Work-Life Conditions as the Primary Determinant of Seafarer Mental Health: An Explainable Machine Learning Analysis

Abstract

Seafarer mental health has become an increasingly critical concern due to its substantial implications for transportation safety and operational performance. Traditional analytical approaches often inadequately capture the complex, non-linear interactions among the multidimensional determinants involved. To bridge this gap, this study proposes an explainable machine learning (ML) framework that integrates a Random Forest classifier with SHapley Additive exPlanations (SHAP) for simultaneous prediction and interpretation. Based on a survey of 500 seafarers, 12 risk factors were preprocessed through label encoding and the dataset was split into training and test sets using stratified sampling. A Random Forest model, optimized via Bayesian hyperparameter tuning, was employed to predict psychological states, with performance evaluated through accuracy, precision, recall, and F1-score. SHAP analysis was then applied to quantify global feature importance and to examine individual prediction mechanisms and interaction effects. The results identify current work-life conditions as the most influential determinant, exhibiting a polarizing effect on psychological states that substantially outweighs other factors. Furthermore, favorable environmental conditions amplify the positive effects of career development and social recognition, whereas high onboard service pressure persistently undermines mental health even under otherwise optimal circumstances. These findings underscore the necessity of prioritizing systemic environmental improvements as a foundation for effective psychological interventions, suggesting that tailored support strategies should be implemented subsequent to such enhancements. This study provides a data-driven, interpretable framework to support precision mental health management in maritime operations. These findings advocate for maritime policymakers and shipping companies to prioritize systemic improvements in onboard living and working conditions as a foundational strategy, complemented by targeted psychological support and enhanced awareness of mental health services.

Keywords

seafarer mental health stressors transportation safety work-life conditions machine learning random forest SHapley Additive exPlanations (SHAP)

Introduction

Seafarers constitute a distinct occupational cohort dedicated to maritime transportation.¹ This specialized workforce consistently faces severe environmental challenges: prolonged isolation in confined spaces, demanding high-risk shift rotations, limited recreational and social facilities, physical separation from family and social support systems, and repeated exposure to hazardous maritime conditions.^2,3 The convergence of these factors elevates seafarers’ vulnerability to mental health disorders.⁴ Empirical studies consistently demonstrate that the prevalence of moderate-to-severe psychological distress among seafarers significantly exceeds rates observed in general land occupations.^5,6 This disparity not only adversely affects individual well-being and navigational safety but also leads to reduced operational efficiency, increased accident risks, and accelerated talent attrition, ultimately compromising both the economic sustainability and safe operations of the global shipping industry.^7,8

Although the importance of seafarer mental health is widely recognized, understanding the numerous and interrelated psychological, social, and environmental stressors that influence it, and developing effective targeted intervention strategies, remains a significant challenge. Conventional mental health research in this domain primarily relies on subjective rating scales, retrospective interviews, and frequency-based statistical analyses.^9,10 While valuable for identifying individual risk factors, these approaches are limited in their ability to capture the complex, nonlinear dynamics among multidimensional determinants, including physiological indicators, behavioral patterns, micro-environmental stressors (such as fatigue induced by irregular working hours, workplace bullying, and burnout), and social connectivity metrics within authentic maritime operational contexts.¹¹ Methodologically, these limitations reflect a broader research paradigm in the social sciences that often prioritizes causal explanation over predictive accuracy.¹² Moreover, traditional statistical methods often yield generalized recommendations that may not account for individual differences among seafarers, limiting their utility for personalized risk assessment and the design of tailored onboard mental health support.¹³ Existing studies indicate that seafarer mental health is characterized by 3 key features: a complex structure influenced by multiple factors, varying in form and severity, and significant inter-individual differences.¹⁴ These features highlight the need for advanced analytical approaches that can overcome the limitations of traditional methods in seafarers’ mental health.

The rapid advancement of machine learning (ML) has enabled increasingly advanced modeling of complex systems, demonstrating considerable potential for predicting mental health risks.¹⁵ However, the inherent opacity of “black-box” models significantly limits their practical adoption in real-world maritime settings. When industrial practitioners, seafarers, and maritime administrators cannot understand the reasoning behind model predictions or trust algorithmic conclusions, evidence‑based decision‑making is hindered.¹⁶ This methodological impasse makes explainable ML not merely beneficial but essential for practical implementation in seafarer health services. By applying predictive modeling alongside SHapley Additive exPlanations (SHAP) analysis, this study identifies and quantifies key risk and protective factors while elucidating their complex interactions, thereby advancing the methodological foundation of seafarer mental health studies toward more actionable insights. This approach effectively demystifies the “black-box” nature of conventional models, providing a translucent, evidence-based diagnostic framework that moves beyond predictive scoring toward a reconceptualization of psychological intervention paradigms for maritime workers.

In this study, seafarer mental health is conceptualized not merely as an individual psychological outcome, but as a multilevel occupational health and preventive health service challenge.^1,13 Within the maritime industry, the early identification of mental health risks, accurate stratification of vulnerable groups, and effective allocation of limited organizational resources require approaches that extend beyond individual clinical assessments toward evidence-based decision support systems.^7,16,17 Accordingly, research on seafarer mental health should not only aim to explain causal mechanisms but also support actionable, predictive, and targeted health service interventions.^9,11 Explainable machine learning approaches respond directly to this need by translating complex risk patterns into transparent and interpretable outputs that can be meaningfully utilized by health service providers, occupational physicians, and maritime administrators.¹⁵ SHAP-based explanations reveal not only which factors contribute to elevated psychological risk, but also under what conditions and for which subgroups these factors exert stronger or weaker effects. This capability is particularly critical for planning mental health services, prioritizing preventive interventions, and designing personalized support strategies in maritime settings characterized by constrained resources and heterogeneous risk profiles. From this perspective, the present study positions explainable machine learning not as a purely analytical technique, but as an applied methodological framework directly relevant to health services research and delivery. By moving beyond generalized risk classification toward individualized and context-sensitive insights, this approach supports the development of precision-oriented mental health services for seafarers and contributes to a more effective and sustainable governance of maritime occupational health.^1,11

Therefore, the primary objective of this study is to develop and apply an explainable ML framework to identify, quantify, and interpret the critical determinants of seafarers’ psychological states. In this context, “critical determinants” are operationalized as those features that exert the strongest average marginal impact on model predictions, as quantified by SHAP values-a method grounded in cooperative game theory. This approach fundamentally differs from conventional correlation-based or regression analyses by capturing complex non-linear interactions and providing both global and local interpretability. Consequently, this research contributes by: (1) establishing a transparent, data-driven model for predicting seafarer mental health risks; (2) to quantify the relative importance and interaction effects of the multifaceted determinants affecting seafarers; and (3) generating actionable, individualized insights to inform precision intervention strategies in maritime settings.

The remainder of this paper is structured as follows: Section 2 reviews the literature on seafarer mental health and machine learning applications in this field, culminating in the identification of the research gap. Section 3 details the materials and methodology, encompassing data description, preprocessing, model selection, evaluation, and the interpretable SHAP framework. Section 4 presents the results and analysis, including sample characteristics, the identification of key determinants and their interaction effects. Section 5 discusses the findings and their theoretical and practical implications. Finally, Section 6 concludes this study by summarizing the key findings, highlighting contributions, and outlining limitations alongside future research directions.

Literature Review

Research on Seafarer Mental Health

Empirical evidence consistently demonstrates that mental health disorders are highly prevalent among seafarers, largely resulting from occupational and environmental stressors.¹⁷ A cross-sectional study reported that 38.8% to 56.7% of seafarers exhibit clinically significant psychological deterioration compared to the general population.¹⁸ Seafarers exhibit elevated levels of somatization, obsessive-compulsive tendencies, depression, and anxiety. Vulnerability is particularly pronounced among seafarers under 35 and contractual personnel, for whom psychosexual distress also represents a notable concern.¹⁹ Methodologically, research in this field has evolved from basic epidemiological surveillance toward systematic deconstruction of underlying determinants. Contemporary studies identify 3 primary categories of risk factors. First, environmental stressors, including physical adversities, informational isolation, and limited social support.¹⁴ Second, occupational hazards, encompassing high-risk operations, prolonged working hours, family separation, and diminished professional identity.²⁰ Third, systemic deficiencies in maritime governance, such as inadequate crew rotation protocols, gaps in regulatory implementation, and interpersonal friction.²¹ These findings underscore the need for advanced analytical frameworks capable of capturing complex interactions among these multifaceted determinants. To further investigate the impact of these risk factors, this study applies quantitative modeling and systematic intervention frameworks.

Building on the identification of multifaceted determinants, existing research has made significant contributions to understanding seafarers’ mental health. Path analysis using structural equation modeling has empirically validated the causal mechanisms through which environmental stressors precipitate occupational strain. These analyses also clarify how multifactorial crises emerge from synergistic interactions.²² Emotional intervention frameworks developed for crew members have further standardized clinical protocols and demonstrated internationally validated efficacy rates.²³ Although the COVID-19 pandemic severely disrupted seafarers’ rest entitlement provisions, it simultaneously accelerated revisions to the Maritime Labour Convention (MLC) 2006 and promoted improvements in emergency response mechanisms.^18,24 These findings highlight the importance of empirically quantifying the relative contributions and interaction effects of determinants on individual psychological states. They also underscore the need to systematically identify novel individual risk profiles that transcend conventional categorical boundaries and to generate personalized intervention recommendations based on transparent model explanations for different risk typologies. Collectively, these methodological contributions establish a new paradigm for precision mental health management in maritime contexts. This approach provides the foundation for the data-driven methods described in the subsequent sections on “ML Applications for Mental Health Prediction” and “Research Contributions.”

ML Applications for Mental Health Prediction

ML applications in mental health face dual challenges stemming from their inherent black-box characteristics: model interpretability limitations and dynamic adaptability constraints. Model interpretability limitations arise from decision-making mechanisms within neural network hidden layers, where complex interdependencies exceed human cognitive interpretability. Meanwhile, dynamic adaptability constraints reflect the fundamental divergence between models’ continuous self-updating capabilities and the static nature of conventional medical interventions.²⁵ Crucially, this critique does not diminish ML’s substantive value, particularly given the ongoing scholarly debate regarding the operational boundaries of “explainability” itself. To advance this field, explainable AI research must address 3 pivotal challenges: establishing essential definitional parameters for explainability, formulating domain-agnostic principles, and adapting Turing-test paradigms to accommodate algorithmic opacity constraints. The transformation from opaque to transparent systems requires a multifaceted approach: developing embedded explanatory architectures within existing models, optimizing human-computer interaction design, implementing institutionally grounded socio-ethical assessment frameworks, and codifying objective evaluation metrics.²⁶

Within healthcare diagnostics, conventional validation frameworks are increasingly inadequate due to ambiguities in algorithmic metrics and variations in dynamic prediction rules. This necessitates the development of a “clinician-algorithm co-decision mechanism” that reconciles population-level algorithmic precision with patient-specific clinical contingencies.²⁷ This technological paradigm has catalyzed jurisprudential innovation across the social sciences, making data-driven decision support more actionable. Despite the inherent sparsity of big data, ML demonstrates robust capability in processing high-dimensional unstructured datasets to achieve predictive accuracy. Nevertheless, even with advances in explainable ML, 4 persistent methodological challenges remain. First, the complex and multi-layered nature of social phenomena makes it difficult to accurately measure key variables, leading to potential issues in measurement validity and increasing the risk of omitted variables in the analysis. Second, predictive generalizability remains constrained, particularly when moving beyond standardized policy evaluation to address emergent systemic disruptions. Third, procedural vulnerabilities emerge from non-standardized practices in algorithm selection, feature engineering, and hyperparameter optimization, introducing researcher subjectivity and potential methodological misuse. Finally, limitations in causal inference persist, requiring supplementary analytical frameworks despite ML’s correlational predictive capabilities.²⁸

These challenges can be advanced through 3 strategic pathways: First, defining the appropriate scope of application for predictive models through tripartite assessment of problem typology, data fidelity, and regulatory permissibility. Second, developing protocol-driven training pipelines that ensure substantive validity thresholds and replicability standards. Finally, pioneering latent applications in counterfactual causal analysis to advance jurisprudential methodology.

Research Contributions

This study advances methodological innovation and theoretical development in seafarer mental health research through the application of the SHAP interpretability framework. The analysis reveals a complex, non-linear interplay between occupational conditions, living environments, and key psychological variables, characterized by distinct nonlinear relationships and context-dependent dynamics. This provides a novel theoretical perspective for understanding multifactorial interactions in mental risk formation. The analysis shows a significant interaction, where favorable environmental conditions are linked to a stronger association between positive psychological factors and better mental health. This insight addresses an empirical gap concerning person–environment interactions and offers substantive evidence for ecological model development.

The theoretical contributions are threefold. First, methodological integration: combining ML with conventional statistical techniques identifies complex interaction patterns that transcend the limitations of linear modeling. Second, contextual grounding: working conditions are established as central to mental health promotion, with environmental optimization demonstrating greater marginal utility than isolated psychological interventions. Third, predictive refinement: A Bayesian-optimized risk assessment framework is developed, achieving diagnostic accuracy superior to traditional scale-based evaluations through cross-validated modeling.

From a practical standpoint, the identified nonlinear interactions inform the design of precision intervention protocols. It is recommended that shipping companies implement “environment–individual dual-track strategies,” prioritizing infrastructural improvements before delivering tailored psychological services aligned with cohort-specific needs. Furthermore, the SHAP-based analytical framework demonstrates transferable utility for other high-risk occupational groups, offering a methodological advance for risk assessment. Together, these advances promote a paradigm shift from generalized interventions toward precision prevention in maritime mental health governance. To empirically address the gaps and leverage the opportunities outlined above, this study employs a survey-based methodology. Data collected from 500 seafarers are analyzed using a Random Forest model, whose predictions are subsequently interpreted through the SHAP framework to uncover the critical determinants and their complex interplay regarding seafarer mental health. The following section details the materials and methods of this approach. To ensure clarity between empirical observation and theoretical inference, this study presents descriptive results in Section 4 and reserves all interpretative discussion for Section 5, following the conventions of rigorous medical and social science reporting.

Materials and Methodology

Dataset Description

The data for this study were derived from a specialized questionnaire survey on seafarer mental health. The questionnaire was designed based on the existing literatures and the theoretical framework of occupational health in maritime settings.^3,14,17,18 Participation was entirely voluntary, and no incentives were provided. It encompassed 13 core variables (12 predictor variables and 1 target variable), covering 4 dimensions: demographic characteristics, occupational status, environmental perceptions, and psychological assessment (variable definitions are detailed in Supplemental Appendix A). It should be noted that this questionnaire was a self-developed tool constructed based on a comprehensive review of existing literatures on seafarer mental health and occupational stress.^3,14,17,18 All variables were measured using structured questions and Likert scales to ensure data standardization and comparability.²⁹

The survey obtained ethical approval from the Ethics Committee of Dalian Maritime University and was subsequently distributed through cooperating shipping companies and seafarer training institutions, which were selected based on established research partnerships and their willingness to facilitate broad, anonymous survey distribution to their seafarer cohorts, between June and September 2025, utilizing an anonymous online questionnaire format. Participants were fully informed about the study’s purpose and their rights, and they voluntarily completed the survey after providing informed consent. A total of 527 questionnaires were collected. After excluding invalid responses (eg, incomplete or patterned answers), 500 valid samples were retained, resulting in a valid response rate of 94.9%. The sample covered multiple shipboard departments, including the deck, engine, and catering departments, providing a certain degree of occupational representativeness. The design and reporting of this cross-sectional study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.³⁰

Data Pre-Processing

To meet the input requirements of machine learning algorithms, systematic numerical encoding was applied to the collected categorical variables. All ordinal categorical variables (eg, “Status of current work-life conditions,” “Presence of stress during onboard service”) were encoded using label encoding, sequentially mapped to integers from 0 to k-1 to preserve their inherent ordinal relationships. Nominal variables (eg, “Gender,” “Occupational position on board”) were also assigned integer codes. Given the insensitivity of tree-based algorithms like Random Forest to the encoding of nominal variables, this treatment does not affect model performance.

The target variable “Frequently experienced psychological state” is a 5-level ordinal variable, corresponding to different psychological levels from “Very Healthy, Optimistic” to “Very Bad, Frequently Depressed,” and was similarly encoded in an ordinal manner (0-4). All mapping relationships during encoding were recorded in a coding table to ensure reproducibility of the results.

Following encoding, the dataset was structured into a 500 × 12 feature matrix and a corresponding target vector. To further evaluate the model’s generalization capability, the data were split into a training set (n = 400) and a test set (n = 100) using stratified sampling at an 80%:20% ratio, ensuring that the distribution of each psychological state category in both sets remained consistent with the original dataset.³¹ Given the scale invariance of Random Forest, no further feature standardization was performed.³²

ML Model Selection

In ML practice, algorithm selection necessitates systematic evaluation across 3 critical dimensions: problem ontology, data characteristics, and research objectives.^29,33,34 For the psychological state prediction task addressed in this study, the selection of appropriate algorithms was guided by several specific considerations. First, composition of the dataset requires methods capable of effectively handling categorical and ordinal features while capturing potential nonlinear interactions. Second, further demands algorithms with inherent regularization properties to prevent overfitting and ensure generalization capability. Finally, the core research objective of achieving interpretability goes beyond mere predictive accuracy, requiring functional compatibility with explanation frameworks.

Synthesizing these requirements, Random Forest was employed as the primary classification algorithm. This ensemble method demonstrates strong alignment with both the data structure and research imperatives.³⁵ This ensemble method has been widely adopted in health and transportation safety research due to its proven ability to handle complex, real-world data while maintaining robust predictive performance. The selection is supported by 3 key advantages: (i) Intrinsic resilience to high-dimensional data and multicollinearity enables effective processing of all predictor variables; (ii) Distribution-free architecture captures complex nonlinear interactions without requiring prescriptive assumptions; (iii) Critically, ensemble diversification through feature randomization substantially mitigates overfitting risks. These characteristics prove particularly advantageous for medium-sized datasets, enhancing both model robustness and generalizability.^35,36 Consequently, the adoption of Random Forest represents a methodologically sound approach that has been successfully applied in diverse fields including occupational health and transportation safety.^37-39

The Random Forest algorithm implements bootstrap aggregation (Bagging) to construct robust ensembles from multiple decision trees.⁴⁰ Its nomenclature derives from 2 stochastic processes employed during model construction.

Step 1: Bootstrap Sampling

In the bootstrap sampling phase, multiple training subsets are created by randomly selecting n observations with replacement from the original N observations, typically with n = N. This process repeats k times to generate k distinct training subsets. Such a sampling regimen guarantees that every training subset encompasses only a fraction of the original data, while the remaining out-of-bag samples serve for internal validation and generalization error estimation.

Step 2: Feature Subspace Randomization

During the construction of each decision tree, stochasticity in feature selection is deliberately injected. For every training subset, the algorithm trains an autonomous decision tree, typically a Classification and Regression Tree (CART). Whenever a split is contemplated at any given node, the search for an optimal split is not conducted across the full set of M features. Instead, a small subset of m features (m ≪ M) is sampled at random from the entire pool, and the optimal feature within this restricted subset is selected to perform the split. The splitting process continues recursively until nodes become indivisible or predefined stopping criteria are met.

This dual randomization ensures low inter-tree correlation.⁴¹ For classification tasks, final predictions derive from majority voting across all trees in the forest, formally expressed as equation (1):

y = mode {y_{1}, y_{2}, \dots, y_{i}}

(1)

where $y$ denotes the predicted class determination, $mode {.}$ represents the statistical mode operation, $y_{i}$ signifies the predictive output of the i-th decision tree. This ensemble decision mechanism markedly fortifies the model’s stability and accuracy, effectively insulating it against the overfitting propensity of any solitary decision tree. Figure 1 depicts the schematic architecture of the Random Forest algorithm.

Figure 1.

Schematic diagram of the structure of a Random Forest.

Model Evaluation and Optimization

This study employed a standardized set of performance evaluation metrics to assess the predictive validity of the classification models and utilized advanced optimization algorithms to enhance model performance. Specifically, the metrics encompass Accuracy, Precision, Recall, and the F1-Score. Concurrently, a Bayesian optimization approach integrated with cross-validation was adopted to systematically identify the optimal model configuration.

The computations of these metrics are based on the 4 fundamental components derived from the confusion matrix. True Positives (TP) are the instances where the model correctly predicts positive cases as positive. True Negatives (TN) are the instances where the model correctly predicts negative cases as negative. False Positives (FP) are the instances where the model incorrectly predicts negative cases as positive. False Negatives (FN) are the instances where the model incorrectly predicts positive cases as negative.^42,43

Accuracy refers to the proportion of correctly predicted samples out of the total predictions. It measures the model’s overall capability to classify input data correctly.⁴³ The formula is given as equation (2):

Accuracy = \frac{TP + TN}{TP + TN + FP + FN}

(2)

Precision refers to the proportion of true positive cases among all samples predicted as positive. This metric quantifies the accuracy of the model when it predicts a sample as belonging to the positive class.³¹ The formula is given as equation (3):

Precision = \frac{TP}{TP + FP}

(3)

Recall denotes the proportion of true positives identified out of all actual positive samples. It measures the model’s ability to detect all genuine positive instances in the dataset.³¹ The formula is given as equation (4):

Recall = \frac{TP}{TP + FN}

(4)

F1-Scoreis defined as the harmonic mean of precision and recall. It integrates these 2 metrics into a single value ranging from 0 to 1, serving as a measure of the model’s balance between precision and recall.⁴³ The formula is expressed as equation (5):

F 1 - Score = 2 \times \frac{\Pr ecision \times Recall}{\Pr ecision + Recall}

(5)

To enhance the performance of the Random Forest model within a reasonable computational budget, this study systematically optimized its hyperparameters. Hyperparameters are parameters set prior to the training process, which directly influence the model’s learning behavior and predictive performance. This study focused on several key hyperparameters known to significantly affect model outcomes, including n_estimators, max_depth, min_samples_split, min_samples_leaf, max_features and bootstrap. Furthermore, to efficiently explore the optimal combination of these hyperparameters, a Bayesian optimization algorithm integrated with 5-fold cross-validation was employed.³¹ This approach conducts an intelligent search within predefined parameter spaces, maximizing the F1-score during cross-validation to effectively identify configurations that enhance the model’s generalization capability. The complete set of performance evaluation metrics, along with the hyperparameter search spaces and the ultimately selected optimal values, are systematically documented in Supplemental Appendix B.

Interpretable ML Methods

Although the Random Forest model demonstrated strong performance in predictive tasks, its inherent “black-box” nature hinders a deeper understanding of the underlying drivers influencing seafarers’ mental health.⁴⁴ To address this opacity and transform the model’s predictive capability into meaningful scientific insights, this study adopted the SHAP framework, a game theory-based post-hoc interpretation technique recognized for its robust theoretical foundations.⁴⁵ Due to its model-agnostic nature and strong theoretical guarantees, SHAP has been widely applied across various fields including clinical diagnosis, occupational health risk assessment, and transportation safety analysis.^46-48 A key advantage of SHAP lies in its ability not only to evaluate the global importance of features, as conventional methods do, but also to quantify the specific contribution of each feature to every individual prediction, thereby achieving theoretical consistency between local and global interpretability.⁴⁹

The theoretical foundation of SHAP lies in the Shapley value from cooperative game theory. It treats each predictive feature as a “player” and the output of a single model prediction as the total “payout” from a cooperative game.⁴⁸ This payout is then fairly distributed among all “players.”⁵⁰ For any given feature $i$ in the model, its SHAP value ϕ_i precisely quantifies the magnitude and direction of that feature’s contribution to a specific prediction.⁵¹ This value is determined by calculating the weighted average of the feature’s marginal contribution across all possible feature subsets, as defined by equation (6):

ϕ_{i} = \sum_{S \subseteq N ∖ {i}} \frac{| S |! (| N | - | S | - 1)!}{| N |!} [f (S \cup {i}) - f (S)]

(6)

where N denotes the complete set of features, $S$ represents an arbitrary feature subset excluding feature $i$ , $f (S)$ indicates the model’s prediction output when using only the feature subset $S$ . This formula ensures that the allocation of SHAP values is both equitable and unique, satisfying the 3 desirable properties of local accuracy, missingness, and consistency.⁵²

In this study, the SHAP interpretability framework serves 2 primary analytical objectives that collectively enhance understanding of seafarers’ mental health determinants. First, the analysis identifies key risk factors and quantifies their global importance through systematic computation of SHAP values across all instances in the dataset. The resulting feature importance plots visually rank features by their mean absolute SHAP values, clearly illustrating which factors exert the most substantial influence on seafarers’ psychological status. This approach provides robust evidence for understanding universal patterns in seafarers’ mental health at a macroscopic level, establishing a foundation for prioritizing intervention targets. Second, the investigation explores complex non-linear relationships and individual heterogeneity using SHAP dependence plots. These visualizations delineate how variations in specific feature values correspond to changes in SHAP values, thereby revealing non-linear patterns that conventional linear methods typically fail to capture. The color-coding scheme further enables detection of interaction effects between primary features and secondary variables, offering insights into the contextual nature of risk factors.

All SHAP analyses were implemented in Python 3.9.13 using the package. Through this comprehensive interpretability framework, this study transforms the ML model from a mere predictive tool into an advanced analytical instrument capable of yielding profound insights into the complex intrinsic mechanisms governing seafarers’ mental health.

Results and Analysis

Demographic and Occupational Characteristics of the Sample

This study included a total of 500 seafarers as valid samples. Supplemental Appendix C presents the frequency distribution of key demographic, occupational, and psychological variables among the study participants (N = 500). The sample is predominantly male and young, with senior maritime roles (eg, Chief Officer, Captain) representing the majority. Participants generally reported positive mental states and recognized lifestyle influences on mental health, while also reflecting variation in perceived stress, career-related concerns, and awareness of organizational mental health support. The distributions of demographic attributes, occupational profiles, and psychological perception variables are visually summarized in Figure C of Supplemental Appendix C. This section elaborates on key descriptive statistics to establish an empirical foundation for subsequent interpretable analyses.

The sample exhibited distinct demographic patterns characteristic of the maritime industry. As shown in Supplemental Figure C(a), gender distribution demonstrated marked skewness, with male seafarers comprising 99.2% (n = 496) of participants, which aligns with global maritime workforce demographics.⁵³ Regarding age structure, as shown in Supplemental Figure C (b), the sample displayed a relatively young profile, with over 80% (n = 435, 87.0%) of seafarers being aged 40 or below. Correspondingly, in terms of seafaring experience, as shown in Supplemental Figure C(c), more than half (n = 265, 53.0%) of the seafarers had less than 5 years of professional experience, indicating a substantial representation of individuals in the early stages of their careers. Recent research confirms that junior seafarers and those with shorter service experience significantly higher stress levels, highlighting the vulnerability of this subgroup.⁵⁴ This sample composition provides crucial context for examining mental health patterns in contemporary maritime populations.

Analysis of occupational characteristics revealed comprehensive coverage of major shipboard departments within the sample, as shown in Supplemental Figure C(d). The deck department (including captains and officers) constituted the largest proportion (n = 337, 67.4%), followed by the engine department (n = 153, 30.6%) ensuring representative inclusion of core maritime operational roles. Regarding welfare perceptions, as shown in Supplemental Figure C(k), “Evaluation of company welfare and benefits” indicated that most seafarers rated their level as “Average” (n = 216, 43.2%) or “Good” (n = 133, 26.6%). However, a notable concern emerged regarding awareness of mental health services: a significant subset (n = 156, 31.2%) reported being “Unfamiliar” or “Completely Unaware” of available mental health support services, as shown in Supplemental Figure C (l). This pattern suggests potential deficiencies in the dissemination and accessibility of mental health support systems within maritime organizational structures.

The core dependent variable “Frequently experienced psychological state” was numerically encoded as 5 classes for model input: 0 (Average, Relatively Stable), 1 (Not Good, Sometimes Poor), 2 (Generally Healthy, Overall Good), 3 (Very Bad, Frequently Depressed), and 4 (Very Healthy, Optimistic). As shown in Supplemental Figure C(f), the distribution reveals a concerning pattern with only approximately one-third (n = 164, 32.8%) of seafarers describing their mental condition as “Generally Healthy, Overall Good” (Class 2). In contrast, nearly one-fifth of respondents (n = 90,18.0%) reported psychological states categorized as “Not Good/Sometimes Poor” or “Very Bad/Frequently Depressed,” underscoring the critical need to address mental health challenges within this occupational group.

Further examination of psychological perceptions revealed significant contrasts. An overwhelming majority of seafarers (n = 441, 88.2%) acknowledged a “Very Strong” or “Quite Strong” link between lifestyle habits and mental health, as shown in Supplemental Figure C (g), while nearly 80% (n = 380, 76.0%) self-assessed their stress resilience as “Quite High” or “Very High,” as shown in Supplemental Figure C (i). These results indicate widespread health awareness and psychological resilience among seafarers. In contrast, high stress prevalence during service was ubiquitous, as shown in Supplemental Figure C (h), and career development concerns were pervasive, with 63.8% (n = 319) indicating seafarers feel worried “Sometimes” “Often” or “Always,” as shown in Supplemental Figure C (j). Compounding these challenges, a pronounced deficit in perceived societal recognition was evident, with 80.0% (n = 400) of respondents rating public recognition of their profession as “Moderate,” “Low,” or “Very Low.” Recent scholarship conceptualizes maritime suicide risk through a 5-factor matrix including the experience of being at sea, the distinct nature of maritime society, health systems, onboard community dynamics, and closed/distant relationships-all reflected in the psychosocial stressors identified in this sample.⁵⁵ These findings collectively depict a complex psychological environment where persistent professional stressors and limited social recognition potentially undermine seafarers’ mental well-being, despite the population’s demonstrated health awareness and self-assessed resilience.

In summary, the sample depicts a contemporary seafaring cohort characterized by youth dominance, high occupational mobility, and robust health consciousness, yet simultaneously experiences substantial occupational pressures, career uncertainties, and a profound sense of societal undervaluation. These attributes form a paradoxical combination, where high self-reported resilience coexists with acute stressors, establishing a unique context for understanding mental health outcomes in this population. Consequently, subsequent analyses will employ advanced ML methods to precisely identify and quantify the key determinants underlying psychological state variations within this complex occupational population.

Identification and Quantification of Key Determinants for Seafarer Mental Health

To identify the key factors influencing seafarers’ psychological states from a global perspective, this study employed the SHAP framework to calculate the mean absolute impact of each feature on the model’s predictions, as visualized in Figure 2. The plot ranks all predictive variables by their Mean |SHAP value|, where the total length of each bar represents the average magnitude of influence of that feature on the overall model predictions. The stacked colored segments within each bar further reveal the distribution of contributions across different psychological state categories (Class 0-4), providing nuanced insights into how each variable differentially affects various mental health outcomes.

Figure 2.

Global feature importance (mean |SHAP|).

The SHAP-based feature importance analysis identified “Status of current work-life conditions” as the predominant predictor of seafarers’ psychological states, achieving a mean absolute SHAP value of approximately 0.50. This value substantially exceeded all other features, with nearly twice the impact of the second-ranked factor, “Concern or uncertainty regarding seafaring career development,” demonstrating the decisive superiority of immediate environmental factors in mental health states. The maritime environment presents a uniquely confined setting where living spaces, working conditions, interpersonal dynamics, and recreational facilities collectively shape seafarers’ experiential reality. Consequently, the quality of current work-life conditions emerges as the primary determinant directly dictating psychological states. Systematic review evidence confirms that social isolation, separation from family, fatigue, stress, and long working hours are the main causes of mood disorders among seafarers, all of which are intrinsically linked to work-life conditions.⁵⁶ The dominant influence of work-life conditions supersedes personal characteristics, professional experience, and other individual factors, highlighting environmental intervention as the cornerstone of effective mental health management in maritime operations.

Further examination of the feature importance ranking reveals a clear hierarchical structure among mental health determinants. The most influential predictors coalesce into 3 primary dimensions: Work environment factors (Status of current work-life conditions), Career development factors (Concern or uncertainty regarding seafaring career development) and sociopsychological factors (Perceived societal recognition of the seafaring profession, Awareness of mental health services provided by the company). Notably, these high-impact variables constitute modifiable environmental and cognitive attributes, whereas traditional demographic characteristics, including Gender, Age, and Occupational position on board, ranked lowest with SHAP values approaching zero. Furthermore, the relative positioning of specific psychological factors reveals meaningful insights. The higher ranking of “Presence of stress during onboard service” (fifth) compared to “Self-assessed stress tolerance capacity” (eighth) indicates a fundamental psychological mechanism. Objective stressors exert greater predictive power than subjective resilience.⁵⁷ This insight informs stress management strategies, suggesting that reducing environmental stressors may prove more effective than exclusively focusing on enhancing individual coping capacities in mental health intervention strategies.

Examining the stacked color segments within feature bars elucidates differential drivers across psychological states. For the top predictor (“Status of current work-life conditions”), color distribution exhibited marked asymmetry. The orange segments (Class 3: Very Bad, Frequently Depressed) and the red segments (Class 4: Very Healthy, Optimistic) dominate the feature’s explanatory profile, indicating its pronounced role in discriminating between extreme psychological conditions. In contrast, the green (Class 2: Generally Healthy, Overall Good) and purple segments (Class 0: Average, Relatively Stable) contribute minimally to the feature’s predictive influence.

This polarized impact pattern signifies that work-life conditions disproportionately propel seafarers toward psychological extremes (optimal or critical states) rather than moderate conditions. Such findings carry vital preventive and interventional implications: improving work-life conditions not only mitigates severe mental health risks but also elevates well-being to optimal levels. Similarly, “Concern or uncertainty regarding seafaring career development” demonstrates comparable polarization, with orange and red segments again predominating, confirming career instability as another significant driver of psychological extremity. Conversely, “Perceived societal recognition of the seafaring profession” demonstrated balanced color distribution across classes, suggesting its function as a foundational psychological need whose absence uniformly impairs all mental states.⁵⁸

Interaction Effects of the Core Driver: Analyzing the Moderating Role of “Status of Current Work-Life Conditions”

To examine the moderating mechanisms through which the core determinant “Status of current work-life conditions” (identified as the most influential variable in Section 4.2) influences seafarer mental health, this study employs SHAP dependence plot analysis to systematically investigate its interaction effects with 4 critical covariates. Based on the established feature importance hierarchy from Section 4.2, this study examines its moderating effects on 4 carefully selected covariates that demonstrated significant predictive validity in the previous global feature analysis. These include “Concern or uncertainty regarding seafaring career development,” “Perceived societal recognition of the seafaring profession,” “Awareness of mental health services provided by the company” and “Presence of stress during onboard service.” This analytical approach enables the identification of complex nonlinear relationships and conditional dependency patterns within the seafarer mental health determinant framework, thereby providing empirical evidence for understanding the multifactorial mechanisms underlying mental health risk formation among seafarers.

Figure 3a illustrates the interaction pattern between “Status of current work-life conditions” and “Concern or uncertainty regarding seafaring career development.” The analysis results reveal a significant non-linear relationship between work-life conditions and their predictive contribution to mental health, which is further moderated by the level of career development concerns. In regions characterized by unfavorable work-life conditions (lower range of the X-axis), all observed SHAP values remain confined to the negative range, indicating consistently adverse impacts on mental health predictions regardless of career concern levels. During this stage, the limited dispersion of SHAP values across different levels of career concerns (color-coded from blue to red) demonstrates a relatively concentrated distribution pattern, suggesting a limited moderating effect. As work-life conditions improve to a moderate level (central range of the X-axis), the moderating effect of career development concerns becomes discernible. The color-coded distribution demonstrates significantly elevated SHAP values for individuals without career development concerns compared to those experiencing frequent concerns. Moreover, red points begin to exhibit positive SHAP values. This pattern indicates that seafarers with lower career development concerns experience more favorable mental health predictions when exposed to comparable working conditions. Under favorable work-life conditions (higher range of the X-axis), the moderating effect of career development concerns is maximized. The disparity in SHAP values between individuals without career concerns (red points) and those with frequent concerns (blue points) becomes substantially more pronounced under favorable work-life conditions, indicating that optimal work environments amplify the differential impact of career development factors on mental health outcomes.

Figure 3.

SHAP dependence plots for work-life conditions interactions.

Figure 3b presents the interaction pattern between “Status of current work-life conditions” and “Perceived societal recognition of the seafaring profession.” The analysis demonstrates that perceived societal recognition significantly moderates how work-life conditions influence mental health predictions. Within the domain of unfavorable work-life conditions (lower X-axis range), all observations yield negative SHAP values. However, the SHAP values for the red points (indicating high perceived societal recognition) are relatively higher than those for the blue points (indicating low perceived societal recognition). This pattern suggests that strong societal recognition can partially mitigate the detrimental effects of poor working environments. When work-life conditions improve to a moderate level (central range of the X-axis), a notable shift occurs in the distribution of color-coded points. High-recognition cases (red points) begin to demonstrate positive SHAP values earlier, while low-recognition cases (blue points) remain predominantly negative. This pattern indicates that perceived societal recognition influences the threshold at which work-life improvements translate into positive mental health contributions. Under favorable work-life conditions (higher range of the X-axis), high-recognition cases (red points) exhibit both significantly higher SHAP values and greater value dispersion compared to low-recognition cases (blue points). This demonstrates that elevated societal recognition amplifies the positive effects of satisfactory working conditions on mental health predictions. Systematic review evidence identifies poor social support as a significant risk factor for adverse mental health outcomes among seafarers, reinforcing the protective role of supportive onboard relationships-a pattern directly observed in the present study.⁵⁹

Figure 3c delineates the conditional moderating effects between “Status of current work-life conditions” and “Awareness of mental health services provided by the company.” The analysis reveals that mental health service awareness exerts context-dependent moderation on work-life conditions’ predictive contributions. Under adverse work-life conditions (left region of the X-axis), SHAP values display concentrated distributions across all awareness levels, primarily occupying negative intervals with minimal differentiation by color. This pattern indicates a limited moderating effect of service awareness in suboptimal environmental conditions. As environmental parameters improve (central transition region), chromatic stratification emerges. Red points (high familiarity with services) shift toward positive SHAP values, while blue points (null awareness) remain predominantly in negative or neutral territories. This divergence signifies the activation of awareness-based moderation mechanisms. Under favorable work-life conditions (right region), red points exhibit significantly elevated SHAP values with broader dispersion ranges compared to constrained distribution of unaware cases. This interaction pattern demonstrates that heightened service awareness amplifies the positive contribution of satisfactory working environments to mental health predictability.

Figure 3d delineates the most intricate moderating pattern between “Status of current work-life conditions” and “Presence of stress during onboard service.” The analysis reveals that onboard service stress levels significantly moderate how work-life conditions influence mental health predictions. Under suboptimal work-life conditions (left region of the X-axis), SHAP values for all stress levels remain predominantly negative, with color-coded distributions demonstrating high concentration. This pattern indicates limited moderating effects of stress factors in adverse environmental conditions. As work-life conditions improve (central region), stratification emerges in the color-coding distribution. Cases experiencing negligible stress (red points) transition toward positive SHAP values, while those reporting extreme stress (blue points) persist within negative domains. This divergence demonstrates stress-dependent variability in how environmental improvements affect mental health predictions. Under favorable work-life conditions (right region), low-stress cases (red points) exhibit substantially elevated SHAP values compared to high-stress counterparts (blue points). Notably, even under optimal conditions, cases with extreme stress (blue points) maintain depressed SHAP values, indicating persistent negative moderating effects of high stress levels on mental health predictions. These interpretations do not imply causal relationships; SHAP analyses only illustrate the predictive contributions of variables and their conditional dependency patterns. Accordingly, terms such as “amplifier effect,” “regulatory network,” or “risk formation mechanisms” should be understood in terms of associations and interactions within the model, rather than as evidence of causality.

Discussion and Implications

Seafarer mental health is shaped by multiple factors arising from the specific nature of maritime work and its pressures. While recent global crises have demonstrated this cohort’s remarkable adaptive resilience, persistent occupational stressors continue to exert profound physiological and psychological impacts despite shipping enterprises’ multilateral initiatives to mitigate crew-change challenges during the pandemic.^18,24 As the industry emerges from pandemic constraints, novel challenges stemming from geopolitical conflicts and supply chain volatility further compound these existing pressures. Enhancing sectoral resilience consequently necessitates establishing comprehensive seafarer health support infrastructures. The findings underscore that among multifaceted determinants, the work-life environment exerts a predominant influence, a result which resonates with core tenets of occupational health theories while shifting focus from static demographic factors to modifiable situational conditions.^60,61 The predominant influence of work-life conditions challenges the conventional emphasis on individual-level risk factors and suggests a paradigm shift toward environment-centered intervention strategies.^58,62 In practical terms, this means that shipping companies should invest first in improving the living and working environment on board, as such improvements are likely to yield the greatest marginal benefit for mental health outcomes.

The research on occupational mental health delivers significant theoretical and practical value.⁶³ This study, while providing theoretical models to understand the psychological development of seafarers, informs the development of early-warning systems for occupational psychological risks. Recent studies highlight that mental and physical health issues elevate maritime accident risk and impair cognitive function, while the prevalence of digital fatigue underscores the evolving nature of contemporary maritime work environments.^64,65 The SHAP analysis has the highest predictive contribution to seafarers’ mental health. Therefore, prioritizing onboard living and working conditions constitutes the primary evidence-based strategy for both preventing psychological distress and enhancing well-being. Regulators and industry bodies can standardize mental health metrics and incorporate psychosocial risk assessments related to work-life quality into routine safety audits. SHAP analyses also demonstrate that societal recognition and social support significantly influence seafarers’ mental health, providing an empirical basis for targeted interventions. Building on these findings, regulatory and rule-making frameworks, including Maritime Labour Convention, International Labour Organization Conventions, Standards of Training, Certification, and Watchkeeping (STCW), as well as IMO guidelines and recommendations, can guide improvements to onboard living and working conditions and support the implementation of measures aimed at enhancing these conditions. This highlights a necessary paradigm shift: from viewing well-being as an individual attribute to recognizing it as an outcome of modifiable environmental and organizational conditions. Ultimately, safeguarding seafarer mental health requires an integrated strategy combining evidence-based corporate practices, targeted clinical support, and proactive industry-wide governance. In practice, this can be summarized into 3 focused recommendations: (1) prioritizing improvements to onboard living and working conditions, (2) implementing targeted screening and support protocols based on key predictive variables, and (3) integrating psychosocial risk assessments into routine safety audits.

Conclusion

This study establishes an innovative analytical framework for seafarer mental health assessment by integrating Random Forest algorithms with SHAP explainability techniques. Through systematic analysis of a 500-seafarer cohort, this study successfully identifies and quantifies key determinants while revealing their hierarchical importance and complex interaction dynamics. The findings demonstrate that “Status of current work-life conditions” emerges as the most influential factor, substantially outweighing other variables in impact magnitude. This finding establishes clear intervention priorities and underscores the fundamental importance of environmental optimization in maritime mental health management.

This study pioneers SHAP-based interpretability techniques to deconstruct multifactorial complexity in seafarer psychology. By examining interaction effects between core drivers and key covariates, significant moderating effects emerge between work-life conditions and 4 critical factors: (i) career development concerns, (ii) perceived social recognition, (iii) mental health service awareness, and (iv) onboard pressure levels. These interactions exhibit marked nonlinear characteristics and context dependency, providing empirical validation for multifactorial patterns of risk association. A crucial insight reveals that favorable environmental conditions amplify positive psychological impacts, whereas adverse settings constrain the efficacy of moderating factors. This pattern underscores the necessity of implementing foundational environmental improvements prior to deploying targeted psychological interventions.

Notwithstanding these contributions, 4 methodological limitations merit consideration. First, the limited cohort size and pronounced gender imbalance may constrain the generalizability of the findings. In particular, this composition may limit the applicability of the results across different ranks (eg, senior officers), geographic regions, and vessel types (eg, cruise ships), highlighting a need for more stratified sampling in future research. Second, the cross-sectional design precludes the analysis of longitudinal trajectories and causal inferences. Prospective cohort tracking is needed to examine temporal dynamics and developmental pathways. Third, the reliance on questionnaire-based assessments introduces a risk of mono-method bias. Integrating multimodal data, such as physiological biomarkers, behavioral recordings, and organizational metrics, would enable a more comprehensive and objective evaluation of psychological well-being. Finally, the questionnaire employed in this study was a researcher-constructed instrument based on a review of the relevant literature. It did not undergo formal psychometric validation or pilot testing, which may introduce measurement error and risks of common method bias. Future research should conduct scale validation studies or employ established, validated scales for verification. Furthermore, this study did not conduct formal sample size estimation or power analysis. Consequently, the statistical power and generalizability of the findings may be constrained by the specific sample size and class distribution. Addressing these limitations would enhance both the ecological validity and the methodological rigor of future research on seafarer mental health.

Supplemental Material

sj-pdf-1-inq-10.1177_00469580261438708 – Supplemental material for Work-Life Conditions as the Primary Determinant of Seafarer Mental Health: An Explainable Machine Learning Analysis

Supplemental material, sj-pdf-1-inq-10.1177_00469580261438708 for Work-Life Conditions as the Primary Determinant of Seafarer Mental Health: An Explainable Machine Learning Analysis by Yongwei Jiang, Zhendong Tang, Hua Liu, Wenjie Cao, Zhiwei Zhao, Özkan Uğurlu and Xinjian Wang in INQUIRY: The Journal of Health Care Organization, Provision, and Financing

Supplemental Material

sj-pdf-2-inq-10.1177_00469580261438708 – Supplemental material for Work-Life Conditions as the Primary Determinant of Seafarer Mental Health: An Explainable Machine Learning Analysis

Supplemental material, sj-pdf-2-inq-10.1177_00469580261438708 for Work-Life Conditions as the Primary Determinant of Seafarer Mental Health: An Explainable Machine Learning Analysis by Yongwei Jiang, Zhendong Tang, Hua Liu, Wenjie Cao, Zhiwei Zhao, Özkan Uğurlu and Xinjian Wang in INQUIRY: The Journal of Health Care Organization, Provision, and Financing

Footnotes

ORCID iDs

Wenjie Cao

Zhiwei Zhao

Özkan Uğurlu

Xinjian Wang

Ethical Considerations

This study obtained ethical approval from the Ethics Committee of Dalian Maritime University (Approval No. DLMU-KJLL-2026-006, June 1, 2025). The research was conducted in accordance with the ethical standards of the institutional and national research committee.

Consent to Participate

Informed consent was obtained from all individual participants included in the study. Participants were informed about the purpose of the research, the voluntary nature of their participation, and the confidentiality of their responses.

Author Contributions

Yongwei Jiang: Conceptualization, Methodology, Formal analysis, Writing—original draft. Zhendong Tang: Formal analysis, Funding acquisition, Writing—review & editing. Hua Liu: Investigation, Data curation. Wenjie Cao: Methodology, Software, Validation. Zhiwei Zhao: Resources, Project administration. Özkan Uğurlu: Visualization, Writing—review & editing. Xinjian Wang: Conceptualization, Writing—review & editing, Supervision.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Data Availability Statement

Data will be made available on request.*

Supplemental Material

Supplemental material for this article is available online.

References

Zhou

Yuen

KF.

A systematic review on seafarer health: conditions, antecedents and interventions. Transp Policy. 2022;122:11-25. doi:10.1016/j.tranpol.2022.04.010

Cui

Liu

Wang

Yang

Fan

Shu

The impact of marine engine noise exposure on seafarer fatigue: a China case. Ocean Eng. 2022;266:112943. doi:10.1016/j.oceaneng.2022.112943

Zhao

Tang

, et al. Fatigue at sea during and after the COVID-19 pandemic: a comparative study of two matched samples of seafarers. Mar Pol. 2023;155:105730. doi:10.1016/j.marpol.2023.105730

Xiao

Zhang

Discussion on crew psychological quality and shipping safety. J Saf Sci Technol. 2011;7(6):184-189.

Yang

Chen

Nie

, et al. A survey on the international crew’s phychological health and the influence factors. Chin J Front Heal Quar. 2016;39(02):132-135. doi:10.16408/j.1004-9770.2016.02.015

Zhang

Gao

Zhou

, et al. Study on the overall mental health status of Chinese ocean-going seafarers. Navig China. 2005;(3):72-77. (in Chinese).

Shao

Liu

Qiu

The impact of occupational stress and environmental factors on seafarers’ mental health and analysis of countermeasures. China Maritime Saf. 2022;(8):28-30. (in Chinese).

Wang

Yuan

Fang

Zhang

Wang

A novel causal inference method of exit choice behaviour analysis for passenger ships during emergency evacuation. Reliab Eng Syst Saf. 2026;272:112489. doi:10.1016/j.ress.2026.112489

Shi

Dai

Chen

Yang

Preliminary development of the occupational stressor scale for Chinese seafarers. Chin Ment Heal J. 2017;31(11):902-906.

10.

Cui

Liu

Wang

, et al. The evaluation of seafarer fatigue as a performance-shaping factor in the maritime HRA method. ASCE-ASME J Risk Uncert Eng Syst Part A: Civil Eng. 2023;9(4):04023034. doi:10.1061/AJRUA6.RUENG-1092

11.

Sun

Gan

Chen

Gao

Sleep quality of ocean-going crew and its influencing factors. J Environ Occup Med. 2024;41(8):892-897. doi:10.11836/JEOM24099

12.

Hofman

Sharma

Watts

DJ.

Prediction and explanation in social systems. Science. 2017;355(6324):486-488. doi:10.1126/science.aal3856

13.

Jia

Xie

Lin

Seafarer fatigue and its influencing factors. Navig China. 2018;41(4):94-100.

14.

Svetina

Perkovič

Yang

, et al. Factors impacting seafarers’ mental health and career intentions. Inquiry. 2024;61:469580241229617. doi:10.1177/00469580241229617

15.

Liu

Yang

, et al. Neurophysiological mechanisms and interventions of subthreshold depression by integrating machine learning techniques. Adv Psychol Sci. 2025;33(6):887-904.

16.

Guidotti

Monreale

Ruggieri

Turini

Giannotti

Pedreschi

A survey of methods for explaining Black Box models. ACM Comput Surv. 2019;51(5):1-42. doi:10.1145/3236009

17.

Zhao

Tang

Fatigue during the COVID-19 pandemic: the experiences of Chinese seafarers. Mar Pol. 2023;153:105643. doi:10.1016/j.marpol.2023.105643

18.

Sivri

Uğurlu

Blanco-Davis

Şenbursa

Wang

Mapping mental health of seafarers post-COVID-19: a Gaussian graphical model of depression, anxiety, and maritime working conditions. Inquiry. 2025;62:469580251371386. doi:10.1177/00469580251371386

19.

Liu

Zhang

Mental health status of Chinese seafarers under 35 years old. J Phys Educ. 2002;9(5):56-58.

20.

Zhao

Wang

Cao

Maritime occupational accidents analysis: a data-driven Bayesian network approach. Ocean Coast Manag. 2025;269:107785. doi:10.1016/j.ocecoaman.2025.107785

21.

Zhang

Discussion on mental health work for seafarers. Navig China. 2001;(2):92-94. (in Chinese).

22.

Yang

Discussion on the relationship between seafarers’ psychological quality and navigation safety. Tianjin Navig. 2021;(3):53-55. (in Chinese).

23.

Survey on crew members’ mental health and recommendations for crew management. Mar Technol. 2014;(6):55-58. (in Chinese).

24.

Cao

Research on seafarers’ right to rest under the COVID-19 pandemic. Res Chin Maritime Law. 2022;33(3):25-32. doi:10.3969/j.issn.1003-7659.2022.03.003

25.

Nzenwata

Ilori

Tai-Ojuolape

, et al. Explainable AI: a systematic literature review focusing on healthcare. J Comput Sci Appl. 2024;12(1):10-16. doi:10.12691/jcsa-12-1-2

26.

Wei

Adaptive representational theory of artificial intelligence cognition. J Univ Electron Sci Technol China (Soc Sci Ed). 2025;27(02):1-19. doi:10.14071/j.1008-8105(2025)-1001

27.

Price

WN.

Big data and black-box medical algorithms. Sci Transl Med. 2018;10(471):eaao5333. doi:10.1126/scitranslmed.aao5333

28.

Peng

Guo

Quantitative methods in social work research and their applications. J Xi’an Jiaotong Univ (Soc Sci). 2022;42(1):46-57. doi:10.15896/j.xjtuskxb.202201006

29.

Jiang

Fan

Seafarer mental workload assessment using a hybrid deep learning model. TransNav Int J Mar Navig Saf Sea Transp. 2025;19(1):55-60. doi:10.12716/1001.19.01.07

30.

von Elm

Altman

Egger

Pocock

Gøtzsche

Vandenbroucke

JP.

The strengthening the reporting of observational studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol. 2008;61(4):344-349. doi:10.1016/j.jclinepi.2007.11.008

31.

Feng

Wang

Chen

, et al. Prediction of the severity of marine accidents using improved machine learning. Transp Res E Logist Transp Rev. 2024;188:103647. doi:10.1016/j.tre.2024.103647

32.

Leightley

Williamson

Korbacz

Curcin

Marshall

Greenberg

Effectiveness of AI-based interventions in workplace mental health: a systematic review and narrative synthesis. Br Med Bull. 2026;157(1):ldag007. doi:10.1093/bmb/ldag007

33.

Hasan

Sultana

Ahmed

Momen

Early detection of occupational stress: enhancing workplace safety with machine learning and large language models. PLoS One. 2025;20(6):e0323265. doi:10.1371/journal.pone.0323265

34.

Mammeri

Dehimi

Bencherif

Pezzimenti

Paths towards high perovskite solar cells stability using machine learning techniques. Sol Energy. 2023;249:651-660. doi:10.1016/j.solener.2022.12.002

35.

Mallala

Ahmed

AIU

Pamidi

Faruque

. Forecasting global sustainable energy from renewable sources using random forest algorithm. Results Eng. 2025;25:103789. doi:10.1016/j.rineng.2024.103789

36.

Adamopoulos

Valamontes

Tsirkas

Dounias

Predicting workplace hazard, stress and burnout among public health inspectors: an AI-driven analysis in the context of climate change. Eur J Investig Heal Psychol Educ. 2025;15(5):65. doi:10.3390/ejihpe15050065

37.

Kumar

Burns

GN.

Determinants of safety outcomes in organizations: Exploring O*NET data to predict occupational accident rates. Pers Psychol. 2024;77:555-594. doi:10.1111/peps.12560

38.

Berhanu

Schröder

Wodajo

Alemayehu

Machine learning for predictions of road traffic accidents and spatial network analysis for safe routing on accident and congestion-prone road networks. Results Eng. 2024;23:102737. doi:10.1016/j.rineng.2024.102737

39.

Georganos

Grippa

Niang Gadiaga

, et al. Geographical random forests: a spatial extension of the random forest algorithm to address spatial heterogeneity in remote sensing and population modelling. Geocarto Int. 2021;36(2):121-136. doi:10.1080/10106049.2019.1595177

40.

Breiman

Random forests. Mach Learn. 2001;45(1):5-32. doi:10.1023/A:1010933404324

41.

Fang

Xin

Zhan

Newton

MAA

Yan

Jin

Classification of wool and cashmere fiber based on LBP and DWT features: performance comparison of random forest, AdaBoost, and KNN classifiers. J Text Inst. 2025;116(11):2663-2675. doi:10.1080/00405000.2025.2450851

42.

Wang

Zhang

Feng

A novel feature engineering method for severity prediction of marine accidents. J Mar Eng Technol. 2025;1-16. doi:10.1080/20464177.2025.2469336

43.

Munim

Sørli

Kim

Alon

Predicting maritime accident risk using automated machine learning. Reliab Eng Syst Saf. 2024;248:110148. doi:10.1016/j.ress.2024.110148

44.

Shen

Wang

Yang

Interpretable model for rockburst intensity prediction based on Shapley values-based Optuna-random forest. Undergr Space. 2025;21:198-214. doi:10.1016/j.undsp.2024.09.002

45.

Popa

Buzea

Gurzu

, et al. An integrated AI framework for occupational health: predicting burnout, long COVID, and extended sick leave in healthcare workers. Healthcare. 2025;13(18):2266. doi:10.3390/healthcare13182266

46.

Tang

Liu

Cui

Explainable machine learning for stroke risk prediction: a comparative study with SHAP-based interpretation. Front Neurol. 2025;16:1716984. doi:10.3389/fneur.2025.1716984

47.

Wang

Chong

Rajendra

Skitmore

Fine-Kinney fuzzy-based occupational health risk assessment for workers in different construction trades. Autom Constr. 2024;168(A):105738. doi:10.1016/j.autcon.2024.105738

48.

Cao

Wang

Feng

Zhou

Yang

Improving maritime accident severity prediction accuracy: a holistic machine learning framework with data balancing and explainability techniques. Reliab Eng Syst Saf. 2026;266:111648. doi:10.1016/j.ress.2025.111648

49.

Liu

Wang

, et al. Investigating the causal effects of anthropogenic factors on urban streams and lakes water quality by integrating causal inference with interpretable machine learning. J Clean Prod. 2025;488:144626. doi:10.1016/j.jclepro.2024.144626

50.

Agheli

Aghabayk

Sadeghi

Das

E-scooter crash severity in the United Kingdom: A comparative analysis using machine learning techniques and random parameters logit with heterogeneity in means and variances. IATSS Res. 2025;49(2):155-168. doi:10.1016/j.iatssr.2025.03.004

51.

Cao

Wang

Shu

, et al. An integrated method of advanced optimisation and adaptive ensemble learning for ship fuel consumption prediction. Transp Res Part C: Emerg Technol. 2026;188:105659. doi:10.1016/j.trc.2026.105659

52.

Zhang

Lyu

Nazir

Mao

Wan

. A multi-source physiological data-driven method for evaluating the mental workload of operators in remote control environment. Saf Sci. 2025/03/26;187:106864. doi:10.1016/j.ssci.2025.106864

53.

Narayanan

Emad

Fei

Key factors impacting women seafarers’ participation in the evolving workplace: a qualitative exploration. Mar Pol. 2023;148:105407. doi:10.1016/j.marpol.2022.105407

54.

Strukcinskiene

Razbadauskas

Grigoliene

, et al. Occupational stressors and workplace challenges faced by seafarers in Lithuania: a cross-sectional study. Healthcare. 2025;13(11):1334. doi:10.3390/healthcare13111334

55.

Baumler

Carrera-Arce

Suicide at sea: navigating in restricted visibility. Marit Stud. 2025;24:42. doi:10.1007/s40152-025-00422-3

56.

Nittari

Gibelli

Bailo

Sirignano

Ricci

Factors affecting mental health of seafarers on board merchant ships: a systematic review. Rev Environ Health. 2024;39(1):151-160. doi:10.1515/reveh-2021-0070

57.

Johnsen

Hystad

SW.

Hardiness and mental health during naval deployment: the relation is mediated by social processes and not by self-regulatory processes. Eur Rev Appl Psychol. 2023;73(3):100817. doi:10.1016/j.erap.2022.100817

58.

Seah

Wang

Yuen

KF.

Investigating the role of sociotechnical factors on seafarers’ psychological capital and mental well-being. Technol Soc. 2022;71:102138. doi:10.1016/j.techsoc.2022.102138

59.

Brooks

Greenberg

Mental health and psychological wellbeing of maritime personnel: a systematic review. BMC Psychol. 2022;10:139. doi:10.1186/s40359-022-00850-4

60.

Tong

Exploring the interplay between teachers’ emotions, personal traits, environmental factors and psychological well-being. Eur J Educ. 2025;60(1):e12903. doi:10.1111/ejed.12903

61.

Pronk

Whitsel

Ablah

Anderson

Imboden

Building a culture of healthy living in the workplace. Prog Cardiovasc Dis. 2025;90:38-44. doi:10.1016/j.pcad.2025.02.006

62.

Arican

Ünal

AU.

The effect of working conditions of sefarers working in internal waters on work performance as a result of leading the work. Int J Soc Humanit Sci Res. 2023;10(98):2057-2068. doi:10.5281/zenodo.8306949

63.

Fan

A review of research on mental health issues in occupational groups. Psychol Mon. 2025;20(1):229-232.

64.

Uğurlu

Sana

Uğurlu

Şenbursa

. Analysis of the impact of officers of the Watch’s mental and physical health issues on collision and grounding accidents. INQUIRY. 2025;62:469580251371875. doi:10.1177/00469580251371875

65.

Arıcan

OH.

The effects of digital fatigue on mental health in seafarers: Delphi and multiple regression analysis. Turk J Marit Mar Sci. 2025;11(1):46-67. doi:10.52998/trjmms.1566728