The effect of resilience training on resilience and stress in medical students: a systematic review and meta-analysis

Abstract

Background

Medical students are known to have high rates of stress, burnout, anxiety, depression, and suicidality. Resilience training has been shown to be beneficial for stress among medical students. This systematic review and meta-analysis evaluated the literature on resilience interventions in medical students and assessed the impact of those interventions on both stress and resilience outcomes.

Methods

A comprehensive search was completed including multiple databases, bibliographic, and hand-searching. Twenty-one studies were included, 15 with resilience outcomes, nine with stress outcomes, and three with both. The standardized mean difference effect sizes were calculated using a random-effects model.

Results

Resilience interventions had a small-to-moderate effect on resilience (d = 0.39, k = 17, n = 15, 95% PI [-0.10 – 0.87], p < 0.001), and a moderate effect on stress (d = 0.53, k = 10, n = 9, 95% PI [-0.96 – 2.02], p = 0.01). Sub-group analyses found that clinical students displayed higher resilience than preclinical students post-intervention. Significant heterogeneity was present in both models. Funnel plots, Egger’s regression intercept, and Duval & Tweedie’s trim and fill tests revealed at least mild publication bias. Risk of bias in the primary studies was high.

Conclusions

This meta-analysis suggests that resilience interventions are moderately effective at reducing stress and mildly effective at increasing resilience among medical students. Resilience intervention may have more of an impact on stress than resilience, and clinical students may benefit more with regards to resilience, but not stress. Further research would be beneficial to examine specific factors that influence resilience most.

Trial registration

https://www.crd.york.ac.uk/PROSPERO/view/CRD42024501653

Supplementary Information

The online version contains supplementary material available at 10.1186/s12909-025-08171-x.

Background

Chronic stress in medical students is not a new phenomenon, but research indicates that their mental health is worsening [50]. Close to 30% of medical students report signs and symptoms of depression, up to 75% report burnout, and more than 10% report suicidal ideation [50]. Chronic stress in healthcare professionals has been worsening for many years [34, 41, 51, 52], but it is thought that physician burnout has its origins in medical school [1]. In viewing the looming healthcare shortages and high rates of healthcare workforce attrition [10], interventions aimed at preventing and reducing burnout and chronic stress as early as possible are of paramount importance, and resilience interventions have established improvements in the mental health of health profession students [21, 29].

Resilience is the ability to adapt to life changes and challenges, requiring mental, behavioral, and emotional flexibility [4]. Resilience has been shown to be a modifiable trait [53], and one which can assist with medical student well-being by reducing burnout [29], stress [21, 22, 54], anxiety [21, 22, 54], and depression [21]; as well as increasing empathy [11], self-awareness [44], and emotional regulation [46]. Many medical students have been found to have difficulty adjusting to change, which is associated with anxiety, depression, and suicidal ideation [54]. However, resilience can be enhanced with training on growth mindset and flexibility, among other skills [43, 53].

Interventions for resilience among medical students have included curricular [6, 44] and extra-curricular workshops [54],a Mindfulness, Affective Reflection, Impactive Experiences, and Supportive Environment training (MaRIS model; [11], a Training for Awareness, Resilience, and Action (TART; [21]); an Everyday Resilience program run by peer mentors [24]; and the Penn Resilience Program (PRP; [46]). There have been several recent systematic reviews and meta-analyses about the positive effects of resilience in the general population [28, 37, 38], as well as in nurses [67] and physicians [3]. The results of these meta-analyses show a small but significant improvement in resilience overall. To our knowledge, there are no meta-analyses assessing the summary effect of resilience interventions among medical students, or any which differentiate summary resilience vs. stress outcomes from resilience training. Our preliminary review of primary studies on resilience interventions found resilience and stress to be the most common outcomes. Other outcomes, such as post-traumatic stress disorder, depression, and empathy, do not appear to be consistently measured in the literature, and as such were unlikely to provide sufficient data to assess in a meta-analysis. Therefore, this systematic review and meta-analysis addresses the following research questions: What are the effects of resilience-building interventions on medical students’ resilience levels and/or stress levels? Which specific interventions or participant characteristics change the degree of effectiveness of resilience training?

Methods

This systematic review and meta-analysis followed the Preferred Reporting Items for Systematic Reviews and MetaAnalyses (PRISMA) guidelines [45]. It was pre-registered at PROSPERO (CRD42024501653). IRB approval was not required as this study involved only previously published material. Clinical trial number: not applicable. Guidelines in “Enhancing rigor in quantitative meta-analyses for mindfulness research: A comprehensive guide” [55, 56] were followed with adaptions for resilience interventions.

Inclusion & exclusion criteria

Inclusion criteria for articles were: (1) must contain an intervention which explicitly aims to improve resilience or has the word ‘resilience’ in the name of the intervention; (2) addressed the medical student population, including programs training medical doctors (MD), doctors of osteopathy (DO), or international students (MBBS); (3) any study design; (4) written in or able to be translated into English; (5) reporting on resilience or stress as an outcome measure, including measures of ‘academic stress,’ ‘perceived stress’ or ‘distress;’ and (6) peer-reviewed. Gray literature was acceptable.

Exclusion criteria including: (1) studies with a combined population without separately reported outcomes, i.e., medical students and dental students; (2) studies which did not have a specific measure of resilience or stress; (3) studies without sufficient data to calculate an effect size, including qualitative studies; (4) non-primary studies, e.g., systematic reviews, letters to the editor, companion studies. No restrictions were placed on study location or date.

Search strategies

A research librarian was consulted to maximize the rigor of the search. The search included database, bibliographic, and hand searches. English language filters were used. The following 12 databases were searched: EBSCOhost (including Academic Search Elite, Alt Healthwatch, CINAHL, ERIC, Health Source – Nursing/Academic Edition, Primary Search, OpenDissertations, APA PsycINFO, MEDLINE); EBSCO Discovery, in the categories (1) Complementary & Alternative Medicine, (2) Health & Medicine, (3) Psychology, and (4) Public Health; PubMed; Trip; Web of Science; Sociological Abstracts; SveMed +; and Education Full Text. Publication Finder was used to search in the following three journals: Journal of Adversity and Resilience Science, Journal of Mind and Medical Sciences, and the International Journal of Medical Students, as they were likely to contain research on resilience in medical students. Research registries were also searched including ClinicalTrials.gov and ISRCTN.

A structured search strategy was used, beginning with Boolean terms. When necessary, these were then refined to MeSH terms. Search terms included “resilience,” “resiliency,” “stress,” and “medical students.” A bibliographic search was done on the systematic reviews [3, 33]. We estimated a minimum of approximately 10–15 studies to perform the meta-analysis for each outcome (resilience and stress), due to the expectation that heterogeneity would be significant [8].

Data management

Zotero and Excel were used for data management. An Excel spreadsheet was used for tracking the search date, source of the article, search terms, and reason for exclusion.

Data extraction and coding

A comprehensive codebook was developed per guidelines in Cooper [15]. The codebook was refined by two researchers. Coders trained in meta-analytic methods independently extracted data from all eligible studies with two researchers coding each article,after coding was complete, codes were compared and any discrepancies discussed and resolved [15]. Categories of variables included the study title, authors, year of publication or in the case of unpublished research the year of completion, research design, funding existence and source, publication source, population characteristics and demographics (medical program type, country, students’ mean age, gender, race/ethnicity), who delivered the intervention (e.g., psychiatrist, psychologist, faculty member, clinical instructor, peers), measures of reliability and validity (blinding, randomization, allocation, attrition reporting), intervention details (name, length of intervention, duration of sessions, total sessions, location of sessions, delivery mode including in-person or virtual, synchronous or asynchronous), outcomes (name of outcome measure, time points, direction of effect), and results (sample size, raw means, standards deviations, standard error, difference in means, and direction of effect for the intervention). All timepoints were coded, pre- and post-intervention. For the meta-analysis, the most proximal post-interventional time points were used.

Data analysis

Data analysis was conducted using Comprehensive Meta-Analysis Software (CMA). Eligible studies included participants and interventions with a wide variety of characteristics, necessitating the use of the random effects model, which accounts for between-study and within-study variance [7, 8]. Standardized mean difference effect sizes using Cohen’s d were calculated in CMA; Cohen’s d was chosen over Hedges' g as g can result in more biased meta-estimation [36]. Cohen’s d is divided into small effect sizes (~ 0.2), moderate effect sizes (~ 0.5) and large effect sizes (~ 0.8; Cohen, 2013). A positive effect size was chosen to indicate improved resilience or stress, while a negative effect size indicates reduced resilience or worsened stress.

Heterogeneity of the effects was assessed using both prediction and confidence intervals; the prediction interval gives the range of effect sizes across the population while the confidence interval estimates the accuracy of the mean effect size [8]. The Knapp-Hartung adjustment was used as it provides a wider and more accurate interval demonstrating the range of dispersion of effect sizes based on the t distribution [8]. Cochrane’s I² is reported as the proportion of variance in observed effect sizes compared to the variance in the true effects (T²) [8]. The I² value is useful to determine if the observed effect sizes’ variance is representative of the population or not [8]. T² (variance of true effects rather than sampling error) and T (the standard deviation of true effects) are used to calculate the prediction interval [8], so both are reported here. The Q value and p of Q are not reported for the full study (as they only apply to a fixed-effect model in which all studies share a common effect size; [8]). However, Q and p of Q are reported for meta-regressions (see Moderator Analyses).

Moderator analyses

Moderator analyses were undertaken to further explore the data when there were enough primary studies reporting the desired data [8]. Subgroup analyses were used to assess possible moderating effects of dichotomous variables, and meta-regression for possible moderating effects of continuous variables. It should be noted that statistically significant differences found between subgroups and in meta-regressions are observational, not causal [8].

Subgroup analyses included preclinical vs. clinical students; study design (randomized control trial (RCT) vs. two-group non-randomized vs. single-group pre-post studies); U.S. vs. international students (due to age and cultural differences); elective vs. mandatory curricular delivery; interventions led by a licensed professional (e.g., psychologist, psychiatrist) vs. those led by other individual(s) or peers; in-person vs. online delivery; mindfulness practices vs. not; and didactic lectures vs. not. Subgroup analyses on racial, ethnic, age group, and socioeconomic differences were planned but not able to be completed due to insufficient data. Subgroup analyses were done with a random-effects model to account for expected heterogeneity [8]. The variance of true effects (T²) was calculated within each subgroup and then pooled to give an overall T² value for all subgroups, which is necessary to explain variance within subgroups given that between-group variance is explained by the subgroup membership [8]. CMA runs a pairwise test to determine if there is a statistically significant difference in the effect sizes in each subgroup [8]. We report the statistical significance or lack thereof, the magnitude of the difference, and the confidence interval for subgroup analyses [8].

Meta-regressions were completed on total minutes of resilience intervention, total number of sessions, average duration of the sessions in minutes, total duration of the interventions in weeks, and mean age. The random-effects model was again used to ensure the expected heterogeneity was accounted for. The Z value and p of Z were assessed to reveal whether or not the covariate of interest had a statistically significant effect on resilience or stress or not [8]. A Q value and p of Q were assessed for each meta-regression as a test of the model; the Q value and its significance tests the hypothesis that none of the covariates are related to effect size; if p is significant, it reveals that at least one covariate has an influence on effect size [8]. The goodness-of-fit test for each meta-regression model was evaluated using I²,T²,T, Q, and p of Q: T² and T divulge the estimated variance of true effects and the standard deviation of true effects, respectively; the I² statistics tells the percent of observed variance of the covariate which is likely due to true effects and not random sampling error; and finally, the Q and p of Q statistics show if the variance of true effects could be entirely due to sampling error [8].

Primary studies’ risk of bias

Cochrane’s Risk of Bias version 2 (RoB2) tool was used to evaluate the risk of bias in the RCTs [58]. Bias in RoB2 is considered across multiple domains and results in a final assessment of “low risk of bias,” “some concerns,” or “high risk of bias” [58]. The five domains which must be rated are biases arising from (1) the randomization process, (2) deviations from the intended intervention, (3) missing outcome data, (4) measurement of the outcomes, and (5) selection of the reported results [58]. Each domain has between three and seven questions which the researcher answers as ‘Yes,’ ‘Probably Yes,’ ‘Probably No,’ ‘No,’ or ‘No Information’ [58].

For non-randomized studies, the Risk of Bias in Non-randomized Studies—of Interventions (ROBINS-I) was used [59]. Final judgements are “Low risk of bias,” “Moderate risk of bias,” “Serious risk of bias,” or “Critical risk of bias” [59]. Domains of assessment include pre-intervention bias such as how participants are selected,at-intervention bias such as how the intervention is classified; and post-intervention bias including missing data or deviations from the intended protocol [59].

Publication bias

Publication bias was assessed to determine if it potentially influenced the conclusions of the meta-analysis. Multiple tests are used to improve accuracy [55, 56]. The Trim & Fill test for publication bias was used as it tests for the presence of bias and corrects for it by giving an adjusted mean by adding hypothetical studies to the “open spaces” in the funnel plot [35]. The funnel plot was used to visually observe the data [31]. Publication bias theoretically can be seen if there are “empty spaces,” or areas around the pyramid which are lacking studies, specifically in the lower left corner which are studies with smaller effect sizes, no effect, or a negative effect (the control group improved more than the intervention group), as studies with larger, more positive effects are more likely to be published and more likely to be found by researchers in their literature search [7].

Results

The search identified 2,631 records. After removal of duplicates, 2,438 citations were screened. See Fig. 1 for the PRIMSA flow diagram for study selection [45]. The 15 studies with resilience outcomes and nine studies with stress outcomes are shown in Table 1; three of the studies had both resilience and stress outcomes and so were included in both meta-analyses [18, 57, 60], giving a final total of 21 studies. The number of data sets analyzed was k = 23 as two studies had multiple cohorts [18, 46]. The total sample size for all studies was 1,956.

Fig. 1.

PRISMA flow diagram for study selection [45]

Table 1.

Study characteristics and outcomes

Study/Location	Design	Intervention	Sample Characteristics	Outcome Measure/Effect Size
[5] United States	One-arm pre-post	Shared meal and resiliency skills course, in-person Facilitation: NR Duration: NR Frequency: NR Total sessions: 8	n = 32 Age: 77.5% aged 20–24 % Female: 82.5 Non-binary: 2.5% Race: White 52.5%, Asian 27.5%, multiracial 10% Ethnicity: 10% Hispanic	Resilience: BRS d = 0.46
[6] United States	Two-arm non-random	Resilience curriculum during clerkships, in-person: taught managing expectations, coping with interpersonal interactions, how to find meaning in daily work, dealing with the unexpected; reflection, mindfulness, didactics, and peer discussion Facilitation: Faculty Duration: 1 h Frequency: NR Total sessions: 4	n = 144 Mean age: NR % Female: NR	Resilience: CD-RISC d = 0.43 (Note that there was no data for a second cohort which had “no significant change” in resilience outcomes)
[9] United Kingdom	Two-arm non-random	SeRenE – Stoic rEflection for ResiliiENce and Empathy, online: acceptance of things out of your control, mindfulness, reflection, empathy Facilitation: online, developed by two psychotherapists, a psychologist, and three faculty Duration: 15–20 min Frequency: Daily Total sessions: 12	n = 24 Mean age: 21.0 % Female: 62.5 Race: 62.5% White, 12.5% Black, 12.5% Asian; 79.2% heterosexual	Resilience: BRS d = 0.50
[11] Australia	One-arm pre-post	Supportive, structured development of communication skills with Standardized Patients (SPs), mindfulness exercise at end of each session, reflective writing; in-person Facilitation: psychologists, counselors, nurses, pharmacists/mixed Duration: 3 h Frequency: NR Total sessions: 21	n = 55 Mean age: 22.4 % Female: 52.7	Resilience: MIQ d = 0.50
[12] United States	One-arm pre-post	LAVENDER (Leveraging Affect and Valuing Empathy for Nurturing Doctors’ Emotional Resilience) based on positive psychology; emotional awareness, gratitude, self-compassion, etc.; 3rd year students mandatory curriculum, in-person small groups Facilitation: researchers Duration: 35–40 min Frequency: Monthly Total sessions: 4	n = 157 Mean age: 25.88 (2.25) % Female: 49	Stress: PSS-4 d = −0.04
[18] United States	One-arm pre-post	Mindfulness-based stress management mandatory curriculum focused on intentional awareness, attention, attitude, non-judgement; 1 st year students in-person Facilitation: mixed; social worker, psychologist, physicians Duration: 30 min – 4 h Frequency: NR; year-long Total sessions: 8	n = Cohort 1 = 44 (entering 2014) and Cohort 2 = 22 (2015) Age: 77.1%/79.5% under age 25 % Female: 58.3/59.1	Resilience: CD-RISC Stress: PSS d = −0.22
[19] United States	One-arm pre-post	Multifaceted wellness curriculum including seminars, programs, and activities, many proposed by students, including health coaching, biofeedback training, mindfulness elective, tai chi, soccer league, exercise equipment check-out, cooking night, etc.; 1 st and 2nd year students in-person Facilitation: mixed Duration: variable Frequency: variable; offerings throughout year Total sessions: variable	n = 29 all timepoints (pre, 6-mo, 12-mo) Age: 65.5% between 24–29 % Female: 48.3	Stress: PSS d = 0.84
[21] Sweden	One-arm pre-post	Training for Awareness, Resilience, and Action (TARA), includes synchronized slow movements, breath-work, psychoeducation, compassion, preclinical years Facilitation: mixed, psychologist/psychiatrist and mindfulness expert Duration: 90 min Frequency: Weekly Total sessions: 12	n = 23 Mean age: 25.38 % Female: 78.3	Stress: PSS d = 0.15
[25] United States	One-arm pre-post	Emotional Intelligence -Resilience Elective, positive thinking, reframing, optimism, mindfulness, self-care, self-compassion, written reflections; in-person; 2nd year students Facilitation: NR Duration: 2 h Frequency: NR Total sessions: 6	n = 30 Mean age: NR % Female: NR	Stress: Bar-on EQ-i 2.0, Stress Management Composite score d = 0.53
[26] United Kingdom	One-arm pre-post	Reboot (Recovery-Boosting Training) coaching program based on cognitive-behavioral principles, live online small groups of clinical students Facilitation: clinical psychologist Duration: 2 h Frequency: Weekly Total sessions: 3 (2 workshops, 1 coaching call) over 4 weeks	n = 115 Mean age: 27.0 % Female: 80.9 Race: 49.6% White, 23.5% Asian, 11.3% Black	Resilience: BRS d = 0.50
[30] Indonesia	RCT	“Transition and Adaptation towards Resilience” with content on university environment transitioning, stress management, symptoms of mental health disorders, and help-seeking; website/independent learning modules, Zoom sessions online, small groups Facilitation: Faculty Duration: 60–90 min, 4 weeks Frequency: twice per week Total sessions: 8 online discussion sessions	n = 105 (intervention 52, control 53) Age: 59.6%/81.1% age 18 % Female: 53.9/67.9	Resilience: CD-RISC d = 0.29
[42] New Zealand	RCT	Mindfulness meditation in small groups in-person with home practice CD; 2nd and 3rd year students Facilitation: Elected, trained peers Duration: NR Frequency: Weekly Total sessions: 19 + 2 social events	n = 232 (intervention 111, control 121) Mean age: 20.9 % Female: 53	Resilience: RS15 d = 0.08
[46] China	RCT	Penn Resilience Program (PRP), focuses on cognitive behavioral therapy, emotional regulation, connecting thoughts and emotions; divided pre-intervention into high- and low-resilience groups, in-person Facilitation: “group leaders” Duration: 90 min – 2 h Frequency: Weekly Total sessions: 10	n = 60 (30 each intervention and control) Mean age: 19.8 (0.77) % Female: 26.7/33.3	Resilience: CD-RISC d = 0.81
[47] United States	Two-arm non-random	Use of creative arts including visual arts, film, and literature related to a mental state such as depression, trauma, dreams, etc. with guided discussions, in-person Facilitation: “instructor” Duration: 90 min Frequency: Weekly Total sessions: 6	n = 27 (5 intervention, 22 control) Mean age: 20.4 % Female: NR	Resilience: CD-RISC d = 1.43
[48] India	Two-arm non-random	Stress intervention lectures on content like defining stress, causes and effects of stress, sleep hygiene, stress diary, stretching, yoga and breath-work; all years of students Facilitation: NR Duration: 20–90 min over 3 months Frequency: Up to three times per week including short sessions to reinforce content in-between Total sessions: 60	n = 526 (intervention 229, control 240) Mean age: 19.98 % Female: 62.2	Stress: PSS (resilience was tested only pre-intervention so is not included) d = 1.42
[57] United States	Two-arm non-random	Psychoeducation process groups including discussion-based emotional awareness, emotional reactivity, discomfort tolerance, boundaries, empathy, psychodynamics Facilitation: Psychiatrist Duration: 60 min each, one academic year Frequency: Weekly Total sessions: 25	n = 16 (8 each intervention and control) Mean age: 27.0 % Female: 68.8	Resilience: CD-RISC-25 Stress: PSS-14 d = 0.65
[60] United Arab Emirates	One-arm pre-post	“Systematic Assessment for Resilience (SAR)” training session on how to use the framework on a daily basis, online; 4th year students Facilitation: Peers as trained coordinators Duration: NR Frequency: NA Total sessions: 1	n = 78 Mean age: 21.9 % Female: 71.8	Resilience: MeRS (global score) Stress: ARS part of the MSSQ d = 0.45
[63] United States	One-arm pre-post	Hyper-realistic surgical simulation training for military students, in-person Facilitation: NR Duration: NR Frequency: Daily Total sessions: 6 days	n = 68 Mean age: 27.04 % Female: 25.0	Resilience: HRG d = 0.50
[64] United States	One-arm pre-post	“Promoting Resilience in Medicine (PRIMe)”, mindfulness, biofeedback, art, journaling, school-life balance, in-person; 2nd year students Facilitation: NR Duration: 2 h Frequency: Weekly Total sessions: 11	n = 24 Mean age: NR % Female: 70.8	Stress: PSS-10 d = −0.03
[65] China	RCT	“Mindfulness-Based Stress Reduction (MBSR)” including present moment attention, nonjudgement, acceptance, yoga; military students with low resilience scores in-person; all years Facilitation: physicians Duration: 2 h Frequency: Weekly Total sessions: 8	n = 88 (52 intervention, 36 control) Mean age: 19.98/19.61 % Female: 36.5/44.4	Resilience: CD-RISC-10 d = 0.44
[66] China	Two-arm non-random	“Mindfulness-Based Cognitive Therapy (MBCT)” including awareness, focusing the wandering mind, acceptance, self-care, etc., and homework; in-person with MP3 lectures Facilitation: psychotherapist & TAs Duration: 3 h Frequency: Weekly Total sessions: 8	n = 57 (30 intervention, 27 control) Mean age: 20.37/20.74 % Female: 20/18.5	Resilience: CD-RISC d = 0.38

BRS Brief Resiliency Scale, CD-RISC Connor & Davidson Resilience Scale, EQ-I 20.0 Bar-On Emotional Quotient Inventory (Stress Management Composite), HRG Hardiness Resilience Gauge, MIQ MaRIS Impact Questionnaire, MeRS Medical Professionals Resilience Scale, MSSQ Medical Students Stressor Questionnaire, NR not reported, RS15 “25-item resilience questionnaire”

Study and participant characteristics

Studies were published between 2014 and 2025 (see Table 1). Study designs included 11 one-arm pre/post studies, six two-arm non-randomized studies, and four RCTs. Sample sizes ranged from 16 to 526. Three studies reported race and/or ethnicity; these details along with study location, description of intervention, and further participant demographics (age, % female) can be seen in Table 1.

Risk of bias

There were significant concerns of bias in the four RCTs (Fig. 2) and 17 non-randomized one- or two-arm studies (Fig. 3a/b). Many studies failed to report protocol adherence and/or had high drop-out rates. All studies had self-reported outcomes. No studies commented on controlling confounding factors.

Fig. 2.

Risk of bias in the RCTs using RoB2 [58]. Key: D1: randomization process; D2: deviations from intended protocol; D3: missing outcome data; D4: measurement of outcome; D5: selection of the reported result.

Fig. 3.

a Risk of bias in the non-randomized studies using ROBINS-I [59]. b Summary plot of ROBINS-I [59]

Interventional characteristics

Characteristics of the different interventions were coded for subgroup analysis, and though many did not meet a minimum number needed for statistical analysis, reviewing the many modalities may be useful for future study designs. Many of the studies used mindfulness practices [6, 9, 11, 18, 19, 25, 42, 57, 64–66], including the standardized mindfulness curricula Mindfulness-Based Stress Reduction (MBSR, [65]) and Mindfulness-Based Cognitive Therapy (MBCT; Yin et. al., 2024[66]), while others were more casual or borrowed from several different traditions. Yoga is part of MBSR and was also included in another study [48], slow movements for nervous system regulation were used [21], as was Tai Chi [19]. Five studies reported using didactic education [6, 21, 30, 47, 48] or psychoeducation [21, 57]. Five studies asked for reflections or journaling [6, 9, 11, 25, 64]. Two studies employed coaching [19, 26],three taught breathwork [21, 48, 57],three explicitly discussed self-care [25, 30, 66],and two used biofeedback [19, 64]. One study focused on communication skills with Standardized Patients (SPs [11],), and one had faculty and students sharing a meal [5]. Many included concepts such as empathy, acceptance, self-compassion, reframing, emotional regulation, and other stress management techniques [9, 12, 21, 25, 30, 46, 48, 57, 65, 66]

Outcome measures

Outcome measures of resilience and stress were included in this review. Seven different resilience outcome measures were used in the studies which measured resilience, most commonly the Connor-Davidson Resilience Scale (CD-RISC; [14]), used in eight studies, and the Brief Resilience Scale (BRS), used in three studies (see Table 1). In the studies that measured stress, the Perceived Stress Scale (PSS) of various (or unstated) versions was used in six studies. All studies used self-reported outcomes. In addition, Kaligis et al. measured salivary cortisol, which showed a significant decrease in the intervention group (MD = −3.96, p < 0.001) but no significant difference between intervention and control groups [30].

Summary effect sizes

Two meta-analyses were done, one on the effect of resilience interventions on resilience, and one on the effect of resilience intervention on stress. Three studies had both resilience and stress outcomes and are included in both meta-analyses [18, 57, 60]. Study designs were included as a sub-group analysis to determine differences in effect size between one-arm pre/post studies and RCTs, for example.

For the effect of resilience interventions on the outcome of resilience, the summary effect size was d = 0.39, k = 17, n = 15, 95% PI [−0.10–0.87], p < 0.001, indicating a small-to-moderate positive effect on resilience interventions increasing resilience among medical students compared to controls or pre-intervention (Fig. 4). Study effect sizes ranged from −0.22 to 1.43. There was one study with a negative effect size. There was significant heterogeneity of effects, as expected (Q = 37.8, I² = 62.9, T = 0.2, p = 0.001).

Fig. 4.

Forest plot of resilience outcomes in medical students. KEY: Cohen’s d = Std diff in

For the effect of resilience interventions on the outcome of stress, the summary effect size was d = 0.53, k = 10, n = 9, 95% PI [−0.96–2.02], p = 0.01, indicating a moderate positive effect on resilience interventions decreasing stress among medical students compared to controls or pre-intervention (Fig. 5). Individual effect sizes ranged from −0.04 to 1.42, with two studies which had negative effect sizes. There was again significant heterogeneity of effects, as expected (Q = 137.9, I² = 94.2, T = 0.6, p < 0.001).

Fig. 5.

Forest plot of stress outcomes in medical students. KEY: Cohen’s d = Std diff in means: 0.2 small effect size, 0.5 moderate effect size, 0.8 large effect size

Publication bias was assessed visually with funnel plots [35] and statistically with Egger’s regression intercept [20] and Duval and Tweedie’s trim and fill test [17]. Multiple methods are recommended to provide more accuracy in estimating publication bias [55, 56]. Funnel plots can be seen in Figs. 6 (resilience outcomes) and 7 (stress outcomes). Both plots exhibit a lack of study representation in the lower left, indicating that both meta-analyses are likely to have publication bias due to fewer studies with neutral or negative results. Egger’s regression intercept for resilience outcomes was 1.67 (1.14), 95% CI [−0.08–4.14], p = 0.08, and for stress outcomes 1.06 (3.33), 95% CI [−6.80–8.93], p = 0.38. An intercept of 0 indicates no publication bias, so these tests again signify publication bias in both meta-analyses; however, neither test was significant, so publication bias may not be severe. In the Duval and Tweedie’s trim and fill test, the point estimates for resilience outcomes were slightly different (observed value = 0.36, adjusted value = 0.30), and for stress outcomes point estimates were the same (0.49), suggesting there is slightly more publication bias in the meta-analysis of resilience outcomes. Due to the likely presence of publication bias in both meta-analyses, effect sizes may be inflated. A one-study-removed sensitivity analysis was performed on each meta-analysis. Neither produced a change in results, indicating no study had undue influence on the summary effect size.

Fig. 6.

Funnel plot of studies with a resilience outcome

Fig. 7.

Funnel plot of studies with a stress outcome

Moderator analyses

Both sub-group analyses and meta-regression were completed on each meta-analysis to evaluate heterogeneity and determine if certain characteristics were more effective at improving resilience or reducing stress among medical students (see Table 2; see Supplement for forest plots of subgroup analyses and scatterplots of meta-regressions). One finding was significant in the sub-group analyses for resilience outcomes. Students in the clinical years were found to have significantly higher resilience outcomes (d = 0.46, p < 0.001) compared to preclinical students (d = 0.32, p = 0.02, Q = 9.22, p of Q = 0.03) after resilience interventions. Insignificant sub-group analyses in the meta-analysis of resilience outcomes included studies completed in the United States (d = 0.40, p = 0.03) vs. those delivered in other countries (d = 0.41, p < 0.001, Q = 0.0, p of Q = 1.0); study design with single-arm pre-post (d = 0.35, p = 0.004), two-group non-randomized (d = 0.49, p < 0.001) and RCTs (d = 0.35, p = 0.02, Q = 1.25, p of Q = 0.53); in-person (d = 0.39, p = 0.001) vs. Zoom or online (d = 0.44, p < 0.001, Q = 0.15, p of Q = 0.70); facilitation of sessions by a mental health professional (d = 0.46, p < 0.001) vs. peers, faculty, or other (d = 0.38, p < 0.001, Q = 0.37, p of Q = 0.83); inclusion of mindfulness (d = 0.30, p = 0.01) vs. not (d = 0.48, p < 0.001, Q = 1.96, p of Q = 0.16); and didactic lecture (d = 0.48, p = 0.01) vs. none (d = 0.37, p < 0.101, Q = 0.26, p of Q = 0.61).

Table 2.

Moderator analyses statistics

Subgroup Analysis
Resilience	Cohen’sd	k	p value	95% CI	Qbetween	pvalue forQ
International	0.41***	10	< 0.001	0.29–0.52	0.00	1.00
vs. U.S	0.40*	7	0.03	0.04–0.77
RCT	0.35**	7	0.004	0.11–0.58	1.25	0.53
vs. 1-arm	0.49***	5	< 0.001	0.31–0.67
vs. 2-arm	0.35*	5	0.02	0.06–0.63
Preclinical	0.32*	8	0.02	0.06–0.58	9.22*	0.03
vs. clinical	0.46***	4	< 0.001	0.33–0.59
Online	0.44***	5	< 0.001	0.32–0.57	0.15	0.70
vs. in-person	0.39**	12	0.001	0.17–0.62
Mental health professional	0.46***	3	< 0.001	0.25–0.66	0.37	0.83
vs. other	0.38***	13	< 0.001	0.19–0.56
Mindfulness	0.30*	9	0.01	0.07–0.52	1.96	0.16
vs. not	0.48***	8	< 0.001	0.36–0.59
Stress
International	0.65	3	0.13	−0.20–1.50	0.20	0.66
vs. U.S	0.44*	7	0.02	0.06–0.82

Meta-regressions
Resilience	Coefficient	k	pvalue	95% CI
Duration	−0.00	9	0.85	−0.00–0.00
# sessions	0.00	9	0.87	−0.02–0.02
Total mins	−0.00	9	0.97	−0.00–0.00
Mean age	0.00	9	0.96	−0.05–0.05
Stress
Duration	−0.00	5	0.82	−0.02–0.01
# sessions	0.02	5	0.66	−0.12–0.17
Total mins	0.00	5	0.79	−0.00–0.00
Mean age	−0.10	5	0.37	−0.42–0.21

*p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001

In studies with stress outcomes, one sub-group analysis was able to be done which was not significant: U.S. studies (d = 0.44, p = 0.02) were not significantly different from studies completed elsewhere (d = 0.65, p = 0.13, Q = 0.20, p of Q = 0.66). Although the studies outside of the U.S. had a higher effect size, there were only three of them with stress outcomes and the effect size was not significant, so accuracy is not ensured. Study design, preclinical vs. clinical, in-person vs. virtual, and facilitator type were all unable to be assessed due to low sub-group numbers.

Meta-regressions were completed on number of sessions, duration of sessions, total number of minutes (sessions x duration), and mean age. No meta-regression was significant in either meta-analysis.

Discussion

This study presents a systematic review and meta-analysis examining the effect of resilience interventions on medical students. Significant findings include a moderate effect of resilience interventions on reducing stress (d = 0.53) and a small-to-moderate effect on improving resilience (d = 0.39). Resilience interventions showed a significantly stronger effect on resilience among clinical students as compared to preclinical students. There was significant heterogeneity across studies, with varying intervention activities, durations, facilitators, and delivery methods. There was also some evidence of publication bias, indicating results should be interpreted with caution.

Results suggest that resilience interventions may be more effective at reducing stress than improving resilience across all students, although both stress and resilience had significant improvements. As we included any primary study which targeted resilience in medical students, there were many interventions which also targeted other outcomes, including stress, even if they didn’t specifically measure it. It's possible that stress is more malleable or quicker to change among medical students compared to resilience, which requires skill-building [23]. More research would be beneficial to investigate the relationship between resilience and stress or if there are specific aspects of resilience training that most effect stress management.

Results also suggest clinical students may respond better to resilience training than preclinical students, gaining more resilience. Research supports that resilience declines over the course of medical training [27, 39, 57] but appears to rebound at some point after transitioning to independent practice (i.e., post-residency [62]). Although mean age was not significant in the meta-regressions, a recent study has suggested a predictive relationship between age and resilience in health profession students,that study found a 0.02 score increase on the Brief Resilience Scale for every 1-year increase in age [40]. As clinical years are later in training, students tend to be several years older,though more investigation is needed, this may assist in explaining the difference between clinical and preclinical resilience scores post-intervention.

Physician Chris Watling wrote, “Medicine’s historical tendency to build stress-resistant individuals rather than to build wellness-supporting environments may reflect the values of the profession” [61]. ‘Resilience’ as it is used in the Western world refers to personal resilience, which puts the onus on the individual to learn resilience techniques and skills, as well as placing the blame on the individual if they are not adept at managing stress [16], [56]. The focus on the individual ignores the very real impact of the healthcare environment, likely missing the full picture of patient, organizational, systemic, and cultural factors that contribute to stress.

Medical student suggestions for resilience typically include changes to systems as opposed to additional resources, such as modifying curriculum to be more consistent over time, prioritizing need-to-know clinical information, having flexible time off, and having a pass/fail grading system [32]. Students recognize the need for personal resilience but often have very little time for health-promoting activities including exercise, social connection, or even the recommended amount of sleep [1, 2]. While many interventions have been trialed, the quality of evidence remains low and many trials are still in the feasibility stage [56]. However, there is some evidence that training on perspective, empathy, self-compassion, communication skills, letting go of negative emotions, and managing setbacks is beneficial [5, 6, 11]. Providing free mental health services may also be helpful (Edmonds et al., 2020) [21, 57]. Mindfulness interventions, while common, are not always measured alongside resilience, but do generally provide evidence of improved resilience [6, 9, 11, 19, 25, 57, 65, 66].

There are several important limitations in this meta-analysis that should be noted. Firstly, there is a high risk of bias in the majority of the primary studies, giving less confidence in the results of the meta-analysis. There is also significant heterogeneity in both meta-analyses, particularly among stress outcomes, and while this was expected, widely variable interventions make it difficult to distinguish which parts are effective and which may not be. High heterogeneity not explained by moderator analyses also suggests it may be random, and further research is needed to determine the effect of different durations and number of sessions, for example. A relatively small number of primary studies were available for meta-analysis, suggesting a repeat of this study in the future may be useful, particularly as publication bias was also present at least to some degree. Included studies were in English, which may have caused us to miss studies which should have been included. Moderator analyses also suggest areas of further research, such as clinical vs. preclinical students, session duration, total number of sessions, and facilitator type, among others. Future meta-analyses may have enough studies to follow recommendations to separate study design types; with the low number of eligible studies in the current study, we elected to instead do subgroup analyses to distinguish effect size differences between study design types.

Conclusions

Resilience interventions were found to have a moderate effect on stress reduction and a small-to-moderate effect on improving resilience among medical students. Students in their clinical years may be primed to gain more resilience skills than those in preclinical years. More research is indicated to assess resilience beyond the individual and to compare organizational vs. resources-based interventions, follow medical students into residency and professional practice to assess long-term resilience strategies, and evaluate methods of change in the wider healthcare system which could support our health professionals.

Authors’ contributions

ELS contributed to conceptualization, data curation, formal analysis, methodology, project administration, software, supervision, validation, and writing the original draft. BM, EG, AL, PK, and ZCM contributed to data curation, investigation, and reviewing and editing. All authors read and approved the final manuscript.

Funding

No funding was received for the preparation of this manuscript.

Data availability

All data generated or analyzed during the current study are included in this published article and the primary articles used and cited in the manuscript.

Declarations

Ethics approval and consent to participate

As a systematic review and meta-analysis, no human subjects were involved in the completion of this manuscript.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.