Venous thromboembolism (VTE) is an important global health problem and the third most prevalent cardiovascular disorder (1, 2). VTE has a high incidence rate worldwide. In North America, Western Europe, southern Latin America (Argentina), and Australia, the yearly incidence rates of VTE vary from 0.75 to 2.69 per 1,000 people (3, 4). Compared to African-American populations, the overall VTE incidence in Asians and Asian Americans may be lower (5). However, with lifestyle changes, longer life spans, widespread use and increased sensitivity of imaging examination techniques (6), VTE incidence is increasing in Asian countries (5, 7).

In China, previous studies reported that the total yearly incidences of PE and DVT were 3.9 and 17.1 per 100,000 population, respectively, in 2001 (8). In 2011, the total yearly incidences of PE and DVT increased to 8.7 and 30.0 per 100,000 population, respectively, in China (7). Another study also found that the rates of VTE grew sharply in the south and north of China (9). In addition, Chinese patients’ hospitalization rate of VTE has gone up with time, rising from 3.2 per 100,000 people to 17.5 per 100,000 people (9). In total, VTE is becoming a common issue in China.

Venous thromboembolism is a preventable disease. Recommended preventive measures of guidelines help to lower the rates of VTE (10). Common preventive measures include pharmacological prevention such as the use of anticoagulant agents (low molecular weight heparin, etc.) and mechanical prevention such as the use of compression stockings and so on (10). Nevertheless, the prophylaxis rate nowadays is rather low (11). According to research, many hospitalized patients did not receive CHEST-recommended (12, 13) VTE prophylaxis in China, and only 14.3% of patients at risk got some prophylaxis, with about 10.3% obtaining the necessary prophylaxis. This means that physicians frequently fail to follow guidelines, thus indicating a huge gap between current Chinese clinical practice and mostly Western consensus guidelines (14).

Many interventions were introduced to help the implementation of thromboprophylaxis, including computerized tools (e.g., electronic reminders), education interventions (e.g., courses), audit and feedback strategies, paper-based tools (e.g., preprinted orders, posters), and multifaceted interventions (combination of interventions such as education, alert, and audit and feedback) (15–17). Studies showed that computerized tools intervention were useful for increasing VTE prophylaxis (17, 18). The computerized tool to prevent and control VTE in this paper refers to the Artificial Intelligence Clinical Assistant Decision Support System (AI-CDSS), which is a big data governance system based on natural language processing, machine learning, knowledge mapping, and other technologies. It is integrated into the hospital information system, and interfaces with laboratories, imaging detection and electronic medical records systems. For clinicians, it is a system which could help them to make a preliminary VTE risk assessment and intelligently provide suggestions on preventive measures based on the assessment results and the latest clinical practice guidelines and expert consensus. For nurses, it is a system which could help them automatically assess the VTE risk of patients, prompt them to implement clinicians’ instruction in time and offer them the patient personalized education materials about VTE (19).

However, when introducing a new artificial intelligence system into the hospital where did not use before, the acceptance of the system among nurses and clinicians is hindered by many factors (20, 21). It is vital to understand nurses’ and clinicians’ willingness to use the AI-CDSS and explore the influencial factors, thus helping the improvement of system design, promoting the implementation of the system, and helping the prevention of VTE.

Previous studies mainly focused on the effect of VTE intervention measures, including education interventions (22), information-based interventions (18, 23) and so on. However, few studies explored the acceptance of intervention measures among healthcare workers. In addition, there are some studies on the intention of healthcare workers to adopt the new system (24), but few studies on the acceptance of VTE prevention and control information systems among healthcare workers and even fewer studies on the acceptance of VTE information systems based on the extended the Unified Theory of Acceptance and Use of Technology (UTAUT) model among healthcare workers. Therefore, the purpose of this study is to explore the factors impacting nurses’ and clinicians’ acceptance of AI-CDSS to prevent and control VTE, and to evaluate the correlations among different factors based on the extended UTAUT model.

The UTAUT model, Information System Success Model (ISSM), and the top management support variable were used in this research. UTAUT can adapt to different scenarios of medical care by supplementing contextual constructs (25–27). However limited studies have attempted to add external factors such as system quality and information quality in ISSM into the original UTAUT model, and discuss the relationship among different variables. Furthermore, a study suggested that the antecedents of the main variables could better explain the acceptance of technology (28). However, there is little known about the influence of top management support as an antecedent on core variables in UTAUT. Therefore, this study adopted top management support as an antecedent in extending UTAUT to study further. Therefore, these three variables are added to extend the UTAUT model. It is the first attempt to integrate the top management support variable, ISSM, and UTAUT model.

Participants including nurses and clinicians were from three grade-A tertiary hospitals in Shanxi province, where the AI-CDSS for VTE prevention has been introduced or intended to be implemented. Grade-A tertiary hospitals were evaluated and classified by the local provincial health administrative department. Eligible participants must: (1) work in the hospital for at least 1 year; (2) have a registered license; (3) be volunteers. We excluded nurses and clinicians in the internship or advanced training period.

Participants were recruited through convenient sampling. The survey was conducted both online and offline from 15th October to 10th November 2023. We briefly introduced the research to participants and acquired their informed permission at the beginning. Participants completed surveys online through the survey platform Wenjuanxing¹. We requested hospital administrators to send questionnaires to the working contact group for nurses and clinicians. Simultaneously, some participants accepted the survey by printed questionnaires.

We distributed 700 questionnaires for the study. After a manual exclusion, including incomplete, duplicate, and missing questionnaires, 510 valid questionnaires eventually remained (valid response rate = 72.86%, 510/700).

The questionnaire was to investigate what factors were influencing healthcare workers’ acceptance of AI-CDSS for VTE prevention. The questionnaire contained two sections. The first section included basic information, involving age, gender, education, occupation, length of work, and so on. The second section contained 24 items covering 7 variables based on the extended UTAUT model: top management support, information quality, system quality, social influence, and effort expectancy, performance expectancy, behavioral intention. Each item was assessed on a 5-point Likert scale varying from 1 indicating strong disagreement to 5 indicating strong agreement. The measurement items of research variables are shown in Supplementary material.

A questionnaire needed several steps to guarantee its quality. To develop the instrument, we adapted scales from previous studies (29, 34, 38, 42, 46, 47, 51–53), and modified expressions to match the context of this study. Before launching the survey, 10 professionals from top-level positions within their hospitals were interviewed to evaluate the questionnaire’s clarity, contextual relevance, logical consistency, and terminology to improve the adapted items. Based on their feedback, additional modifications and adjustments were made to improve the instrument. Then, the revised questionnaires were sent to 75 clinicians and nurses for the pre-test. We made several modifications and ensured the final questionnaire. The pre-test findings suggested Cronbach’s α is in the range of 0.861 to 0.983, which is ≥0.7, showing the reliability of the items (54).

In UTAUT (29, 51–53), the performance expectancy variable had four items, the effort expectancy variable had three items, the social influence variable had four items, and the behavioral intention variable contained three items. The measuring constructs in ISSM (34, 38, 42) included information quality (three items), and system quality (four items). Finally, top management support (46, 47) was measured using three items.

The data was analyzed using structural equation modeling. This study followed the two-stage analytical method (55). We assessed the validity and reliability of the measurement model in the first phase. We assessed reliability by the composite reliability (CR), Cronbach’s alpha, and item-loadings of each item, convergent validity (CV) by the Average Variance Extracted (AVE), and discriminant validity (DV) by the square roots of AVEs (55, 56). The second phase involved examining the relationships between the constructs and evaluating structural equation model fitting. We did a analysis on the whole healthcare workers and a multi-group analysis on clinicians and nurses, respectively. We assessed the goodness-of-fit index (GFI), the incremental fit index (IFI), and so on for the fit of the model (55, 56). We applied AMOS 21.0 software to evaluate the significance of all hypotheses by analyzing T-statistics, p-values, and path coefficients. Data were imported into AMOS21.0 for analysis. p-value ≤ 0.05 indicated statistical significance (two-tailed).

In total, 510 participants were valid responses, including 96 male and 414 female. The sample comprised 186 clinicians (36.4%) and 324 nurses (63.6%). About half (46.7%, n = 238) were aged 26 to 35 years, followed by 36 to 45 years (26.3%, n = 134). Most participants (68.4%, n = 349) obtained an undergraduate degree, followed by a master’s degree or higher (29.2%, n = 149), and a junior college degree or lower (2.4%, n = 12). Concerning the working experience of participants in this study, 31.8% (n = 162) worked less than 5 years, 18% (n = 92) worked 5–10 years, 28.4% (n = 145) worked 10–15 years, 7.1% (n = 36) worked 15–20 years, and 14.7% (n = 75) worked more than 20 years. The detailed results are shown in Table 1.

Reliability assesses the internal consistency and stability of the questionnaire (56). CR, Cronbach’s α, and standardized factor loading were used to evaluate all variables’ reliability. Cronbach’s coefficients varied from 0.901 to 0.973, all of which were greater than 0.7 (56). The lowest CR value in all constructs was 0.915, which was more than 0.7 (57). The lowest standardized factor loading in all items was 0.593, greater than 0.5 (54). CV was measured by AVE and item loading in this study (57). AVEs of all constructs were greater than 0.5 (57). Therefore, reliability and CV were ensured in this model. Table 2 shows the detailed results.

The square roots of AVE could evaluate DV (56). The results showed that all the correlation coefficients of different variables were lower than the square roots of AVEs. Thus, DV is ensured. The statistics results are shown in Table 3.

We first analyzed the whole healthcare workers’ intention and its influencing factors based on the structural equation modeling. Figure 2 and Table 4 indicate hypothesis measuring results of whole healthcare workers, the model’s fit statistics revealed a good model fit (χ²/df = 2.599, RMSEA = 0.056, IFI = 0.976, GFI = 0.910, CFI = 0.976, and TLI = 0.972).

Specifically, performance expectancy (β = 0.45, P <0.001), effort expectancy (β = 0.21, P < 0.001), system quality (β = 0.14, P < 0.001), and top management support (β = 0.30, P < 0.001) positively influenced behavioral intention. Additionally, Social influence positively influenced performance expectancy (β = 0.52, P < 0.001), and information quality positively influenced system quality (β = 0.65, P < 0.001). Third, top management support was the crucial antecedent of performance expectancy (β = 0.41, P < 0.001), social influence (β = 0.57, P < 0.001), effort expectancy (β = 0.61, P < 0.001), system quality (β = 0.15, P < 0.001) and information quality (β = 0.59, P < 0.001).

Then, we further did the multi-group analysis of clinicians’ and nurses’ intentions. However, the results showed that the differences were significant among different occupations [P (structural weights) = 0.00 < 0.05, P (Structural residuals) = 0.00 < 0.05, P (measurement residuals) = 0.00 < 0.05]. Therefore, there were occupation differences in the model. Table 5 shows the multi-group analysis results of clinicians, Table 6 shows the multi-group analysis results of nurses.

According to Tables 5, 6, H1, H2, H4, H7, H8, H9, H10, H11, and H13 were also supported in multi-group models. However, H3, H5, H6, and H12 were different in multi-group models. In H3, social influence did not influence nurses’ behavioral intention (β = 0.06, P = 0.376), but negatively influenced clinicians’ behavioral intention (β = −0.19, p < 0.001). In H5, system quality positively influenced nurses’ behavioral intention (β = 0.16, p < 0.001), but did not affect clinicians’ behavioral intention in the clinicians’ model (β = 0.09, P = 0.187). In H6, information quality did not influence nurses’ behavioral intention (β = −0.02 , p = 0.679), but positively influenced clinicians’ behavioral intention (β = 0.15, p = 0.025). In H12, top management support influenced system quality in the clinicians’ model (β = 0.26, P < 0.001), but the relation was not obvious in the nurses’ model (β = 0.09, p = 0.093).