Chronic diseases-including hypertension, diabetes, dyslipidemia, and obesity-constitute a major public health burden worldwide. Most of these conditions are noncommunicable diseases (NCDs), which frequently coexist and interact with one another [1]. Although diverse in etiology, many remain asymptomatic for long periods, leading to prolonged treatment and a low probability of complete recovery.

In Korea, 86.1% of older adults report having at least one chronic disease lasting more than three months, and 63.9% are multimorbid, with an average of 2.2 concurrent conditions [2]. Chronic diseases such as malignant neoplasms, cardiovascular disease, cerebrovascular disease, and diabetes remain the leading causes of mortality. Furthermore, the high prevalence of dyslipidemia (32.0%), hypertension (23.4%), and diabetes (10.2%) among adults younger than 49 years indicates that the burden extends across all age groups [3]. Healthcare expenditures related to chronic diseases have risen sharply, from 71 trillion KRW in 2020 to 90 trillion KRW in 2023, accounting for 84.5% of total medical costs and imposing a substantial socioeconomic burden [4].

Multimorbidity, defined as the presence of two or more chronic conditions irrespective of type [5], is associated with increased preventable complications, hospitalizations, and emergency department visits [6]. Polypharmacy further elevates the risk of adverse drug reactions, accelerates functional decline, and increases healthcare costs. Collectively, these consequences limit daily functioning, exacerbate physical, psychological, and socioeconomic challenges, and ultimately reduce quality of life (QOL) [7-9]. QOL refers to an individual's subjective perception of life in the context of cultural and personal values, encompassing goals, expectations, and concerns [1]. It is shaped by physical health, psychological well-being, social relationships, and environmental factors [10]. Health related quality of life (HRQoL) is a more specific construct reflecting how disease and treatment experiences influence physical, psychological, and social functioning [11].

Andersen's behavioral model (1997) provides a comprehensive framework for explaining HRQoL by categorizing determinants into predisposing, enabling, and need factors [12]. Predisposing factors include inherent individual characteristics such as age, sex, and educational level; enabling factors represent socioeconomic conditions such as income, insurance coverage, and access to resources; and need factors reflect illness severity, functional decline, and subjective health perceptions that directly drive healthcare utilization. Prior Korean studies have identified sociodemographic, socioeconomic, behavioral, and psychological variables associated with HRQoL [13-16]. However, most were limited to older adults or specific subgroups (e.g., older women's leisure activities, nutrition counseling), thereby restricting generalizability [16,17]. Even studies including younger and middle-aged adults were hampered by classification errors due to the absence of medical records and insufficient consideration of interactions among multiple conditions [18]. These limitations highlight the need for innovative approaches that incorporate multidimensional and interactional effects.

Machine learning (ML) has emerged as a powerful tool for identifying patterns and enhancing prediction in healthcare research [19]. Among these methods, Random forest (RF) reduces overfitting, improves predictive accuracy, and provides estimates of variable importance for interpreting influential factors [19,20]. By combining bootstrap aggregation with random feature selection, RF enhances model stability and generalizability through repeated resampling [21,22]. In Korea, HRQoL studies employing RF have been limited by narrow variable selection and heterogeneous sample distributions [23]. Importantly, grounding ML analyses in Andersen's behavioral model allows variables to be systematically structured and predictive outcomes to be aligned with a well-established theoretical framework. This approach strengthens interpretability, as determinants of HRQoL can be explicitly mapped onto predisposing, enabling, and need factors. It also addresses the frequent criticism of ML as a "black box," wherein models may achieve strong predictive performance but provide little theoretical rationale or mechanistic insight. By anchoring both variable selection and interpretation within Andersen's model, this study mitigates such concerns, ensuring that predictive outcomes remain theoretically coherent and clinically meaningful.

Taken together, integrating the predictive strengths of RF with the explanatory framework of Andersen's model offers methodological rigor and clinical relevance. This combination enhances predictive power, supports evidence-based decision-making, and generates actionable insights for psychiatric nursing practice.

The Korean National Health and Nutrition Examination Survey (KNHANES) provides nationally representative data on health, lifestyle, and nutrition, encompassing variables that reflect Andersen's behavioral model. The 8th wave (2019~2020) includes updated demographic and health information, making it suitable for analyzing HRQoL in multimorbid populations. Therefore, this study applied RF to KNHANES data to identify key predictors of HRQoL within Andersen's framework and to construct a predictive model. Findings are expected to guide psychiatric nursing practice by enabling early identification of high-risk groups, informing tailored interventions, and supporting evidence-based resource allocation to improve HRQoL among patients with multimorbidity.

This study analyzed data from the 8th KNHANES (2019~2020) to identify factors influencing HRQoL among patients with multimorbidity, guided by Andersen's behavioral model. Machine learning techniques (RF and SHAP) were applied to identify and interpret key predictors. The discussion addresses interpretation of findings, comparison with previous research, integration with Andersen's model, implications for psychiatric nursing, strengths and limitations of the study, and directions for future research.

The participants were predominantly older adults (mean age over 65 years), with low educational attainment and limited participation in economic activity. Many perceived their health negatively, and a substantial proportion reported obesity, high stress, and depressive symptoms. Risky health behaviors such as smoking and drinking were frequent, whereas engagement in moderate or vigorous physical activity was limited. These findings suggest that patients with multimorbidity constitute a vulnerable population characterized by aging, socioeconomic disadvantage, and psychological burden.

The predictive performance of the model was modest. Although prediction errors were relatively low, the explanatory power was limited, indicating that the model accounted for only a small proportion of variance in HRQoL. The discrepancy between the training dataset and the test dataset further suggests that the model's generalizability was restricted and that its explanatory capacity diminished in unseen data. This pattern highlights the inherent complexity of HRQoL, which is influenced by multifaceted biological, psychological, and social factors not fully captured in the available dataset. Accordingly, the results should be interpreted conservatively, and future research should incorporate additional psychosocial and service-related variables (e.g., social support, resilience, and healthcare quality) while also evaluating the model's transportability across cohorts and periods.

The analysis revealed that among need factors, poor self-rated health status was the most critical predictor of lower HRQoL. Health behavior factors, such as the number of days walked per week, also contributed positively, indicating that greater physical activity was associated with better HRQoL. In terms of predisposing factors, younger age was linked to higher HRQoL, underscoring the vulnerability of older adults. In addition, enabling factors such as higher monthly income and private health insurance demonstrated protective effects. These findings align with Andersen's theoretical assumption that enabling resources may buffer the adverse effects of health needs.

The present results are consistent with prior literature. The central role of self-rated health in determining quality of life has been repeatedly reported in both Korean and international studies [8,13]. The protective role of socioeconomic resources, such as income and private health insurance, also parallels the findings of Hong (2022). Unlike prior research that often focused on older adults or specific subgroups (e.g., women's leisure activities, nutritional counseling)[17,22], this study encompassed a broader age range and systematically applied Andersen's model to organize predictors. Furthermore, the application of machine learning (RF and SHAP) allowed the detection of nonlinear and interaction effects, providing novel insights. For example, the combined influence of age, walking frequency, and income revealed patterns of HRQoL variation that would be difficult to detect using traditional regression models.

This study offers several implications for psychiatric nursing. First, the perceived health status as the most influential predictor underscores the importance of incorporating regular assessments of subjective health status into chronic disease management. Patients reporting poor self-rated health should be prioritized for comprehensive case management and supportive interventions. Second, the finding that more frequent walking was associated with better HRQoL highlights the role of nurse-led physical activity counseling, including goal setting, structured walking programs, and promotion of daily movement, as practical strategies to improve quality of life. Third, the association of younger age with higher HRQoL emphasizes the particular vulnerability of older adults, suggesting that targeted nursing interventions addressing mobility, fall prevention, and social participation are needed for this group. Fourth, the protective influence of higher monthly income and private medical insurance demonstrates the importance of socioeconomic resources. Psychiatric nurses should advocate for improved access to financial and healthcare support, and collaborate with social welfare and mental health networks to provide resource linkage and counseling. Fifth, the application of explainable artificial intelligence (RF with SHAP) offers opportunities for psychiatric nurses to use patient-level explanations in clinical decision-making. SHAP results can be integrated into multidisciplinary care discussions to identify modifiable factors contributing to poor HRQoL and to design tailored interventions.

Although depressive symptoms and stress did not emerge as the top predictors in this model, they remain clinically important and have been consistently reported in prior psychiatric nursing research. Therefore, they should continue to be included in screening and intervention programs, but resource prioritization may need to be balanced with other stronger predictors identified in this analysis.

This study also has several strengths. First, the use of nationally representative data enhances the generalizability of findings across Korean adults with multimorbidity. Second, combining a RF algorithm with SHAP analysis provided both predictive performance and interpretability, addressing the common limitation of "black-box" machine learning models. Third, structuring predictors based on Andersen's behavioral model allowed for theory-driven interpretation, linking findings to a well-established conceptual framework and highlighting interactions among predisposing, enabling, need, and health-behavior factors.

However, several limitations should be acknowledged. First, the cross-sectional design precludes causal inference, underscoring the need for longitudinal studies. Second, the modest explanatory power of the model and the discrepancy in R² values between the training and test datasets suggest that the model may not fully capture the multidimensional determinants of HRQoL. Nevertheless, the relatively low error rates (MAE and RMSE) indicate that the model offers clinically useful predictions, even if its explanatory scope remains limited. Third, self-reported data are subject to recall and reporting bias, particularly in relation to mental health indicators. Fourth, the KNHANES dataset does not include clinical diagnoses, disease severity, or treatment history, limiting the ability to capture detailed health and psychiatric profiles. Accordingly, these findings should be interpreted conservatively.Future research should incorporate longitudinal data and clinical diagnostic information to track changes in HRQoL over time. The integration of diverse data sources-including electronic medical records (EMRs), wearable devices, and ecological momentary assessment (EMA)-could enhance predictive precision by capturing patients' psychosocial and behavioral contexts in real time. Comparative studies of different machine learning algorithms, particularly explainable approaches such as gradient boosting with SHAP or other XAI methods, are warranted to optimize both accuracy and interpretability. Finally, psychiatric nursing interventions targeting the key predictors identified in this study should be developed and tested in randomized controlled trials (RCTs), thereby advancing evidence-based practice to improve HRQoL among patients with multimorbidity.

This study analyzed KNHANES (2019~2020) data using Andersen's behavioral model and a random forest algorithm with SHAP analysis to identify determinants of HRQoL in adults with multimorbidity. Self-rated health was the strongest predictor, followed by walking frequency, age, monthly income, and private medical insurance.

From a psychiatric nursing perspective, these findings highlight the importance of routine assessment of subjective health, promotion of physical activity, targeted support for older adults, and advocacy for socioeconomic resource accessibility. The appflication of explainable machine learning further suggests its potential utility in clinical decision-making and individualized intervention planning.

Although the model's explanatory power was modest, its low prediction error supports clinical usefulness for risk stratification. Future studies should adopt longitudinal designs and integrate diverse data sources to enhance predictive accuracy and intervention development.