Artificial Intelligence in Questionnaire-Based Research: Quality of Life Classification Across Different Population Groups

Abstract

This interdisciplinary study presents a novel questionnaire analysis methodology using Artificial Intelligence (AI) and Machine Learning (ML). The framework is broadly applicable to all areas of research using questionnaire data analysis, including health sciences and physical education. Our predictive modeling was based on the XG-Boost algorithm, which classified individuals into three distinct groups—employees and two cohorts of retirees—based on their demographic profiles and responses to the WHOQOL-BREF survey. In order to ensure the credibility and reliability of the predictions, the model building process used the implementation of cross-validation. This procedure produced a model with a resultant accuracy of 0.8038 ($95\%$ confidence interval: 0.7551–0.8908). To go beyond conventional performance metrics, we implemented the SHapley Additive exPlanations (SHAP) method, providing a transparent and detailed interpretation of the model’s decision-making process. This explainable AI analysis clarifies both the magnitude and direction of the impact of key factors such as age and various predictors of quality of life, providing detailed, data-driven insights into what differentiates groups.

1. Introduction

Building upon our previous work on questionnaire data analysis [1,2,3,4,5,6], this study employs advanced Machine Learning (ML) techniques to analyze data collected using the World Health Organization Quality of Life-BREF (WHOQOL-BREF) questionnaire [7].

1.1. Concept of Quality of Life

Quality of life (QoL) research has gained significant prominence in recent years as part of a holistic and interdisciplinary approach to understanding the individual’s situation in medical and social sciences. The primary objective of QoL research is to comprehend a person’s subjective well-being and to identify the factors that decisively influence it. As a multidimensional and interdisciplinary concept, defining and evaluating quality of life has proven challenging. Consequently, numerous definitions and diverse methodologies exist to measure this elusive construct. The philosophical roots of QoL discussions trace back to antiquity, where ancient philosophers contemplated the foundations of a good and happy life.

Modern QoL research emerged in the 1960s within social and economic sciences. In medicine, interest in patients’ quality of life surged in the latter half of the 20th century. Initially, the focus was on oncology patients due to the substantial impairment of daily functioning caused by both the disease and its treatment. In the late 1940s, Karnofsky investigated the well-being and self-care capabilities of cancer patients [8]. In 1960, Zubrod et al. concentrated on functional impairments in cancer patients using performance status as a metric [9]. The term “quality of life” entered medical nomenclature in the 1970s, becoming a keyword in the Medline bibliographic database in 1975 and being included in the Index Medicus two years later. Schipper formally introduced the definition of Health-Related Quality of Life (HRQoL) into medical science in 1990, defining it as the impact of disease and its treatment on functioning and life satisfaction from the patient’s perspective [10]. According to Schipper, QoL is determined by factors such as physical condition, mobility, economic circumstances, and somatic sensations like pain.

The 1990s marked a period of flourishing QoL research. The International Society for Quality of Life Research (ISOQOL) was established in 1993 and now boasts over 1200 members from more than 40 countries, with 50% residing in North America, 35% in Europe, and the remaining 15% across the globe [11]. The society’s mission is to advance research on health-related quality of life, enhance healthcare quality, and promote healthy lifestyles.

Research into quality of life has continued to expand, with a vast body of literature available. Recent studies illustrate the breadth of this field. Baday-Keskin and Ekinci examine the link between kinesiophobia and HRQoL in rheumatoid arthritis patients [12]. Moitra et al. explore the influence of the physical environment on HRQoL in chronic obstructive pulmonary disease [13]. Gao et al. investigate the relationship between childhood maltreatment and quality of life [14]. Abdelsalam et al. study the association between depression and oral health-related quality of life in people who inject drugs [15]. Tooths et al. examine family contextual effects on the connection between screen time, behavior, and HRQoL in child siblings [16]. Redondo-Tebar et al. analyze differences in HRQoL between typically developing children and those with developmental coordination disorder, comparing perceptions of children and their parents [17]. Haggart et al. compare HRQoL outcomes in gay and bisexual men following prostate cancer treatment [18].

The conceptual understanding and measurement of QoL are continually refined. Long-Meek et al. reviewed the often inconsistently defined concepts of community attachment, community satisfaction, and quality of life, highlighting the need for theoretical precision in community-based research [19]. Similarly, Wicaksana and Hertanti conducted a concept analysis of Diabetes-Related Quality of Life (DRQoL), identifying its critical attributes as general health, diabetes-related satisfaction, diabetes impact, and diabetes self-management [20]. Pandarakutty and Arulappan performed an evolutionary concept analysis of HRQoL in children and adolescents with sickle cell disease, identifying key attributes such as its multidimensional nature, coping, pain management, and treatment burden [21]. For school-aged children with Attention-Deficit Hyperactivity Disorder (ADHD), Saputra et al. emphasized that QoL encompasses the subjective perceptions of both the children and their parents across physical, emotional, social, and academic domains [22].

Significant efforts are underway to define core concepts for HRQoL in specific patient populations. Raymundo et al. have conducted mixed-methods studies to generate concepts for a core domain set to assess HRQoL in patients with Cutaneous T-Cell Lymphoma (CTCL), including Mycosis Fungoides and Sézary Syndrome [23,24]. Their systematic review identified frequently measured concepts like itching, pain, and fatigue, but also highlighted missing concepts such as hair loss, temperature dysregulation, and treatment burden, underscoring the need for comprehensive, patient-centered outcome measures [24]. In a similar vein, Remon et al. investigated disease burden, QoL, and treatment experience in patients with hypophosphatasia and their caregivers [25].

The application of QoL concepts is also evolving in novel contexts. Bunn et al. explored in a qualitative proof-of-concept study how the Adult Social Care Outcomes Toolkit (ASCOT) could help care home staff minimize the impact of infection control measures on resident QoL during infectious outbreaks [26]. In architecture, Alsolami et al. analyzed the challenges of implementing WELL building standards, which focus on human health and well-being, to enhance QoL in residential apartment design in Saudi Arabia [27].

The measurement of QoL itself is a subject of ongoing research to ensure cultural relevance. Ding et al. used concept mapping to assess how well the EuroQol 5-Dimensional (EQ-5D) and EuroQol Health and Well-Being (EQ-HWB) instruments capture QoL perceptions from a Chinese perspective, finding that while both fit, cultural differences influenced domain prioritization [28]. Miyazaki et al. [29] and Suzukamo et al. [30] tracked changes in the general public’s concept of QoL over time through national surveys, examining both awareness and conceptual structure.

Furthermore, the concept of QoL is being integrated into new frameworks. Ramos et al. reviewed integrating animal welfare and HRQoL concepts for preweaning dairy calves, adapting a human HRQoL framework to illustrate the interconnection of behavior, mental state, and health [31]. In nursing, Galiana et al. validated the Spanish version of the Nursing Self-Concept Instrument and found a significant relationship between clinical practice, nursing self-concept, and professional quality of life among students [32].

Clinical research continues to demonstrate the importance of QoL across conditions. Brabcová et al. found that quality of life, academic self-concept, and mental health in children with epilepsy are significantly worse for those with comorbidities like learning disabilities or ADHD [33]. Romagna et al. compared microsurgery and radiosurgery for vestibular schwannoma, finding differences in HRQoL outcomes post-treatment [34]. Zaglauer et al. showed that a multimodal pain therapy concept significantly improved pain perception and QoL in patients with chronic back pain [35]. Jost-Engl et al. provided proof-of-concept that a high-intensity exercise intervention during chemotherapy could enhance cardiorespiratory fitness and QoL in patients with high-grade glioma [36].

Innovative care models are also being evaluated for their impact on QoL. Huang et al. demonstrated that an empowerment education concept combined with humanistic care improved moods, reduced adverse reactions, and enhanced QoL in lung cancer patients undergoing chemotherapy [37]. Chen and Gong found that combining the Enhanced Recovery After Surgery (ERAS) concept with Roy’s Adaptation Model improved perioperative mental health, QoL, and recovery in prostatectomy patients [38]. Chen et al. showed that graded nursing based on a risk early warning concept reduced pressure injuries and improved QoL in long-term bedridden patients [39].

In pediatric cardiology, Fuhrmeister et al. discussed how electronic Patient-Reported Outcome Measures (ePROMs) can systematically capture HRQoL, providing a holistic view of well-being for children with congenital heart disease that goes beyond clinical parameters [40]. Finally, technological advancements are aiding QoL assessment. Allgaier et al. developed a cross-lingual framework using Linguistic Linked Open Data (LLOD) to detect HRQoL concepts from French online health communities, demonstrating the potential of text analytics for understanding patient-reported outcomes [41].

1.2. The WHOQOL-BREF Instrument

WHOQOL-BREF questionnaire stands as a cornerstone in the field of quality of life assessment, renowned for its cross-cultural applicability and robust validation across a myriad of populations and research contexts [42,43,44,45,46,47,48,49]. Its brevity and comprehensive nature make it an ideal tool for large-scale studies aiming to capture the multifaceted essence of human well-being.

The structure of the WHOQOL-BREF is meticulously designed to provide a holistic yet efficient evaluation. The questionnaire consists of 26 items. The initial two items are global indicators, designed to capture the respondent’s overarching perception of their Overall Quality of Life and General Health. These questions serve as a foundational self-assessment. The remaining 24 items are systematically distributed across four core domains, each representing a fundamental pillar of human existence:

• Physical health—This domain explores aspects such as pain, discomfort, energy, fatigue, mobility, sleep, and activities of daily living.
• Psychological health—Items here assess positive and negative feelings, self-esteem, bodily image, spirituality, and cognitive functions like concentration.
• Social relationships—This dimension focuses on personal relationships, social support, and sexual activity.
• Environmental health—A broad domain that evaluates safety, physical environment, financial resources, access to health services, opportunities for recreation, and transportation.

The selection of this instrument is justified by its extensive and diverse application in contemporary research. Recent studies underscore its versatility. For example, Barwais utilized it to establish a positive correlation between physical activity levels and various QoL dimensions among Saudi female university students [42]. In the clinical realm, Yu et al. conducted a rigorous psychometric validation of the WHOQOL-BREF among non-dialysis chronic kidney disease patients in Taiwan, employing both Rasch and confirmatory factor analyses, thereby confirming its reliability in a specific patient population [43]. Similarly, its application among medical students in Jordan, Hamad et al. revealed key factors influencing their well-being [44], while a study in rural Bangladesh highlighted significant gender differences in domain-specific quality of life among older adults [45].

1.3. Psychometric Robustness and Global Norms

The psychometric integrity of the WHOQOL-BREF is a subject of ongoing research, ensuring its continued relevance. Kangwanrattanakul and Kulthanachairojana performed a modern psychometric evaluation of the Thai version, including the development of shorter forms, for patients on warfarin, demonstrating its adaptability [46]. Furthermore, establishing interpretability, Pietrzykowski et al. calculated the Minimal Clinically Important Difference (MCID) for the WHOQOL-BREF in adults with neurofibromatosis, providing clinicians with a benchmark for meaningful change [47]. The instrument’s utility for establishing population-level benchmarks is exemplified by the work of Bat-Erdene et al., who provided normative data for the general population of Mongolia [48]. Finally, a systematic review by Gagliardi et al. consolidated evidence on the health-related quality of life of refugees using the WHOQOL-BREF, affirming its value in vulnerable and displaced populations [49].

1.4. Research Context: The Retirement Threshold as a Key Application

To compellingly demonstrate the practical utility and predictive power of our machine learning framework, we apply it to a complex and well-studied socio-economic and public health issue: the retirement threshold problem. The transition from active employment to retirement represents one of the most significant life course changes, with profound, multifaceted implications for an individual’s daily routine, social identity, financial security, and physical and psychological health. The moment of crossing this threshold triggers a dynamic process of adaptation that has been the focus of extensive scholarly investigation [50,51,52,53,54,55,56,57,58,59].

The studies reveal that the impact of retirement is not monolithic but varies considerably based on individual circumstances, prior occupation, and social context. A key area of research involves tracking changes in health behaviors. For instance, studies like the Finnish Retirement and Aging Study (FIREA) have meticulously documented how daily physical activity patterns evolve across the retirement transition, noting variations by occupational class and pre-retirement modes of commute [53,55]. The broader health consequences of this life shift are a central theme, with researchers debating its net effect; some studies report improvements in mental well-being and a reduction in work-related stressors, while others highlight risks associated with decreased structure and potential loss of purpose [54,57].

Beyond physical health, the psychological and social dimensions are paramount. Research has shown that the maintenance and formation of social group memberships post-retirement are critical buffers against decline in physical health and act as a safeguard against loneliness [52]. The subjective experience of the transition is also explored qualitatively; for example, Lund describes the journey at the “threshold of retirement” as a move from all-absorbing work relations towards a new phase of self-actualization and identity reformation [56]. Furthermore, the concept of enjoyment in everyday activities undergoes a shift, as retirees often report increased satisfaction with leisure and domestic activities, a phenomenon captured by the evocative title of the study by Olds et al.: “Everybody’s Working for the Weekend.” [58].

1.5. Defining the ‘Threshold’ in Multidisciplinary Contexts

The very concept of a “threshold” in retirement research is multifaceted. In a financial context, Bär et al. explore it as an optimal control point in asset allocation strategies, where utility maximization is dependent on crossing specific wealth thresholds [51]. In a completely different domain, the term is used metaphorically to describe decision-making under risk, such as in the study of athletes’ perceived concussion retirement thresholds [50]. A critical aspect often studied is the role of external factors, such as the loss of employer-sponsored private insurance before and after early retirement, which Kail identified as having significant consequences for both mental and physical health outcomes [59].

This rich, multidisciplinary scholarly discourse, encompassing epidemiology, sociology, psychology, and economics, establishes the retirement transition as an ideal, real-world laboratory for testing our ML model. It is a period of concentrated change where the latent patterns within quality of life data are likely to be most pronounced and interpretable. By applying our model to this problem, we aim to move beyond traditional correlational analyses and uncover predictive insights that can identify individuals at risk of a decline in well-being or, conversely, those likely to thrive, thereby informing targeted interventions.

2. Materials and Methods

2.1. Study Population and Design

This cross-sectional study involved a cohort of 437 (after final verification) participants aged 50 and above from Bydgoszcz, Poland. The cohort was strategically selected to capture the transition from employment to retirement and the potential moderating role of intellectual and social activity in later life. It comprised two primary groups: 179 actively working individuals (119 women and 60 men) and 258 retirees (165 women and 93 men). To investigate the impact of structured social and educational engagement post-retirement, the retiree group was further stratified based on their affiliation with the University of the Third Age (U3A), a prominent institution for lifelong learning. The U3A affiliation was used as a class label based on established research indicating that participation in Universities of the Third Age is correlated with specific psychosocial profiles and measurable outcomes in quality of life. As an organized, peer-learning community, U3A attracts individuals with distinct characteristics, often higher educational backgrounds, proactivity, and social engagement, which differentiates them from the broader retiree population [60,61,62]. Furthermore, numerous studies have directly characterized U3A members and demonstrated a positive correlation between U3A participation and enhanced quality of life metrics, including improved mental health, physical well-being, social connectedness, and overall life satisfaction [63,64,65]. The provided references substantiate this characterization and established correlation. This resulted in the following three distinct groups for our analysis:

• employees: Individuals currently engaged in employment ($n=179$).
• retirees1: Retired individuals not participating in U3A activities ($n=167$; 98 women, 69 men).
• retirees2: Retired individuals who are active students at the U3A ($n=91$; 67 women, 24 men).

This design allows for a nuanced comparison not only between workers and retirees but also between different retirement lifestyles, enabling the model to identify patterns associated with both retirement status and active versus passive post-retirement engagement.

2.2. Data Collection and Features

The primary instrument for data collection was the Polish version of the World Health Organization Quality of Life-BREF (WHOQOL-BREF) questionnaire. Our analysis considered 24 questions from this instrument, aggregated into their respective domains: Physical Health (Domain 1), Psychological (Domain 2), Social Relationships (Domain 3), and Environment (Domain 4). In addition to these QoL metrics, key demographic variables—specifically, age and gender—were included as features in the model. The complete feature set used for predictive modelling thus consisted of the four domain scores, age, and gender, providing a multidimensional profile of each participant.

2.3. Machine Learning Approach: The XGBoost Algorithm

As a pivotal preliminary analysis, we developed a classification model utilizing the eXtreme Gradient Boosting (XGBoost) algorithm [66,67]. XGBoost is a sophisticated and highly efficient implementation of the gradient boosting framework [68].

The core principle of XGBoost is ensemble learning, where predictions are made by combining the outputs of a collection of simpler, weaker models, typically decision trees. It operates in an iterative, additive manner. In each iteration, a new tree is constructed to correct the residual errors made by the ensemble of all previous trees. The “gradient” in its name refers to the use of gradient descent optimization to minimize a defined loss function (e.g., multi-class log-loss for classification) when adding new models.

The selection of XGBoost for this study was motivated by several key advantages. It efficiently handles missing values and sparse data, which is particularly beneficial for real-world questionnaire data. A central strength is its incorporation of L1 (Lasso) and L2 (Ridge) regularization directly into the objective function, which helps to control model complexity, prevent overfitting, and leads to better generalization on unseen data [66]. Additionally, the algorithm is known for its computational speed and scalability, even with large datasets. Finally, XGBoost provides built-in metrics for assessing feature importance, which is crucial for interpreting the model and identifying which quality of life domains or demographic factors are most predictive of group membership [67].

Model Implementation and Training Protocol

To ensure a robust and generalizable model, a detailed training protocol was established. The dataset was divided into a training set (≈70%, $n=307$) and a test set (≈30%, $n=130$) using a stratified split via the createDataPartition procedure from the caret package [69]. This approach preserves the original distribution of the target variable in both subsets, which is essential for the reliability of the model validation process.

The modeling methodology involved preprocessing with one-hot encoding for categorical variables and the direct use of continuous variables. The model was then trained using a 10-fold cross-validation procedure implemented directly in the XGBoost package [70]. This method provides a reliable estimation of the model’s performance on unseen data by repeatedly splitting the training set, thereby minimizing the risk of optimistic bias. An early stopping rule was integrated into the cross-validation to automatically determine the optimal number of boosting iterations and prevent overfitting.

In terms of parameter configuration, the objective parameter was set to multi:softprob for multi-class classification with probability outputs, and the eval_metric was defined as multi-class logarithmic loss (logloss). All other hyperparameters were utilized at their default values. After the cross-validation process identified the optimal settings, a final model was refit on the entire training dataset using these parameters for subsequent evaluation on the hold-out test set.

3. Results and Discussion

The performance and generalization capability of the final XGBoost model were first assessed by analyzing the cross-validation results. The result of this process is presented in Figure 1, which shows selected parameters such as the optimal number of iterations and the Logloss values for both the Validation Folds and the Training Folds.

The final model was evaluated on an independent test set, which was not used in any of the previous stages of model building or tuning. The model’s fit quality and its classification effectiveness were assessed using a confusion matrix (Figure 2). This tool provides detailed insight into the nature of the errors made by the model, enabling the identification of classes that are most frequently confused and the calculation of key performance metrics, such as precision, recall, and the F1-score.

In the conducted analysis, the classification model achieved very good and balanced results, characterized by a precision of 0.8016, a recall of 0.7799, and an F1-score of 0.7856. This means that the model not only accurately identifies positive cases, avoiding a large number of false alarms, but also maintains a balance between these two aspects, as confirmed by the value of the F1-score, which is their harmonic mean. The model’s accuracy 0.8308 is confirmed by the accuracy metric, whose $95\%$ confidence interval ranges from 0.7551 to 0.8908, indicating a precise estimation of its true performance.

The obtained results allow us to conclude that the model significantly (significance level $\alpha =0.05$) outperforms random classification, as confirmed by the extremely low p-value (<$2\times {10}^{-16}$) of the test comparing its accuracy to the no-information rate. Furthermore, a Cohen’s Kappa statistic value of 0.7324 indicates a moderate to good agreement between the predictions and reality. Conversely, the non-significant result (p-value = 0.2276) of McNemar’s test suggests that the model does not systematically make one particular type of error, which may indicate its stability.

While the model’s overall performance was strong, results varied significantly between classes. For the retirees2 group (≈21%) of the sample), the recall value was 0.5185, compared to 0.9811 for employees and 0.8400 for retirees1. The poor model performance for retirees2 is not only due to methodological errors but also reflects real, more elusive characteristics of this group in the study population. The retirees2 are a subgroup with unique characteristics, for example, they have a slightly lower average age than retirees1. For an in-depth analysis of the role of individual variables in the model, three key feature importance metrics offered by XGBoost were calculated. The Gain metric measures the average improvement in prediction accuracy brought by a feature across all splits where it is used, with a higher value indicating a greater contribution of the feature to improving the model’s quality. Cover specifies the number of observations in the training set that are influenced by a given feature across all splits made, reflecting the scope of the feature’s impact on the learning process. Frequency, in turn, is a simple counter that determines how many times a given feature was used for a split across all trees in the ensemble.

It is important to note, however, that these metrics provide information only about the global feature importance, indicating which variables are generally significant for the entire model, but they do not explain how they influence individual predictions. To gain detailed insight into the model’s mechanism at the level of single observations, the SHapley Additive exPlanations (SHAP) method, based on game theory, was applied. The SHAP method allows for the quantification of each feature’s contribution to the prediction for a specific instance, explaining both global and local patterns of the model’s behavior, identifying the direction of a feature’s influence on the final prediction, and understanding the interactions between features and their combined effect on the output. This approach enables a comprehensive model interpretation, combining the analysis of the overall structure with a detailed understanding of individual classification decisions.

Using the available functions and procedures of the computational packages [71], the values of the mentioned metrics and the SHAP values were calculated and compiled in

Table 1. SHAP and Traditional Importance Comparison.

	Variables	Men	Domain 1	Domain 2	Domain 3	Domain 4	Age
Importance Measures
SHAP Global Importance		0.0556	0.1215	0.1044	0.0714	0.1402	1.2920
SHAP Importance employees		0.0386	0.0662	0.0127	0.0903	0.0455	1.9376
SHAP Importance retirees1		0.0248	0.1582	0.1438	0.0896	0.1938	0.9356
SHAP Importance retirees2		0.1033	0.1400	0.1565	0.0344	0.1813	1.0028
Gain		0.0166	0.0758	0.0556	0.0314	0.0667	0.7538
Cover		0.0296	0.1250	0.1383	0.0705	0.1544	0.4823
Frequency		0.0443	0.1907	0.1508	0.1131	0.1663	0.3348

. Analysis of the data contained in this table unequivocally indicates that the dominant variable in terms of importance, according to both SHAP measures and the key metrics provided by the XGBoost algorithm, is age. This variable demonstrates a decisively greater influence on the model, surpassing all other factors in this regard. Subsequently, the values of the variables Domain 4 and Domain 1 influence the global model importance. A detailed comparison showed that Domain 1 ranks second in terms of magnitude for metrics such as Gain and Frequency. Conversely, Domain 4 is identified as the second most important variable in the context of parameters such as Global Importance, Importance retirees1, Importance retirees2, and Cover, confirming its significant role in shaping the model’s predictions.

Figure 3, Figure 4 and Figure 5 present summary charts for each predicted class. In each chart, the y-axis displays the values of the classification variables, the x-axis shows the identification numbers of individuals from the training set, and point colors correspond to their SHAP values [72,73].

Let us focus on the bottom panel of Figure 5 concerning the “age” variable. The plot displays 307 points corresponding to individuals from the training set, with their original Subject IDs from the complete dataset placed on the X-axis. The arrangement of points and their color, corresponding to SHAP values, reveals an interesting tendency in the model’s learning: for individuals with original IDs 1–258 (encompassing both the actual retirees2 and retirees1 groups) “ages” 65–75 show positive SHAP values, meaning that being in this age range increases the probability of being assigned to the predicted retirees2 group compared to other age groups. Conversely, “ages” below 60 or above 80 within the same ID group show negative SHAP values, reducing this probability. For individuals with IDs 259–437 (actual employees group), predominantly negative SHAP values are observed, correctly reflecting the low probability of classifying working individuals as members of the predicted retirees2 group. Nevertheless, as shown by the confusion matrix (cited earlier in the text), the age effect alone is insufficient for reliable classification of the retirees2 group, whose recall is only 0.5185.

Building upon the SHAP-based insights, we implemented targeted modifications to address the identified limitations. The XGBoost hyperparameters were adjusted to reduce over-reliance on age and improve minority class performance: max_depth = 4 (shallower trees), gamma = 0.5 (more conservative splitting), eta = 0.05 (slower learning rate), and min_child_weight = 5 (requiring larger leaf samples). These changes aimed to discourage the model from developing overly specific age-based rules while promoting consideration of other features. Additionally, post-hoc threshold adjustment was applied specifically for the retirees2 class, multiplying its posterior probability by a factor of 1.5 before final classification.

The optimized model demonstrates substantially improved performance for the retirees2 class (see resulting confusion matrix in Figure 6). The recall for retirees2 increased to 0.7778, compared to 0.5185 in the original configuration. This improvement was achieved while maintaining strong overall classification accuracy (0.8462) and preserving excellent performance for the employees class (recall: 0.9623). The retirees1 class shows a recall of 0.7600, indicating balanced performance across all groups.

These results confirm that the previously identified limitations in retirees2 classification were addressable through appropriate model configuration. The combination of constrained model complexity and targeted threshold adjustment effectively leveraged the age-related patterns identified by SHAP analysis while mitigating their insufficient impact in the original model.

Additional feature importance analysis for the optimized model is presented in

Table 2. Feature Importance Analysis: Optimized Model.

	Variables	Men	Domain 1	Domain 2	Domain 3	Domain 4	Age
Importance Measures
SHAP Global Importance		0.0830	0.2040	0.1095	0.0588	0.0848	1.3431
SHAP Importance employees		0.1263	0.1769	0.0331	0.0045	0.0096	2.2182
SHAP Importance retirees1		0.0000	0.2372	0.1371	0.0821	0.1087	0.8731
SHAP Importance retirees2		0.1225	0.1981	0.1583	0.0897	0.1360	0.9380
Gain		0.0188	0.0767	0.0416	0.0286	0.0403	0.7941
Cover		0.0408	0.1548	0.1185	0.0574	0.1083	0.5201
Frequency		0.0474	0.1565	0.1654	0.0920	0.1160	0.4228

. Similarly to Table 1 (pertaining to the initial model), age remains the most influential variable, particularly for employees classification. The Domain 1 maintains consistent importance across all retiree groups, while Domain 2 and Domain 4 show elevated importance specifically for the retirees2 class. The influence of gender is marginal, except in a few class-specific contexts. The obtained metrics confirm that despite the dominant role of age, the optimization process enhanced the model’s ability to utilize domain variables in differentiating retirement subgroups.

Figure 7, Figure 8 and Figure 9 present analogous summary charts for the optimized model—one for each predicted class. Following the same format, each visualization displays training set individuals by their identification numbers on the x-axis, shows the collected values of classification variables on the y-axis, and uses point colors to represent SHAP values calculated from the improved model configuration.

4. Conclusions

The impact of features on quality of life is a classic approach and has been widely explored in the literature, including our previous work [1,2,3,4,5,6]. The primary motivation for this study was to investigate a different question: Can we identify a distinct “psychosocial profile”—based on quality of life and demographic data—that characterizes different post-employment lifestyles (working, retired, retired+U3A)? Our goal was to allow the model to freely discover the most predictive patterns for these group memberships from our data.

This study assessed the predictive capacity of demographic and quality-of-life variables for classifying individuals into three distinct categories: employed individuals (employees), retired persons not involved in structured educational activities (retirees1), and retirees participating in University of the Third Age programs (retirees2). Our optimized XGBoost model, configured with constrained complexity (max_depth = 4, min_child_weight = 5) and enhanced regularization (gamma = 0.5), achieved an overall accuracy of 0.8462 on the test set. The model demonstrated particular success in classifying employees (recall: 0.9623) and retirees1 (recall: 0.7600), while showing substantially improved but still challenging identification of retirees2 (recall: 0.7778 after targeted threshold adjustment).

SHAP analysis revealed that age serves as the dominant predictive feature across all classes, with a particularly strong influence on the identification of employees. For retiree subgroups, quality-of-life domains showed differentiated importance patterns: Domain 1 (Physical Health) maintained consistent relevance across both retiree groups, while Domains 2 and 4 (Psychological and Environment) exhibited increased relative importance specifically for retirees2 classification. The gender variable demonstrated minimal predictive contribution in most contexts.

These computational findings should be interpreted within important methodological constraints. First, the observed age patterns (particularly the association between “ages” 65–75 and retirees2 classification) may reflect cohort characteristics, sampling distributions, or demographic realities rather than causal relationships or optimal intervention targets. Second, while SHAP values indicate which features contribute to model predictions, they do not establish causality nor identify modifiable intervention targets—especially when the predictive features are themselves outcome measures. Third, the persistent classification challenge for retirees2, despite optimization efforts, suggests this group may share substantial characteristics with other retiree subgroups or that the available features provide limited discriminatory power for this specific distinction.

Methodologically, this study demonstrates how explainable machine learning techniques can illuminate complex predictive patterns in this type of data. The substantial improvement in retirees2 recall following model optimization (from 0.5185 to 0.7778) illustrates how algorithmic adjustments can address classification disparities identified through interpretability analysis. However, the continued performance gap across classes underscores the inherent challenges in distinguishing subtle subgroup differences within retiree populations using primarily quality-of-life measures.

Our future studies will focus on identifying additional discriminatory features beyond those measured here, validating these patterns in independent cohorts, and exploring alternative modeling approaches that might better capture the nuanced distinctions between different retiree lifestyles. Additionally, investigations combining predictive modeling with causal inference methods would help distinguish correlational patterns from potentially actionable intervention targets.