Machine learning-based prediction of dissatisfaction after occupational injury: a retrospective cohort study using the nationwide Korean workers’ compensation insurance database Authors: Youjin Lee, Taeyeon Kim, Dahyeon Koo, Dougho Park Published: 2025 ABSTRACT AbstractObjectivesTo develop a machine learning (ML)-based predictive model to determine the key predictors of dissatisfaction after occupational injury (OI).DesignA retrospective cohort study.SettingNationwide 5-year panel data (2018–2022) from the Panel Study of Workers’ Compensation Insurance in South Korea.ParticipantsA total of 2298 workers who completed compensation-related medical care in 2017.MethodsPredictive modelling was conducted with extreme gradient (XG) Boost, light gradient boosting machine (GBM), CatBoost and random forest. SHapley Additive Explanations (SHAPs) analysis was conducted to interpret the feature importance. Further, logistic regression was conducted for comparison.Primary outcome measuresThis study evaluated postinjury satisfaction among workers using survey items associated with satisfaction levels. We adopted a 5-year follow-up period.ResultsOf the 2298 participants, 570 were dissatisfied. The logistic regression model indicated that dissatisfaction was significantly associated with unemployment (adjusted OR (aOR) 1.701; 95% CI: 1.296 to 2.233), lack of private health insurance (aOR 1.347; 95% CI 1.042 to 1.741) and lower perceived socioeconomic status (aOR 2.097; 95% CI 1.109 to 3.965). Among the ML models, light GBM exhibited the highest area under the receiver operating characteristics curve (0.770 (95% CI 0.718 to 0.819)), followed by CatBoost (0.768 (95% CI 0.718 to 0.815)), random forest (0.766 (95% CI 0.715 to 0.814)) and XGBoost (0.765 (95% CI 0.717 to 0.811)). The SHAP analysis demonstrated the total number of household members, extent of pain interference with daily life, perceived health status before injury and financial factors as the strongest predictors.ConclusionThis study developed and demonstrated robust predictive performance of an ML-based model for determining dissatisfaction after OI. The key features included employment status, financial stability, chronic pain and cognitive function, highlighting the multifaceted nature of worker satisfaction.Keywords: MENTAL HEALTH, Artificial Intelligence, EPIDEMIOLOGY, Occupational Health Services STRENGTHS AND LIMITATIONS OF THIS STUDY This study used a large, nationally representative panel dataset of injured workers drawn from the Korean Workers’ Compensation and Welfare Service. This study used a large, nationally representative panel dataset of injured workers drawn from the Korean Workers’ Compensation and Welfare Service. A standardised 5-year follow-up enabled consistent long-term assessment of postinjury satisfaction. A standardised 5-year follow-up enabled consistent long-term assessment of postinjury satisfaction. Employing multiple machine learning algorithms with the SHapley Additive Explanation-based interpretability, internal cross-validation and independent test dataset evaluation supports methodological robustness and transparency. Employing multiple machine learning algorithms with the SHapley Additive Explanation-based interpretability, internal cross-validation and independent test dataset evaluation supports methodological robustness and transparency. The use of a standardised government-managed dataset minimised heterogeneity in data collection but limited the inclusion of detailed clinical or psychosocial variables. The use of a standardised government-managed dataset minimised heterogeneity in data collection but limited the inclusion of detailed clinical or psychosocial variables. As the analysis relied on self-reported survey responses during a period that partly overlapped with the COVID-19 pandemic, potential recall bias and contextual influences on satisfaction cannot be entirely ruled out. As the analysis relied on self-reported survey responses during a period that partly overlapped with the COVID-19 pandemic, potential recall bias and contextual influences on satisfaction cannot be entirely ruled out. Introduction Occupational injuries (OIs) have long-term consequences not only on workers’ physical and mental health but also on their financial stability and overall life satisfaction.1 Workers receiving no appropriate rehabilitation and compensation postinjury may experience physical disabilities, psychological distress, job insecurity and social isolation.2 3 Therefore, assessing and predicting workers’ long-term satisfaction post-OI is crucial for guiding compensation policies and assessing their effectiveness. Several studies have attempted to identify factors associated with postinjury satisfaction among occupationally injured workers. For example, Park et al reported that lower satisfaction with life was associated with a higher risk of OI, suggesting that psychosocial factors, including subjective well-being, may play a role in workplace safety and injury prevention.4 Similarly, O’Neil-Pirozzi et al identified that self-rated health and pain interference were strong predictors of life satisfaction 10 years after traumatic brain injury in their multicentre cohort study.5 O'Toole et al reported that patient satisfaction was primarily determined by physical function, pain, depression and return to work, rather than by demographic, injury or treatment characteristics after severe lower-limb injuries using multivariable regression.6 While these studies provided valuable insights, most relied on traditional regression-based approaches to identify independent predictors, limiting their ability to capture complex, non-linear relationships or interactions between socioeconomic, psychological and clinical variables.7 Furthermore, existing models were primarily explanatory rather than predictive, making it difficult to apply them in real-world decision-making or policy-level risk stratification. Consequently, the existing approaches may not universally represent the multifaceted nature of satisfaction outcomes and their determinants. Machine learning (ML) algorithms are used to predict long-term satisfaction among injured workers. ML techniques can effectively capture intricate variable interactions, efficiently process large datasets and model non-linear associations that traditional statistical approaches may overlook.8 Further, ML enables hyperparameter tuning to optimise predictive performance as well as the comparison of multiple algorithms to determine the most suitable model.9 Using these capabilities, ML improves the accuracy of satisfaction predictions and provides a more data-driven approach to determining high-risk individuals. Moreover, existing ML techniques, such as random forests, gradient boosting machines (GBMs) and support vector machines, consistently outperform traditional logistic regression in predicting return-to-work and rehabilitation success.10 11 Hence, this study aimed to develop an ML-based predictive model to forecast 5-year postinjury satisfaction among workers who have completed their compensation-related medical care post-OI. Further, we determined key features that affect satisfaction, thereby providing information that can improve policy design and targeted interventions for injured workers. Materials and methods Data source and participants This retrospective cohort study used the 5-year panel data from the second cohort of the Panel Study of Workers’ Compensation Insurance (2018–2022), provided by the Labor Welfare Research Institute of the Korea Workers’ Compensation and Welfare Service.12 The initial sample involved 3294 workers who had completed industrial accident compensation insurance treatment in 2017 after OI. The participants’ baseline characteristics were collected from the first survey conducted in 2018, and follow-up data obtained from 2019 to 2022 were used to assess satisfaction at the 5-year postinjury mark, which served as the primary outcome variable of this study. This study excluded 430 participants with missing baseline characteristics in 2018 and 566 participants who did not respond to the survey in 2022 (online supplemental table 1). Hence, the final analytical sample comprised 2298 participants (figure 1). Flowchart of participants’ inclusion. D group, dissatisfied group; S/N group, satisfied and neutral group. The Institutional Review Board (IRB) of Pohang Stroke and Spine Hospital reviewed and approved this study (PSSH0475-202410-HR-018–01). The IRB waived the requirement for informed consent due to the retrospective study design and the anonymised dataset format. Outcome definition This study evaluated postinjury satisfaction among workers using survey items associated with satisfaction levels. We adopted a 5-year follow-up period, as the Panel Study of Workers’ Compensation Insurance provides standardised longitudinal data collected annually from 2018 to 2022, allowing consistent long-term assessment. Online supplemental table 2 presents the specific survey questions and response scales. To ensure consistency, all responses—initially measured on a 1–4 or 1–5 point scale—were standardised to a 1–5 point scale. Negatively framed items, where higher scores denoted lower satisfaction, were reverse-coded so that all higher scores consistently represented greater satisfaction. Positive items underwent a linear transformation, whereas negative items were linearly transformed and reverse-coded to maintain interpretability. Each participant’s average satisfaction score was calculated after standardisation by summing all survey item scores and dividing by the number of items. The satisfaction groups were then defined according to the quartiles of the satisfaction score distribution. The dissatisfied group (D group) included participants who scored below the first quartile (Q1: 3.23), whereas the neutral group involved those scoring between Q1 and the third quartile (Q3: 3.94), and the satisfied group consisted of those scoring at or above Q3. The groups were further consolidated for predictive modelling into the satisfied and neutral group (S/N group) and the D group (figure 2). Distribution of satisfaction scores among workers with occupational injuries. This histogram illustrates the distribution of postinjury satisfaction scores among the study participants. The red, green and orange dashed lines indicate the first quartile (Q1: 3.23), median (3.60) and third quartile (Q3: 3.94), respectively. Participants were classified into satisfaction groups according to these quartile thresholds. Those with scores below Q1, between Q1 and Q3 and above Q3 were categorised into the dissatisfied (D group), neutral and satisfied groups, respectively. Covariates This study included multiple covariates to consider potential factors that affect postinjury satisfaction. These covariates were categorised into demographic characteristics, personal lifestyle factors, individual health and functional status and work-related injury variables. Online supplemental table 3 presents detailed descriptions of each variable. Statistical analysis We analysed the continuous variables using mean±SD and compared the groups using an independent t-test. We assumed normality of continuous variables based on the central limit theorem and additionally checked the skewness and kurtosis of each variable (online supplemental table 4). Categorical variables were expressed as frequencies (proportions), and we conducted χ2 tests for trends to assess group differences. To predict dissatisfaction (D group) 5 years postinjury, we conducted a binary logistic regression analysis with backward elimination for variable selection. Results were reported as adjusted ORs (aORs) with 95% CIs. We calculated the area under the receiver operating characteristics curve (AUC) to evaluate the overall model performance. We employed a complete-case analysis approach and set statistical significance at p values of <0.05. R software V.4.4.1 (R Foundation for Statistical Computing, Vienna, Austria) was used for all analyses. ML prediction All ML predictions were conducted in a Python V.3.12.3 environment, and the entire code for the prediction models is available in the online supplemental content. We developed an ML-based predictive model using the scikit-learn column transformer for data preprocessing. Categorical variables underwent one-hot encoding through One Hot Encoder, whereas continuous variables were standardised with StandardScaler, which normalises values according to their mean and SD. To train and assess the model, we divided the dataset into training (80%) and testing (20%) sets. To address the class imbalance in the training data, we applied the synthetic minority oversampling technique (SMOTE) for nominal and continuous variables. This technique generates synthetic samples for the minority class while considering both categorical and continuous variables. In this study, the sampling strategy was set to 0.5, ensuring that the minority class was augmented to 50% of the majority class size for a more balanced dataset. For a sensitivity analysis, we also established prediction models without SMOTE (ie, under the original imbalanced data condition). We trained four ML models: extreme gradient (XG) boost, light GBM, CatBoost and random forest. The entire pipeline, including data preprocessing, oversampling and model training, was integrated using the scikit-learn and imbalanced-learn pipelines. The model was validated using 10-fold cross-validation, and the hyperparameter was tuned with a combination of random and grid search. Online supplemental table 5 presents the optimal hyperparameters. Model performance was internally validated within the training set, and the final model was subsequently evaluated on an independent test set. We used AUC, accuracy, precision and recall to assess the model performance. The final model was selected according to the highest AUC among the trained models. Further, we analysed the receiver operating characteristic curves to visually compare the model performance. To improve model interpretability, we used SHapley Additive Explanations (SHAPs) to evaluate feature importance. SHAP TreeExplainer was applied to XGBoost, light GBM and CatBoost, whereas the same method was used for random forest to ensure consistency in feature importance assessment. SHAP values were visualised using shap.summary_plot, providing information about the contribution of each feature to model predictions. Patient and public involvement None. Results Baseline characteristics Among 2298 participants, 1728 and 570 were included in the S/N and D groups, respectively. Significant differences were found in several demographic and clinical characteristics between the two groups. The D group demonstrated a higher proportion of older participants, with 36.0% aged ≥60 years compared with 27.5% in the satisfied group (p=0.001). Further, the D group exhibited a lower educational level, with only 11.6% having >12 years of educational experience compared with 19.9% in the S/N group (p<0.001). Further, employment duration was significantly different, as a greater proportion of dissatisfied participants had <1 year of employment preinjury (68.1% vs 59.0%, p=0.001). The S/N group consisted of a higher proportion of participants with no disability (20.4% vs 15.1%, p<0.001). Further, the D group reported longer recovery periods, with 26.1% requiring >12 months compared with 12.6% in the S/N group (p<0.001). Finally, the D group demonstrated a lower perceived socioeconomic status, with 56.8% rating their status as low, compared with 33.2% in the S/N group (p<0.001). Online supplemental table 6 shows the detailed results of the entire baseline characteristics of participants. Stepwise logistic regression Table 1 shows the logistic regression model for predicting dissatisfaction. Table 1Stepwise logistic regression model for the dissatisfied groupVariableAdjusted OR (95% CI)P valueAge groups <40Reference 40–491.177 (0.734 to 1.887)0.498 50–590.910 (0.566 to 1.463)0.697 ≥600.666 (0.390 to 1.140)0.138Sex MaleReference Female0.594 (0.410 to 0.860)0.006Education level ≤6 yearsReference 7–12 years0.679 (0.494 to 0.934)0.017 >12 years0.528 (0.331 to 0.842)0.007Regions Seoul/GangwonReference Gyeonggi1.456 (1.006 to 2.106)0.046 Daejeon/Sejong/Chungcheong1.176 (0.716 to 1.929)0.522 Gwangju/Jeolla0.855 (0.515 to 1.417)0.543 Daegu/Gyeongsangbuk1.278 (0.809 to 2.021)0.293 Busan/Ulsan/Gyeongsangnam0.980 (0.651 to 1.476)0.925Disability rating Mild to moderate (1–9)Reference Severe (10–14)0.671 (0.489 to 0.922)0.014 No disability0.683 (0.447 to 1.044)0.079Current employment status EmployedReference Unemployed1.701 (1.296 to 2.233)<0.001Marital status at injury Not-marriedReference Married0.861 (0.560 to 1.323)0.494 Others1.321 (0.832 to 2.097)0.238Surgery during injury YesReference No1.309 (1.019 to 1.682)0.035Private health insurance coverage, n (%) YesReference No1.347 (1.042 to 1.741)0.023 Unknown0.983 (0.409 to 2.361)0.969Job classification Managers and professionalsReference Clerical and service1.205 (0.630 to 2.303)0.573 Agriculture, forestry and fisheries0.422 (0.163 to 1.088)0.074 Skilled workers and technicians1.25 (0.718 to 2.175)0.431 Manual labourers1.773 (0.995 to 3.158)0.052Total workforce at the workplace <5 peopleReference 5–19 people0.876 (0.648 to 1.184)0.389 20–99 people0.807 (0.566 to 1.152)0.238 100–9990.785 (0.457 to 1.350)0.382 ≥10000.665 (0.202 to 2.187)0.501Extent of pain interference with daily life NoneReference Moderate0.856 (0.543 to 1.350)0.503 Severe or constant1.567 (0.973 to 2.523)0.064Perceived health status before injury HealthyReference Average1.710 (1.257 to 2.327)<0.001 Unhealthy1.599 (1.124 to 2.275)0.009Learning, remembering and concentrating DifficultReference Moderate0.875 (0.525 to 1.457)0.607 Intact0.477 (0.309 to 0.737)<0.001Dressing, bathing and household tasks DifficultReference Moderate0.622 (0.386 to 1.004)0.052 Intact0.537 (0.352 to 0.820)0.004Perceived socioeconomic status HighReference Middle1.430 (0.764 to 2.678)0.263 Low2.097 (1.109 to 3.965)0.023 Monthly income (per 10 000 KRW)0.999 (0.998 to 1.000)0.102 Weekly working out (per day)0.962 (0.913 to 1.013)0.139 Total household members (per person)0.880 (0.778 to 0.996)0.043KRW, Korean won. Stepwise logistic regression model for the dissatisfied group KRW, Korean won. Among demographic factors, females demonstrated a lower risk of dissatisfaction compared with males (aOR 0.594; 95% CI 0.410 to 0.860; p=0.006). Further, educational level was considered a significant predictor, with higher educational level associated with a reduced risk of dissatisfaction. Compared with participants with ≤6 years of educational experience, those with >12 years exhibited an aOR of 0.528 (95% CI 0.331 to 0.842; p=0.007), indicating a protective effect of higher education. Regarding employment status, being unemployed was significantly associated with dissatisfaction (aOR 1.701; 95% CI 1.296 to 2.233; p<0.001). Among the health-related factors, participants with severe disabilities (rating: 10–14) demonstrated a lower likelihood of dissatisfaction compared with those with mild to moderate disabilities (aOR 0.671; 95% CI 0.489 to 0.922; p=0.014). Conversely, individuals who underwent no surgery during injury reported a higher risk of dissatisfaction (aOR 1.309; 95% CI 1.019 to 1.682; p=0.035). Further, socioeconomic factors played a role. Individuals without private health insurance demonstrated a significantly higher likelihood of dissatisfaction (aOR 1.347; 95% CI 1.042 to 1.741; p=0.023). Participants who perceived their preinjury health as ‘average’ or ‘unhealthy’ were more likely to report dissatisfaction (aOR 1.710; 95% CI 1.257 to 2.327; p<0.001 for ‘average’ and aOR 1.599; 95% CI 1.124 to 2.275; p=0.009 for ‘unhealthy’). Moreover, limitations on daily activities were associated with dissatisfaction. Participants with intact learning, remembering and concentrating abilities reported a lower risk of dissatisfaction (aOR 0.477; 95% CI 0.309 to 0.737; p<0.001). Similarly, individuals with intact abilities in dressing, bathing and household tasks exhibited a reduced likelihood of dissatisfaction (aOR: 0.537; 95% CI 0.352 to 0.820; p=0.004). Finally, participants with a lower perceived socioeconomic status were at a higher risk of dissatisfaction. Compared with those who perceived their status as high, participants in the low socioeconomic status group reported an aOR of 2.097 (95% CI 1.109 to 3.965; p=0.023), indicating a significant association between financial hardship and dissatisfaction. ML performance Table 2 summarises the predictive performance of the four ML models in the test dataset: XGBoost, light GBM, CatBoost and random forest. Among these models, light GBM exhibited the highest AUC (0.770 (95% CI 0.718 to 0.819)), followed by CatBoost (0.768 (95% CI 0.718 to 0.815)), random forest (0.766 (95% CI 0.715 to 0.814)) and XGBoost (0.765 (95% CI 0.717 to 0.811)) (figure 3). Moreover, light GBM achieved the highest accuracy of 0.787 (95% CI 0.750 to 0.826) and a balanced precision (0.820 (95% CI 0.780 to 0.858)) and recall (0.919 (95% CI 0.889 to 0.946)), indicating a robust performance in predicting dissatisfaction among injured workers. Random forest demonstrated the highest recall of 0.934 (95% CI 0.906 to 0.958), indicating its effectiveness in capturing dissatisfied cases; however, it exhibited a lower precision of 0.803 (95% CI 0.763 to 0.842), revealing a higher false positive rate. XGBoost and CatBoost showed competitive performance across all metrics but were slightly behind light GBM regarding AUC. These results indicate that the light GBM was the most effective model for predicting dissatisfaction 5 years postinjury. The confusion matrix of each algorithm’s prediction is provided in online supplemental table 7. Finally, models trained without SMOTE consistently demonstrated similar predictive performance compared with those trained with SMOTE (online supplemental table 8). Table 2Prediction performance of each algorithm in the test datasetAlgorithmsPrecision(95% CI)Recall(95% CI)AUC(95% CI)Accuracy(95% CI)XGBoost0.824(0.785 to 0.862)0.893(0.858 to 0.925)0.765(0.717 to 0.811)0.776(0.739 to 0.813)Random forest0.803(0.763 to 0.842)0.934(0.906 to 0.958)0.766(0.715 to 0.814)0.778(0.741 to 0.815)Light GBM0.820(0.780 to 0.858)0.919(0.889 to 0.946)0.770(0.718 to 0.819)0.787(0.750 to 0.826)CatBoost0.818(0.778 to 0.856)0.899(0.866 to 0.929)0.768(0.718 to 0.815)0.774(0.733 to 0.811)AUC, area under the receiver operating characteristics curve; GBM, gradient boosting machine; XG, extreme gradient. Prediction performance of each algorithm in the test dataset AUC, area under the receiver operating characteristics curve; GBM, gradient boosting machine; XG, extreme gradient. Receiver operating characteristic (ROC) curves for the machine learning models: extreme gradient (XG) Boost, random forest, light gradient boosting machine (GBM) and CatBoost. The area under the ROC curve values for all models is similar, indicating comparable predictive capabilities. AUC, area under the receiver operating characteristics curve. Feature importance Across all models, total household members appeared as the most crucial feature for model performance, followed by the extent of pain interference with daily life, perceived health status preinjury and total household income. Further, employment-related variables, such as current employment status, job classification and workplace size, significantly predicted dissatisfaction. Socioeconomic factors, including perceived socioeconomic status, private health insurance coverage and total monthly household expenditure, were consistently related to dissatisfaction when categorised. Further, lifestyle and health factors, such as weekly working out days, smoking status and treatment duration adequacy, contributed to the model’s predictions (table 3). The SHAP summary plots further illustrate that higher perceived socioeconomic status and stable employment were related to lower dissatisfaction risks. In contrast, financial strain, health deterioration and lack of insurance increased the probability of dissatisfaction (online supplemental figure 1). Table 3Top-five predictors across machine learning modelsOrdersXGBoostRandom forestLight GBMCatBoost1Total household membersTotal household membersTotal household membersTotal household members2Extent of pain interference with daily life (severe or constant)Extent of pain interference with daily life (severe or constant)Extent of pain interference with daily life (severe or constant)Perceived health status before injury (healthy)3Perceived health status before injury (healthy)Current employment status (employed)Perceived health status before injury (healthy)Perceived social economic status (low)4Total monthly household expenditurePerceived social economic status (low)Sex (male)Extent of pain interference with daily life (severe or constant)5Total household incomePerceived health status before injury (healthy)Perceived social economic status (low)Regions (Gyeonggi)GBM, gradient boosting machine; XG, extreme gradient. Top-five predictors across machine learning models GBM, gradient boosting machine; XG, extreme gradient. Discussion This study aimed to develop an ML-based predictive model to determine the factors influencing dissatisfaction among workers 5 years after OI. The results revealed that the light GBM achieved the highest AUC, indicating its effectiveness in predicting dissatisfaction. The key predictors identified across the models included total household members, the extent of pain interference with daily life, perceived health status preinjury and socioeconomic factors such as income and employment status. Similarly, the logistic regression analysis revealed unemployment, lower educational levels, lack of private health insurance and financial instability as significant risk factors for dissatisfaction. Notably, the cohort used in this study can be considered nationally representative of injured workers in Korea. All OI cases in Korea are centrally managed under the Ministry of Employment and Labour through the Korea Workers’ Compensation and Welfare Service. The Panel Study of Workers’ Compensation Insurance dataset, constructed through stratified random sampling of individuals who received compensation-related medical care, provides a comprehensive and population-representative sample of injured workers across diverse regions and industries,12 thereby demonstrating the robustness of our dataset. This study presented several crucial methodological and practical advancements in predicting dissatisfaction among injured workers. Unlike previous studies that primarily depended on traditional statistical methods, our study is among the first to apply multiple ML algorithms to predict dissatisfaction 5 years postinjury. Using ML models, we captured complex, non-linear interactions between variables that conventional logistic regression may overlook. Furthermore, by applying multiple ML algorithms and optimising the hyperparameters, we determined the most effective model with the highest predictive power. However, the lack of model interpretability is a major limitation of many ML-based studies.13 Therefore, we incorporated SHAP analysis to address this, which can be conducted to quantify and visualise feature importance.14 This approach not only improved the explainability of our models but also provided practical information about key features affecting dissatisfaction, thereby helping clinicians and policy makers prioritise intervention strategies. Moreover, this study provided an in-depth comparison between ML and traditional regression models. The logistic regression model determined employment status, socioeconomic status and educational level as significant predictors of dissatisfaction. ML models highlighted total household members and financial factors as crucial contributors. This difference may originate from the ability of ML models to capture complex, non-linear interactions that traditional regression methods overlooked. Further, ML models provided the ranked importance of features, whereas logistic regression estimated the direct effect of each variable. The discrepancy underscored the advantage of using ML for predictive performance while retaining logistic regression for explainability. The results of this study are congruent with those of previous studies that identified key factors affecting satisfaction after OI. In particular, previous studies have revealed that self-esteem and subjective health status significantly affected overall satisfaction,15 16 consistent with our results. However, notable differences appeared in the relative importance of certain features. Previous studies emphasised the return to work as a crucial determinant of life satisfaction,17 18 whereas our ML models assigned it lower importance than other factors. In contrast, our results indicate that psychosocial factors, such as self-esteem and social support, may affect satisfaction more than employment status alone. This discrepancy may have originated from methodological differences—traditional regression analyses assess the independent effect of each feature. In contrast, ML models capture complex interactions between multiple factors while emphasising the possibilities of ML models to uncover hidden associations among complex features. This study provided a data-driven approach for predicting long-term satisfaction after OI and determining high-risk individuals more at risk of experiencing dissatisfaction. By integrating ML models, this predictive framework supports the development of personalised intervention strategies. In particular, early identification of workers at risk of low satisfaction enabled targeted psychological support programmes, such as self-esteem enhancement initiatives and workplace reintegration mentoring systems. Further, economic assistance programmes, including job retraining and reemployment support, can be optimised to address the specific requirements of at-risk workers, thereby ensuring a smoother transition back into the workforce. Furthermore, this study reinforces the effectiveness of existing policy interventions while providing information for more precise and evidence-based modifications. The prominence of psychosocial factors as significant determinants of satisfaction indicates that interventions should extend beyond financial compensation to include mental health support and the strengthening of social support networks. These results emphasise the need for holistic policies that integrate economic security and psychosocial well-being, thereby ensuring a more sustainable recovery for injured workers. It is also important to acknowledge that part of the study period (2020–2022) overlapped with the COVID-19 pandemic, which may have influenced workers’ satisfaction and socioeconomic conditions. Previous studies have reported that the pandemic led to increased psychological distress, job insecurity and changes in work environments, potentially affecting perceived health and satisfaction levels among workers.19 20 Although the Panel Study of Workers’ Compensation Insurance surveys were government-driven and standardised annually, some residual effects of the pandemic on satisfaction responses cannot be completely ruled out. Future studies using postpandemic data could help disentangle these effects and further validate the generalisability of our findings. Despite this study’s strengths, we should acknowledge its several limitations. First, as a retrospective cohort study, it could not establish causal associations between predictive features and outcomes. ML models effectively identified patterns and associations; however, they did not infer causation similar to experimental or longitudinal studies.21 Second, this study analysed satisfaction levels at a single time point rather than obtaining longitudinal changes. Data were available from 2018 to 2022, whereas the analysis focused on a specific endpoint, thereby limiting the ability to evaluate the evolution of satisfaction. Future research is recommended to explore time-series modelling or repeated-measures analysis to better understand the long-term trends in satisfaction after OI. Third, SHAP analysis was used to improve model interpretability, whereas ML models are inherently limited in precisely quantifying each variable’s effect.22 Fourth, we used a complete-case analysis for predictive modelling, necessitated by the nature of ML algorithms, and acknowledge that if non-response were also dependent on the unobserved satisfaction outcome itself. In other words, this method could be subject to selection bias. Finally, a substantial portion of the dataset relied on self-reported survey responses, which increased the risk of measurement bias.23 Participants’ subjective perceptions of satisfaction, health status and economic conditions may have introduced recall or reporting bias, which could have influenced the accuracy of the model predictions. Future studies should consider incorporating objective clinical and socioeconomic indicators to improve data reliability. Resolving these limitations in future research will help refine predictive models and policy applications, ultimately improving the satisfaction of injured workers. In conclusion, this study developed an ML-based predictive model that demonstrated reliable and robust performance in identifying dissatisfaction among injured workers 5 years after compensation-related medical care. The key features included employment status, financial stability, chronic pain and cognitive function, highlighting the multifaceted nature of worker satisfaction. This study provides a data-driven, interpretable framework for the early identification of at-risk individuals by integrating ML and SHAP analysis, thereby providing actionable information for targeted rehabilitation planning, policy interventions and workforce reintegration programmes. Further investigation into personalised intervention strategies, including job retraining programmes and mental health support, is warranted to ultimately improve long-term worker satisfaction and well-being. Supplementary material Acknowledgements The authors would like to thank the Korea Workers’ Compensation and Welfare Service (Labor Welfare Research Institute) for providing the raw data from the Panel Study of Workers’ Compensation Insurance. The paper’s contents are solely the responsibility of the author and do not represent the official view of the Labor Welfare Research Institute. Footnotes Funding: The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors. Prepublication history and additional supplemental material for this paper are available online. To view these files, please visit the journal online (https://doi.org/10.1136/bmjopen-2025-103321). Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Data availability free text: Data are available on reasonable request. Patient and public involvement: Patients and/or the public were not involved in the design, conduct, reporting or dissemination plans of this research. Ethics approval: The institutional review board (IRB) of Pohang Stroke and Spine Hospital reviewed and approved this study (PSSH0475-202410-HR-018-01). The IRB waived the requirement for informed consent due to the retrospective study design and the anonymised dataset format. Data availability statement Data are available upon reasonable request. References Supplementary Materials Data Availability Statement Data are available upon reasonable request.