich, Switzerland via advertisements during lectures and posts on the official online learning platform. We also distributed flyers and made announcements via mailing lists, student association newsletters, and chat groups. Interested students were informed about the study procedures and screened via a phone interview to ensure they met the following inclusion criteria: (1) 5th year medical student about to start their internship year, (2) physically and mentally healthy (i.e. no current psychopathology), (3) fluent in German, (4) MRI-compatible (e.g. not pregnant, no metal in the body). We thus recruited a total of 115 medical students, N = 105 of whom participated in this study (72% female, aged between 21 and 35 years, M = 23.9, SD = 1.8; see Supplementary Fig. 1 for a flow chart). The sample size was determined for the overall project based on an a priori power analysis (see the project preregistration: osf.io/e5ksv).

Procedure

Two cohorts were collected: The first started their internships in 2022 (n = 21) and the second started in 2023 (n = 84). Students were invited to the lab for the baseline (T0) assessment, ideally within 1 month prior to internship start. This assessment involved obtaining informed consent, blood collection, video interview, fMRI assessment, and questionnaires with concomitant saliva collection. From the start of the internships, students completed EMA bursts of 14 days and wore a fitness tracker for ~6.5 months. After 3 months (T1), the students were invited back to the lab for blood and saliva collection, filled out follow-up questionnaires in the lab or remotely, and completed another 14 days of EMA. This was repeated after another 3 months (T2). Finally, after 12 months of starting the internship (T3), students completed a final questionnaire remotely. Here, we report on select questionnaires and EMA data (other results will be reported elsewhere). For procedural details outside the scope of the current study, please see the OSF project (osf.io/umtpr). Ethics approval was obtained from the Cantonal Ethics Commission of Zurich (KEK 2022-01169).

Questionnaires

Questionnaires were completed via online survey at T0, T1, T2, and T3. The exact timing of the testing was adjusted where possible to ensure the students had been working in an internship for at least 1 month prior to questionnaire completion. They were compensated CHF 12.50 per questionnaire.

Ecological momentary assessment

EMA was conducted with the SEMA3 app43. Students were sent five survey prompts per day for 14 days, starting the day before first internship start, and again after T1 and T2. The 14-day EMA periods were individually adjusted where possible so that data was collected during internship placements. At T0, students received a handbook that provided details of the EMA questions and schedule. They were compensated up to CHF 35 depending on compliance, with the full amount given for the completion of 70% or more surveys.

Measures

All measures that were acquired by questionnaire and EMA are listed in Supplementary Data 1.

Statistics and reproducibility

Statistical analyses were performed in R version 4.4.144. Model assumptions were assessed using the performance package45, and transformations were applied where indicated. Tables were generated with the sjPlot46 and gtsummary47 packages, and Holm-Bonferroni post-hoc p value adjustment was performed where specified with the emmeans package48. All p values reported are based on two-tailed statistical tests with an alpha significance threshold of 0.05.

The stressful internship events and psychological safety questionnaires were added after the first 20 participants had completed the T1 assessment. They were thus missing these data due to procedural changes, as opposed to e.g. participant characteristics. Hence, we think it unlikely that missings were systematically related to our variables of interest and sensitivity analyses in a smaller dataset without imputed values showed that, while effects changed slightly in size, the results remained unchanged. We conducted multiple imputation by chained equations (MICE) using the mice package49. Missing values were imputed by predictive mean matching using information from our T0 and other T1 predictors, as well as stressful internship event exposure and psychological safety questionnaire values for T2. The burnout questionnaire was added only for the second cohort; however, since this was an outcome variable, we did not impute these missing values. In addition, career motivation and open reports were only collected from the second cohort.

Ecological momentary assessment

To investigate whether students reported differing affect in moments when they were working versus when they were off duty, we fit multilevel models with the lmer function of the lme4 package50, using Satterthwaite degrees of freedom. Negative and positive affect were entered separately as dependent variables with the binary variable of being on/off duty as the independent variable and gender and burst (T0/T1/T2) as covariates. We included by-participant random intercepts and random slopes for being on/off duty. Negative affect was log transformed to remedy a violation of the assumption of normality. The estimates for the random slopes were extracted using the coef() function for each individual and denoted NA_slope and PA_slope, respectively. We related these slope estimates to average work dissatisfaction reported at T1 and T2 in linear regression models, including gender as a covariate. NA_slope and PA_slope were then entered into elastic net regressions to predict our outcome variables.

We investigated the impact of the social environment by including an interaction of feeling comfortable in company and being on/off duty into the multilevel models predicting negative and positive affect. For ease of interpretation, feeling comfortable in company was within-person mean centred. Again, burst was included as a covariate alongside gender. Although a maximal random effects structure is generally recommended51, lack of model convergence led us to follow recommendations by Barr52 and we removed the random slopes for being on/off duty and feeling comfortable in company, keeping only the interaction term.

In order to test whether feeling comfortable in company mediates the positive relationship (see Supplementary Fig. 3) between being on/off-duty and momentary negative affect, we conducted an exploratory multilevel structural equation model (SEM) using the lavaan package53. Being on/off-duty was entered as the independent variable, momentary negative affect as the dependent variable, and feeling comfortable in company as the mediating variable, all at level 1. Burst and gender were dummy coded and entered as covariates at level 1 and level 2, respectively. The participant id was used as the clustering variable with random intercepts and fixed slopes since random slopes are not supported in lavaan. Maximum likelihood estimation with robust (Huber-White) standard errors (MLR) estimation was employed to account for non-normality of standard errors.

Multiple Regression

We conducted multiple regression analyses to investigate theory-derived factors associated with our outcome variables (skills improvement at T2, the subscales of burnout at T1 and T2, well-being at T1 and T2, and career motivation at T3). Gender and prior experience were included as covariates in all models to control for differences in mental health outcomes54 as well as internship expectations and skills55, respectively. Specifically, for skills improvement, the predictor set comprised practical involvement (average of T1 and T2) and psychological safety (average of T1 and T2). For burnout subscales and well-being at T1, the predictor set was made up of: internship events (assessed at T1 with imputation), psychological safety (T1, imputed), work dissatisfaction (T1), stress reactivity (T0) and self-efficacy (T0). Accordingly, to predict burnout subscales and well-being at T2, we entered internship events (T2), psychological safety (T2), work dissatisfaction (T2), stress reactivity (T0), and self-efficacy (T0). In the models predicting well-being at T1 and T2, we added well-being at T0 as a covariate. Lastly, career motivation was regressed on burnout subscales (average of T1 and T2), well-being (average of T1 and T2), and skills improvement (T2). We used the lm function for the continuous outcome variables and an ordinal logistic regression model in the polr() function of the MASS package56 for career motivation change.

Elastic net regression

To complement the multiple regression analyses, we applied elastic net regression predicting the same outcomes (skills improvement at T2, the burnout subscales at T1 and T2, well-being at T1 and T2, and career motivation at T3). Elastic net models are better able to account for a larger number of predictors without overfitting and avoid problems of multicollinearity by allowing correlated predictors to coexist and separating their individual contributions57. They thus balance feature selection and shrinkage to produce the most parsimonious model. This method allowed us to expand the predictor sets used in the multiple regression analyses and include derived variables, namely experiences and corresponding prediction errors (experiences minus expectations; PE) of support, stressful situations, and responsibility and contribution and slope estimates of momentary affect. To minimise bias, we merged the predictor sets from the multiple regressions, incorporating all possible predictors in all models, unless the predictor was assessed chronologically after the outcome. The predictor set for each outcome can be seen in the corresponding result figure. The α parameter was set to 0.557,58 and standardisation was performed within the function though the coefficients are given on the original data scale.

For the continuous outcome variables, we used functions from the glmnet package59. Ten-fold cross-validation was performed using cv.glmnet to produce an optimal estimate for the regularisation parameter, λ. We used λ1se in order to obtain the most parsimonious model while maintaining accuracy within one standard error of the best model as generated by λmin as recommended by refs. 60,61. This cross-validation and coefficient estimation procedure was repeated 5000 times to minimise the influence of random chance (as in ref. 58) and the average coefficient values for each predictor across all repeats were extracted. Model fit was assessed with the deviance ratio, which is one minus the deviance of the model divided by the deviance of the null model including only the intercept59.

As career motivation was measured on an ordinal scale, we performed ordinal regression with elastic net penalty using the ordinalNet package62. Ten-fold cross-validation was performed with the ordinalNetCV function based on 100 possible λ values. The λ with the best average out-of-sample log-likelihood was returned for each fold and the mean across all folds was calculated to determine a single λ value, which was used to calculate predictor coefficients. This process was repeated 500 times and the average coefficient value for each predictor was extracted. Model fit was assessed with percentage deviance explained, defined as one minus the log-likelihood of the model divided by the log-likelihood of the null model62.

Topic modelling

We explored students’ open reports about their internship experiences with the aim of capturing factors not otherwise considered and complementing our quantitative analyses. We conducted unsupervised topic modelling to automatically extract common themes. First, we translated reports to English using the Deepl API via Python63, verified accurate translation, and reviewed reports on overall experiences, stressful and helpful aspects, and recommendations. Based on this, we decided to focus on analysing what students found most helpful and their suggestions for reducing internship stress and increasing resilience, combining these conceptually overlapping reports to form one input corpus. Second, we removed invalid reports (e.g. ‘I don’t know’, ‘na’) and split up multiple sentences into separate inputs or documents, so they could be assigned different topics. Third, we employed machine learning to automatically assign a topic to each document which is considerably faster than manual clustering as well as unbiased and reproducible. Specifically, we conducted BERTopic modelling64, version 0.16.4, based on the Bidirectional Encoder Representation from Transformers (BERT) language model65 in Python, version 3.11.566. BERTopic identifies meaningful topics in a corpus of documents by first creating vector representations of the texts that capture semantic meaning, grouping similar documents using clustering techniques, and then analysing word patterns within each group to extract keywords for easy interpretation of topics. We applied BERTopic modelling with the following six steps: (1) transforming documents into numerical representations (embeddings) using the all-mpnet-base-v2 sentence transformer, (2) reducing the dimensionality of the embeddings with Uniform Manifold Approximation and Projection (UMAP), (3) clustering the reduced embeddings using Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), (4) tokenising documents with CountVectorizer from scikit-learn, incorporating unigrams, bigrams, and trigrams, while removing stop words, (5) computing class-based term frequency-inverse document frequency (c-TF-IDF) with BM25 weighting to reduce frequent words, and (6) fine-tuning the topic representation with KeyBERT. A document including multiple topics is assigned to one topic only (single-membership model), but such cases were rare in our dataset. In contrast to traditional clustering approaches that rely on word frequencies, BERTopic clusters semantically similar texts—even if worded differently—into coherent topics without the need for predefined labels or training data. As such, it is a useful method for unbiased, exploratory analysis of open-ended survey data that can also handle short texts well67,68. Consequently, BERTopic has increasingly been applied in health psychology, see e.g.69,70,71. To ensure interpretability and semantic coherence of the extracted topics, we conducted hyperparameter tuning, adjusting key model parameters (e.g. minimum topic size, embedding model, and vectoriser settings) for optimal results. After modelling, we manually verified that the resulting topics adequately represented the data, merged topics if appropriate, and applied custom labels. Finally, two independent raters (L.M. and E.M.) labelled outliers which could not be assigned by BERTopic.

Reporting summary

Further information on our research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Descriptives

Supplementary Data 2 summarises demographic, internship-related, and EMA variables, zero-order correlations are presented in Supplementary Table 1. Shortly before the start of their internship year, 46% of the students reported little (0–3 months) or no prior medical work experience. Over the course of the internship year, students were required to work a minimum of 9 months in various departments at different locations. A 3-month rotation in internal medicine and 2 months in a surgical department were obligatory. Students reported experiencing some stressful situations (M = 3.84 on a scale of 1–5, SD = 1.11) and internship events (M = 7.11 (SD = 3.77) at T1 and M = 7.21 (SD = 4.27) at T2, out of a possible maximum of 44) during the internship year. On average, there was a decrease in well-being from T0 to T1 (b = −11.61, t(200) = −6.86, padj = 2.57e-10), but no change from T1 to T2 (b = 0.548, t(199) = 0.318, padj = 0.75). At the T0 assessment, 15 (14%) students met the ≤50 WHO-5 criteria of depression72, at T1, it was 30 (30%), and at T2, it was 32 (33%). Twelve months after starting the internship year, 10% of the students who provided reports stated that they were less likely to work as a doctor, 42% said it did not change their career choice, and 42% felt encouraged in their career choice.

Students’ momentary affect when on vs. off duty

Participants received 210 prompts overall (5 prompts × 14 days × 3 bursts) and completed around half (M = 48.43%, SD = 22.49%). Two participants were excluded from analysis as they had completed fewer than one prompt on or off duty. We minimised exclusions since multilevel models are generally good at handling unbalanced data. This left n = 103 participants with an average of M = 27.24 (SD = 14.31, range = 1–64) prompts while on duty and M = 77.08 (SD = 34.92, range = 5–154) prompts while off duty across all three bursts.

We conducted multilevel models predicting negative affect, momentary stress, and positive affect on duty compared to when off duty (Fig. 1), controlling for burst and gender. Negative affect was higher on duty on average (b = 0.04, t(93.02) = 4.09, p = 9.09e-05). To explore how stress (one of the components that makes up negative affect) specifically was affected by being on/off duty, we first mean aggregated momentary stress across all prompts within each burst and then performed square-root-transformation to better meet model assumptions. Momentary stress was higher while on duty (b = 0.06, t(464.64) = 4.01, p = 6.96e-05) than when off duty. Positive affect was also higher on duty on average (b = 0.07, t(95.83) = 2.07, p = 0.041). See Supplementary Table 3 for detailed results. We extracted random slope estimates for negative affect (NA_slope; M = 0.04, SD = 0.05, range = −0.14 to 0.16) and positive affect (PA_slope; M = 0.07, SD = 0.18, range = −0.39 to 0.51). NA_slope and PA_slope were negatively correlated (rs(101) = −0.56, p = 9.51e-10), i.e. students who experienced more negative affect at work also experienced less positive affect. NA_slope was positively associated, and PA_slope was negatively associated with average work dissatisfaction, see Supplementary Table 4 for detailed results.

Fig. 1: Momentary affect and stress when on versus off duty.
figure 1

Mean momentary a negative affect, b positive affect, and c stress when on versus off duty. Each colour represents an individual student, and the black dashed line is the mean across all students with error bars showing standard errors. n = 103 participants.

Full size image

To test whether the impact of being on duty on momentary affect was influenced by social company, we first conducted a moderation analysis. Specifically, we included the feeling of being comfortable in current company into the models predicting negative and positive affect, including an interaction with being on/off duty (Table 1 and Supplementary Fig. 2). Feeling comfortable in company was negatively associated with negative affect (b = −0.07, t(77.29) = −10.19, p = 6.04e-16) and positively associated with positive affect (b = 0.22, t(77.28) = 12.16, p < 2.22e-16). When feeling comfortable in company was added, being on/off duty at that moment was no longer significantly associated with negative affect (b = 0.00, t(6245.53) = 0.52, p = 0.602). The interaction term was non-significant for negative and positive affect, such that the effect of feeling comfortable in present company was similar when on duty and off duty.

Table 1 Multilevel model results on (log) negative affect and positive affect
Full size table

To follow up on these findings, we formally tested whether feeling comfortable in company mediated the relationship between being on/off duty and momentary negative affect. A significant total (β = −0.09, z = 4.51, p = 6.53e-06) and indirect (β = 0.10, z = 8.29, p < 2.22e-16) effect and a non-significant direct effect (β = −0.01, z = −0.47, p = 0.635) indicated a full mediation. See Supplementary Fig. 3 for the mediation model and results in more detail.

Students’ expectations and experiences

Two-tailed one-sample t-tests revealed that at T2, students reported experiencing more support (M = 0.20, t(97) = 2.15, p = 0.034) and fewer stressful situations (M = −0.83, t(97) = −7.29, p = 8.20e-11) than they expected at T0, but were on average accurate with respect to responsibility and contribution (M = −0.01, t(97) = −0.12, p = 0.902) (Fig. 2). Support, stressful situations, and responsibility and contribution experienced at T2 correlated positively with the respective PE (support: rs(96) = 0.58, p = 4.08e-10; stressful situations: rs(96) = 0.87, p < 2.2e-16; responsibility and contribution: rs(96) = 0.64, p = 1.18e-12).

Fig. 2: Differences between expected and experienced support, stressful situations and responsibility during the internships.
figure 2

Distributions of prediction errors (experienced at T2–expected at T0) for a support, b stressful situations and c responsibility and contribution. n = 98 participants.

Full size image

Associations of internship experiences with skills improvement, burnout, well-being, and career motivation

Multiple regression models were run on a small set of theory-derived predictors to predict skills improvement, burnout subscales, well-being, and career motivation. See Supplementary Tables 6–11 for results. In further exploratory analyses, we conducted elastic net regression with expanded predictor sets, the results of which are reported below and broadly align with the multiple regression analyses.

Skills improvement

We conducted elastic net regression to predict skills improvement reported at T2. The list of predictors and results are shown in Fig. 3. Experienced responsibility and contribution (M = 0.12, SD = 0.01) and its PE (M = 0.01, SD = 0.01), as well as average practical involvement (M = 0.10, SD = 0.01) and average psychological safety (M = 0.001, SD = 0.001) had positive average coefficient estimates across 5,000 repeats. Thus, higher experience (as well as PE) of responsibility for patients and active contribution, more practical involvement, and higher psychological safety were associated with higher improvement of skills reported at T2. Average work dissatisfaction showed a negative association (M = −0.08, SD = 0.01). The optimal lambda (lambda.1se) over the 5000 repeats was M = 0.23 (SD = 0.03, range = 0.14–0.36) and the average deviance ratio was 0.33 (SD = 0.03, range = 0.20–0.42).

Fig. 3: Prediction of skills improvement by elastic net regression.
figure 3

Distribution of coefficient estimates across 5000 elastic net repeats predicting skills improvement at T2. Boxplots show the first and third quartiles as hinges, 1.5 times the interquartile range as whiskers, and outliers as points. n = 96 participants.

Full size image

Burnout

In the elastic net regression models predicting the burnout scales depersonalisation, emotional exhaustion, and personal accomplishment at T1 and T2, there were no predictors with a consistent association and the deviance ratios were very small (<0.01) (Supplementary Figs. 4, 5).

Well-being

We found experienced support (M = 0.41, SD = 0.23) and responsibility and contribution (M = 0.11, SD = 0.17), as well as self-efficacy (M = 0.11, SD = 0.07) and psychological safety (M = 0.03, SD = 0.02) to be positively associated with well-being at T2 (Fig. 4). The PE for support also had a small positive coefficient on average (M = 0.01, SD = 0.05). Work dissatisfaction showed a negative association on average (M = −3.66, SD = 0.81), this was also found for well-being reported at T1 (see Supplementary Fig. 6). NA_slope had a small and very inconsistent negative coefficient on average (M = −0.01, SD = 0.20). The optimal lambda (lambda.1se) over the 5000 repeats was M = 8.60 (SD = 1.48, range = 4.70–14.35) and the deviance ratio was M = 0.16 (SD = 0.05, range = 0.00–0.30).

Fig. 4: Prediction of well-being at T2 by elastic net regression.
figure 4

Distribution of coefficient estimates across 5000 elastic net repeats predicting well-being reported at T2. Boxplots show the first and third quartiles as hinges, 1.5 times the interquartile range as whiskers, and outliers as points. n = 95 participants.

Full size image

Career motivation

The ordinal elastic net regression (Fig. 5) showed that PA_slope was positively associated with career motivation at T3 (M = 0.88, SD = 0.14), as was skills improvement (M = 0.32, SD = 0.04), experienced responsibility and contribution (M = 0.21, SD = 0.03), average personal accomplishment across T1 and T2 (M = 0.12, SD = 0.03), and the PE for support (M = 0.05, SD = 0.02). Average emotional exhaustion across T1 and T2 was negatively associated with career motivation (M = −0.04, SD = 0.02). The optimal lambda was M = 0.19 (SD = 0.02, range = 0.13–0.26). The percentage deviation had a median value of −0.26 and range = −0.65 to Inf. This suggests that the likelihood of the null model without predictors was higher (less negative) than the likelihood of the fitted model. Please refer to Supplementary Fig. 7 for a visual depiction of how the most important predictors related to career motivation. While greater improvements in skills, higher average personal accomplishment, and lower average emotional exhaustion were negatively associated with intention to change career (no change > less likely) and boosted career motivation (more likely > no change), it seems that PA_slope and support PE were only linked to enhancing career motivation.

Fig. 5: Prediction of change in career motivation by ordinal elastic net regression.
figure 5

Distribution of coefficient estimates across 500 ordinal elastic net repeats predicting career motivation at T3. Boxplots show the first and third quartiles as hinges, 1.5 times the interquartile range as whiskers, and outliers as points. n = 70 participants.

Full size image

Topic modelling of students’ recommendations for building resilience to internship-related stress

At T3, 77 students (92% of the 84 students asked) reported on what they found most helpful and what they would recommend to reduce internship stress and increase resilience. Following splitting into sentences and removal of invalid responses, we had 298 documents. BERTopic modelling revealed 16 topics to which 229 documents were assigned. The remaining 69 documents were labelled as outliers. Following human evaluation, three topics were merged into other topics due to conceptual overlap, resulting in a final 13 topics (see Supplementary Fig. 8 for a visualisation of documents by topic). Topic labels were created through human interpretation. Checks revealed that the topic Work-Life Balance & Self-Care included the subtopic Social Exchange, so we re-applied topic modelling on its documents, setting the number of topics to two. Documents thereby assigned to Social Exchange (n = 10) were merged with Work-related Social Support, forming Work-related Social Support & Exchange. Finally, two independent human raters (L.M. and E.M.) assigned topics to the outlier documents with only 15 remaining unassigned (e.g. ‘Problems can or may be addressed.’). Rater agreement was acceptable (70%).

The identified topics cover distinct but related themes referring to structural aspects (e.g. Organisation and Working Conditions), interpersonal internship aspects (e.g. Team Support & Integration and Open Communication & Feedback), and aspects related to work-life balance and stress reduction techniques (e.g. Stress Management, Self-Calming). Most participants reported organisational improvements, work-related social support and exchange, and a good work-life balance as recommendations to tackle internship-related stress. For instance, participants suggested a nationally regulated training structure, talking difficult situations through with others, and deliberately planning time to exercise or spend with friends during leisure hours. Supplementary Data 3 shows the topics by number of documents, including the number of participants who reported a topic, top five keywords, and representative documents.

Discussion

We investigated Swiss medical students during their first year of clinical internships. We utilised EMA to probe how the internships and social company influenced momentary mood. Regularised regression was used to assess the importance of various predictors in the improvement of skills, burnout, well-being, and career motivation. These analyses were complemented with topic modelling to explore what students found helped them with internship-related stress and what they suggest would increase resilience. We identified three overarching predictors associated with our outcomes, namely work dissatisfaction, social factors, and active involvement. While students also highlighted structural and social factors in open reports, they additionally noted personal factors, suggesting a good work-life balance, self-care, and stress management to improve well-being.

The well-being of the students decreased from pre-internship to 3 months and remained at the same level to 6 months, aligning with a previous study in Swiss medical students73. During the internship, levels of depression were similar to global rates among medical students74. Nevertheless, for the majority, the internship year did not lessen their ambition to continue in the medical career. This is in contrast to a recent nationwide survey in Switzerland where 34% were considering not working as a doctor after the internship year10. Although efforts were taken to reach all students in the academic year, our sample likely reflected a more enthusiastic portion, which may explain this difference.

Both negative and positive affect were significantly, albeit slightly, higher on than off duty. Differences in positive affect when on vs. off duty (individual slope estimates) showed the strongest link with career motivation, while neither differences in negative affect nor work dissatisfaction were associated. Our findings align with prior research emphasising the importance of positive emotions in occupational settings75 and their role in motivating career goals76. While much of the existing literature, e.g. Hockey et al.77 has focused on negative affect and its association with the intention to quit during medical training, our analysis suggests a distinct role for positive affect. Exploratory visualisations (Supplementary Fig. 7) show that it may actively boost motivation rather than merely protect against attrition. Future studies should further investigate the mechanisms by which positive affect influences career development and investigate whether fostering positive affect could enhance resilience and retention across medical training.

Although stressful internship events were reported, few students experienced more stressful situations than they were expecting. This may partly explain the lack of association of stressful events or situations, such as unexpected patient deaths, with mental health. Others have made similar observations19,78,79,80,81, despite links between stressful internship events and post-traumatic stress disorder18. Overall, students’ expectations of the internship were relatively accurate, and PEs rarely provided additional predictive information. One exception was that those who experienced more support than they had expected were more likely to be encouraged by the internship year to continue studying medicine (Supplementary Fig. 7). Such a change in motivation could be driven more by prediction errors than experience as they generally give rise to belief updates82.

Work dissatisfaction was consistently found to be associated with skills improvement and well-being. Furthermore, as the sole predictor associated with depersonalisation and emotional exhaustion in our sample (though only in multiple regression models), it may offer a potential target for early intervention before burnout symptoms manifest. This is crucial for student retention because those with higher emotional exhaustion (but not work dissatisfaction) were less motivated to continue studying medicine, a finding corroborated in a meta-analysis83 and EMA study84. Social factors came out as important in most of our analyses. When students felt more comfortable in their company, they experienced lower negative and higher positive affect, independent of whether on or off duty. Students felt less comfortable when on duty, which fully accounted for the observed higher negative affect on duty. While this exploratory mediation analysis highlights the importance of social company on negative emotions, these results should be replicated with temporally separated variables.

In the prediction analyses, psychological safety was positively related to skills improvement, which aligns with research highlighting the importance of the social environment in learning85,86. Furthermore, support, and to some extent psychological safety, were linked to well-being (as found in Helou et al.87), but not to symptoms of burnout, despite evidence of their importance13. In terms of career motivation, those who felt more supported than they had expected were more likely to want to continue in the medical profession. In contrast, mistreatment can discourage continuation in medical school88 while teaching by humiliation can contribute to stress and burnout34.

Students highlighted the importance of a good team, social support (both from supervisors and peers), and open communication and feedback in the open reports, in line with similar research89. These factors may outweigh patient-related clinical events in importance. For example, when asked about the three most stressful clinical events, medical students were more likely to cite poor team dynamics and difficult encounters with other staff than unanticipated patient deaths or medical errors37. In addition to expecting such stressful situations, the impact of stressful internship events may have been mitigated by the overall high support and psychological safety (Supplementary Data 2) reported in our sample90,91. Further work in larger samples may investigate the role of these factors in explaining mixed results concerning the association between stressful clinical events and poor mental health.

More practical involvement was linked with skills improvement, while active contribution and responsibility for patients related to skills improvement as well as well-being and career motivation. Others have found a preference for hands-on participation in medical students and have highlighted how it promotes professional confidence92. Given that students choose to study medicine due to the meaningful nature of the work10 and that patient experiences often have a positive effect81, active involvement and responsibility for patients may improve fulfilment. From the open reports, we gauged that students were motivated to get involved and were sometimes not given enough tasks (‘Where you can leave when there’s nothing more to do: Don’t just do time.’) or tasks with very low/high demands (‘Good training concepts in the hospitals so that neither boredom nor excessive demands occur.’). Thus, assigning and communicating tasks that meet the skill levels of the students and are hands-on would benefit learning as well as mental health and retention.

Our study has limitations. First, the sample only included Swiss medical students. Sources of medical intern stress have been shown to differ by culture17,93, however meta-analyses find common challenges and contributors to stress and burnout across multiple countries7,19. Second, while we focused on certain structural and social factors, we did not collect quantitative data on working conditions, such as financial burden, mistreatment, and working hours, which have been linked to stress and mental health13,19,94. These factors were mentioned in students’ open reports, with work-life balance and working conditions as frequent topics. Future studies should include such factors and investigate the impact of socio-economic and minority status in a more heterogeneous sample. Third, since the elastic net analyses are data-driven and the sample size was too small for a train-test split, these results are exploratory and should be validated in further studies. Fourth, the model fit of the ordinal elastic net regression was poor and it would be beneficial to investigate factors influencing a more sensitive (ideally continuous) measure of career motivation. Moreover, the measures of expectations and experiences, prior medical experience, and skills improvement were developed for this study and should be validated for use in future studies. Finally, internal consistency of some survey measures (skills, work dissatisfaction, and burnout subscales) was lower than is commonly regarded as acceptable, potentially reflecting a more complex picture in this population. Such low internal consistency in the burnout measures may explain why previously identified variables were not significant predictors. Additionally, 6 months may be too short to see significant changes in burnout or capture associations15,95.

Despite these limitations, our study has several notable strengths. The prospective longitudinal design provides a more robust framework for examination of causal relationships than the more common cross-sectional studies19. In line with recommendations from resilience research, we investigated relevant outcomes in relation to a stressful transition phase, i.e. the internship year. Moreover, our inclusion of momentary affect reports represents an innovative methodological strength. Such real time assessments during students’ work and private life are relatively rare, but have demonstrated considerable potential in predicting mental health outcomes42,96,97. By capturing in-the-moment experiences, our study offers a nuanced understanding of the emotional and psychological factors influencing stress, resilience, and career motivation.

Our results, though observational, provide insights for the development of interventions aimed at improving medical education and clinical practice. These implications are grouped into three themes: personal, social, and structural factors.

In open reports, students highlighted the role of personal resources and adopting proactive coping strategies in resilience against stress. Planning time outside of the internship for friends, sport, and self-care was a popular topic and students noted stress management and self-calming. Despite the stigma around mental health in the medical sector98, it is promising that psychological support was also mentioned (‘I think it might be helpful to have a contact person or centre to go to in extreme stress or situations’). Identifying at-risk individuals and correctly triaging students to appropriate psychological care is vital99. However, it can be challenging to predict how an individual will respond to a new stressor from self-report alone, as shown by the poor predictive ability of self-reported stress reactivity in our sample. More objective measures of stress reactivity (e.g.73) or assessment during stressor exposure may prevent biases and provide more relevant information.

Interventions to boost personal resources (e.g. mindfulness) are generally effective in reducing stress and burnout in medical students and physicians100,101,102,103 and could be offered prior to or during the internship. Training of self-efficacy has shown promise in reducing anxiety and possibly even improving patient satisfaction104,105. While shown to be important in stress and mental health in medical students19, self-efficacy was only weakly linked to well-being in our sample, and not burnout. As discussed previously, this may be due to limitations in the measurement of burnout.

Social factors, including peer support and team environment, present another potential target for interventions. Some students suggested peer-led support groups (‘sharing is caring’) which have been effective elsewhere106,107. Another student highlighted how in-depth discussion of difficult treatment cases with their supervisor helped them process their experiences (‘They gave us the time and space to ask questions and not just have to continue working immediately.’). In line with this, one student recommended feedback meetings (‘Strengthen open communication, feedback culture (e.g. feedback meetings)’), however, some also noted that supervisors themselves (i.e. junior doctors) are often under considerable strain and may lack capacity. Improvements require system-level change along with interventions for the individual. For example, training students on how to respond to and learn from errors as well as improving error culture would reduce associated stress108,109,110.

Fewer studies have investigated structural or organisational interventions111,112, despite evidence for their importance and effectiveness in reducing physician burnout103,113. Students in our sample frequently mentioned organisational challenges, such as structure and planning on the part of the host hospital, and the role of the university in supporting and preparing them before the internship year and throughout. This could take the form of clinical and communications workshops as well as handbooks and introductory meetings, which have been shown to increase confidence114. Working conditions were also a topic. In Switzerland, medical placements vary in financial remuneration, some even being voluntary, and expected working hours. There is currently discussion around fixing working hours to a maximum 46 h per week115. However, reducing working hours alone is unlikely to improve mental health in medical students116.

The internship year marks a pivotal and transformative stage in medical training. Our results confirm that it is a demanding period for medical students whose experiences are shaped by personal, social, and structural factors. Support and team dynamics emerge as critical factors—likely surpassing the impact of patient-related stressors. Notably, active involvement and participation enhance learning, well-being and career motivation, while positive affect plays a central role in sustaining long-term commitment to the medical profession. Crucially, many of the factors are amenable to interventions which should be developed in cooperation with universities and internship host institutions. Future research should adopt a participatory approach, involving medical students at an early stage in the design process to ensure interventions are tailored to their needs and experiences.

Conclusion

This study leveraged quantitative and qualitative data collected at multiple time points during medical students’ first internships, providing a comprehensive assessment of their experiences and how these relate to skills development, mental health, and career outcomes. Our results offer actionable insights, highlighting modifiable aspects of medical education that, if addressed, may decrease stress levels among students, foster resilience and, by extension, improve patient care.