Knowledge Of Artificial Intelligence In Nursing Practice: A Meta-analysis Of Perception, Attitude, And Intention

Summary

This meta-analysis examined differences in nurses' perceptions, attitudes, and intentions regarding the use of artificial intelligence (AI) in patient care. Nurses who knew how AI is used in nursing practice had significantly higher perceptions, attitudes, and intentions than those who did not.

Abstract

This meta-analysis aimed to evaluate differences in perception, attitude, and intention regarding AI use in patient care between nurses with and without knowledge of how AI is applied in nursing practice.

We performed a meta-analysis using a continuous outcome model with either fixed- or random-effects methods to estimate the mean difference (MD) and 95% confidence intervals (CIs) for each outcome. We selected 9 studies with 3648 nurses for this meta-analysis. Nurses who know how AI is used in nursing practice had significantly higher perception (pooled raw MD, 1.43; 95% CI, 0.86–1.99, p < 0.001), attitude (MD, 1.80; 95% CI, 0.81–2.78, p < 0.001), and intention (MD, 2.89; 95% CI, 1.61–4.16, p < 0.001) compared to those who do not know how AI is used in nursing practice. However, heterogeneity was very high for all outcomes (I² = 91–98%), indicating substantial variation across studies. However, because these outcomes were measured using instruments with widely varying scale ranges (e.g., 5-point to 100-point scales), the pooled raw MD does not represent a consistent absolute difference. The consistent direction of effect across all studies (positive) is the primary finding, not the specific MD values.

Nurses who know how AI is used in nursing practice report more favorable perceptions, attitudes, and intentions than those who do not. However, due to high heterogeneity, scale variation, and cross-sectional designs, these findings are hypothesis-generating only. Causal claims are not justified.

Introduction

Particularly in the healthcare sector, artificial intelligence has become a crucial disruptive technology. Healthcare workflows, clinical outcomes, and patient care delivery are all being revolutionized by artificial intelligence¹. Nursing aims to provide compassionate, evidence-based care across clinical settings². Nurses must be ready to embrace any technology that enhances overall patient outcomes because they are frontline healthcare professionals. Artificial intelligence is likely to become integrated into nursing practice. To ensure the successful adoption of artificial intelligence in the healthcare sector, it is necessary to assess nurses' preparedness to embrace it³. Because artificial intelligence has the potential to advance healthcare through precise, individualized, and creative solutions, its significance is increasingly recognized⁴. With its enormous potential to improve clinical outcomes and increase efficiency in healthcare settings, the medical community has been excited about integrating artificial intelligence in its early stages of adaption⁵. However, despite increasing interest, there is little research on stakeholders' viewpoints regarding the application of artificial intelligence in healthcare⁶. Interest in AI implementation in healthcare varies across stakeholders. Notwithstanding the apparent advantages, there are a number of obstacles to implementing artificial intelligence in healthcare, including worries about data privacy, difficulty with regulatory compliance, and ethical dilemmas⁷. Researchers must close the knowledge gap regarding nurses' perspectives on the implementation of artificial intelligence to ensure the acceptance of evidence-based care in this context⁸. To address this gap, the present meta-analysis explicitly compares nurses who report knowing how AI is used in nursing practice with those who do not, across the domains of perception, attitude, and intention regarding AI in patient care. Research has to focus more on the revolutionary potential of artificial intelligence to boost output and develop novel delivery systems, as well as take into account alternative viewpoints regarding its application⁵. Uncertainty regarding the proper applications of artificial intelligence and problems with data quality are further obstacles to its adoption in the healthcare industry⁹. According to economic assessments, artificial intelligence improves medical quality and enables cost-effective methods, particularly in complex domains such as ophthalmology¹⁰. Addressing the obstacles to artificial intelligence's adoption is essential to ensuring that all of its advantages are fully realized. Upskilling the workforce, ethical issues, and the challenge of successfully using artificial intelligence technology in practical contexts are a few of the obstacles noted in earlier research¹¹. These obstacles all have a direct bearing on how prepared medical personnel are to adjust to artificial intelligence. With techniques such as fuzzy expert systems, Bayesian networks, artificial neural networks, and hybrid intelligent systems used in clinical settings to improve care delivery, artificial intelligence is becoming increasingly integrated into the healthcare industry¹². Additionally, sophisticated systems are being developed, such as deep-learning artificial intelligence systems, which may be able to automate processes and detect diseases¹³. As a result, the effects of artificial intelligence in the healthcare industry will be seen across a variety of disciplines and have enormous potential to improve clinical treatment, anticipate risks, and streamline the entire process¹⁴. Integrating artificial intelligence presents significant hurdles for nurses, including ethical issues, the need to adapt to new technology, and the potential impact on nursing responsibilities¹⁵. Nurses must acquire the knowledge and skills required to properly use artificial intelligence tools as their use advances¹⁶. According to research, emotional intelligence enhances nurses' interactions with artificial intelligence, thereby impacting both the quality of patient care and nurses' capacity to adapt to its use¹⁷. A critical factor that may influence how nurses perceive and engage with AI is their existing knowledge of how AI is actually applied in nursing practice. Without a basic understanding of AI's concrete functions—such as clinical decision support, predictive analytics, or automated monitoring—nurses may form opinions based on misinformation or a lack of exposure. Conversely, nurses with practical knowledge of AI's role in nursing practice may exhibit more favorable attitudes, greater perceived utility, and stronger intentions to adopt AI tools. Therefore, we selected "knowledge of how artificial intelligence is used in nursing practice" as the primary exposure variable to test whether this specific form of knowledge is associated with meaningful differences in perceptions, attitudes, and intentions. Therefore, we conducted a meta-analysis to evaluate nurses' perceptions, attitudes, and intentions regarding the use of artificial intelligence in patient care. The study closes important gaps in the literature on the attitudes, abilities, and knowledge required to apply artificial intelligence in patient care settings. The study provides plans to increase the integration of artificial intelligence in nursing by evaluating nurses' readiness and pinpointing critical areas for improvement.

Specifically, this meta-analysis was designed to compare three outcomes—perception, attitude, and intention toward AI in patient care—between two groups of nurses: those who report knowing how AI is used in nursing practice and those who do not.

This study's main goal was to find out how prepared registered nurses were to use artificial intelligence in their patient care procedures.

Protocol

Study design:

This study was a systematic review and meta-analysis conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. The completed PRISMA 2020 checklist is provided as Supplementary File 1. The protocol aimed to quantitatively synthesize evidence on nurses' perceptions, attitudes, and intentions regarding the use of Artificial Intelligence (AI) in patient care, comparing those with prior AI knowledge to those without¹⁸. Figure 1 illustrates the study selection process.

Eligibility criteria
Studies were selected based on the following PICOS framework¹⁹:

Population (P): Registered nurses or nurse managers of any clinical specialty or setting.

Intervention (I)/Exposure: Having knowledge, training, or awareness of AI use in nursing practice. 

Comparison (C): Nurses or nursing students who report not knowing how AI is used in nursing practice.

Outcomes (O): Quantitative measures of perception, attitude,  or intention toward AI in patient care, reported as mean scores with standard deviations (SD) or data convertible to such.

Study design (S): Observational studies (cross-sectional or cohort) and survey-based studies. Interventional studies (e.g., randomized controlled trials) were excluded because the exposure of interest—knowledge of how AI is used in nursing practice—is an existing characteristic rather than a randomized intervention.

Exclusion criteria:

Qualitative studies, review articles, editorials, letters, conference abstracts without full data, and non-peer-reviewed literature were excluded from this study. Studies not providing comparative quantitative data (e.g., mean, SD, sample size) for the two groups (with vs. without AI knowledge) were excluded. Studies where the population was not exclusively or primarily nurses (e.g., mixed healthcare professional groups without separable nurse data) were also excluded. No studies were excluded based on language.

Information sources and search strategy

A systematic search was performed across five electronic databases from their inception until July 31, 2025²⁰: PubMed, Cochrane Library, Embase, OVID, and Google Scholar.
The search strategy combined controlled vocabulary (e.g., MeSH terms) and free-text keywords related to three core concepts: (1) Nurses, (2) Artificial Intelligence, and (3) Perception/Attitude. The Boolean operators "AND" and "OR" were used to link terms within and between concepts. A sample search strategy for PubMed is provided in Table 1. The reference lists of all included studies and relevant reviews were manually screened to identify additional eligible publications²¹. The complete reproducible search strings for each database are provided below. Syntax was adapted to each database's required format, including appropriate field tags, controlled vocabulary (MeSH, EMTREE), and Boolean operators. No language or date restrictions were applied. The search was performed on July 31, 2025.

Study selection process

The study selection process followed the PRISMA flow diagram (see Figure 1). All retrieved records were imported into EndNote X9 (Clarivate Analytics) for deduplication. Two independent reviewers screened titles and abstracts against the eligibility criteria. Studies deemed potentially relevant by either reviewer proceeded to full-text review. The same two reviewers independently assessed the full texts of the shortlisted studies. Disagreements at any stage were resolved through discussion and consensus. No third reviewer was required. The final list of studies meeting all criteria was agreed upon by consensus.

Data extraction and management²²

A standardized, piloted data extraction form was developed in a spreadsheet. The two reviewers independently extracted data from each included study. Extracted data included study characteristics, population details, exposure definition, and outcome data.

Study characteristics: First author, publication year, country, study design, and sample size.

Population details: Nurse type (e.g., clinical, student, manager), clinical setting, mean age, gender distribution.

Exposure definition: How "knowledge of AI use in nursing practice" was defined and measured (e.g., specific training course, self-reported familiarity on a Likert scale).

Outcome data: For each relevant outcome (perception, attitude, intention), the mean score, standard deviation (SD), and sample size (n) for both the "AI-knowledgeable" and "non-AI-knowledgeable" groups were extracted. If means and standard deviations (SDs) were not directly reported, they were calculated from available statistics using methods described in the Cochrane Handbook for Systematic Reviews of Interventions (Version 6.4)¹⁹. Specifically:

From medians and interquartile ranges (IQR): The method of Wan et al. (2014)²³ was used to estimate means and SDs.

From 95% confidence intervals (CIs) and sample sizes: SDs were back-calculated using the formula: SD = √n × (upper CI limit – lower CI limit) / (2 × 1.96).

From p-values and sample sizes: SDs were estimated using the method described in the Cochrane Handbook (Section 6.5.2.3) when t-statistics or exact p-values were available.

Among the 9 included studies, three required data conversion because means and SDs were not reported in the format required for meta-analysis:

Study [Author, Year]: Reported medians and IQRs; converted using Wan et al.²³ method.

Study [Author, Year]: Reported only 95% CIs; SDs back calculated.

Study [Author, Year]: Reported means without SDs but provided p-values for group comparisons; SDs estimated from p-values.

The remaining 6 studies reported means and SDs directly and required no conversion. All converted values were verified by two reviewers independently. 

Risk of bias (quality) assessment²⁴

The methodological quality of the included observational studies was assessed independently by two reviewers using the Joanna Briggs Institute (JBI) critical appraisal checklist for analytical cross-sectional studies²⁵. This tool evaluates domains such as sample representativeness, exposure and outcome measurement, confounding, and statistical analysis. Each item was scored as "Yes," "No," "Unclear," or "Not Applicable." An overall study quality rating (High, Moderate, Low) was assigned based on consensus. Disagreements were resolved as described above.

Data synthesis and statistical analysis

Effect measure: The primary effect measure was the mean difference (MD) with a 95% confidence interval (CI). An MD > 0 indicated a higher score (e.g., more positive attitude) in the group with AI knowledge.

Meta-analysis model: A random-effects model was used for all primary analyses due to anticipated clinical and methodological heterogeneity across studies. A random-effect model was also applied as a sensitivity analysis^26,27.

Heterogeneity assessment: Statistical heterogeneity was quantified using the I² statistic. I² values of 25%, 50%, and 75% were interpreted as low, moderate, and high heterogeneity, respectively. Cochran's Q test (p < 0.10, indicating significant heterogeneity) was also consulted.

Subgroup and sensitivity analysis: Planned subgroup analyses were not feasible due to the limited number of studies (<10) per outcome. Sensitivity analyses were conducted by switching to a fixed-effect model. Sensitivity analyses were planned to assess the robustness of the pooled estimates, including excluding studies rated as low quality and switching to a fixed-effect model. A sensitivity analysis excluding the two studies rated as low quality (JBI score ≤4)^27,28 was performed. The pooled effect estimates for the remaining seven studies (n = 3,161) were similar to the main analysis for all outcomes, and all associations remained statistically significant (p < 0.001). Therefore, all 9 studies were retained in the primary analysis. Because excluding these studies did not materially alter the findings, they were retained in the primary analysis to maximize sample size and generalizability.

Publication bias assessment: Visual inspection of funnel plots for asymmetry was planned for outcomes including ≥10 studies. Because all outcomes included fewer than 10 studies, formal tests such as Egger's regression test are underpowered and typically not recommended. However, an exploratory Egger's regression test was performed for the outcome with the largest number of studies (attitude, n = 9) as a post-hoc analysis. Results should be interpreted with extreme caution due to low statistical power.

Software: All statistical analyses were performed using Review Manager (RevMan) software, version 5.4 (The Cochrane Collaboration, Copenhagen, Denmark).

Results

After examining 2245 pertinent publications, 9 studies that were published between 2021 and 2025 met the inclusion criteria^{26,27,28,29,30,31,32,33,34}. Table 2 summarizes the findings of these studies. A to...

Discussion

For the current meta-analysis, 9 studies with 3648 nurses were studied^{26,27,28,29,30,31,32,33,34}. This meta-analysis compared nurses who report knowing how AI is used in nursing practice with those who do not. Across nine ...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Cochrane Library	Cochrane Library	https://www.cochranelibrary.com/
Embase	Embase	https://www.embase.com/landing?status=grey
EndNote X9	Clarivate Analytics	https://support.clarivate.com/Endnote/s/?language=en_US	Reference management software for deduplication and citation organization
Google Scholar	Google	https://scholar.google.com/
Joanna Briggs Institute (JBI) Critical Appraisal Checklist	Joanna Briggs Institute	https://jbi.global/critical-appraisal-tools	Quality assessment tool for analytical cross-sectional studies
Microsoft Excel	Microsoft Corporation	https://www.microsoft.com/en-us	Data extraction form development; data management; conversion of statistics
OVID	OVID	https://www.ovid.com/
PRISMA 2020 Flow Diagram Template	PRISMA Working Group	https://www.prisma-statement.org/prisma-2020-flow-diagram	Template for study selection flow diagram
PubMed	National Institutes of Health	https://pubmed.ncbi.nlm.nih.gov/
Review Manager (RevMan)	The Cochrane Collaboration	Version 5.4	Software for preparing and maintaining Cochrane reviews; used for meta-analysis (pooling, forest plots, heterogeneity, sensitivity analyses).