Factors underpinning international student mental well-being: a UK and China survey Authors: Jianyiwen Zhao, Rob Saunders, Phoebe Barnett, Siying Li, Yuqi Niu, Stephen Pilling, Peter Fonagy Published: 2026 ABSTRACT AbstractBackgroundInternational students experience a high mental health burden and face barriers to accessing and sustaining support, yet domain-level stress frameworks provide limited guidance on which specific stressors should be prioritised.ObjectiveTo prioritise item-level stressors associated with perceived stress and mental well-being among international students using machine-learning models, and to evaluate cross-site transportability via external testing between Chinese students in the UK and international students in China.MethodsWe conducted a cross-sectional online survey of 481 students across two sites, Chinese students in the UK and international students in China. Item-level stressors, perceived stress (PSS 10) and mental well-being (Warwick–Edinburgh Mental Wellbeing Scale; WEMWBS) were assessed. Elastic net and random forest models were evaluated using nested cross-validation, and cross-site external testing was conducted by training in one site and testing in the other.FindingsAcross the six models, random forest explained approximately 22%–35% of the variance in perceived stress and 25%–27% of the variance in mental well-being. Concerns about mental well-being, sleep difficulties, physical health and family or self-imposed academic expectations consistently ranked highly across sites. Social isolation and restricted social life were strongly associated with lower mental well-being. Context-specific stressors also emerged: in the UK setting, greater time-management pressure, low confidence in interacting with others and loneliness were more strongly associated with poorer outcomes, whereas in the Chinese setting, difficulties balancing study and personal life and not knowing where to seek help played a larger role. Cross-site validation indicated that a substantial portion of the item-level stressor signal generalised across settings.ConclusionsItem-level stressor profiles provide informative, partly generalisable evidence on stressors linked to international students’ mental health. The identified high-impact stressors offer concrete, context-sensitive targets for prevention and intervention in university mental health services.Clinical implicationsThese findings can inform how university mental health services prioritise support and prevention by focusing on core stressors while adapting provision to locally salient challenges. WHAT IS ALREADY KNOWN ON THIS TOPIC International students have been reported to experience a higher mental health burden in many studies and face substantial barriers to accessing and sustaining mental health support. However, most research relies on total scores or broad stress domains, making it difficult to identify which specific stressor items should be prioritised and limiting the development of more precise support and prevention strategies. International students have been reported to experience a higher mental health burden in many studies and face substantial barriers to accessing and sustaining mental health support. However, most research relies on total scores or broad stress domains, making it difficult to identify which specific stressor items should be prioritised and limiting the development of more precise support and prevention strategies. WHAT THIS STUDY ADDS This study provides a clearer item-level prioritisation of stressors relevant to international students’ mental health, moving beyond broad domains to specific stressor items. Using cross-context external testing across two international study settings in the UK and China, we identified a core set of consistently high impact stressors and also distinguished more context-specific stressors, providing evidence to inform both shared priorities and local adaptation. This study provides a clearer item-level prioritisation of stressors relevant to international students’ mental health, moving beyond broad domains to specific stressor items. Using cross-context external testing across two international study settings in the UK and China, we identified a core set of consistently high impact stressors and also distinguished more context-specific stressors, providing evidence to inform both shared priorities and local adaptation. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY These item-level findings offer a more practical basis for universities and related services to prioritise support and prevention under resource constraints. They suggest combining a focus on core stressors that are shared across contexts with targeted adaptation to locally salient academic, social and service navigation challenges, and they provide clear priorities for developing and evaluating more precise support strategies. These item-level findings offer a more practical basis for universities and related services to prioritise support and prevention under resource constraints. They suggest combining a focus on core stressors that are shared across contexts with targeted adaptation to locally salient academic, social and service navigation challenges, and they provide clear priorities for developing and evaluating more precise support strategies. Introduction Poor mental health and reduced well-being are common among university students, with more than one third reporting clinically significant symptoms of anxiety, depression, or sleep disturbance.1 International students may be particularly vulnerable because of the combined demands of academic adjustment, social adaptation, and separation from established support networks.2 As global participation in higher education expands, in many universities international student numbers match or exceed the number of domestic students.3 International students contribute substantially to the academic and financial standing of host institutions, but consistently report high levels of depressive symptoms and lower mental well-being than national peers.4 Despite elevated need, international students are less likely to seek formal psychological support and more likely to discontinue treatment when they do engage, with stigma, language barriers and limited culturally responsive care commonly cited barriers.5 Although universities have expanded provision, gaps in access, cultural fit and perceived usefulness persist.6 These shortfalls underscore the need to culturally adapt interventions specifically for international students. Identifying modifiable targets is fundamental to the effective design of mental health prevention and intervention efforts. However, existing work in international students remains limited. Most studies focus on overall symptom burden or global distress indices, supplemented by broad stress domains (eg, academic, interpersonal, financial).7 While useful for identifying vulnerability, this approach offers limited guidance on what to target because it does not specify which specific stressors within a domain drive risk. For example, ‘academic stress’ may reflect time-management difficulties, family expectations or language barriers, each implying different support targets such as skills-building, targeted psychoeducation or structural support. Much research still relies on total scores or a small number of subscales, even as prevention and treatment increasingly require more fine-grained targeting to specific groups and contexts.8 Domain-level approaches provide only coarse direction for intervention design and limited clarity about actionable leverage points. Few studies have compared the relative importance of individual stressors to inform mental health interventions, and fewer still have tested whether item-level patterns generalise across institutional and cultural contexts. Recent computational advances, the use of machine-learning methods, allow more precise modelling of multiple stressors to inform mental health and well-being, offering a promising framework for developing and refining intervention strategies.9 The aim of this study was to identify which item-level stressors were most strongly associated with perceived stress and mental well-being in international students and to examine whether these associations were consistent across contrasting study environments or primarily reflected context-specific challenges. Perceived stress and mental well-being are related but distinct dimensions of mental health.10 Examining both outcomes may help identify stressors associated primarily with distress, mental well-being or both. To this end, we focused on Chinese students enrolled at a UK university and international students enrolled at a Chinese university as two contrasting study contexts rather than as homogeneous groups. Both groups must adapt to a new educational system, social environment and language but may face different structural and contextual challenges. We fitted models separately within each site and in the pooled sample with a site indicator, followed by bidirectional cross-site validation to examine whether item-level stressor patterns showed predictive transportability across settings. This approach was intended to identify potentially modifiable stress-related experiences that may represent candidate targets for future intervention development. Given the cross-sectional design, findings are interpreted as associations and relative item importance rather than causal effects. Methods Participants and procedures Between April and November 2024, we conducted an anonymous online Qualtrics survey among Chinese students enrolled at a UK university and international students enrolled at a Chinese university. Participants at both sites were recruited through institutional email lists and social media, primarily WeChat. Recruitment advertisements included a link directing interested students to the online survey. As recruitment included participant forwarding of the survey link, the total number of individuals exposed to the invitation could not be determined; consequently, precise response rates were not available. At both sites, eligible participants were aged 18 years or older, currently enrolled at the respective university and did not report a current diagnosis of a psychotic disorder. At the UK site, participants were required to hold Chinese nationality; at the China site, participants were required to hold non-Chinese nationality. Participants at the UK site received a £15 honorarium and participants at the China site received 30 RMB for participation, consistent with local compensation standards and approved by the relevant ethics committees. The study formed part of a broader project on international students’ mental health and adjustment. Sample size reflected the number of eligible participants recruited during the study period. Because analyses focused on predictive modelling rather than hypothesis testing, a conventional power calculation was not performed. The final sample size was comparable to similar machine-learning studies in mental health.11 12 Measures Stressors list A 53-item Stressors List (SL) was developed through literature review, review of existing measures, expert consultation and pilot testing with students at both sites. The SL was designed to capture a broad range of stress-related experiences relevant to international students, including both contextual stressors (eg, financial pressures, social integration difficulties) and more proximal subjective experiences (eg, sleep difficulties and concerns about well-being). The SL was used as an exploratory item pool rather than a psychometrically validated scale. Items were rated on a 7-point scale (1=not stressful, 7=extremely stressful). In this study, stressor items refer to the questionnaire items included in the SL, whereas predictors refer to their statistical role within the machine-learning models, where discussed as potential intervention targets, these interpretations are intended as hypothesis-generating rather than causal. Perceived stress Perceived stress was assessed using the 10-item Perceived Stress Scale (PSS-10),13 widely used in university student populations.14 Items are rated from 0 (never) to 4 (very often) and summed (0–40), with higher scores indicating greater perceived stress. Mental well-being Mental well-being was measured with the 14-item Warwick–Edinburgh Mental Wellbeing Scale (WEMWBS).15 Items are rated from 1 (none of the time) to 5 (all of the time) and summed (14–70), with higher scores indicating better mental well-being. WEMWBS and PSS showed good internal consistency in this study (α≥0.82). Data preparation Participants were included if they completed at least 50% of SL items and provided data on either the PSS or WEMWBS. We screened for careless responding using Mahalanobis distance and long-string indices and removed flagged cases.16 Categorical variables were recoded into binary indicators where required, and near-zero variance features were removed. Missing predictor values were imputed using the median for numeric variables and the mode for binary variables, with additional missingness indicators created for all predictors. Missing item-level data were negligible (<0.1% of all responses; see online supplemental material). Numeric features, including ordinal Likert-scale responses, were standardised for EN models.17 We treated SL items as approximately continuous to support regularised regression and comparability of importance scores. To prevent data leakage (train–test contamination), all preprocessing steps were fit exclusively on the outer training split and applied unchanged to that fold’s held-out outer test data. Modelling approach Given that item-level stressors are numerous and likely intercorrelated, stable estimation of relative predictor importance requires methods that can accommodate multicollinearity. We, therefore, compared two complementary machine-learning algorithms, Elastic net (EN) and random forest (RF), to estimate the relative importance of individual stressor items for mental health outcomes. Both methods are well suited to settings with many correlated predictors; EN is a regularised linear model with embedded feature selection, whereas RF can capture non-linear and interaction effects.18 19 The goal was to identify potentially actionable stressor patterns rather than optimise prediction alone. Analyses were conducted in R (V.4.4.1). Full technical details of both algorithms are provided in the Supplement. EN is a penalised linear regression that combines L1 (lasso) and L2 (ridge) penalties to shrink coefficients and, where appropriate, set some to zero, enabling variable selection under multicollinearity.19 20 Model complexity was controlled by tuning the mixing parameter (alpha) and the regularisation strength (lambda). RF is an ensemble of decision trees built on bootstrap samples with random feature selection at each split, accommodating non-linearities and interactions and performing well in high-dimensional settings.18 Validation and performance metrics Given overfitting risk in small-sample machine-learning settings, we used nested cross-validation to obtain more reliable estimates of generalisation performance and reduce selection bias.21 The outer five-folds were used for performance evaluation, whereas the inner three-folds were used solely for hyperparameter tuning and model selection.22 At each outer iteration, all preprocessing steps were fit only on the outer training split and applied unchanged to the corresponding held-out outer test split; tuning was likewise confined to the outer training split using three-fold inner cross-validation with grid search. The tuned model was then refit on the full outer training data and evaluated on the held-out outer test fold. This procedure was repeated across all five outer folds, and performance was summarised as the median across outer test folds. In pooled analyses, outer folds were stratified by site to preserve the UK-to-China ratio. To provide interpretable benchmarks and quantify the incremental value of item-level stressors beyond demographics, we fit two ordinary least-squares baselines: Baseline A included demographic variables only, and baseline B additionally included the overall stressor score (SLtotal). EN was tuned in inner cross-validation by searching alpha over five values (0, 0.25, 0.5, 0.75 and 1) and selecting lambda using the one-standard-error rule23; demographic covariates were left unpenalised. RF used a standard implementation within the same nested framework, fixing the number of trees at 1000 for stability and tuning the number of candidate predictors at each split (mtry) via cross-validation. Predictive performance was summarised using the median outer-fold R² as the primary metric, with RMSE (Root Mean Squared Error) as a secondary measure. Under cross-validation or external testing, R² can be negative when test error exceeds that of an intercept-only model using the training mean.24 Variable importance For EN, importance was quantified using standardised coefficients: within each outer fold, we took the absolute value of each stressor-item coefficient, rescaled values to 0–100 (maximum=100) and averaged scores across outer folds.11 For RF, we used permutation importance (ranger implementation), defined as the increase in prediction error after permuting a single predictor while holding the trained model and all other predictors fixed.12 25 Fold-specific permutation importances were similarly rescaled to 0–100 and averaged across outer folds to support direct comparison with EN. Because correlated predictors can substitute for one another and attenuate permutation importance,26 we emphasised relative ranking rather than absolute magnitudes and compared RF and EN rankings as a robustness check. External validation To assess generalisability and cross-site transportability, we conducted bidirectional external validation: models were trained on the UK site (with three-fold inner tuning and final fitting) and evaluated blind on the China site, and vice versa. External validation used only the 52 stressor items common to both sites together with demographic covariates (excluding the site indicator) and mirrored the development-stage preprocessing and parameter selection. All imputation and, for EN, standardisation parameters were estimated in the training site and applied unchanged to the test site. Results In the UK site, 253 of 323 respondents met inclusion criteria and were included in the analytic sample. In the China site, 228 of 260 respondents were included. The final analytic sample comprised 481 participants. Among Chinese students in the UK, mean age was 23.3 years (SD=3.48); 78 (30.83%) were male and 173 (68.38%) were female. 70 (27.67%), 120 (47.43%) and 63 (24.90%) were enrolled in undergraduate, master’s and doctoral programmes, respectively. Among international students in China, mean age was 25.1 years (SD=5.36); 97 (42.4%) were male and 129 (56.6%) were female. A total of 127 (55.70%), 39 (17.1%) and 57 (25.0%) were undergraduates, Master’s students and doctoral students, respectively. Among those reporting region of origin, most were from Asia (n=180; 78.94%), followed by Europe (n=19; 8.33%), Africa (n=14; 6.14%), North America (n=7; 3.07%), South America (n=7; 3.07%) and Oceania (n=1; 0.44%). Full demographic characteristics are presented in eTable 1. Across outcomes, models incorporating item-level stressors and demographic covariates performed well on held-out outer test folds. Performance is summarised as the median R² and RMSE across five outer test folds. Median R² ranged from 0.175 to 0.356 across EN and RF models (table 1). Performance was highest for models predicting perceived stress in the China site (R²=0.356 for EN; R²=0.347 for RF). Relative to baseline B, which captures overall stress burden, improved performance indicates that item-level stressor patterns contributed predictive information beyond total stress level (table 1). Table 1Predictive performance of baseline and machine-learning modelsDatasetModelMedian (95% CI)R² with CIRMSE with CIPooled—PSSBaseline A−0.004 (−0.021 to 0.039)6.260 (4.803 to 6.814)Pooled—PSSBaseline B0.139 (0.092 to 0.289)5.506 (4.544 to 6.461)Pooled—PSSElastic net0.252 (0.087 to 0.268)5.154 (4.921 to 6.153)Pooled—PSSRandom forest0.310 (0.112 to 0.356)4.971 (4.496 to 5.927)Pooled—WEMWBSBaseline A−0.004 (−0.078 to 0.001)8.771 (7.687 to 9.353)Pooled—WEMWBSBaseline B0.109 (−0.040 to 0.324)7.841 (7.244 to 8.285)Pooled—WEMWBSElastic net0.185 (0.105 to 0.312)7.414 (6.805 to 8.019)Pooled—WEMWBSRandom forest0.259 (0.189 to 0.361)7.059 (6.747 to 7.740)UK site—PSSBaseline A−0.028 (−0.064 to 0.015)5.822 (5.219 to 6.452)UK site—PSSBaseline B0.173 (0.114 to 0.198)5.224 (4.818 to 5.685)UK site—PSSElastic Net0.175 (0.107 to 0.303)5.119 (4.493 to 5.749)UK site—PSSRandom forest0.224 (0.058 to 0.264)4.966 (4.726 to 5.507)UK site—WEMWBSBaseline A−0.087 (−0.151 to −0.006)8.531 (7.615 to 9.163)UK site—WEMWBSBaseline B0.163 (0.113 to 0.230)7.531 (6.685 to 7.846)UK site—WEMWBSElastic Net0.178 (0.127 to 0.233)7.558 (6.629 to 7.573)UK site—WEMWBSRandom forest0.267 (0.205 to 0.357)6.880 (6.329 to 7.292)China site—PSSBaseline A0.084 (−0.098 to 0.102)6.201 (6.009 to 6.995)China site—PSSBaseline B0.207 (0.124 to 0.366)5.752 (4.566 to 6.061)China site—PSSElastic net0.356 (0.245 to 0.378)5.322 (4.727 to 5.627)China site—PSSRandom forest0.347 (0.222 to 0.501)5.492 (4.049 to 5.712)China site—WEMWBSBaseline A−0.017 (−0.096 to 0.023)8.851 (8.283 to 9.613)China site—WEMWBSBaseline B0.180 (0.103 to 0.206)7.936 (7.501 to 8.714)China site—WEMWBSElastic net0.182 (0.100 to 0.321)7.950 (6.937 to 8.695)CHINA site—WEMWBSRandom forest0.253 (0.131 to 0.339)7.808 (6.873 to 8.553)R² and RMSE values are the median out-of-sample performance from fivefold nested cross-validation; 95% CIs were obtained by bootstrapping the fold-level estimates. Higher R² and lower RMSE indicate better predictive performance. Baseline A models include only sociodemographic predictors; baseline B models additionally include the mean total stressor score. Elastic net and random forest models include all sociodemographic and item-level stressor variables.PSS, Perceived Stress Scale; WEMWBS, Warwick–Edinburgh Mental Wellbeing Scale. Predictive performance of baseline and machine-learning models R² and RMSE values are the median out-of-sample performance from fivefold nested cross-validation; 95% CIs were obtained by bootstrapping the fold-level estimates. Higher R² and lower RMSE indicate better predictive performance. Baseline A models include only sociodemographic predictors; baseline B models additionally include the mean total stressor score. Elastic net and random forest models include all sociodemographic and item-level stressor variables. PSS, Perceived Stress Scale; WEMWBS, Warwick–Edinburgh Mental Wellbeing Scale. Overall, RF outperformed EN in out-of-sample R² in nearly all models. The exception was PSS in the China site, where EN yielded a slightly higher R² than RF (figure 1 & figure 2). We, therefore, interpret item-level stressors primarily using RF, with EN providing complementary information on directionality and ranking consistency. eFigure 1 reports RF permutation importance for each SL item and summarises, for each of the six datasets, the 15 highest ranking stressors. R² comparison: elastic net versus random forest. CV, cross-validation; PSS, Perceived Stress Scale; SL, Stressors List; WEMWBS, Warwick–Edinburgh Mental Wellbeing Scale. RMSE comparison: elastic net versus random forest. CV, cross-validation; PSS, Perceived Stress Scale; SL, Stressors List; WEMWBS, Warwick–Edinburgh Mental Wellbeing Scale. Perceived stress In the pooled sample, RF explained 31.0% of variance in PSS. The highest ranking predictors were ‘concerned about my mental well-being’ (SL46), ‘uncertainty about the future, especially whether to stay here or return home’ (SL51), ‘my family expectations for my academic performance’ (SL50), ‘trouble with falling asleep or staying asleep’ (SL48) and ‘it is unclear where to seek help for problems’ (SL32) (table 2). Table 2Top-ranked item-level predictors in pooled RF models for PSSStressor itemImportanceSL_46 Concerned about my mental well-being100.00SL_51 Uncertainty about future (stay or return home)36.00SL_50 Family expectations for academic performance31.80SL_48 Trouble with falling asleep or staying asleep30.31SL_32 Unclear where to seek help for problems29.41SL_25 Lacking confidence interacting across backgrounds26.31SL_52 Possibility of being judged for mental health issues24.90SL_8 Managing my time effectively22.68SL_45 Worried about my physical health21.68SL_3 Balancing studies with personal life20.85SL_53 Finding suitable job or internship opportunities19.26SL_2 Impact of intense studying on my health18.06SL_5 Meeting my academic expectations16.60SL_43 Guilt about financial strain on family15.39SL_49 Feeling insecure in this environment14.04Relative importance scores are scaled 0–100 within each model; higher scores indicate greater relative contribution to prediction.PSS, Perceived Stress Scale; RF, random forest. Top-ranked item-level predictors in pooled RF models for PSS Relative importance scores are scaled 0–100 within each model; higher scores indicate greater relative contribution to prediction. PSS, Perceived Stress Scale; RF, random forest. Across both sites, and in pooled PSS models (eTables 4.1 & 4.3 & 4.5), six item-level stressors consistently appeared among the top 15 predictors: ‘concerned about my mental well-being’ (SL46), ‘my family expectations for my academic performance’ (SL50), ‘trouble with falling asleep or staying asleep’ (SL48), ‘worried about my physical health (SL45)’, ‘impact of intense studying on my health’ (SL2) and ‘meeting my academic expectations’ (SL5). RF explained 22.4% of variance in the UK site and 34.7% in the China site. Several predictors appeared more context-specific. In the UK site, ‘managing my time effectively’ (SL8), ‘lacking confidence when interacting with people from different backgrounds’ (SL25), ‘competitive environment among peers’ (SL10) and ‘the possibility of being judged by others for mental health issues’ (SL39) ranked among the top predictors but did not appear among the top 15 in the China site model. In the China site, ‘it is unclear where to seek help for problems’ (SL32), ‘uncertainty about the future, especially whether to stay here or return home’ (SL51) and ‘finding suitable job or internship opportunities’ (SL53) ranked among the top predictors but did not appear among the top 15 in the UK site model. Mental well-being In the pooled sample, the RF model explained 25.9% of the variance in WEMWBS. The highest-ranking stressor items were ‘concerned about my mental well-being’ (SL46), ‘trouble with falling asleep or staying asleep’ (SL48), ‘feeling lonely here’ (SL22), ‘feeling bored here’ (SL31) and ‘managing my time effectively’ (SL8) (table 3). Table 3Top-ranked item-level predictors in pooled RF models for WEMWBSStressor itemImportanceSL_46 Concerned about my mental well-being100.00SL_48 Trouble with falling asleep or staying asleep37.33SL_22 Feeling lonely here21.95SL_31 Feeling bored here17.23SL_8 Managing my time effectively16.36SL_33 Social life feels restricted15.65SL_51 Uncertainty about future (stay or return home)14.38SL_23 Lack of sense of community and belonging here12.83SL_3 Balancing studies with personal life11.86SL_27 Forming friendships with local people11.05SL_25 Lacking confidence interacting across backgrounds8.20SL_52 Possibility of being judged for mental health issues8.18SL_50 Family expectations for academic performance7.99SL_49 Feeling insecure in this environment7.89SL_53 Finding suitable job or internship opportunities7.02Relative importance scores are scaled 0–100 within each model; higher scores indicate greater relative contribution to prediction.RF, random forest; WEMWBS, Warwick–Edinburgh Mental Wellbeing Scale. Top-ranked item-level predictors in pooled RF models for WEMWBS Relative importance scores are scaled 0–100 within each model; higher scores indicate greater relative contribution to prediction. RF, random forest; WEMWBS, Warwick–Edinburgh Mental Wellbeing Scale. Across the UK site, China site and pooled WEMWBS models (eTables 4.2 & 4.4 & 4.6), seven items consistently appeared among the top 15 predictors: ‘concerned about my mental well-being’ (SL46), ‘trouble with falling asleep or staying asleep’ (SL48), ‘feeling lonely here’ (SL22), ‘managing my time effectively’ (SL8), ‘my social life feels restricted’ (SL33), ‘uncertainty about the future, especially whether to stay here or return home’ (SL51) and ‘my family expectations for my academic performance’ (SL50). RF explained 26.7% of the variance in the UK site and 25.3% in the China site. In the UK site, feeling bored here (SL31), lacking confidence when interacting with people from different backgrounds (SL25), worrying about physical health (SL45) and financial pressures from tuition and living expenses (SL42) ranked among the top predictors but did not appear among the top 15 in the China site model. In the China site, lack of a sense of community and belonging (SL23), adjusting to the new educational system (SL4), managing romantic relationships (SL24) and lack of opportunities to improve language skills (SL16) ranked among the top predictors but did not appear among the top 15 in the UK site model. To assess algorithm dependence, we also derived item-level importance from EN standardised coefficients. Rankings were largely consistent across methods. In five of six analytic datasets, Spearman correlations between EN and RF importance scores ranged from 0.704 to 0.744, indicating strong agreement. In the China site WEMWBS model, the correlation was more modest (ρ=0.442), consistent with a moderate association between the two importance estimates (eFigure 2). To evaluate cross-institution generalisation, we conducted external validation by training models at one site and testing them at the other. In both directions, RF explained approximately 19.4%–27.5% of variance in outcomes, outperforming EN (approximately 4.3%–15.2%). As expected, R² values were lower than in within-site nested cross-validation. However, the retained explained variance remained substantial, suggesting that item-level stressor patterns showed reasonable portability across the two universities (eTable 6). Discussion This study used RF and EN models to identify item-level stressors associated with perceived stress and mental well-being in international students. Across all six models, item-level approaches outperformed baseline models, indicating that specific stressors provide predictive information beyond overall stress burden. Across outcomes, concern about mental well-being (SL46) and sleep difficulties (SL48) consistently ranked among the strongest predictors. Family academic expectations (SL50) were particularly important for perceived stress, whereas loneliness (SL22) was more strongly associated with mental well-being. The stressors most strongly associated with perceived stress and mental well-being overlapped but were not identical. This pattern is consistent with evidence that perceived stress and mental well-being represent related but distinct dimensions of mental health.27 Some stressors, therefore, appear relevant across both outcomes, whereas others show greater specificity to distress or mental well-being. Perceived stress In the pooled sample, several stressors consistently emerged as key predictors of higher perceived stress. Health-related concerns were among the strongest, including worries about mental and physical health and sleep difficulties, suggesting that an unfamiliar environment may heighten vigilance about health. Other high-impact stressors centred on academic pressure, students’ own performance expectations and family expectations. Given that many participants came from collectivistic cultural backgrounds in which academic achievement is closely tied to family honour this pattern is unsurprising.28 Students who study abroad may be highly academically motivated yet face challenges beyond those encountered by local peers. When these challenges occur alongside internalised family expectations, they may be experienced as especially stressful, consistent with the prominence of these items in perceived stress models. In the UK site, several stressors appeared to be more context-specific. Time-management difficulties (SL8) were particularly salient. Compared with prior educational experiences, UK universities may require greater independent study and self-directed scheduling. Uncertainty about how to organise this may be appraised as a major demand. Social and interpersonal stressors were also prominent. Lack of confidence when interacting with people from different cultural backgrounds (SL25) and feeling lonely in the host country (SL22) ranking highly. These findings suggest that social difficulty itself is a meaningful source of stress rather than reflecting a preference for distance from peers. This supports the need for structured opportunities for intercultural contact and peer networking, alongside active encouragement of participation rather than assuming integration occurs spontaneously. Concerns about stigma related to mental health difficulties also emerged as salient for Chinese students in the UK. Such concerns do not necessarily imply unwillingness to seek professional help. However, language barriers and uncertainty about how to access support may increase the perceived social costs of help-seeking and heighten stress linked to mental health concerns. That competitive environment among peers (SL10) was identified as an important stressor aligns with broader social trends: social media visibility of peer achievements may intensify upward comparison, and uncertainty in graduate labour markets may strengthen perceptions that students must stand out from peers. For international students, high financial investment in study abroad may further amplify this pressure. A shift toward more metric-driven competition in academia and employment may also make performance differences more visible. Together, these factors may convert peer competition from a potentially motivating context into one experienced as overload and reduced control. For international students studying in China, a different set of context-specific stressors emerged. Not knowing where to seek help (SL32) and difficulties finding suitable jobs or internship opportunities (SL53) ranked highly. This pattern suggests the need for clearer signposting to academic, careers and mental health services. Uncertainty about whether to remain in China or return home (SL51) was particularly salient. For many, staying in China raised concerns about employability, visa policy and long-term prospects. Mental well-being For mental well-being, the data further supports dual-continua models. Items such as Concerned about my mental well-being, Trouble with falling asleep or staying asleep and My family’s expectations for my academic performance ranked highly for both outcomes, suggesting a core cluster of health—and achievement-related concerns that cuts across distress and mental well-being. At the same time, some stressors were more specific to reduced mental well-being than to perceived stress. In particular, ‘feeling lonely here’ and ‘my social life feels restricted emerged’ were high-impact predictors of poorer mental well-being even when they were among the strongest contributors to perceived stress. This pattern suggests that limited social connection and fewer opportunities for enjoyable engagement may erode positive affect even when they are not appraised as acute ‘stress’ in a narrow sense. For Chinese students in the UK, the mental well-being models strengthened the picture observed for perceived stress. Items such as ‘forming friendships with local people’ and ‘feeling bored here’ again suggest that these students are not simply disengaged or uninterested in socialising. Instead, they may be motivated for cross-cultural interaction but face linguistic or cultural barriers that prevent this happening. The importance of financial pressures also challenges assumptions that Chinese international students are uniformly affluent. Rather, financial strain appears to be a correlate of reduced mental well-being, consistent with evidence linking economic hardship to lower mental well-being among international students.29 For international students in China, several stressors were salient for mental well-being. Balancing studies with personal life suggests that such an imbalance may erode positive affect and day-to-day functioning and may be detrimental for students already at risk of reduced social contact and limited local support. Lack of a sense of community and belonging here further suggests a strong but unmet desire for integration into the host society. It is important to note that the predictive models in this study captured variables at different levels of proximity to mental health outcomes. Among the highest ranking predictors were both proximal distress-related experiences (eg, trouble with falling asleep or staying asleep, worries about physical health) and distal contextual and adaptation-related stressors (eg, academic pressures, difficulties accessing support, restricted social life). Proximal items are conceptually closer to mental health outcomes and may, therefore, show stronger predictive associations in cross-sectional models. However, it is important to emphasise that these experiences are not merely redundant with distress outcomes. Several, such as sleep difficulties and concerns about physical health, are themselves dynamically involved in the maintenance and amplification of psychological distress and poorer mental well-being over time, with evidence suggesting bidirectional relationships between these experiences and common mental health difficulties.30 At the same time, a number of more clearly external and potentially modifiable stressors also emerged consistently across models, including family academic expectations, social isolation, restricted social life, uncertainty regarding future plans and difficulties navigating support systems. These more distal stressors may be especially clinically informative as potential targets for prevention and intervention, offering an initial empirical basis to inform the design of support tailored to the specific challenges faced by international students. This study extends previous work focused on broad stress domains by identifying specific stress-related experiences associated with perceived stress and mental well-being among international students. Cross-site validation showed that, although performance decreased relative to within-site cross-validation, models retained meaningful predictive signal across sites, suggesting partial transportability of item-level stressor patterns. The larger performance drop for EN than RF suggests that non-linear relationships may be more robust to contextual variation than purely linear associations. These findings provide preliminary evidence that some stressor patterns may generalise across international student contexts, while others remain context-specific. Limitations Several limitations should be noted. Although overfitting risk was reduced through nested cross-validation and cross-site validation, larger and more diverse samples are needed to refine performance estimates and stabilise item-importance rankings. The sample was not designed to be nationally representative: participants were recruited from a single university in each country, and the China-site sample was predominantly Asian with limited representation from several world regions. The two sites also differed in degree level and length of stay, limiting direct comparability. Financial compensation may have contributed to self-selection effects. Consequently, findings should be interpreted as evidence from two contrasting institutional contexts rather than as representative of all international students in the UK or China. Recruitment used open survey links that could be forwarded through social media, meaning that the total number of individuals who viewed the survey invitation could not be determined. Consequently, precise response rates could not be calculated. Relative to the number of individuals directly contacted, approximately 28% of potential participants at the UK site and 36% at the China site accessed the survey; these should, therefore, be regarded as upper-bound estimates rather than true response rates. As with all voluntary online surveys, the possibility of non-response bias cannot be excluded and may limit the generalisability of the findings. Some highly ranked items, such as concern about mental well-being and sleep difficulties, were conceptually close to the outcomes and may partly reflect proximal symptoms or appraisals. Future research should examine whether similar patterns emerge when such items are excluded. In addition, although the SL was developed through literature review, expert consultation and pilot testing, it has not yet undergone full psychometric validation. Future work should evaluate its factor structure, measurement invariance, reliability and predictive validity across international student populations. Finally, the analyses were not preregistered, and the design was cross-sectional. Findings should, therefore, be interpreted as exploratory predictive associations rather than causal effects. Longitudinal and intervention studies are needed to examine whether the identified stressors prospectively predict changes in mental health and whether modifying them improves outcomes. Conclusion Machine-learning models identified item-level stressors associated with perceived stress and mental well-being in international students, distinguishing stressors that showed consistent importance across outcomes from those that appeared more context-specific. Across samples, concerns about mental and physical health, sleep difficulties and academic and family expectations showed the strongest associations with poorer mental health, whereas loneliness and restricted social life were more closely linked to mental well-being than to perceived stress. Several stressors differed in importance between Chinese students in the UK and international students in China, suggesting that support may benefit from adaptation to local academic, social and cultural contexts. These findings provide an empirical basis for future hypothesis generation, service development and longitudinal or intervention research aimed at improving international student mental health. Supplementary material