We followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines to report our systematic review and meta-analysis [23]. Our PRISMA checklist is provided in Multimedia Appendix 1 [23]. The protocol of our study was registered in the PROSPERO (International Prospective Register of Systematic Reviews).

We adhered to the PRISMA-S (PRISMA Search Reporting Extension) checklist [24]. The reviewer (YZ) searched the CENTRAL and CDSR (via the Wiley platform), MEDLINE (via the PubMed platform), CINAHL (via the EBSCO platform), Web of Science Core Collection (via the Web of Science platform), Embase (via the Ovidsp platform), and Chinese Academic Journal (via the CNIK platform) databases from inception to December 31, 2022, with no restrictions on language or year of publication. The search strategy included terms related to hand hygiene and intelligent technology. Our strategy was developed in consultation with a medical research librarian. Textbox S1 in Multimedia Appendix 2 details the search strategies of databases. ClinicalTrials.gov and the World Health Organization International Clinical Trials Registry Platform were searched for ongoing and unpublished trials. In addition, we manually searched the references of the collected articles and systematic reviews.

Articles were eligible for inclusion in the meta-analysis if they met all of the following criteria: (1) were randomized controlled trials (RCTs) or quasi-experimental studies; (2) included HCWs or adult patients (18 years or older) as participants; (3) evaluated the effectiveness of an intelligent technology–related intervention alone or in combination with usual care compared with placebo or usual methods; and (4) reported at least one clinical end point such as HHC rates, HCAIs rates, or multidrug-resistant organism (MDRO) detection rates (Table 1).

Articles were excluded if they met any of the following criteria: (1) failed to provide the full text and the abstract provided insufficient information, (2) had insufficient or incorrect data, or (3) were duplicate studies.

The data management software EndNote X9 (Clarivate Analytics) was used. Two reviewers (XBQ and KL) independently screened the titles and abstracts for eligibility. Articles were retrieved in full upon request from the reviewers. Then, the reviewers independently screened the full texts and resolved disagreements through discussion. If they could not reach an agreement, another author (YYZ) was consulted, and a decision was made by a majority vote.

Data were extracted independently by 2 authors (YZ and YWL) using predetermined forms (Table S1 in Multimedia Appendix 2). The following data were collected: authors, year of publication, country, study design, setting, participants, intelligent technology intervention, data collection period, and study outcomes.

Two reviewers conducted the risk of bias assessment using the Cochrane risk-of-bias methodology [25] for RCTs and the ROBINS-I (Risk Of Bias In Non-randomised Studies—of Interventions) tool for nonrandomized intervention studies [26]. The assessment tools were both developed by the Cochrane Collaboration. In addition, we used the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach to classify the certainty of evidence into high, moderate, low, or very low for each outcome [27].

Meta-analysis was performed using Review Manager (RevMan, Version 5.3; the Nordic Cochrane Centre, the Cochrane Collaboration, 2014, Copenhagen, Denmark) and STATA Version 15.1 (version 15.1.629; StataCorp). Heterogeneity among studies was assessed using the chi-square test, and Ι² values were used to determine heterogeneity across studies. A random- or fixed-effects model was used to calculate the pooled effect sizes and corresponding 95% CIs based on the heterogeneity. If Ι²≤50%, which represented homogeneity, fixed-effects models were selected. If Ι²>50%, which indicated substantial heterogeneity of the effects, random-effects models were applied [28]. For continuous data, the mean difference and 95% CI were assessed for the pooled outcomes, and for dichotomous outcomes, the odds ratio and 95% CI were used in accordance with intent-to-treat principles. A forest plot was generated to represent the meta-analysis results. To gain insight into the sources of substantial heterogeneity, prespecified meta-regression was conducted with the following covariates: article publication year, study design, setting, and intelligent technology intervention (different components of the intelligent technology system). The sensitivity analysis was performed by eliminating studies to assess whether the results were stable [29]. If more than 10 studies were included in the analysis of outcomes, funnel plots were constructed to identify publication bias by Egger tests (with P<.05 considered significant) [29].

The PRISMA flowchart depicts the extensive search process (Figure 1). We identified 16,791 articles and reviewed 8571 unique titles and abstracts (after removing duplicates across databases) and 440 full-text articles, with 36 studies meeting the predefined eligibility criteria [15,20,22,23,30-61]. Study characteristics related to the population, interventions, and outcomes for the 36 included studies are provided in Table S2 in Multimedia Appendix 2 [15,19,21,22,30-60,62]. The 36 unique articles included 2 (6%) RCTs [53,54] and 34 (94%) quasi-experimental studies [15,19,21,22,30-52,55-60,62] (8 non-RCTs and 26 one-group pretest-posttest quasi-experimental designs). Studies were published between 2013 and 2022, apart from 2 studies published in 2008 [48,58]. Twenty-five (69%) studies were published in the past 5 years. The demographic information of participants was provided in only 5 (14%) studies. Most studies (31/36, 86%) recruited HCWs from hospitals or clinics.

Table S3 in Multimedia Appendix 2 [15,19,21,22,30-52,55-60,62] provides a summary of the risk-of-bias assessment for all included nonrandomized intervention studies (n=34) based on the ROBINS-I tool. Nine articles were evaluated as having serious biases, and the other studies were evaluated as having moderate biases. The risk-of-bias assessment for the RCT studies is provided in Figure S1 in Multimedia Appendix 2 [53,54]. Two studies all showed moderate biases [53,54]. Overall, the quality of the studies was deemed as having high bias.

This systematic review appraised evidence from 36 studies evaluating the effects of intelligent technology interventions for HH on the behavior of HCWs, nosocomial infection rates, and MDRO detection rates. All studies, except 2 [48,58], were conducted after 2013, indicating a growing interest in applying intelligent technology for the management of HH. Our synthesized findings from the meta-analysis suggested that intelligent technology interventions for HH had a positive effect on HHC and contributed to the decrease in HCAI rates. This may be because most of the intelligent technology interventions provided reminders and real-time feedback to improve the HH awareness and habits of HCWs [50,61], further reducing infections [63]. However, our study could not determine the sustainability of the impact of intelligent technology interventions on HHC. Studies have shown that after the abolition of intelligent technology interventions, HHC dropped significantly, and the intelligent interaction between equipment and HCWs and direct personal feedback were important methods for improving the sustainability of HHC [64,65]. According to the analysis of the characteristics of the included literature, there were various methods, including instant prompts and feedback. Nevertheless, each type of reminder was associated with specific drawbacks, such as audible reminders that could interrupt a patient’s rest [66,67]. It was also challenging to effectively provide feedback to help HCWs understand the situation according to their needs and different educational backgrounds [68,69].

In contrast, our research focused on the outcomes of intelligent technology interventions for HH and showed no effect on MDRO detection rates. However, these results must be interpreted with caution because of the statistical heterogeneity (>90%), heterogeneity in terms of publication year, study design and interventions delivered (type of components), and unstable sensitivity analysis. At present, there is still controversy about the relationship between the improvement of HH behaviors and the detection rate of MDRO [70]. Studies found that the change in the MDRO detection rate was related to the length of time to improve HHC [71,72]. Improving HHC in a short period of time had no effect on the MDRO detection rate, and there was a delay effect. Studies have pointed out that this may be due to the nonlinear relationship between HHC and MDRO prevalence [72].

We were aware of 4 reviews of intelligent technology interventions for HH on the outcomes of HHC and HCAI rates. Previous systematic reviews led to different and incomplete conclusions. Two studies evaluated published articles indicating that technology systems could significantly improve HHC among health care professionals [65,73], in agreement with our results. However, the review by Srigley et al [61] indicated that 1 study evaluating the impact of electronic and video monitoring systems with a minimal potential for bias presented the smallest effect for HHC. We found only 1 systematic review showing that electronic and video monitoring systems have the potential to prevent HCAIs, but the results are not supported by sufficient data [74].

In addition to the 3 outcomes of this study, we noticed that some studies focused on the barriers to the application of intelligent technology for HH. Two systematic reviews found that usage anxiety, privacy, and confidentiality were key elements influencing the acceptance of intelligent technology interventions by HCWs [65,75]. They were concerned about potential risks posed by intelligent technology such as wearable sensors that could cause hand contamination and radio-frequency interference [76,77]. Some of HCWs perceived that these intelligent technology interventions using video cameras to monitor all 5 moments of HH would invade their and the patient’s privacy [11]. There were studies that suggested that a camera could be placed on the chest of HCWs that was aimed at their hands rather than installing cameras in the environment [78,79]. However, none of the included studies mentioned privacy protection before implementing intelligent technology interventions for HH. Another systematic review, which discussed costs, found that implementing intelligent technology interventions for HH in health care facilities would entail high costs, including equipment installation and maintenance costs, and that it was not realistic to install the camera system in community settings [4].

We assessed the quality of evidence for this study based on the GRADE classification. The quality of the evidence for the outcomes of the 3 studies was low, most of which were downgraded because of high risk of bias and imprecision. Therefore, the quality of evidence in this meta-analysis was low, and the results of the present meta-analysis were not strongly recommended.

Several limitations should be noted in this meta-analysis. First, most of the included studies were quasi-experimental studies. Although only 2 of the studies conducted an RCT, blinding was difficult to implement because of the nature of intelligent technology interventions, which may potentially result in performance bias. Second, high heterogeneity was identified among studies in terms of results for the HHC and MDRO detection rates. Part of the heterogeneity may be due to differentiation in terms of populations and inconsistent inclusion and exclusion criteria among studies. Other sources of heterogeneity may be due to the diversity of intelligent technology interventions for HH and the lack of standardization, and system-related standards based on hardware limitations and WHO recommendations need to be established. Third, the sensitivity analysis showed unstable pooled results of MDRO detection rates. Further studies are needed to examine the effect of intelligent technology interventions for HH on this outcome. Fourth, our study performed no cost-benefit analysis of HH, and the results were unclear with respect to the rate of correct HH steps and the long-term sustainability of intelligent technology interventions. These limitations should be considered in future research.

Nevertheless, this meta-analysis strictly followed the PRISMA statement and applied a rigorous search strategy to identify potential studies in all available databases to ensure the generalizability of the results. Moreover, we included a relevantly large number of studies and sample sizes from various geographic areas, substantially enhancing the internal and external validity of the meta-analysis. This is the first meta-analysis to evaluate the impact of intelligent technology interventions for HH on multiple outcomes, which could provide valuable evidence to encourage intelligent technology application to improve clinical and nursing outcomes.

Improving HH behaviors is an important part of hospital management, and it is of great significance to patients and HCWs. The application of intelligent technology to HH involves the innovation of management methods. This systematic review determined that intelligent technology interventions for HH had an important role in improving HHC and reducing HCAI rates, but it could not be determined whether it had an effect on the MDRO detection rate. However, low-quality evidence and important heterogeneity were observed. Important directions for future work are to further verify the 3 outcomes through high-quality research and conduct more research to evaluate the impacts of intelligent technology interventions on the long-term sustainability, cost-effectiveness, and rate of correct HH.