The cross-sectional design and convenience sampling methods were adopted, with the aim of evaluating the different subtypes of nurses’ innovative behavior by ranking the importance of determinants. We adhered to Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines and methodology in reports of cross-sectional studies (details in Supplementary material).

In the context of this cross-sectional investigation, data were collected from August through December 2024, utilizing the professional survey platform, Questionnaire Star. The participants were recruited from six tertiary public hospitals in Zhejiang province, China.

The criteria for participant inclusion were stringently defined. Inclusion criteria were as follows: (1) holding professional qualification certificates for nurses issued by the People’s Republic of China; (2) working with the head nurse of the department for 1 year or more; (3) engaging in front-line clinical nursing work; (4) participating in this study voluntarily. The exclusion criteria were as follows: (1) nurses who were on vacation at the time of the survey, out for training, sick leave, or business leave; (2) nursing managers.

To enhance the reliability of the latent profile analysis, a sample size of at least 500 is advocated in research practice. Our study, with a sample size of 832, meets and exceeds this recommendation, thereby strengthening the confidence in our analytic results (Nylund et al., 2007).

The investigators coordinated with the hospital management department, and with the consent of the department head, the questionnaires were distributed to the nurses. Prior to commencing the formal survey, the researchers provided participants with a detailed explanation of the survey’s objectives and content. They were explicitly informed of their right to voluntarily participate or decline to answer any questions. Participants were assured that the survey was confidential, with all data collected being used solely for scientific research. In total, 890 questionnaires were distributed, and 832 valid questionnaires were collected (excluding those with obvious errors, incorrect answers to lie tests, or incomplete responses). The effective response rate was 92.47%.

Ethical approval for this study was obtained from the School of Public Health and Nursing, Hangzhou Normal University, Hangzhou, China (ethical review NO. 2024039), based on the principles of the Declaration of Helsinki. All eligible nurses were informed of the study and its ethical principles. Participants were fully informed about the study and voluntarily participated. Specifically, to protect the privacy of the participants, this study anonymized them. The questionnaires did not collect any information that could identify individuals. The collected data was stored in password-protected computers and backed up through encrypted storage devices. The data will be retained for 5 years after the final publication of the paper for possible verification. After the retention period expires, the data will be permanently deleted from all storage devices.

The statistical software M-plus 8.3, IBM SPSS 26.0 and R4.5.1 were used for data analysis.

The M-plus 8.3 software was employed in the LPA process. The LPA was used to identify subgroups of individuals who share a similar profile of scores on nurses’ innovative behavior. The mean variables of nurses’ innovative behavior were measured to assess the trait indicator variables. The Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), adjusted BIC (aBIC), and an Entropy test were used to select the model. The entropy (entropy) is a measure of the accuracy of model classification is an evaluation index of model classification accuracy. The Bootstrap Likelihood Ratio Test (BLRT), Lomond-Dale-Rubin corrected likelihood ratio (LMR) and adjusted LMR (ALMR), were used to compare the difference in fit between the models (Vrieze, 2012). A stepwise approach was used to determine the number of latent profiles that best characterize the data and sample. The best model is characterized by the following criteria: when AIC, BIC, and aBIC in the model are minimized; entropy value is greater than 0.8; the statistical significance of LMR, ALMR, and BLRT is less than 0.05 indicates that the k-category model is better than the k-1-category model; the profiles are both concise and interpretable (Wang and Bi, 2018).

IBM SPSS 26.0 software was employed to conduct statistical analysis. Descriptive statistics were presented as frequency and percentage for categorical data, and mean and standard deviation for continuous data. To identify differences in innovative behavior profiles based on socio-demographic and professional characteristics, Pearson’s χ² test was used for categorical variables, and one-way ANOVA was used for continuous variables. Prior to conducting ANOVA, its assumptions were assessed. The normality of the distribution for all continuous variables (including Likert-scale scores and subscales) was evaluated by examining skewness and kurtosis; In this study, the absolute values were, respectively, lower than 2 and 4, indicating that the data conformed to a normal distribution. Homogeneity of variances was tested using Levene’s test. If the normality assumption was violated, the non-parametric Kruskal-Wallis test was applied as an alternative. For comparisons that met the normality assumption but violated homogeneity of variances (Levene’s test p < 0.05), Welch’s ANOVA was reported.

In order to deeply explore the key predictors affecting different innovation behavior profiles, this study used Random Forest machine learning algorithm with R4.5.1 and R Studio to rank the factors influencing innovative behaviors of different subgroups (C1, C2, and C3) in terms of importance respectively, and assessed their innovative performance and feature importance.

Data preprocessing and sampling strategy: The analysis was based on the three populations that were classified by the latent profile analysis described above. For each profile, we first constructed a dichotomous dataset (e.g., for the “C1” subgroup, labeled as “C1” = 1 and “non-C1” = 0). To effectively address the category imbalance problem, a stratified sampling strategy is used in the model training phase to ensure that the sample sizes of the target and non-target categories in the training set are equal, which can optimize the model’s ability to recognize a small number of categories. The dataset was randomly divided into training set and test set in the ratio of 7:3.

Since there are many variable questionnaires involved in this study, all filled in by the same respondent, there may be common method variance. This study was tested by the Harman univariate test. The results showed that the percentage of variance explained by the first common factor was 31.016%, less than 40%, and there were 11 common factor eigenvalues greater than 1, so there was no serious common method bias problem (Tang and Wen, 2020).

Descriptive and univariate analyses of innovation behavior among nurses were performed based on data collected from 832 participants. The majority of participants in this study were women (n = 794, 95.43%), predominantly aged between 31 and 40 years (n = 401, 48.20%). Other general information is shown in Table 1.

The score of each entry of 832 nurses’ innovative behavior was (3.58 ± 0.48) and the scores of the entries of the three dimensions Generating Ideas, Realizing Ideas, and Obtaining Support were 3.74 ± 0.63, 3.54 ± 0.68, and 3.49 ± 0.63, respectively. In this study, 10 entry scores of nurses’ innovative behavior were used as exogenous indexes, and 1 to 5 models were fitted sequentially, and the results of the fitting indices of each model are shown in Table 2. The AIC, BIC and aBIC indices decreased with the increase of the number of categories, the entropy of categories 2 to 5 were all >0.800, the LMR and LMRT indices of categories 2 and 3 were significant, and when the number of categories was 4, LMR(P) was not significant. The entropy was 0.841 when the number of categories was 3, and the LMR and BLRT indices reached a significant level (p < 0.001). After comprehensively comparing the fitting indexes of each model, Model 3-profile was selected as the best fitting model. We analyzed the posterior probability for the 3 categories, reaching 0.955, 0.929, and 0.915, respectively, which indicated the results of the optimal model obtained are reliable and have a high discriminative ability.

Based on the classification results, the latent profiles of nurses’ innovative behavior were plotted in Figure 1. The horizontal axis represents the 10 items of nurse innovative behavior and the vertical axis represents their corresponding scores. The higher the item score, the better the level of innovative behavior of nurses. Meanwhile, the profiles were named by combining the characteristics of the item scores. From the results in Figure 1, it can be seen that there are 111 nurses in profile C1 (13.34%). The lowest scores on all dimensions indicate that this subgroup tends to be cautious and traditional in its innovative behavior, with an initial awareness of the problem, but a weak willingness and ability to explore and analyze solutions on their own initiative, and the least ability to turn ideas into reality. Hence the name “Innovation-Conservative”; There are 545 nurses in profile C2 (65.50%). This subgroup scores moderately well on all dimensions. It is a stable, reliable group that lacks prominence and is the backbone of the nurse community that is actively experimenting and exploring new approaches. Hence the name “Innovation-Exploratory”. There are 176 nurses in profile C3 (21.16%). This subgroup is much higher than the other two groups on all dimensions. They are a source of new ideas and are extremely good at promoting their ideas, obtaining support and developing solutions, but even so, finalizing the idea is the most challenging part of the innovation process. As such, this subgroup is a key driver and leader of innovation. Hence the name “Innovation-Driven.”

Table 1 lists the socio-demographic and professional characteristics associated with latent profiles of innovative behavior. The results of univariate analysis showed that there were significantly different in the three latent profiles: age (p < 0.001), marital status (p < 0.05), working years (p < 0.001), education level (p < 0.001), professional title (p < 0.001), average monthly income (p < 0.001), satisfaction with scheduling (p < 0.001), experience of innovation (p < 0.01), participation in scientific research training (p < 0.01), participation in academic conferences (p < 0.01), participation in teaching innovation competition (p < 0.01), publication of academic articles (p < 0.01), paradoxical leadership (p < 0.001), creative self-efficacy (p < 0.001), and work engagement (p < 0.001). The scores of all dimensions of nurses’ innovative behavior, paradoxical leadership, creative self-efficacy, and work engagement in the “Innovation-Conservative” were significantly lower than the overall level. Nurses in the “Innovation-Conservative” accounted for a larger proportion in terms of age 18–30, working years <5 years, relatively low professional title, lower monthly income, and lack of innovative experiences. The proportion of nurses in the “Innovation-Exploratory” was the largest and the nurses in this subgroup were mainly composed of 31–40 year-olds, quite satisfied with nurse scheduling, and had worked for 11–20 years. The score of each dimension in the “Innovation-Driven” was the highest. In this subgroup, the composition of the nurses was dominated by those with the title of supervisor nurse or above, more than 20 years of working, higher income levels, and richer experience in innovation. Furthermore, compared to the “Innovation-Conservative” and “Innovation-Exploratory” subgroups, the “Innovation-Driven” had significantly higher scores in paradoxical leadership, creative self-efficacy, and work engagement.

For the random forest model performance test, we measured the accuracy value of 0.90, 0.71, and 0.75, sensitivity value of 0.85, 0.75, and 0.75, and AUC of 0.94, 0.76, and 0.82 for the three latent classes of Innovation-Conservative, Innovation-Exploratory, and Innovation-Driven. The results of these metrics are acceptable (Huang and Ling, 2005).

After training the algorithm, the variables’ feature importance scores were plotted for “Innovation-Conservative,” “Innovation-Exploratory,” and “Innovation-Driven.” The feature importance ranking is demonstrated in Figures 2–4, respectively. The y-axis represents the explanatory variables, while the x-axis represents the feature importance score. A total of 15 variables were ranked from highest to lowest.

Despite differences in specific rankings, creative self-efficacy, work engagement, and paradoxical leadership were consistently identified as the top three most critical features across all subgroups, with differences in subsequent rankings: the variables of working years, average monthly income, and satisfaction with scheduling, while all ranking in the top five of all subgroups, differed in their specific importance rankings.

For the “Innovation-Conservative,” the five most important predictors are listed in descending order: creative self-efficacy (Mean Decrease Gini: 19.82, p < 0.01), work engagement (Mean Decrease Gini: 16.00, p < 0.01), paradoxical leadership (Mean Decrease Gini: 15.77, p < 0.01), working years (Mean Decrease Gini: 5.46, p < 0.01), and average monthly income (Mean Decrease Gini: 3.03, P >0.05). Other significant predictors included professional title (Mean Decrease Gini: 2.87, p < 0.01), age (Mean Decrease Gini: 2.76, p < 0.05), and education level (Mean Decrease Gini: 1.83, p < 0.01), seeing Figure 2.

For the “Innovation-Exploratory,” the five most important predictors were: creative self-efficacy (Mean Decrease Gini: 38.16, p < 0.01), paradoxical leadership (Mean Decrease Gini: 32.33, p < 0.05), work engagement (Mean Decrease Gini: 31.43, p < 0.01), average monthly income (Mean Decrease Gini: 12.86, p < 0.05), and satisfaction with scheduling (Mean Decrease Gini: 11.66, p > 0.05). The predictors working years (Mean Decrease Gini: 9.87, p < 0.01) and age (Mean Decrease Gini: 9.08, p < 0.01) were also statistically significant, seeing Figure 3.

For the “Innovation-Driven,” the five most important predictors were: creative self-efficacy (Mean Decrease Gini: 34.39, p < 0.01), work engagement (Mean Decrease Gini: 18.76, p < 0.01), paradoxical leadership (Mean Decrease Gini: 17.88, p < 0.01), satisfaction with scheduling (Mean Decrease Gini: 6.90, p < 0.05), and average monthly income (Mean Decrease Gini: 6.56, p < 0.05). Significant predictors also included working years (Mean Decrease Gini: 5.88, p < 0.01), age (Mean Decrease Gini: 5.55, p < 0.01), professional title (Mean Decrease Gini: 4.33, p < 0.05), and publication of academic articles (Mean Decrease Gini: 2.30, p < 0.05), seeing Figure 4.

The results of this survey showed that the score of the innovative behavior scale for nurses was (35.79 ± 4.84), which is slightly above the average of previous studies (Wang et al., 2024; Lu et al., 2025; Han and Ma, 2023). Moreover, regarding the latent profile analysis, our findings revealed a distinct distribution of nurses across the three profiles Innovation-Conservative (13.34%), Innovation-Exploratory (65.50%), and Innovation-Driven (21.16%). The proportion of nurses in the Innovation-Conservative subgroup was relatively small, while that in the Innovation-Exploratory subgroup was relatively large. This distribution differs notably from the patterns identified in other studies employing LPA (Li et al., 2025; Fu et al., 2024). The differences in these subgroup distributions may be attributed to variations in hospital levels, and regional socio-economic backgrounds. Compared with previous studies, our research was conducted in public tertiary hospitals in Zhejiang Province, which is a region renowned for its economic development and advanced medical system. Meanwhile, tertiary hospitals are equipped with advanced facilities, advocate an open and inclusive working atmosphere that encourages exploration, and attach great importance to scientific research activities and teaching quality, providing a favorable environment for nurses’ innovative behavior.

Creative self-efficacy is the most significant factor showing a positive association with nurses’ innovative behaviors across all subgroups. Theoretically, it is positively related to intrinsic motivation and willingness to engage in innovative behavior. It is also a critical psychological resource for individuals to realize creative work and carries important practical implications for guiding individual creative work. Studies indicate that nurses with high creative self-efficacy are more likely to actively explore new methods, put forward improvement suggestions, and try innovative strategies to solve problems (Yang et al., 2024). From the perspective of COR theory, innovative behavior does not occur naturally. Essentially, it requires nurses to invest their precious personal resources (e.g., time, energy, and cognitive efforts) while also facing uncertain returns and potential failure risks. Therefore, nurses’ decisions to engage in innovation are linked to their assessment of personal resource status. As a key personal resource, creative self-efficacy represents an important component of nurses’ personal resource pool. Nurses with higher creative self-efficacy may possess greater confidence to consider innovative, high-risk investments, as they are more likely to believe in a higher probability of success, potentially lowering their perceived investment risk. Thus, when nurses have a high level of creative self-efficacy and believe they possess the knowledge and skills to support creativity, they may be more autonomous in accepting work challenges and engaging in creative work practices.

Paradoxical leadership employs balanced thinking within chaotic contexts, flexibly adapting styles to achieve “both-and” outcomes. The personalized care and support characteristics of paradoxical leadership align with the needs of contemporary nurses (Qing et al., 2023). Li and Wang (2021) hold that paradoxical leadership is associated with reduced internal team conflict and carries positive implications for team innovative capacity. Leaders with paradoxical leadership could assist nurses in find a balance among multiple contradictions, such as standardized processes and personalized care, efficiency and safety, strict management, and humanistic care. Specifically, this leadership approach may be consistent with addressing individual nurse needs while maintaining managerial efficiency, supporting creativity within standardized workflows, and facilitating evidence-based innovation initiatives from a practical standpoint. Moreover, paradoxical leadership helps nurses better cope with the complexity and uncertainty in their work, reduces role conflicts and stress, and creates an environment that embraces contradictions, encourages attempts, and allows for failure, which is more likely to lower the risk of potential losses from innovative investments. Under the paradoxical leadership style, nurses might perceive that unsuccessful innovative attempts would not lead to the loss of key resources such as leader recognition or professional reputation, which may in turn be associated with reduced hesitation to engage in innovative behaviors.

Work engagement represents a positive work state distinct from job burnout, characterized by sustained energy and passion toward work (Bakker et al., 2008). Work engagement is negatively associated with negative emotions and passive work behaviors commonly observed among nurses experiencing job burnout and work pressure (Lv et al., 2022). Innovative behavior is a complex behavior, which is a series of attempts with uncertainties and risks, requiring knowledge, motivation, and substantial resource investment. Meanwhile, compared with ordinary nurses, those with such a positive state tend to devote more time and energy to their work and show more tolerance for failure and setbacks. Highly engaged nurses demonstrate greater persistence through challenges, actively seek solutions, and invest discretionary effort. These characteristics of nurses are positively associated with the likelihood of engaging in innovative behaviors among nurses (Janssen, 2000). Furthermore, when nurses possess both abundant internal resources and reliable external safeguards, that is, high creative self-efficacy and paradoxical leadership, they tend to enter a state of abundant resources and a sense of security, which is manifested as a high level of organizational commitment (Ling et al., 2012). In such a state, nurses could be more willing to invest their available personal resources to pursue future work-related benefits, including a sense of achievement at work, professional recognition, and career development. This kind of investment tendency align with a greater likelihood of innovative behavior among nursing staff. Conversely, nurses lacking these resources may exhibit a state of resource defense, striving to protect their already scarce resources, and consequently showing a lower tendency to engage in activities that may consume resources, such as innovation. This pattern of associations is consistent with the observation that the Innovation-Conservative subgroup reported the lowest levels of creative self-efficacy, paradoxical leadership, and work engagement.

Beyond these organizational and psychological factors, the random forest feature importance analysis revealed that each subgroup of innovative behavior has its relatively important influencing factors such as working years, average monthly income and satisfaction with nursing work scheduling. Nurses with long working years and high professional title typically possess extensive clinical experience and robust professional knowledge. They care for a large number of patients, hold experience in handling emergencies, and are well familiar with the strengths and weaknesses of various work processes. Innovation is often not about imagination out of thin air, but rather based on the identification and improvement of existing problems. The acute discovery of problems in clinical practice and the generation of innovative ideas are positively linked to rich professional experience. In addition, nurses with long-term clinical practice accumulation are often better able to assess the feasibility of innovative ideas from a practical perspective, which is associated with lower potential risks in the innovation process and makes it more likely for innovative attempts to succeed. Moreover, nurses with rich working experience typically have accumulated a deep professional reputation and interpersonal relationships in their past work. The innovative ideas they propose may be more likely to receive support from colleagues and leaders, which is conducive to the implementation of innovative ideas.

Average monthly income emerged as another factor associated with differential innovative behavior among the nurse profiles. In the Innovation-Conservative subgroup, nurses are usually in the early stage of their careers, characterized by lower professional titles and thus lower incomes. According to the COR theory, nurses with less resource reserves are more prone to the pressure of resource loss. Moreover, the existence of this pressure is linked to the resources running behind their expenses, potentially associated with gradual declines in available resources and gradually increasing the psychological pressure perceived by individuals. Therefore, nurses with lower monthly incomes may have higher survival anxiety and tend to prioritize actions aimed at preventing resource loss, which may create additional barriers to carry out resource investment activities such as innovation. Furthermore, nurses facing financial constraints typically exhibit lower risk tolerance. For these individuals, failure could be associated with additional pressure and negative evaluations from leaders, which may be linked to greater hesitancy toward innovative attempts. Unlike the Innovation-Conservative subgroup, basic livelihood needs among nurses in the Innovation-Exploratory subgroup are generally met, and income serves as an incentive that aligns with organizational recognition and affirmation of their value. This sense of recognition is positively related to stronger organizational commitment and professional identity, which are associated with greater engagement in innovative and other extra-role behaviors as a form of reciprocal contribution to the organization (Joo and Hong, 2026). Ultimately, within the Innovation-Driven subgroup, a competitive income likely acts as a sustaining factor. It helps to bolster their sense of belonging and high engagement, which may be essential for maintaining their long-term innovative output and preventing burnout.

Collectively, our findings point to the potential importance of optimizing the work environment, particularly work scheduling, in fostering nurses’ innovative behavior. Reasonable and efficient work scheduling could help nurses better balance work and life, ensuring adequate rest to maintain the energy required for innovative efforts. Reasonable work scheduling generally involves a certain degree of flexibility and employee participation, which is associated with higher perceived work autonomy among nurses and is positively related to their intrinsic motivation for innovation. For the nurses in the Innovation-Driven subgroup, they scored very high in both the dimensions of generating ideas and obtaining support. However, they encountered obstacles when implementing their ideas. Their scores in item 8 were relatively low, indicating that nurses are likely to have some difficulties in applying the innovative implementation plan to their work. Meanwhile, random forest analysis demonstrates that the satisfaction with work scheduling is a relatively top-ranked factor among the influencing factors of Innovation-Driven subgroup, suggesting that their demand for work scheduling may reflect a sense of job control when implementing innovative behaviors (Joachim Gerich and Weber, 2020). When translating ideas into practice, nurses tend to benefit from greater autonomy in time arrangement, which supports the continuous progression of innovation initiatives. Nurses with a higher sense of control over their work tend to show greater enthusiasm and initiative towards innovative behaviors.

In the random forest analysis, variables reflecting prior innovative experiences such as experience of innovation and publication of academic articles consistently ranked in the lower half of feature importance across all three innovation subgroups. This pattern aligns with COR, which posits that individuals prioritize behaviors supported by immediately available resources. Personal resources like creative self-efficacy, energy resources reflected in work engagement, and conditional resources provided by paradoxical leadership represent immediate, active assets that directly support present innovation efforts. In contrast, past innovative experiences may not constitute presently available resources. Completing a publication or innovation project typically consumes substantial time, cognitive energy, and emotional investment at the time of occurrence. This suggests that the resources associated with past innovative activities may show a weaker correlation with current innovative behavior when not supported by ongoing contextual resources, aligning with their lower predictive importance in our model. Consequently, when proximal resource indicators are included in predictive models, historical experiential variables contribute minimally to explaining current innovation levels. Practically, clinical nursing environments prioritize real time problem solving under time pressure. Furthermore, workload pressures and institutional structures may frequently limit nurses’ opportunities to accumulate publication or competition experiences, making these metrics less representative of frontline innovative capacity. Consequently, experiential indicators demonstrate weaker predictive relevance in this context.

This study provides valuable insights into the factors that influence innovative behaviors among nurses. However, in order to contextualize the findings and guide future research, it is important to recognize some limitations. Firstly, this study used a cross-sectional design that involved surveys at a specific point in time. To explore the influencing factors of innovative behaviors of nurses in different subgroups, it is necessary to analyze these variables at different time frames. Future research should use longitudinal studies to empirically analyze the causal relationships between mechanisms. Secondly, the use of a convenience sample from six tertiary hospitals in Zhejiang Province introduces a risk of selection bias, which limits the generalizability of our findings. The results may not be fully representative of nurses working in other regions or in different levels of healthcare institutions. Future studies should employ more diverse and representative sampling strategies across various cultural and healthcare settings to verify the universality of the conclusions.