Support Vector Machine-Based Prediction Model for Healthcare Workforce Transition Success Under Decentralization

Study design and population

This was a cross-sectional study that focused on quantitative analysis, complementing a previously published qualitative investigation from our larger project on Fringe Benefits, Welfare, and Career Paths of Personnel in Health Promotion Hospitals under Provincial Administrative Organization (published elsewhere), conducted between March to October 2023. The study aimed to develop predictive models using Support Vector Machine (SVM) analysis to identify factors influencing workforce transition success during Thailand’s healthcare decentralization process.

Sampling strategy

Eight provinces were strategically selected from Thailand’s 77 provinces, representing three levels of healthcare decentralization implementation: low (less than 50% of districts within the province had transferred healthcare facilities to local organizations), moderate (50-99% of districts had completed transfers), and full implementation (all districts within the province had completed transfers to local organizations). These levels were represented by four, two, and two provinces respectively. Sample size was calculated using population proportion estimation (95% confidence interval, ±5% precision) with 15% adjustment for non-response, yielding 430 participants.

Survey instrument

A validated structured questionnaire was developed comprising two main sections. The first section collected demographic and organizational characteristics, including participants’ age, marital status, education level, monthly income, work experience, current position, facility staff headcount (pre- and post-decentralization), number of registered nurses, and facility capacity classification. The second section assessed satisfaction across benefits (19 items) and welfare (11 items) domains, evaluating aspects such as compensation, career advancement, and professional development opportunities using a 5-point Likert scale (1 = very dissatisfied to 5 = very satisfied). The instrument demonstrated strong psychometric properties, with an Item-Objective Congruence Index of 0.8-1.0 for content validity and a Cronbach’s alpha coefficient of 0.96 from pilot testing with 30 non-study healthcare facilities.

Ethical considerations

The study protocol was approved by the Ethics Committee for Human Research of PCKCN (approval number: REC No. 13/2566, dated March 23, 2023). All participants provided informed consent prior to data collection.

Support vector machine model development

The study used SVM analysis in R statistical software (version 4.0.2, e1071 package) to develop a predictive model for workforce transition success. Model development included data preprocessing through standardization of 37 predictor variables spanning demographic factors, benefits, and welfare domains. Four kernel functions (linear, radial basis function, polynomial, and sigmoid) were evaluated to determine optimal model performance.

Outcome variable definition

The binary outcome “successful transition” was operationalized using a composite satisfaction score approach;

The threshold of 3.5 was substantiated by a preliminary ROC analysis (not presented here), which demonstrated that this cutoff offered an appropriate balance between sensitivity and specificity for classifying transition success. Notably, 3.5 also represents the midpoint between ‘neutral’ and ‘satisfied’ on the Likert scale, rendering it both a statistically and conceptually significant threshold. The final distribution of outcomes was as follows: successful transition (n=195, 45.3%) and unsuccessful transition (n=235, 54.7%).

The linear kernel function was selected based on comparative performance metrics:

Model performance was assessed using three metrics: accuracy for measuring overall correct classification rate, sensitivity for assessing true positive rate of successful transitions, and specificity for evaluating true negative rate of unsuccessful transitions. Feature importance analysis was subsequently performed using the weight vectors of the selected kernel model to identify key predictors of successful transitions.

Class imbalance assessment and handling

To address the identified moderate class imbalance (successful transitions: unsuccessful), we implemented the Synthetic Minority Oversampling Technique (SMOTE) during the training phases. The SMOTE was employed to balance the dataset by generating synthetic examples of the minority class. This process, which involves interpolating between existing successful transition cases, was conducted to enhance the model’s sensitivity for successful outcomes.

Comparative model assessment

To validate the SVM selection, we compared the cross-validated performance with that of alternative machine learning algorithms.

All models underwent identical preprocessing and cross-validation procedures to ensure fair comparison.

Performance evaluation metrics

Model performance was assessed using standard classification metrics: accuracy for the overall correct classification rate, sensitivity for the true positive rate of successful transitions, specificity for the true negative rate of unsuccessful transitions, F1-score for balanced precision-recall performance, and Area Under the Curve (AUC) for overall discriminative ability.

Demographic characteristics

Of the 430 healthcare personnel studied, the majority were female (78.60%, n=338) with a bimodal age distribution peaking at 25-35 years (34.88%, n=150) and over 45 years (33.49%, n=144). More than half were married (56.51%, n=243), and nearly three-quarters held bachelor’s degrees (71.16%, n=306). Professional experience was substantial, with approximately one-third having over 20 years of service (32.79%, n=141). The workforce composition primarily comprised public health officers (23.02%, n=99) and registered nurses (10.47%, n=45), with most personnel (58.14%, n=250) serving in medium-sized sub-district health promoting hospitals.

Predictive factors for workforce transition success

Cross-validation analysis of feature importance revealed stable rankings for the top predictors, with a coefficient of variation <0.15 for the five highest-weighted features, confirming the robustness of the identified determinants. The target variable was defined as successful transition based on improvements in personnel satisfaction across benefits, welfare, and career advancement domains after transferring to work under the Provincial Administrative Organization.

Feature importance analysis of the 37 predictor variables (10 demographic/organizational, 16 benefits, and 11 welfare variables) using the linear kernel SVM model revealed the relative importance of predictors as shown in Table 1.

Analysis of feature weights derived from the SVM model identified ten key predictors of workforce transition success ( Table 1), with coefficients ranging from 0.427 to 0.190. Financial considerations demonstrated the strongest predictive power, with competitive compensation and benefits (Benefits5, coefficient=0.427) emerging as the primary determinant. Career development opportunities (Welfare30, coefficient=0.358) ranked as the second most influential predictor, suggesting that successful transitions are driven by both immediate financial incentives and long-term professional growth prospects. Among demographic characteristics, educational qualification (coefficient=0.236) emerged as a significant predictor, highlighting the role of individual capacity in transition outcomes. The hierarchical distribution of feature weights provides evidence-based guidance for prioritizing workforce management interventions in decentralized healthcare systems.

Kernel performance summary

Based on the five highest-ranked predictors (feature weights 0.427-0.252) identified through SVM analysis, we evaluated classification performance using four different kernel functions (linear, RBF, polynomial, and sigmoid). These key predictors encompassed competitive compensation (Benefits5), career development opportunities (Welfare30), promotion processes (Benefits4), hazardous work compensation (Benefits15), and educational opportunities (Welfare20). Table 2 presents the comparative performance metrics, where the linear kernel demonstrated superior cross-validated performance with optimal accuracy and balanced sensitivity-specificity trade-off. While the RBF kernel showed comparable results, the linear kernel’s combination of performance and simplicity made it the preferred choice for our workforce transition prediction model.

Cross-validation performance and model comparison

Cross-validation analysis demonstrated stable model performance across the different data partitions (Table 3). The linear SVM achieved a cross-validated accuracy of 68.5±3.8%, representing approximately 3% degradation from single-fold performance, which is typical for datasets of this size and complexity.

Class imbalance analysis and SMOTE implementation

The outcome distribution revealed a moderate imbalance, with successful transitions comprising 45.3% (n=195) and unsuccessful transitions 54.7% (n=235) of the cases. Implementation of SMOTE during training phase improved sensitivity from 46 % to 54 % while maintaining reasonable specificity of 78%. This improvement in balanced performance provides more clinically relevant metrics for workforce management decision-making (Table 4).

The performance of the SMOTE-enhanced optimal SVM model was evaluated using the confusion matrix presented in Table 4. The model’s accuracy was 67.0%, calculated as (TP+TN)/Total. Further metrics were derived, including a sensitivity (recall) of around 54% (TP/(TP+FN)), a specificity of 79% (TN/(TN+FP)), a precision of 68% (TP/(TP+FP)), and an F1-Score of 60%.

Methodological considerations and model performance

Our cross-validated accuracy of 68.5±3.8% (Table 3) represents a conservative and realistic assessment of model performance, falling within the acceptable range for ML-based clinical prediction and management models reported in recent systematic reviews (65-85% accuracy range) (Lee et al., 2022; Maghami et al., 2023). This performance demonstrates an adequate discriminative ability for healthcare workforce prediction applications. The application of SMOTE to address class imbalances demonstrates methodological sophistication in handling real-world healthcare data challenges. The improvement in sensitivity from 46.1% to 54.3% provides more balanced performance metrics, which is crucial for practical deployment in workforce management decisions, where identifying successful transitions is essential for resource allocation and policy planning. The stability of the feature importance rankings across the validation folds (coefficient of variation <0.15) strengthens the confidence in the identified predictors. The consistent emergence of competitive compensation and career development opportunities as top predictors aligns with established workforce retention research, demonstrating the significant positive effects of compensation and career development on employee retention rates (Houssein et al., 2020), while providing quantitative validation through machine learning approaches.

Model selection justification and comparative analysis

Comparative cross-validation analysis confirmed the appropriateness of the linear SVM selection over the alternative algorithms. While Random Forest showed competitive performance (65.8% accuracy), the linear SVM’s combination of superior accuracy (68.5%) and interpretable feature weights provided optimal value for healthcare administrative decision-making. The linear kernel’s performance suggests that workforce transition success can be effectively modeled through linear combinations of organizational and individual factors, supporting the interpretability of our findings. The superior performance of SVM compared to traditional logistic regression (66.2% accuracy) validates the application of machine learning approaches in healthcare workforce management. This advancement aligns with recent trends in healthcare analytics, where sophisticated algorithms increasingly outperform conventional statistical methods (Bikku, 2020).

Comparative cross-validation analysis confirmed the appropriateness of the linear SVM selection over the alternative algorithms. While Random Forest showed competitive performance (65.8% accuracy), the linear SVM’s combination of superior accuracy (68.5%) and interpretable feature weights provided optimal value for healthcare administrative decision-making. The linear kernel’s performance suggests that workforce transition success can be effectively modeled through linear combinations of organizational and individual factors, supporting the interpretability of our findings. The superior performance of SVM compared to traditional logistic regression (66.2% accuracy) validates the application of machine learning approaches in healthcare workforce management. This advancement aligns with recent trends in healthcare analytics, where sophisticated algorithms increasingly outperform conventional statistical methods (Bikku, 2020).

Practical implications for workforce management

These findings offer generalizable lessons for several countries implementing healthcare decentralization, specifically in predicting and managing successful personnel transfers. The predictive modeling approach enables the proactive identification of personnel at risk of unsuccessful transitions, allowing for targeted interventions and support mechanisms. Our findings align with Thailand’s National Health Security Act of 2019, which emphasizes local autonomy in healthcare management while maintaining service quality standards. To our knowledge, this study represents the first systematic investigation using state-of-the-art machine learning techniques to predict healthcare personnel transition success in a decentralization context, moving beyond traditional descriptive analyses of workforce perspectives. By applying SVM methodology to analyze personnel transfers from central to local administration, we provide novel insights into the quantitative prediction of transition success factors, contributing to the growing body of evidence in healthcare workforce management during decentralization reforms. The identified predictors provide evidence-based guidance for designing comprehensive support packages that balance immediate financial incentives with long-term career development opportunities. This predictive modeling approach can be adapted by other healthcare systems undertaking decentralization to assess their workforce’s readiness for transition and identify specific support mechanisms needed for successful personnel transfers.

Limitations and considerations

Several methodological limitations should be considered when interpreting our findings. First, while our cross-validation approach provides robust performance estimates, the single-country nature of our study limits its external generalizability. Future multi-country validation studies should assess the transferability of models across different healthcare systems and decentralization policies. Second, our cross-sectional design prevented the assessment of temporal causality between predictors and transition outcomes. Longitudinal studies tracking personnel across multiple transition phases would enable dynamic prediction modeling and stronger causal inference. Third, despite implementing SMOTE to address class imbalance, the moderate cross-validated sensitivity (around 54%) suggests that there is room for improvement in identifying successful transitions. Advanced ensemble methods or deep learning approaches may achieve better balanced performance, particularly given recent advances in healthcare prediction modeling (Bikku, 2020; Bikku et al., 2025). Fourth, the predominance of bachelor’s degree holders may have influenced the apparent predictive importance of education. Future studies with stratified sampling across educational levels could clarify this relationship and reduce the potential sampling bias effects on SVM classification. Finally, our linear kernel assumption implies a linear relationship between the predictors and outcomes. Nonlinear kernel exploration or alternative ML approaches may reveal complex interaction patterns in workforce transition dynamics, potentially improving predictive accuracy beyond our current cross-validated results.

Ethical approval and consent

The study protocol was approved by the Ethics Committee for Human Research of Prachomklao College of Nursing (PCKCN), Praboromarajchanok Institute, Ministry of Public Health, Thailand (approval number: REC No. 13/2566, dated March 23, 2023). All participants provided written informed consent prior to data collection. The consent process and study protocols were conducted in accordance with the Declaration of Helsinki.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the author(s) used Claude 3.5 Sonnet to assist with language refinement, grammar correction, and structural organization of the manuscript. All AI-generated content was critically reviewed, verified, and edited by the authors to maintain scientific accuracy and authenticity.