A constructive reading of Mudallal, Mrayyan, and Kharabsheh (2025)
Creativity among nurses has become increasingly central to discussions of healthcare quality, innovation, and adaptive capacity. In complex and resource-constrained health systems, the ability of nurses to generate practical solutions, improve workflows, and respond flexibly to emerging problems is not a peripheral issue. It is closely tied to service quality and organizational resilience. For that reason alone, attempts to study nurse creativity empirically deserve attention. When such an effort is combined with machine learning, the work becomes even more noteworthy. Mudallal, Mrayyan, and Kharabsheh’s study, published in BMC Nursing, makes exactly such an attempt, and it should be recognized as a valuable and ambitious contribution to a still-developing area of inquiry.
The study draws on data from 191 registered nurses working in eight hospitals in Jordan. Its aims are notably broader than those of many conventional nursing studies. First, it seeks to identify the factors associated with nurse creativity. Second, it attempts to move beyond explanation toward prediction by developing a decision support perspective using machine learning. In this sense, the article is not merely descriptive. It tries to build a bridge between nursing management, organizational behavior, and computational modeling. That interdisciplinary ambition is one of its most important strengths.
The authors use the self-rated creativity scale developed by Zhou and George (2001) as the dependent variable. Among the explanatory variables are humble leadership, psychological safety, and a set of individual and workplace characteristics. Their results suggest that humble leadership, psychological safety, professional experience, hospital quality initiatives, and educational level are significant predictors in the regression model. In the machine learning phase, Naïve Bayes and K-Star reportedly achieved the strongest performance on selected classification metrics. These findings, taken together, support the broader claim that managerial and organizational factors may play a meaningful role in fostering creative behavior among nurses.
This is, in several respects, a commendable study. Creativity remains underexamined in nursing research compared with more established topics such as burnout, job satisfaction, or turnover intention. By centering creativity, the article opens a useful conversation. It also deserves credit for attempting to integrate conventional statistical analysis with machine learning methods rather than treating the two as mutually exclusive. Even under relatively limited sample conditions, the use of cross-validation reflects an awareness of overfitting risks. Ethical procedures also appear to have been handled carefully, which strengthens the credibility of the work.
At the same time, the article leaves several important questions unresolved. These do not negate its contribution. Rather, they indicate where the next generation of research could deepen, test, and refine the insights offered here.
One major issue concerns causality. Because the design is cross-sectional, the direction of the relationship between humble leadership and nurse creativity cannot be established with confidence. It is plausible that humble leadership encourages creativity, but it is also plausible that more creative nurses perceive or respond to leadership differently. Longitudinal or experimental designs would be much better suited to clarifying temporal direction and causal mechanisms. This is not an unusual limitation, but it is an important one, especially when leadership effects are discussed in practice-oriented terms.
A second issue is the unexpected absence of a significant relationship between humble leadership and psychological safety. Prior literature has often reported a positive association between these constructs, so this null finding is potentially interesting in its own right. Yet the discussion appears to move past it rather quickly. A fuller interpretation could have enriched the article considerably. For example, contextual explanations related to organizational culture, hospital type, hierarchy, or national setting might have been explored. When a finding runs against the dominant literature, that is not necessarily a weakness. Sometimes it is where the most interesting theoretical work begins. But it needs deeper reflection than simple acknowledgment.
Another limitation is that creativity is assessed only at the individual level through self-report. In actual hospital settings, creativity often emerges collectively, through teamwork, informal coordination, and shared problem-solving. A nurse may be individually creative, but whether that creativity translates into service innovation depends heavily on team climate and institutional support. Future studies could therefore benefit from measuring team-level creativity, supervisory assessments, or observable innovation outcomes rather than relying solely on self-perceptions.
Relatedly, the study does not directly connect nurse creativity to patient outcomes. This is perhaps the most important missed opportunity from a health management perspective. If creativity is to be treated as a strategically relevant organizational capability, its downstream effects should be examined more explicitly. Does higher nurse creativity relate to lower error rates, better patient satisfaction, stronger problem resolution, or improved care coordination? These are not minor extensions. They are central to the practical significance of the concept.
The contextual scope of the model also remains somewhat narrow. Variables such as organizational culture, accreditation environment, staffing adequacy, resource availability, and power distance are not incorporated into the analysis. In healthcare organizations, creativity is rarely a purely individual phenomenon. It is enabled or constrained by institutional conditions. Even within the same country, hospitals may differ substantially in terms of managerial philosophy, professional autonomy, and quality infrastructure. Greater attention to these contextual dimensions would likely strengthen future modeling efforts.
The article’s discussion of decision support systems is also more conceptual than operational. The study states that a decision support orientation is developed, yet no actual software prototype, user interface, usability assessment, or real-world implementation test is presented. In that sense, the DSS component remains more aspirational than demonstrated. This does not mean the idea lacks value. On the contrary, it is a promising direction. But the distinction between a predictive modeling exercise and a clinically usable decision support system should be kept clear.
Methodologically, the study also raises several issues that deserve careful consideration. One of the most important concerns is common method bias. The dependent variable and the core independent variables are all derived from the same self-reported source. In such cases, shared measurement context can inflate or distort associations. Standard diagnostic steps such as Harman’s single-factor test or marker-variable approaches are not reported. Given how frequently this issue is discussed in organizational and behavioral research, its omission weakens confidence in the observed relationships.
The use of stepwise regression is another point that merits caution. Stepwise procedures are still seen in applied research, but they remain controversial because they may capitalize on sample-specific variation and increase the risk of unstable model selection. A theory-driven hierarchical regression or a full-model specification would generally provide a stronger basis for inference. The reported adjusted R² of approximately 30 percent is not trivial, but it also means that a substantial proportion of the variance remains unexplained. That does not invalidate the model, though it does suggest that creativity is shaped by a broader and more complex set of influences than those captured here.
There also appears to be a terminology issue in the presentation of machine learning results. Reporting numerical values under the heading “ROC” can create confusion, because ROC technically refers to the curve itself, whereas the scalar summary is usually the AUC or AUC-ROC. This may seem like a small point, but in predictive modeling, clear metric labeling matters. Ambiguity in terminology can make it harder for readers to judge whether the analysis was interpreted and reported correctly.
Class imbalance is another concern that deserves explicit treatment. Based on the reported mean creativity score and the threshold used for defining high creativity, it seems likely that a large share of the sample fell into the high-creativity class. Under such conditions, recall and precision values can become misleading if the class distribution is not addressed carefully. A very high recall for one classifier may partly reflect majority-class dominance rather than genuine predictive strength. Methods such as resampling, class weighting, or at minimum a fuller discussion of class proportions would have made the machine learning evaluation more transparent.
The feature selection process also appears insufficiently documented. Some predictors that were meaningful in the regression stage were not retained in the machine learning stage, yet the rationale for this transition is not fully explained. Readers are left without a clear sense of which feature selection logic was used, how variable importance was ranked, or whether sensitivity analyses were conducted. In predictive studies, reproducibility depends not only on the final model, but also on how features were chosen, transformed, and compared.
Finally, the use of Naïve Bayes introduces a theoretical issue of its own. The algorithm assumes conditional independence among predictors, an assumption that may be difficult to defend in a model containing leadership perceptions, psychological safety, age, and professional experience. In real organizational settings, these variables are often interrelated. That does not mean Naïve Bayes cannot perform well empirically. It often does. But when its assumptions are likely violated, readers benefit from some acknowledgment of that fact and from a discussion of how it may affect interpretation.
Despite these limitations, the article should not be read as a failed attempt. That would be an unfair conclusion. It is better understood as an important exploratory study that takes a meaningful risk in an area where the literature is still relatively thin. Its real value may lie not only in the specific findings it reports, but also in the methodological and conceptual agenda it sets for others. It invites nursing and health management scholars to think more seriously about how creativity can be measured, how leadership and safety climates shape it, and how predictive analytics might eventually support managerial decision-making in healthcare settings.
In that sense, the study performs a useful scholarly function. It opens the door. Future work now has the opportunity to go further by using longitudinal designs, incorporating contextual variables, testing multiple sources of creativity assessment, linking creativity to patient and organizational outcomes, and developing decision support tools that can actually be piloted in practice. The article’s contribution is therefore not diminished by critique. On the contrary, thoughtful critique is one of the clearest signs that a study has entered an important conversation.
Reference: Mudallal, R. H., Mrayyan, M. T., & Kharabsheh, M. (2025). Use of machine learning to predict creativity among nurses: A multidisciplinary approach. BMC Nursing, 24, Article 539. https://doi.org/10.1186/s12912-025-03151-4
Note: For machine learning algorithms to produce reliable results, the minimum sample size remains a matter of debate in the literature. Even so, commonly cited thresholds suggest that N = 191 is still a modest basis for the kind of predictive claims advanced in this study. Vabalas et. al. (2019)*. for example, propose a practical rule of at least 50 to 100 observations per variable. In the present study, six classification variables were used together with 10-fold cross-validation. Under that setup, each test fold would contain only about 17 to 19 observations. That is a narrow empirical base, and one that is clearly vulnerable to statistical instability.
More importantly, when discussing sample adequacy, the authors refer to the formula N ≥ 10(k) + 50. That formula belongs to regression-based sample size logic; it is not a machine learning criterion. Using the same formula to justify both regression and machine learning reflects a conceptual conflation between two fundamentally different analytical traditions. Regression seeks parameter estimation and inferential interpretation. Machine learning, at least in its predictive form, is primarily concerned with generalization performance under repeated unseen data conditions. Treating the sample size requirements of these two approaches as interchangeable weakens the methodological foundation of the study.
The more important question, then, is not simply whether machine learning was used, but whether it was the most suitable option for the study’s aims under this sample condition. Several alternatives would likely have served the research question more effectively while also preserving interpretability.
A first alternative is PLS-SEM. In a context where the regression model explains about 30 percent of the variance, Partial Least Squares Structural Equation Modeling would have offered a more coherent framework for testing latent variable relationships and indirect paths simultaneously. It is particularly well suited to relatively small samples and theory-oriented models. In this case, it could have examined the humble leadership → psychological safety → creativity pathway directly, while also estimating measurement quality and mediation effects in one integrated model. With N = 191, bootstrapping in SmartPLS would have provided adequate inferential support. More importantly, the null finding for the relationship between humble leadership and psychological safety could have been positioned more meaningfully within such a structural framework rather than appearing as an isolated and weakly interpreted result.
A second option is Bayesian regression. In small samples, frequentist methods often suffer from reduced power and unstable estimates. Bayesian approaches can partly address this problem by incorporating prior knowledge into the model. That is especially appropriate in studies like this one, where the literature already provides a substantive basis for directional expectations. A Bayesian model would not only allow uncertainty to be represented more honestly, but would also produce probability-based interpretations of coefficients that are often more informative than dichotomous significance testing.
A third alternative would have been logistic regression or elastic net penalized regression. Given a dataset with only six predictors and a likely class imbalance problem, these methods are considerably more suitable for N = 191 than more ambitious classification algorithms. Both methods include built-in mechanisms that help reduce overfitting in small datasets, and both preserve interpretability to a degree that is valuable in applied health management research. If the purpose is to inform decision-making rather than merely to showcase technique, interpretability is not a luxury. It is a requirement.
A fourth and perhaps even more practical option concerns the decision support system the authors claim to have developed. If the aim was truly to build a usable decision aid, a rule-based decision tree would have provided a more transparent foundation. A single CART tree, for instance, is easier to explain, easier for clinicians and nurse managers to interpret, and often better aligned with real-world decision environments than a Naïve Bayes classifier. Especially in a modest sample, transparency may matter more than nominal algorithmic sophistication.
There is also a broader methodological lesson here. For a study that seeks to introduce machine learning into nursing research, the statistically honest route would have been to build a solid regression model first, then present machine learning only as an exploratory supplementary analysis, and finally discuss the limitations of both approaches side by side. Instead, machine learning was placed at the center of the study in a way that the data structure does not fully support. High recall values were then presented almost as if they constituted proof of methodological success. Yet the reported 99 percent recall most likely reflects inflation driven by the majority class rather than genuine predictive strength. In that sense, the metric creates an impression of robustness that the study design may not be able to sustain.
The main lesson from this article is therefore quite clear. When machine learning is chosen not because it is the best tool for the analytical problem, but because it appears modern, innovative, or impressive, the result is often methodological overreach. In such cases, findings can become more misleading, and the method itself begins to overshadow the study’s actual contribution. This is unfortunate, because the real contribution of the article lies elsewhere: in drawing attention to nurse creativity as a serious management issue and in encouraging more analytically ambitious research in nursing. That contribution is real. But it would have been stronger had the methodological choices remained more closely aligned with the scale and structure of the data.
*Vabalas, A., Gowen, E., Poliakoff, E., & Casson, A. J. (2019). Machine learning algorithm validation with a limited sample size. PLoS ONE, 14(11), e0224365. https://doi.org/10.1371/journal.pone.0224365