This review is reported according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) statement [23].

We conducted a systematic search of nine health science databases: PubMed, EMBASE, Web of Science, Cumulative Index to Nursing and Allied Health Literature (CINAHL), Cochrane Library, China National Knowledge Infrastructure (CNKI), the Wanfang database, the VIP database, and the China Biomedical Literature Database (CBM). We used Medical Subject Headings (MeSH) terms, Emtree terms, subject headings, and free text associated with the concepts of ML and PI. Searches were performed in June 2020. We also undertook a manual search of the reference list of all potentially eligible studies. Textbox 1 shows the search strategy that was used.

This review included studies that met the following criteria: (1) used a method related to ML technologies (including support vector machine, k-nearest neighbor [KNN], decision tree [DT], convolutional neural network, Bayesian network model, and logistic regression) in PI management, and (2) was published in English or Chinese. We excluded studies that met any of the following criteria: (1) review papers, opinion papers, editorials, discussion papers, dissertations, or conference abstracts; (2) papers on PI education; (3) papers about PI in animals; (4) papers lacking an outcome; and (5) papers without explicit algorithms.

Two independent investigators screened titles and abstracts using the eligibility criteria. They then obtained full-text versions of all potential articles and scrutinized the full texts independently. Any discrepancies about study inclusion were resolved through discussion or by referral to a third investigator.

Data were extracted from all identified studies using a predefined format. Variables included the first author, year of publication, country, aim, subject, algorithm used, study outcomes, performance of the algorithm, and findings. One investigator extracted the information into a standard data extraction sheet and a second investigator cross-checked the entries. Any disagreements were resolved via discussion.

The methodological quality of the included studies was assessed independently by two investigators using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) [24]. Disagreements were resolved by discussion. The PROBAST was designed to assess the risk of bias and applicability of diagnostic and prognostic prediction model studies, and it includes 20 signaling questions to judge the risk of bias from four domains (participants, predictors, outcome, and analysis). The risk of bias is judged as low, high, or unclear. If one domain is found to have a high risk of bias, the overall risk of bias is judged as high. Similarly, if one domain is assessed as unclear, the overall risk of bias is judged as unclear even if all other domains are assessed to have a low risk of bias.

Our initial search retrieved 2207 published articles, of which 269 were duplicates. After screening titles and abstracts, the full texts of 48 articles were obtained and assessed for potential eligibility. Of those 48 articles, 16 did not fulfill the inclusion criteria. The reasons for studies being ineligible were as follows: (1) lacking a clear algorithm (n=5); (2) lacking a result (n=4); (3) review studies (n=4); (4) studies in pigs (n=2); and (5) study on PI education (n=1). Finally, a total of 32 studies were eligible for our research (see Figure 1).

The articles that were included in our analysis were published between 2007 and 2020 and were undertaken in the United States [25-35], China [36-44], Spain [45-50], Japan [51,52], Italy [53,54], Korea [55], and Greece [56]. According to the applied area of the included studies, we divided the articles into three components: predictive model (12 studies), posture recognition (11 studies), and image analysis (9 studies). The characteristics of the included studies are presented in Multimedia Appendix 1.

Figure 2 shows the roles of the three components in the PI management process:

The performance indicators of ML algorithms include sensitivity, specificity, precision, accuracy, F score, positive predictive value, negative predictive value, geometric mean, false-positive rate, run time, and so on. Multimedia Appendix 2 shows the detailed results of the included studies.

Twelve studies explored PI risk factors by data mining from the electronic health records (EHRs) of patients. The patients included in the studies were from a variety of settings: ICU (3 studies); operating room (2 studies); long-term care facilities (1 study); acute care hospital (1 study); orthopedic department (1 study); oncology department (1 study); end-of-life care (1 study); medical-surgical, critical care, and step-down units (1 study); and with mobility-related disabilities (1 study). The number of EHRs ranged from 147 to 125,213. The identified risk factors were different due to diverse input variables. In the majority of included studies, the PI percentage (the number of patients with PI/the number of total patients) of the data sets analyzed was imbalanced, and the minimum was 0.6% (51/8286). The accuracy ranged from 63.0% to 90.0%, the sensitivity ranged from 47.8% to 84.8%, and the specificity ranged from 70.3% to 94.7%. The DT algorithm was a typical data mining approach.

Eleven studies were concerned with posture identification by analyzing the pressure distribution of the body to achieve a robust assessment. Regarding the subjects of posture recognition, one study focused on wheelchair users [38], while the others looked at bed bound patients. The number of sensors was between 4 and 8192, and the number of subjects ranged from 2 to 58. Of the 11 studies, 10 studies detected and classified different postures or movements of a person and one study classified the bed inclination [31]. The common postures detected were supine, right lateral, and left lateral.

All articles reported on accuracy, which ranged from 49.1% to 100%. The difference in run times among different algorithms was quite large, from 0.04 seconds to 320.34 seconds. No articles reported on specificity. The sensitivity ranged from 62.0% to 100%, and the precision ranged from 65.0% to 100%. All eight studies applied the KNN algorithm in the processing of pressure sensor data.

Nine studies conducted PI wounds’ tissue segmentation and measurement using ML algorithms. We included studies that only analyzed PI images and excluded those involving the wound images of diabetes foot ulcers or venous leg ulcers. The number of digital images ranged from 14 to 193. Three articles were written by Veredas et al [46,48,49] using the same 113 color images to achieve tissue classification. Because different algorithms were used, we considered these three articles as independent research. Furthermore, the number of tissue segmentations ranged from 3 to 6. The most common PI wound tissue classifications were granulation, slough, and necrosis. One study developed an image processing algorithm that automatically measured the PI size [30]. The accuracy ranged from 78.3% to 92.0%, the sensitivity ranged from 61.7% to 99.9%, and the specificity ranged from 93.9% to 99.8%. Convolutional neural network algorithms, as deep learning architectures, were often used in medical image analysis in recent years.

The PROBAST was used to assess the risk of bias of the predictive model studies from four domains (participants, predictors, outcome, and analysis). However, the PROBAST was not suitable for the posture recognition and image analysis studies; to the best of our knowledge, there is still no appropriate tool to assess these engineering articles. The overall risk of bias of all of the predictive model studies was judged as high, and there was no low risk in the analysis domain (Figure 3).

Our systematic review provided a broad overview of the ML technologies applied to PI management. After study selection, we were able to categorize these technologies into three components: predictive model, posture recognition, and image analysis. We discuss these different components in detail below.

PI management should be a holistic process, but the current research in these three components is separate. We’ll use the case of a patient admitted to hospital to illustrate. First, according to the predictive model, we rated the patient as low risk. The repositioning schedule was implemented as the low risk required. Unfortunately, the patient developed PI, so we needed to assess the PI wound. The ML technologies on the predictive model and posture recognition need feedback from the PI wound image analysis to improve their performance. However, the research in these three components was conducted in different populations in different locations at different times. This point should be explored in future research.

The results on the risk of bias, surprisingly, were far from satisfactory. Similar to the research of Nagendran et al [59], the analysis domain was the major deficiency. More attention needs to be paid to the methodological quality of predictive model studies. The participants in posture recognition studies were healthy volunteers and the subjects in image analysis studies were images, so we could not judge these types of articles as medical research. There is a growing literature on interdisciplinary research such as in the fields of engineering and medicine. It is essential to develop a tool to assess the methodological quality of the relevant articles.

In summary, ML technologies furnish new alternatives to PI management. Given the global shortage of professional nurses and PI-related knowledge deficit, ML technologies will significantly reduce the burden on frontline clinicians and help to improve the quality of care, as Obermeyer and Emanuel [20] pointed out in 2016. However, because the current technologies only cover three components of PI management, there is a marked lack of novel technologies to assess potentially healthy skin, to achieve better skin care, to manage nutrition status, and to create intelligent support surfaces. Besides, IBM has discovered that its powerful technology is no match for the messy reality of today’s health care system [60]. There is still a long way to go to integrate ML technologies into clinical care practices.

It is important to acknowledge some limitations. First, we only include articles published in English and Chinese. It will be better to include other language research for representing the current evidence. Second, due to the various aims and outcomes of the included studies, the quantitative synthesis has not been performed to obtain a direct result. Third, the aim of our review was to survey the current status of ML algorithms applied in PI management, so the eligibility criteria were defined broadly. After study selection, we found the related research can be divided into three components. We have no specific criteria for one component. Hence, under the guidance of our findings, future research can define detailed eligibility criteria.

The study results from various laboratory settings show an array of ML technologies with potential uses in PI management. Future research should apply these technologies on a large scale with clinical data to verify their effectiveness, enhance their performance, and improve methodological quality.