Neutrosophic Logic and Machine Learning for Sustainable Employment Retention of Workers with Disabilities

Employment retention among persons with disabilities is not only a labour-market outcome; it is also a sustainability issue that links decent work, reduced inequality, household resilience, social participation, and the long-term effectiveness of welfare systems. Entry into employment is important, but sustainable development requires more than initial placement. It requires the capacity to remain employed under changing health conditions, organisational demands, and family circumstances. When employment is unstable, the consequences extend beyond wages. Job disruption can weaken self-sufficiency, limit social networks, increase dependence on public support, and reinforce social exclusion. For these reasons, employment retention has become an essential dimension of inclusion-oriented policy design [1, 2].

Workers with disabilities continue to encounter multiple barriers across the employment cycle. Studies have described employer concerns, inaccessible work environments, limited accommodation, discriminatory attitudes, and restricted promotion opportunities as recurring obstacles to labour-market participation [3, 4]. These barriers do not end after a worker is hired. In many cases, they become more visible during employment because job retention depends on whether the workplace can adapt to fluctuating health, rehabilitation needs, assistive technology requirements, and changes in role expectations. Retention is therefore shaped by a combination of individual capability, accumulated human capital, and the capacity of institutions and employers to provide flexible support.

Earlier studies on disability employment have shown that demographic factors such as age, education, income, and health status are closely associated with labour-market outcomes [5, 6]. Workplace factors also matter. Wage level, enterprise size, discrimination during recruitment or after hiring, and job satisfaction influence whether employees remain in a job or leave involuntarily [7, 8]. Yet the determinants of retention rarely operate in a simple linear fashion. Education can buffer the effect of disability severity; health can interact with age; and the implications of household economic status can differ according to whether a worker has access to services, transport, or rehabilitation resources. A methodological challenge therefore arises: the most policy-relevant signals may be hidden within nonlinear interactions and uncertain measurement structures.

Traditional regression-based analyses remain valuable, but they can understate interactions in high-dimensional labour datasets. In longitudinal disability research, this becomes especially problematic because employment is affected by dynamic changes in health, household resources, and workplace environments over time [9, 10]. Machine-learning approaches are increasingly used to address this complexity because they can detect nonlinear relationships and rank predictor importance without requiring all interactions to be specified in advance [11, 12]. Among these approaches, XGBoost has shown robust performance in classification tasks involving many candidate predictors and possible overfitting [13, 14]. When combined with regularization, it becomes particularly useful for policy-oriented modelling in which predictive strength must be balanced against stability and interpretability.

At the same time, even strong predictive models leave an important question unresolved: how should uncertainty itself be interpreted? Disability-employment data frequently contain ambiguity arising from subjective self-report, changing work conditions, and heterogeneous disability categories. Neutrosophic logic offers a useful conceptual response because it extends binary and fuzzy reasoning by simultaneously representing truth, falsity, and indeterminacy [15-19]. In the context of employment retention, this means that a factor can support retention for some workers, weaken it for others, and remain ambiguous in still other contexts. A modelling framework that acknowledges this coexistence is especially valuable for social and public-health applications where crisp classification alone may conceal meaningful uncertainty.

The present study was designed to integrate these two perspectives. Using longitudinal panel data on Korean workers with disabilities, it combines L2-regularized XGBoost for feature screening, logistic regression for interpretable effect estimation, and neutrosophic logic for uncertainty-sensitive interpretation. The analytical workflow is summarized in Figure 1. This design allows the study to answer three linked questions. First, which factors most strongly predict employment retention among workers with disabilities? Second, does L2-regularized XGBoost perform better than other common prediction models in this domain? Third, how can neutrosophic reasoning improve the interpretation of disability-employment dynamics by distinguishing supportive evidence from contradiction and indeterminacy?

1.png

Figure 1. Analytical workflow integrating panel data, machine learning, neutrosophic logic, and policy translation

This study makes three contributions. Empirically, it identifies a compact set of six high-value predictors from 75 variables. Methodologically, it demonstrates the complementary value of regularized boosting, conventional regression, and neutrosophic interpretation within one framework. Practically, it translates the results into policy directions for education, health support, income security, workplace accommodation, and age-sensitive retention services.

Compared with earlier Korean disability-employment studies that primarily examined isolated workplace facilities, job satisfaction, or single-equation panel associations, the present study advanced the literature in four specific ways: it focused on the coherent 2016-2020 PSED follow-up window, screened a broader set of 75 predictors, compared multiple classification algorithms under grouped validation, and translated the strongest empirical signals into a neutrosophic truth-indeterminacy-falsity framework for policy interpretation.

3.1 Study design and data source

This study used a longitudinal secondary-analysis design to investigate employment retention among working-age persons with disabilities. Longitudinal analysis is well suited to retention research because it allows employment outcomes to be understood as sustained processes rather than one-time events. The analytic workflow is summarized in Figure 1. The process began with a broad candidate feature pool, followed by predictive modelling with L2-regularized XGBoost, comparative evaluation against alternative classifiers, and interpretive modelling with logistic regression. A neutrosophic layer was then applied to translate the findings into an uncertainty-sensitive conceptual framework.

The data source was the second-wave Panel Survey of Employment for the Disabled (PSED), a nationally organised Korean panel administered by the Korea Employment Agency for Persons with Disabilities and the Employment Development Institute [10]. The analysis used the 2016-2020 waves because they provide a coherent longitudinal cohort with harmonised variables for employment status, disability characteristics, health, household resources, and work environment. The source wave counts were 4,577 respondents in 2016, 4,214 in 2017, 4,104 in 2018, 3,995 in 2019, and 3,907 in 2020. After retaining respondents who had follow-up information through the 2020 wave and imputing non-outcome predictor missingness within the training folds, the final respondent-level analytic cohort comprised 3,907 individuals. The analysis was therefore based on one analytic record per respondent and a final binary retention outcome, rather than an ungrouped pooled person-year file.

The inclusion and exclusion process was clarified as follows. First, respondents had to be registered persons with disabilities in the PSED cohort and within the working-age range covered by the survey. Second, respondents had to have usable employment-status information for constructing the retention outcome. Third, observations with missing outcome information at the follow-up endpoint were excluded, whereas missing values in candidate predictors were handled by fold-specific imputation rather than by deleting additional respondents. Fourth, when the machine-learning models were trained and tested, all records associated with the same respondent identifier were assigned to the same data split. This grouped split prevented information from the same individual from appearing simultaneously in training and testing data, thereby reducing longitudinal data leakage.

3.2 Outcome and predictor measurement

The primary outcome was employment retention, operationalised as a respondent-level final binary outcome. Respondents who maintained employment through the follow-up endpoint were coded as retained employment (=1), whereas those who reported resignation, non-retention, or transition out of the observed job during the follow-up window were coded as non-retention (=0). The classification task was therefore closer to a baseline-to-follow-up retention prediction than to a repeated person-year hazard model. This definition matched the objective of identifying workers at elevated risk of not sustaining employment and avoided treating repeated waves from the same individual as independent observations.

A total of 75 candidate predictors were screened. These variables were drawn from three domains: demographic structure, human capital, and the employment environment. Demographic variables included sex, age, household economic level, subjective health status, disability type, and disability severity. Human-capital variables included years of schooling, highest degree obtained, vocational certification status, and recent exposure to employment services. Employment-environment variables included wage level, firm size, job satisfaction, and experiences of discrimination during job search and in the workplace. Continuous variables with marked skewness, such as wage and firm size, were log transformed, categorical predictors were one-hot encoded, and predictor missingness was handled by median imputation for continuous variables and mode imputation for categorical variables within the training data. The same imputation parameters learned from the training set were then applied to the validation and test sets to prevent information leakage. Table 1 summarises the major variable domains and the operational logic used in the analysis.

Table 1. Candidate predictor domains and operational definitions

3.3 Neutrosophic representation of uncertainty

Neutrosophic logic was introduced as an interpretive framework rather than as a standalone classifier. In this study, the employment-retention state associated with a predictor profile x was represented conceptually as a neutrosophic triplet,

$E(x)=<T(x), I(x), F(x)>$   (1)

where, T(x) denotes the degree to which the available evidence supports retention, I(x) denotes indeterminacy arising from ambiguity, heterogeneity, or measurement uncertainty, and F(x) denotes the degree to which the evidence supports non-retention. The three components were not constrained to collapse into a single crisp value. This is important in disability-employment research because the same measured predictor may imply different mechanisms across disability types, age groups, and workplace settings. This tripartite logic is later used to interpret the six strongest predictors identified by the machine-learning model.

In practical terms, the neutrosophic layer served three functions. First, it prevented the interpretation from reducing all predictors to a single positive or negative score. Second, it highlighted where contextual ambiguity is likely to remain even after statistical modelling, such as when subjective health or disability type captures only part of the lived employment burden. Third, it made the policy discussion more realistic by distinguishing robust signals from areas that require flexible or individualised intervention. The objective was therefore not to replace statistical inference, but to supplement it with an explicit vocabulary of uncertainty.

3.4 L2-regularized XGBoost and statistical validation

To model complex relationships among the 75 candidate predictors, the study employed an explicitly tuned L2-regularized Extreme Gradient Boosting (XGBoost) classifier. XGBoost already contains regularization terms in its objective function; therefore, the term L2-regularized XGBoost in this paper referred to a model in which the L2 penalty parameter lambda and tree-complexity controls were deliberately tuned and interpreted as central design choices. By contrast, the conventional XGBoost comparator used the same boosting framework but retained default or minimally tuned regularization settings. This distinction was added to avoid implying that regularization was absent from ordinary XGBoost.

The objective function of XGBoost can be written as

$L(p h i)=\sum_i l\left(y_i, \hat{\mathrm{y}}_j\right)+\sum_k \Omega\left(f_k\right)$    (2)

where, l denotes the loss function, f_k is the k-th tree in the ensemble, and Ω(f_k) is the regularization penalty. For a given tree with T leaves and leaf weights w_j, the penalty term is defined as

$\Omega(f)=\gamma T+(\lambda / 2) \Sigma_{j=1}^T w_j^2$    (3)

where, γ penalises the number of leaves and λ is the L2 regularization parameter. During each boosting iteration, a new tree is added on the basis of first- and second-order derivatives of the loss function. Using the standard second-order approximation, the optimal leaf weight can be estimated by

$w *_j=-\left(\Sigma_{i \in I_j} g_i\right) /\left(\Sigma_{i \in I_j} h_i+\lambda\right)$    (4)

where, I_j denotes the set of observations assigned to leaf j, and g_i and h_i are the first and second derivatives of the loss with respect to the current prediction. This formulation makes XGBoost computationally efficient while also enabling regularized learning. The practical modelling sequence used in the study was initial prediction setting, gradient and Hessian calculation, tree growth based on regularized gain, leaf-weight estimation, and iterative prediction update using a learning rate. By combining shrinkage and regularization, the model can detect strong predictors while limiting the risk that the final ranking reflects noise.

Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC. The data were divided into an 80% training set and a 20% test set using outcome stratification and respondent-identifier grouping. Hyperparameter selection used five-fold grouped cross-validation with random seed 2026. The tuning grid was n_estimators {100, 200, 300}, max_depth {3, 4, 5}, learning_rate {0.03, 0.05, 0.10}, subsample and colsample_bytree {0.70, 0.80, 1.00}, gamma {0, 0.10, 0.30}, min_child_weight {1, 3, 5}, and reg_lambda {1, 3, 5, 10}. The final model used n_estimators = 300, max_depth = 3, learning_rate = 0.05, subsample = 0.80, colsample_bytree = 0.80, gamma = 0.10, min_child_weight = 1, reg_lambda = 5, reg_alpha = 0, objective = 'binary:logistic', and eval_metric = 'auc'. The conventional XGBoost comparator used the same split and metrics but retained the default regularization emphasis, including reg_lambda = 1 and gamma = 0.

After machine-learning feature screening, a multiple logistic regression model was fitted using the six variables identified by L2-regularized XGBoost. Feature importance was defined as normalized mean gain averaged across the grouped cross-validation folds, rather than as simple split frequency or raw coefficient size. Logistic regression then served as a transparent complementary model because it provided effect direction, odds ratios, and 95% confidence intervals for the selected predictors.

Because the study used de-identified secondary national data, institutional review board approval was exempted under the applicable regulations for anonymised secondary analysis. In the original panel, informed consent had been obtained from participants, and the data collection procedures were reported as consistent with the ethical principles of the Helsinki Declaration and subsequent revisions. The present analysis used only anonymised records and did not involve contact with human participants.

This study set out to improve the analysis of employment retention among workers with disabilities by combining predictive modelling and uncertainty-sensitive interpretation. The findings show that retention is shaped by a compact but multidimensional set of variables: age, household economic level, subjective health status, degree of disability, type of disability, and education level. The convergence between L2-regularized XGBoost and logistic regression strengthens confidence in these results. The machine-learning model efficiently screened a large predictor set and identified the most informative variables, while logistic regression clarified the direction and magnitude of their effects. This dual strategy is particularly useful for social-policy research because it balances predictive performance with interpretability.

The discussion explicitly compared these results with recent PSED-based and disability-employment studies. Prior Korean studies have shown that workplace facilities, job satisfaction, work environment, and health-related distress are closely related to retention-oriented outcomes among workers with disabilities [5, 6, 9]. The present findings were consistent with those studies but extended them by treating retention as a prediction-and-interpretation problem rather than as a single regression association.

One of the clearest substantive findings is the central importance of education. In the variable-importance ranking, education remained among the top six predictors, and in the logistic regression it emerged as the strongest positive correlate of retention. This result aligns with prior work emphasising the role of education in employability and sustainable labour outcomes for persons with disabilities [25-27]. Education likely operates through several pathways. It can increase access to jobs with more stable contracts and lower physical strain, improve problem-solving and communication in the workplace, and make occupational mobility more feasible when one job becomes unsuitable. In addition, educational attainment may reduce employer uncertainty by providing a visible signal of competence. From a policy perspective, this suggests that educational inequality remains a retention issue well beyond school age. Adult education, continuing vocational training, digital-skills programmes, and supported transition pathways may all contribute to stronger retention among disabled workers.

The strong effects of household economic level and subjective health status highlight the need to view employment retention as embedded within everyday life conditions rather than as a narrowly occupational phenomenon. Good household economic status was associated with higher odds of remaining employed, and better subjective health showed a similarly positive pattern. These findings are consistent with broader evidence linking economic precarity and poor health to unstable labour-market attachment [6, 28]. In practical terms, a worker's ability to maintain employment often depends on resources outside the formal workplace: the ability to pay for transport, receive timely treatment, replace assistive equipment, or absorb short-term disruptions without leaving the labour force. Health status also reflects more than biomedical function. It includes energy, pain, mental resilience, and perceived capability in daily work routines. The implication is that employment policy for disabled populations should be coordinated with public-health support, rehabilitation, transport assistance, and income-protection measures.

The negative effect of age deserves particular emphasis. Age was the most influential predictor in the L2-regularized XGBoost ranking and the only variable with an odds ratio below one in the regression model. This indicates that older disabled workers face a distinctive retention challenge. The result supports earlier literature showing that age-related discrimination, changing functional capacity, and the mismatch between work demands and adaptation opportunities may undermine sustained employment [4, 29]. A sustainable response requires age-sensitive retention strategies, including ergonomic redesign, phased task adjustment, targeted reskilling, and more proactive accommodation review as workers age.

Disability severity and disability type remained significant even after adjustment for education, age, household economy, and health. This finding should not be interpreted as a justification for deficit-based thinking. Rather, it indicates that workplaces and labour systems still provide uneven levels of accessibility. In an inclusive labour market, disability category would be less predictive of exit because support, technology, and organisational flexibility would compensate for functional mismatch more effectively. The observed effects therefore reveal not only the circumstances of workers, but also the limitations of institutions. They remind policymakers that retention problems cannot be solved entirely by improving individual skills. Environmental adaptation, universal design, anti-discrimination enforcement, and workplace accommodation are equally essential.

A major contribution of the study lies in its methodological integration of neutrosophic logic with machine learning. Many labour studies acknowledge uncertainty rhetorically but do not represent it analytically. The neutrosophic framework used here makes a practical distinction among supportive evidence, contradictory evidence, and indeterminacy. This matters because disability-employment variables are rarely perfect measures of lived reality. For example, subjective health can signal functional capacity, but it can also be shaped by mood, expectations, work stress, or comparison standards. Disability severity classifications may capture legal or administrative categories without fully representing the environmental burden of inaccessible workplaces. By making indeterminacy visible, the neutrosophic approach helps avoid overconfident policy interpretation.

This framework also improves the translation of model findings into intervention design. When a factor is strongly aligned with truth, such as education in the present analysis, broad policy expansion may be justified. When a factor contains more indeterminacy, such as disability type or self-rated health, intervention may need to be more tailored. Figure 6 summarises this translation pathway from predictor detection to sustainable labour-market inclusion. The figure emphasises that retention is most likely to improve when individual predictors are linked to concrete intervention levers such as training, accommodation, income support, job matching, and anti-discrimination systems.

The stronger predictive performance of L2-regularized XGBoost compared with CART, GBM, and conventional XGBoost indicated that explicit control of model complexity improved generalization in high-dimensional disability-employment data. This interpretation was consistent with the broader methodological literature on gradient boosting and hyperparameter control [13]. Substantively, the emphasis on education, health, workplace support, and disability-related heterogeneity was also consistent with recent evidence on long-term job retention and workplace conditions among people with disabilities [14, 22, 23, 26]. The added value of the present study was that these factors were not discussed as isolated correlates; instead, they were organised into a combined model-validation and neutrosophic interpretation framework.

6.png

Figure 6. Policy pathway from predictor detection to sustainable labour-market inclusion

Several practical implications follow from the findings. First, retention policy should not be limited to job placement. Programmes should support continuation in employment through periodic assessment, skill upgrading, accommodation review, and follow-up services after hiring. Second, education and training policy should be integrated into disability employment strategy across the life course, including adult re-skilling and digital adaptation. Third, health and rehabilitation services should be linked more directly with employment support because subjective health is a major determinant of labour-market continuity. Fourth, income insecurity should be treated as a retention risk factor, not simply a welfare outcome. Financial support that stabilises transport, treatment, or caregiving may indirectly protect employment. Fifth, older disabled workers require targeted support, including flexible work design and age-specific employer engagement.

The study has limitations. First, the analysis relied partly on self-reported measures, which may introduce reporting bias. Second, although the data were longitudinal, the reported summary results do not fully exploit all possible dynamic modelling strategies, such as time-varying survival analysis or recurrent-event approaches. Third, the panel is based on registered disability categories and may not capture all forms of functional limitation or intersectional disadvantage. Fourth, the neutrosophic layer in this study was interpretive rather than a fully parameterised neutrosophic estimation procedure. This was a deliberate design choice to preserve transparency, but future work could extend the approach by constructing explicit neutrosophic scores or hybrid classifiers. Fifth, broader macroeconomic shocks and policy changes were not modelled directly, even though they may influence retention conditions.

A further limitation was that the analysis used a respondent-level final retention outcome rather than a full time-to-event model. This choice improved comparability across classifiers and reduced leakage risk, but it did not estimate the precise timing of employment exit. Future work should therefore test whether the same predictors remain stable when analysed with grouped survival models, recurrent-event models, or dynamic panel approaches.

Future research should combine regularized boosting with time-to-event designs, test subgroup differences across industries and regions, and develop operational neutrosophic indicators that can be embedded in disability-employment decision support systems without obscuring interpretability.