ABSTRACT
This study investigated the relationship between clinician assessments and the AI‐generated scores, highlighting how correlations vary based on clinician expertise. It also explored the proportion of tissue types identified by clinicians relative to AI assessments and assess the inter‐clinician agreement in quantifying tissue types, identifying variations based on clinician experience. A cross‐sectional survey used purposive, non‐random sampling to recruit 50 wound care clinicians. Participants reported their specialisation and experience level before identifying and quantifying granulation, slough, eschar, and epithelialisation in nine wound images. An AI model analysed the same images for comparison. Experienced clinicians and wound care specialists reported higher confidence in assessments. Inter‐clinician agreement was moderate–good for granulation and slough (ICC: 0.763–0.762) and moderate–excellent for eschar (ICC: 0.910), but moderate–poor for epithelialisation (ICC: 0.435). Clinicians strongly correlated with AI for granulation, slough, and eschar (r = 0.879, 0.955 and 0.984, respectively). Epithelialisation was more challenging, with a 60% identification rate and moderate correlation with AI (r = 0.579). AI‐generated scores aligned with clinician assessments for granulation, slough, and eschar. However, epithelialisation, which is crucial for objectively measuring healing progress, showed greater variability, suggesting that AI could improve the reliability of its assessment, potentially leading to more consistent wound evaluation to guide treatment decisions.
Keywords: AI‐driven wound care, epithelialisation assessment, granulation tissue quantification, inter‐clinician agreement, slough and eschar identification
Summary.
This cross‐sectional study evaluated how consistently clinicians identify and quantify wound tissue types—specifically granulation, slough, eschar, and epithelialization—and how their assessments compare to an AI‐driven tissue segmentation tool (SmartTissue) embedded in Swift Medical's wound care platform. Fifty clinicians assessed nine AI‐captured wound images, estimating the percentage of each tissue type. Inter‐rater agreement between clinicians for granulation, slough, and eschar was moderate to excellent, whereas agreement for epithelialization was lower, Additionally, the study also revealed that the estimates for granulation, slough, and eschar made by the AI appear to have good relationships with those estimates made by clinicians. However, epithelialization had only a moderate relationship.
Further, while the experienced clinicians' assessments of tissue types were generally comparable, variability was noted among participants especially for epithelialization. Even among professionals trained in tissue type differentiation, reliable assessment of epithelial tissue was challenging, perhaps because of the subjective element, and the complexity of the image, and may be why there was lesser agreement with epithelialization. These results provide additional support for the use of artificial intelligence systems such as SmartTissueTM to standardize assessments, decrease inter‐rater reliability variability, and provide a more objective means of assessing wounds in a clinical situation.
1. Introduction
Wound healing is a dynamic and complex process that involves a coordinated interaction between numerous cell types, signalling molecules, extracellular matrix components and the vascular system [1]. There are four primary tissue types (i.e., granulation, slough, eschar and epithelialisation) that may occur during wound healing. Optimal wound healing response is identified with the presence of granulation and epithelialisation tissue [2]. Granulation tissue is a highly vascular connective tissue that forms during the proliferative phase of wound healing [3]. It appears as red, granular tissue and serves as a scaffold for new blood vessel formation [3]. Granulation tissue also forms the foundation for epithelialisation to occur [2]. Islands of epithelial tissue can form over granulation tissue as the final step in wound closure. It appears as pink, thin, translucent tissue that grows from the wound edges inward [4, 5]. Distinguishing between granulation and early epithelialisation can be challenging, particularly in the initial stages of epithelial formation [4]. When wound healing is compromised during the proliferative phase, slough and eschar may occur [5]. Slough is formed when there is increased wound inflammation causing the production of yellow, devitalised tissue that can be thick and adherent to the wound bed [5]. Eschar results from tissue death and appears as black, necrotic tissue [5]. Early identification and treatment of slough and eschar are essential as they both prevent wound healing and increase the risk of wounds progressing to be chronic non‐healing [3, 4].
Accurate wound tissue assessment is critical for effective treatment, but clinicians' agreement remains variable, leading to delayed interventions, increased healthcare costs and potentially poorer patient outcomes [6]. Studies have shown that epithelialisation is the most difficult tissue type to identify accurately, with a very poor Krippendorff value (measurement of the reliability of agreement between raters) for inter‐rater variability [7]. This difficulty likely stems from the subtle appearance of early epithelial tissue, which can blend into surrounding tissue and easily be overlooked [7]. While studies do not explicitly state that eschar and slough are the easiest to identify, they do suggest that these tissue types may be more readily recognised due to their distinct colour characteristics [8, 9]. Research has shown that outside of specialised dermatology and wound care clinics, consistency of wound assessments often demonstrates only moderate agreement [8, 9].
Traditional methods often used in wound assessments rely on individual clinicians' visual assessment, which can introduce imprecise, inconsistent measurements between clinicians [10]. This inconsistency may be influenced by the level of training, years of experience, or number of wounds cared for per day [11]. Efficient wound care is reliant on accurate identification of wound tissue types to obtain an informed comprehensive assessment of a wound [10]. Inaccurate wound assessment can result in delayed intervention, misguided treatment and potentially poor patient outcomes [10]. Accurate assessment ensures appropriate and earlier treatment, contributing to improved health outcomes, fewer complications and decreased hospitalisations [12]. Standardising wound assessment may also contribute to significant reductions in resource utilisation and can enhance clinical documentation, contributing to enhanced workflow [12].
The integration of artificial intelligence (AI) into clinical practice offers a unique opportunity to introduce a standardised approach to wound assessment [13]. AI, through objective and consistent wound assessment [14, 15], has the potential to enhance patient care, leading to more informed treatment plan formulation by healthcare providers [13]. The ability of AI to identify various tissue types (i.e., granulation, slough, eschar and epithelialisation) introduces the potential for it to serve as a tool to inform healthcare providers' clinical assessments [13]. This insight may contribute to more tailored treatment plans, enhancing patient outcomes [13].
AI has been shown in the literature to enhance wound care assessment and management [16, 17]. Evaluations of the usability and effectiveness of AI for wound assessment and management have found that its integration resulted in enhanced objectivity, encouraged shared wound care planning between healthcare professionals, and increased patient adherence [16]. AI has also been shown to demonstrate high sensitivity and specificity in the identification of diabetic foot ulcers, with high levels of inter‐ and intra‐rater reliability for clinician agreement on the ability of AI to identify these ulcers [17].
Wound care serves as a large healthcare expense and thus warrants investigation on initiatives or advances in clinical practice that have the potential to contribute to streamlined workflow, improved resource allocation, and shorter treatment courses for patients [18]. Reduced healing time is a hallmark of effective treatment; however, inaccurate wound assessment can inform inadequate treatment plans, contributing to prolonged healing times and increased financial burden on the healthcare system [19]. Accurate, reliable, and fast wound assessment through AI offers the opportunity to not only enhance the evaluation of a patient's wound, but also liberate time for clinicians to see more patients, thus decreasing wait times and increasing clinic productivity [19, 20].
A tissue segmentation feature (SmartTissue) was added to Swift—a digital wound‐care solution powered by AI using a deep learning convolutional neural network based on a modified U‐Net architecture. This algorithm, originally developed and validated by Ramachandram et al. [7], was trained on 17 000 wound images labelled by expert reviewers and validated against a dataset of 465 000 wound images to identify the percentage of granulation tissue, slough, eschar, and epithelialisation within a wound [7]. The AI achieved a high pixel‐wise segmentation accuracy of 94% and excellent discriminatory capability. This was demonstrated with the values of 0.97 for granulation tissue, 0.98 for slough, 0.99 for eschar, and 0.92 for epithelialisation in the area under the receiver operating characteristics curve (AUROC) [7]. The ability of this AI to objectively assess wound tissue segmentation has significant implications for supporting clinical decision‐making through serving as an objective wound assessment tool and image‐based wound tracking device [7]. This AI has the potential to optimise clinical evaluation through fast and efficient wound assessment, potentially contributing to improved inter‐clinician agreement, enhanced documentation and tailored patient care plans [7].
The version of SmartTissue employed in this study was the same as reported by Ramachandram et al. [7]. This study's primary aim was to investigate the relationship between clinician assessments and the generated scores from Swift AI‐ SmartTissue tool, highlighting how correlations vary based on clinician expertise. It is secondary aim was to explore the proportion of tissue types identified by clinicians relative to AI assessments, and assess the inter‐clinician agreement in quantifying granulation, slough, eschar, and epithelialisation using intraclass correlation coefficients (ICCs), identifying variations based on clinician experience and practice patterns.
2. Methods
2.1. Study Design and Recruitment
This study is a descriptive cross‐sectional survey that targeted clinicians actively involved in wound care assessment within their practice settings. We employed a purposive non‐random sampling approach to recruit clinicians at a Canadian wound care conference. To expand recruitment, advertisements for the study and survey links were disseminated via the authors' social media accounts and professional networks. Through the various recruitment methods used, clinicians were encouraged to share the survey link within their professional networks for a wider dissemination among providers. Participants accessed the survey link through SurveyMonkey—an online survey platform that collects anonymous responses.
Upon clicking on the survey link, participants were presented with a note detailing the purpose of the feedback collection. No identifiable information was collected, ensuring that all responses remained anonymous. The survey took about 10–15 min to complete. Participation was entirely voluntary, and participants had the choice to skip over questions should they prefer not to respond to them.
To enhance participation, a reminder post was sent about a month after the initial posts to the above‐described network of clinicians. Another reminder post was shared 2 months later. The survey link remained active for 4 months, starting from September to December 2024, allowing for an extended time to gather the required sample size.
2.2. Participant Characteristics
The participants involved in the study were identified as clinicians. The clinician group was comprised of registered nurses (RNs), nurse practitioners (NPs), physician assistants (PAs), nurse specialists in wound care (NSWs) and physicians (MDs). The MDs had education in wound care training from accredited Canadian programmes (e.g., the International Interprofessional Wound Care Course [IIWCC] or another formal Clinical Training or wound management certification). Clinicians identified as wound care specialists are those who met at least one of the following criteria: an identified wound care certification such as Certified Wound Specialist (CWS), Certified Wound Ostomy Continence Nurse (CWOCN), Certified Wound Care Nurse (CWCN) or another recognised certification; advanced wound care training programme; or involvement in wound management in the clinical area for 2 years or more.
2.3. Selected Wound Images From the AI‐Database for Survey Inclusion
The selection of wound images for this study was conducted following a structured process of data abstraction to encourage clinical diversity, representativeness, and adherence to ethical guidelines. The study began with a major screening of a wide range of wound images from a de‐identified database of images collected from participating organisations that collect AI‐enabled wound imaging documentation under consent during routine clinical practice. Random stratified sampling targeting certain diversity variables was applied: various tissue compositions, various wound types, various skin tones and varied stages of healing. The inclusion criteria prioritised images that:
Have a variety in tissue compositions (e.g., granulation, slough, eschar and epithelialisation).
Illustrate various wound types (e.g., diabetic wounds, pressure injuries, arterial and venous ulcers).
Captured wounds on various skin tones for assessment across diverse patient populations.
Captured different stages of healing, where an assessment will be made in terms of how well clinicians can appreciate evolving tissue changes.
Two wound‐certified wound care specialists independently reviewed the randomly screened and selected images as appropriate and clinically relevant before confirming the final nine images for inclusion in the survey (Figure 1). All images were completely de‐identified prior to screening, with no link to the patient metadata or links to their healthcare settings. The AI system took these pictures only of the wound site, and no identifiable background about the patient was included (e.g., face, tattoos or key identity markers). The selection and utilisation of the images within the research were consistent with the ethical principles outlined by the Declaration of Helsinki on data integrity and IRB exemption in research and adherence to safe principles during the conduct of the study.
FIGURE 1.
AI‐wound captured images included in the survey.
2.4. Sample Size Calculation
Sample size for this study was estimated by performing two power analyses: one for inter‐rater reliability among clinicians using ICC and the other for the correlation between clinicians' assessments and AI‐generated scores. For the inter‐rater reliability analysis, sample size estimation was based on an F‐test for ICC, assuming a moderate agreement (ICC = 0.6), a 5% significance level (α = 0.05) and 80% power. Using this method, the minimum number of clinicians needed for the desired statistical power was estimated at 8.
In the correlation analysis of AI‐clinician, sample size estimation was calculated based on the Pearson's correlation coefficient, expecting a 0.6 correlation with a 5% significance level and 80% power. By using Fisher's z‐transformation method, the estimated minimum number of clinician‐AI comparisons needed was 33 to detect significant differences in data if one exists.
We have 50 clinicians participating in the study, in excess of the minimum sample size required for both analyses; therefore, the number of responses was sufficient to assure reliable statistical inference.
2.5. Survey Instrument
Two wound‐certified clinicians independently performed a qualitative review of the AI‐generated assessments for accuracy and clinical relevance. This included verification of AI‐identified tissue regions in relation to wound characteristics to ensure that segmentation and quantification were appropriate and consistent with clinical expectations. Clinicians evaluated if the AI had appropriately differentiated tissue types, especially in wounds that were heterogeneous or had subtle transitions between granulation and epithelialisation. This review provided contextual insights that helped in determining whether AI outputs were clinically interpretable and usable in real‐world settings. The high level of agreement between AI and expert clinicians supported the AI model's reliability in its scores before its use in the larger study.
The survey tool was designed to explore clinicians' experience and confidence level in accurately identifying and quantifying different wound tissue types. It consisted of three main sections: demographic data, self‐assessment of experience and confidence in the identification of wound tissue, and assessment of nine wound images, each including different tissue types.
Demographic characteristics gathered data on the participants' professional roles, practice settings, wound care specialisation, advanced training, years of experience and number of wounds cared for per week.
The second section explored whether clinicians utilised technology in wound assessment, and if they had encountered any barriers when they identified tissue types using a multiple‐choice format question. This section also assessed the level of clinicians' confidence in identifying and quantifying tissue types within the wound bed using a 5‐point Likert scale that ranged from ‘Extremely confident’ to ‘Not at all confident’.
In the wound images assessment section, participants were provided with nine different images of wounds. Clinicians were asked to identify the presence or absence of each one of the four tissue types within the wound images: granulation, slough, eschar, and epithelialisation. They were also asked to provide the percentage of each tissue type present within the wound bed. The clinicians performed indirect quantification by estimating tissue percentages rather than directly delineating tissue areas. The responses have been used to assess clinician agreement on tissue identification and quantification, and to compare their assessment scores with the AI‐generated wound tissue quantifications.
2.6. Ethics Approval and Consent to Participate
Our study did not include any identifiable patient or clinician information and was fully compliant with ethical standards pertaining to confidentiality. Participation of clinicians in this survey was strictly voluntary, and their answers were made anonymously, leaving no identifiable information linked to any participant. Completing the survey was considered as implied consent, as they were informed of the study's objective and methods before taking the survey.
As for the wound images used in the survey, they were randomly selected from the AI‐driven wound care database, which were de‐identified before selection and were not associated with any patient‐related metadata, personal data and identifiable features such as faces, tattoos or distinguishing marks when pulled from the repository, ensuring complete anonymity. Furthermore, images are strictly focused on the wound itself and no patients' identities or healthcare settings and organisations from which the images were taken were known to the researchers or the survey participants. Moreover, this research was a quality improvement initiative exempt from ethics since it was meant to measure clinician agreement in wound assessment, and not human subject research. The use of de‐identified data and uncoded images adhered to ethical guidelines for quality improvement studies, and as a result, individual patient consent was not needed for this study.
This study was considered a quality improvement study and had been granted an exemption from ethics review by Pearl IRB LLC, an independent institutional review board (ID: 2023‐0100).
2.7. Statistical Analysis
The inter‐rater agreement in estimating the wound tissue types (granulation, slough, eschar and epithelialisation) was determined through the ICCs that present a measure of reliability and consistency among raters. The following fixed cut‐off points for ICC were established for interpretation: poor (< 0.50), moderate (0.50–0.75), good (0.75–0.90) and excellent (> 0.90) agreement [21]. Subgroup analysis was done to explore variations in agreement according to clinical experience and practice patterns (e.g., years of experience, specialisation in wound management and number of managed wounds).
The Pearson correlation coefficients (r) were calculated for each tissue type in order to evaluate the correlation between clinicians' assessments and those produced by AI tissue quantification. The correlation was interpreted as weak (r < 0.30), moderate (r = 0.30–0.59), strong (r = 0.60–0.79) and very strong (r ≥ 0.80) [22]. The differences in the strength of correlation were checked across clinician subgroups (e.g., level of experience and specialisation) to determine whether the agreement between AI and clinician depended on expertise.
Paired t‐tests were performed to compare the average clinicians' scores of tissue types with the AI‐generated quantifications.
3. Results
3.1. Clinician Characteristics, Technology Use and Confidence in Wound Tissue Assessment
Among the 50 participants, the majority consisted of RNs (60%). The more experienced clinicians (> 10 years) had a slightly higher proportion of physicians (MDs) (15.4%), while those with < 10 years of experience had more NSWs (12.5%) (Table 1).
TABLE 1.
Clinicians' characteristics and experience levels.
| All participants (N = 50) | Experience > 10 years (N = 26) | Experience < 10 years (N = 24) | > 20 wounds/week (N = 21) | < 20 wounds/week (N = 26) | Wound care specialists (N = 36) | |
|---|---|---|---|---|---|---|
| N (%) | N (%) | N (%) | N (%) | N (%) | N (%) | |
| Role | ||||||
| MD | 6 (12.0%) | 4 (15.4%) | 2 (8.3%) | 1 (4.8%) | 5 (19.2%) | 4 (11.1%) |
| Resident | 2 (4.0%) | 0 (0.0%) | 2 (8.3%) | 0 (0.0%) | 1 (3.8%) | 0 (0.0%) |
| NP | 3 (6.0%) | 1 (3.8%) | 2 (8.3%) | 3 (14.3%) | 0 (0.0%) | 3 (8.3%) |
| NSW | 5 (10.0%) | 2 (7.7%) | 3 (12.5%) | 2 (9.5%) | 3 (11.5%) | 5 (13.9%) |
| PA | 4 (8.0%) | 3 (11.5%) | 1 (4.2%) | 2 (9.5%) | 1 (3.8%) | 4 (11.1%) |
| RN | 30 (60.0%) | 16 (61.5%) | 14 (58.3%) | 13 (61.9%) | 16 (61.5%) | 20 (55.6%) |
| Do you use any technology to assist in wound assessment and documentation? | ||||||
| Yes | 29 (58.0%) | 14 (53.8%) | 15 (62.5%) | 11 (47.0%) | 17 (65.4%) | 21 (58.3%) |
| No | 21 (42.0%) | 12 (46.2%) | 9 (37.5%) | 10 (52.0%) | 9 (34.6%) | 15 (41.7%) |
| Have you encountered any barriers in accurately assessing wound bed tissue? | ||||||
| Yes | 30 (60.0%) | 17 (65.4%) | 13 (54.2%) | 12 (42.9%) | 15 (57.7%) | 25 (69.0%) |
| No | 20 (40.0%) | 9 (34.6%) | 11 (45.8%) | 9 (57.1%) | 11 (42.3%) | 11 (31.0%) |
| Confidence identifying tissues | ||||||
| Extremely confident | 8 (16.0%) | 6 (23.1%) | 2 (8.3%) | 5 (23.8%) | 3 (11.5%) | 8 (22.2%) |
| Very confident | 31 (62.0%) | 14 (53.8%) | 17 (70.8%) | 12 (57.1%) | 17 (64.4%) | 26 (72.2%) |
| Somewhat confident | 9 (18.0%) | 5 (19.2%) | 4 (16.7%) | 4 (19.1%) | 5 (19.2%) | 2 (5.6%) |
| Not so confident | 1 (2.0%) | 1 (3.8%) | 0 (0.0%) | 0 (0.0%) | 1 (3.8%) | 0 (0.0%) |
| Not at all confident | 1 (2.0%) | 0 (0.0%) | 1 (4.2%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
| Confidence quantifying tissues | ||||||
| Extremely confident | 7 (14.0%) | 6 (23.1%) | 6 (23.8%) | 5 (23.8%) | 2 (7.7%) | 7 (19.4%) |
| Very confident | 20 (40.0%) | 7 (26.9%) | 7 (33.3%) | 7 (33.3%) | 12 (46.2%) | 17 (47.2%) |
| Somewhat confident | 16 (32.0%) | 9 (36.6%) | 6 (28.6%) | 6 (28.6%) | 9 (34.6%) | 10 (27.8%) |
| Not so confident | 6 (12.0%) | 4 (15.4%) | 3 (14.3%) | 3 (14.3%) | 3 (11.5%) | 2 (5.6%) |
| Not at all confident | 1 (2.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
Abbreviations: MD, Doctor of Medicine; NP, nurse practitioner; NSW, nurse specialist in wound care; PA, physician assistant; RN, registered nurse.
A chi‐square test was conducted to explore the distribution of roles among participants. A statistically significant difference (p value = 0.00069) among participants' roles. Pairwise comparisons using standardised residuals indicated that the role of residents was significantly underrepresented, while registered nurses were markedly overrepresented in the sample.
No significant differences between all participants vs. wound care specialists, experience (< 10 years vs. > 10 years) and wounds per week (< 20 vs. > 20, p > 0.05).
Technology use in practice was reported among 58% of participants, including wound care specialists (58.3%), with slightly higher use among participants with < 10 years of experience (62.5%) and those who care for < 20 wounds per week (65.4%). Barriers to tissue identification were reported by 60% of all participants, with higher rates among experienced clinicians (65.4%) and wound care specialists (69%) (Table 1).
Overall, confidence was high in identifying tissue types in the wound bed, with 78% of participants stating they were very confident/extremely confident. Confidence among wound care specialists was the highest, with 94.4% very confident/extremely confident (Table 1).
There was less confidence in estimating the percentage of tissue types, with only 54% of all participants being extremely confident/very confident. Confidence among wound care specialists was the highest, with 66.6% very confident/extremely confident (Table 1).
3.2. Clinician vs. AI Identification of Wound Tissue Types
When comparing the presence or absence of tissue types identified by clinicians with those identified by AI across nine wound images, there was a strong alignment and consistency in identifying granulation tissue, slough and eschar, achieving average identification rates of 82%, 84% and 87%, respectively. However, there were exceptions: in Wound 1, where only 18% of clinicians accurately identified the presence of slough, and in Wound 5, where just 36% accurately identified the absence of eschar (Table 2, Figure 2).
TABLE 2.
Comparison of clinician‐identified vs. AI‐identified wound tissue types.
| Granulation, N (%) | Slough, N (%) | Eschar, N (%) | Epithelialisation, N (%) | |
|---|---|---|---|---|
| Wound 1 AI‐identify | Present | Present | Absent | Present |
| Wound 1 clinicians | ||||
| Matched identification | 47 (94.0%) | 9 (18.0%) | 49 (98.0%) | 17 (66.0%) |
| Did not match | 3 (6.0%) | 41 (82.0%) | 1 (2.0%) | 33 (34.0%) |
| Wound 2 AI‐identify | Present | Absent | Absent | Present |
| Wound 2 clinicians | ||||
| Matched identification | 49 (98.0%) | 44 (88.0%) | 48 (96.0%) | 41 (82.0%) |
| Did not match | 1 (2.0%) | 6 (12.0%) | 2 (4.0%) | 9 (18.0%) |
| Wound 3 AI‐identify | Absent | Absent | Present | Absent |
| Wound 3 clinicians | ||||
| Matched identification | 50 (100.0%) | 42 (84.0%) | 49 (98.0%) | 44 (88.0%) |
| Did not match | 0 (0.0%) | 8 (16.0%) | 1 (2.0%) | 6 (12.0%) |
| Wound 4 AI‐identify | Present | Present | Absent | Absent |
| Wound 4 clinicians | ||||
| Matched identification | 14 (28.0%) | 50 (100.0%) | 50 (100.0%) | 38 (76.0%) |
| Did not match | 36 (72.0%) | 0 (0.0%) | 0 (0.0%) | 12 (24.0%) |
| Wound 5 AI‐identify | Present | Present | Absent | Present |
| Wound 5 clinicians | ||||
| Matched identification | 50 (100.0%) | 50 (100.0%) | 18 (36.0%) | 28 (56.0%) |
| Did not match | 0 (0.0%) | 0 (0.0%) | 32 (64.0%) | 22 (44.0%) |
| Wound 6 AI‐identify | Present | Present | Present | Absent |
| Wound 6 clinicians | ||||
| Matched identification | 40 (80.0%) | 44 (88.0%) | 43 (86.0%) | 32 (64.0%) |
| Did not match | 10 (20.0%) | 6 (12.0%) | 7 (14.0%) | 18 (36.0%) |
| Wound 7 AI‐identify | Present | Present | Absent | Present |
| Wound 7 clinicians | ||||
| Matched identification | 47 (94.0%) | 49 (98.0%) | 38 (76.0%) | 24 (48.0%) |
| Did not match | 3 (6.0%) | 1 (2.0%) | 12 (24.0%) | 26 (52.0%) |
| Wound 8 AI‐identify | Present | Present | Absent | Present |
| Wound 8 clinicians | ||||
| Matched identification | 43 (86.0%) | 46 (92.0%) | 48 (96.0%) | 18 (36.0%) |
| Did not match | 7 (14.0%) | 4 (8.0%) | 2 (4.0%) | 32 (64.0%) |
| Wound 9 AI‐identified | Present | Present | Absent | Absent |
| Wound 9 clinicians | ||||
| Matched identification | 32 (64.0%) | 44 (88.0%) | 48 (96.0%) | 29 (58.0%) |
| Did not match | 18 (36.0%) | 6 (12.0%) | 2 (4.0%) | 21 (42.0%) |
| Average total matched | 82% | 85% | 87% | 60% |
FIGURE 2.
Percentage of clinicians matching AI‐identified presence or absence of tissue types across nine wound images.
In contrast, identifying epithelialisation seemed to be more challenging, with an average identification rate of only 60% for both its presence and absence. Clinicians particularly struggled in reaching consensus with the AI when identifying epithelialisation in wounds, with agreement rates at 66% for Wound 1, 56% for Wound 5, 36% for Wound 8% and 48% for Wound 7. Furthermore, the agreement rates for recognising the absence of epithelialisation were also notably low in Wounds 6 and 9, with only 64% and 58% alignment, respectively (Table 2, Figure 2).
3.3. Inter‐Clinician Agreement in Quantifying Wound Tissue Types by Experience Levels
Overall, the findings of single‐measure ICCs highlight good–excellent agreement for slough and eschar, moderate–good for granulation, and poor for epithelialisation, with variations based on experience (Table 3, Figure 3a,b).
TABLE 3.
Inter‐clinicians' agreement in quantifying tissue types by experience and practice pattern.
| ICC value Single measure, N | ICC 95% CI | p | Mean tissue percentage (% ± SD) | Range (min–max) (%) | |
|---|---|---|---|---|---|
| All clinicians | |||||
| Granulation | 0.763 | 0.590–0.920 | < 0.001 | 26.80 ± 14.3 | 4.22–41.11 |
| Slough | 0.762 | 0.570–0.930 | < 0.001 | 39.40 ± 24.4 | 21.25–63.75 |
| Eschar | 0.910 | 0.805–0.900 | < 0.001 | 23.39 ± 10.9 | 12.85–34.24 |
| Epithelialisation | 0.435 | 0.242–0.700 | < 0.001 | 9.27 ± 16.6 | 1.25–26.25 |
| Clinicians > 10 years' experience | |||||
| Granulation | 0.720 | 0.529–0.900 | < 0.001 | 26.50 ± 12.9 | 4.22–39.44 |
| Slough | 0.750 | 0.555–0.900 | < 0.001 | 38.61 ± 18.3 | 21.37–56.87 |
| Eschar | 0.931 | 0.850–0.900 | < 0.001 | 20.49 ± 9.4 | 11.25–29.87 |
| Epithelialisation | 0.466 | 0.257–0.700 | < 0.001 | 9.15 ± 17.1 | 1.25–26.25 |
| Clinicians < 10 years' experience | |||||
| Granulation | 0.807 | 0.646–0.900 | < 0.001 | 27.16 ± 13.9 | 7.22–41.11 |
| Slough | 0.735 | 0.545–0.900 | < 0.001 | 44.05 ± 23.7 | 18.88–67.77 |
| Eschar | 0.809 | 0.771–0.900 | < 0.001 | 20.67 ± 10.6 | 11.25–31.25 |
| Epithelialisation | 0.392 | 0.206–0.700 | < 0.001 | 8.96 ± 16.1 | 1.11–25.00 |
| Care for > 20 wound/week | |||||
| Granulation | 0.744 | 0.557–0.900 | < 0.001 | 26.90 ± 13.1 | 4.22–40.00 |
| Slough | 0.761 | 0.579–0.900 | < 0.001 | 44.50 ± 16.1 | 18.88–60.55 |
| Eschar | 0.890 | 0.812–0.900 | < 0.001 | 17.89 ± 13.2 | 10.00–31.11 |
| Epithelialisation | 0.392 | 0.242–0.700 | < 0.001 | 6.64 ± 8.4 | 1.11–15.00 |
| Care for < 20 wound/week | |||||
| Granulation | 0.802 | 0.639–0.900 | < 0.001 | 27.37 ± 13.7 | 15.55–41.11 |
| Slough | 0.814 | 0.646–0.900 | < 0.001 | 38.42 ± 25.3 | 21.37–63.75 |
| Eschar | 0.921 | 0.822–0.900 | < 0.001 | 24.13 ± 10.2 | 12.85–34.28 |
| Epithelialisation | 0.469 | 0.259–0.700 | < 0.001 | 10.56 ± 15.7 | 1.87–26.25 |
| Wound care specialist | |||||
| Granulation | 0.802 | 0.641–0.900 | < 0.001 | 27.40 ± 12.6 | 10.22–40.00 |
| Slough | 0.797 | 0.634–0.900 | < 0.001 | 44.29 ± 23.5 | 18.88–67.77 |
| Eschar | 0.912 | 0.822–0.900 | < 0.001 | 19.01 ± 12.1 | 10.00–31.11 |
| Epithelialisation | 0.553 | 0.348–0.800 | < 0.001 | 7.35 ± 17.7 | 1.66–25.00 |
Note: Bolded values indicate statistically significant (p < 0.05). Non statistically significant (p < 0.05).
FIGURE 3.
(a) Inter‐clinicians' agreement in quantifying tissue types by experience and practice pattern. (b) Mean tissue percentages identified by clinicians, stratified by experience and practice pattern.
Single‐measure ICC values indicated that clinician agreement related to tissue types varies according to practice skills. For example, eschar consistently demonstrated the highest ICC agreements across all clinicians with different experience levels, suggesting good–excellent agreement by clinicians in identifying and quantifying this tissue type. Similarly, the ICC values for granulation tissues and slough were consistently moderate–good across the different groups (Table 3, Figure 3a,b).
In contrast, consistency on epithelialisation was the lowest across all groups, with ICC values ranging between 0.392 and 0.553, indicating a moderate–poor agreement in quantifying this tissue type. Wound care specialists showed the highest ICC for epithelialisation (0.553), though still relatively low (Table 3, Figure 3a,b).
3.4. Correlation Between Clinician and AI‐Generated Wound Tissue Assessment Scores by Experience Levels
For granulation tissue, slough and eschar, all groups demonstrated a similar level of alignment and minimal variation, with a positive very strong correlation with the AI‐generated scores across all groups (Table 4, Figure 4–c).
TABLE 4.
Correlation between clinician and AI measurements by experience levels.
| AI‐scores | Average clinicians' score | Correlation with AI scores (r and p values) | Average clinicians' score specialist (mean) | Correlation with AI scores (r and p values) | Average clinicians' score > 10 years' experience (mean) | Correlation with AI scores (r and p values) | |
|---|---|---|---|---|---|---|---|
| All participants (mean) | |||||||
| Wound 1 granulation | 44 | 85.16 | r = 0.879 | 87.36 | r = 0.881 | 81.96 | r = 0.879 |
| Wound 2 granulation | 35.11 | 63.92 | p = 0.002 | 68.12 | p = 0.002 | 62.07 | p = 0.002 |
| Wound 3 granulation | 0 | 0 | 0 | 0 | |||
| Wound 4 granulation | 2.9 | 2.42 | 1.75 | 3.33 | |||
| Wound 5 granulation | 13.25 | 21.08 | 21.63 | 22 | |||
| Wound 6 granulation | 19.74 | 7.92 | 7.18 | 9.03 | |||
| Wound 7 granulation | 30.64 | 26.44 | 26.03 | 25.96 | |||
| Wound 8 granulation | 12.96 | 13.02 | 10.15 | 11.44 | |||
| Wound 9 granulation | 27.03 | 21.32 | 24.66 | 20.33 | |||
| Wound 1 slough | 0.3 | 4.34 | r = 0.955 | 5.36 | r = 0.958 | 6.55 | r = 0.961 |
| Wound 2 slough | 0 | 1.58 | p < 0.001 | 1.78 | p < 0.001 | 1.28 | p < 0.001 |
| Wound 3 slough | 0 | 4.74 | 0.6 | 0.55 | |||
| Wound 4 slough | 97.06 | 93.52 | 93.93 | 91.96 | |||
| Wound 5 slough | 41.22 | 60.2 | 59.75 | 58.7 | |||
| Wound 6 slough | 39.34 | 27.26 | 27.57 | 28.7 | |||
| Wound 7 slough | 47.94 | 63.78 | 64.48 | 61 | |||
| Wound 8 slough | 85.07 | 79.44 | 83.3 | 82.26 | |||
| Wound 9 slough | 72.97 | 60.26 | 61.54 | 58.55 | |||
| Wound 1 eschar | 0 | 0.18 | r = 0.984 | 0.27 | r = 0.984 | 0.33 | r = 0.989 |
| Wound 2 eschar | 0 | 0.2 | p < 0.001 | 0.3 | p < 0.001 | 0.29 | p < 0.001 |
| Wound 3 eschar | 100 | 95.92 | 98 | 98.22 | |||
| Wound 4 eschar | 0 | 0 | 0 | 0 | |||
| Wound 5 eschar | 40.92 | 9.64 | 11.33 | 9.07 | |||
| Wound 6 eschar | 0 | 57.36 | 57.78 | 54.77 | |||
| Wound 7 eschar | 0 | 2.61 | 2.3 | 3.55 | |||
| Wound 8 eschar | 0 | 0.48 | 0.72 | 0.88 | |||
| Wound 9 eschar | 1.42 | 0.3 | 0.37 | ||||
| Wound 1 epithelial | 55.5 | 7 | r = 0.579 | 4.69 | r = 0.576 | 5.74 | r = 0.556 |
| Wound 2 epithelial | 64.89 | 32.98 | p = 0.028 | 30.93 | p = 0.031 | 31.33 | p = 0.045 |
| Wound 3 epithelial | 0 | 1.1 | 1.21 | 1.11 | |||
| Wound 4 epithelial | 0 | 2.82 | 2.15 | 2.25 | |||
| Wound 5 epithelial | 13.25 | 7.74 | 6 | 7.71 | |||
| Wound 6 epithelial | 0 | 6.02 | 5.15 | 8.11 | |||
| Wound 7 epithelial | 21.42 | 4.7 | 3.18 | 4.92 | |||
| Wound 8 epithelial | 1.96 | 7.4 | 2.72 | 6.14 | |||
| Wound 9 epithelial | 0 | 8.28 | 6.15 | 11.07 |
Note: Bolded values indicate statistically significant (p < 0.05). Non statistically significant (p < 0.05).
FIGURE 4.
(a) AI vs. clinicians' scores and correlation between clinician and AI measurements for all participants for each tissue type across wounds. (b) AI vs. specialists' scores and correlation between specialists and AI measurements for each tissue type across wounds. (c) AI vs. experienced clinicians' scores and correlation between scores of clinicians with > 10 years' experience and AI measurements for each tissue type across wounds.
In contrast, average scores of clinicians compared to the AI‐generated rates for epithelialisation showed a moderate correlation, representing less alignment between clinicians and AI‐generated assessment scores for this tissue type (Table 4, Figure 4a–c).
Paired t‐test was conducted and indicated no statistically significant differences between the AI‐generated scores and clinicians' mean estimates for all four tissue types. For granulation tissue, p values were 0.318 for all participants, 0.308 for specialists and 0.334 for experienced clinicians with> 10 years. For slough, p values were 0.745 for all participants, 0.657 for specialists, and 0.788 for experienced clinicians. The p values for eschar were 0.187 for all participants, 0.266 for specialists and 0.124 for experienced clinicians, while for epithelialisation, they were 0.220, 0.320 and 0.226, respectively. Non‐significance suggests the alignment of the AI scores and mean estimates determined by clinicians, confirming that AI‐generated scores provide comparable values.
4. Discussion
This study explored the proportion of tissue types—specifically granulation, slough, eschar, and epithelialisation—identified by clinicians compared to AI‐generated assessments and evaluated the inter‐clinician agreement in classifying wound tissue types, examining variations based on clinician experience. It also explored the relationship between clinician evaluations and AI‐generated scores, assessing the consistency of tissue identification across various levels of clinical expertise. The findings of this study showed variability in clinicians' confidence and accuracy in identifying and quantifying wound tissue types. While most clinicians (78%) were quite confident in identifying the various tissue types (62% very confident and 16% extremely confident), there was a sharp decrease in their reported confidence when asked about quantifying the percentage of each tissue type in a wound (54%), with only 14% stating that they felt extremely confident and 40% very confident.
This discrepancy aligns with previous studies that have pointed to subjectivity in wound assessment as an ongoing challenge in everyday clinical practice, particularly in terms of quantifying the distribution of tissues [23]. This is because wound assessments often rely on clinician‐reported observations, which can be influenced by individual experience, training and biases [24]. Traditionally, wound tissue assessment relies on manual methods such as visual inspection, ruler‐based measurement and digital planimetry, with a clinician's subjective assessment of wound and tissue composition [25]. However, such an approach lacks consistency and accuracy, as assessment ratings can vary widely between clinicians [15]. In addition, chronic wounds have irregular shapes, poorly defined margins and heterogeneous colouration, thereby making reliable assessment of tissue extremely difficult [14]. While the red‐black‐yellow colour scale is classic, it lacks the precision required in wound assessment for standardised wound monitoring [9]. Blood, fascia and granulation all can have presentations in the red colour spectrum, making the colour classification approach inappropriate for assessing wound healing [9]. These manual, experience‐dependent methods are both time‐consuming and inconsistent, and introduce variability in the treatment decision‐making process [15].
Notably, our study found that wound care specialists were the most confident among clinicians in identifying (94.4%) and quantifying (66.6%) different tissue types. This is consistent with the findings of McCluskey and McCarthy [26], who suggested that self‐rated competence is highly correlated with the clinician's experience. Education and training play a role in developing reflective skills and self‐awareness [27]. Also, this could be due to their higher exposure to complex cases, so wound care specialists may critically reflect on their clinical practice, which would enhance their confidence of accurate assessment of the wound bed.
Our study further examined clinicians' ability to assess specific wound tissue types. For the assessment of granulation tissue and slough within the wound bed, clinicians correctly identified the presence/absence of granulation tissue in (82%) of cases on average and slough in (84%) across the nine wounds. Agreement among clinicians was very consistent across all groups, attesting to good inter‐rater reliability across the different experience levels. Furthermore, when examining the rates at which clinicians quantified granulation tissue and slough relative to their AI‐generated scores, clinicians showed a strong positive correlation with the AI‐generated scores.
Similarly, for eschar, clinicians successfully identified its presence or absence in 87% of cases, on average, across the nine wounds. There were minimal differences in agreement between clinicians across all groups, and their average quantifications were in close alignment with the AI scores. Therefore, these findings suggest granulation tissue, slough, and eschar can be reliably assessed and quantified across the participating clinicians with diverse levels of expertise. Furthermore, the strong agreement and alignment between clinician assessments and the AI scores further validated the consistency of such assessments, highlighting the potential for AI to standardise the identification of wound tissues across different clinical backgrounds.
Previous studies have explored the variability in clinician assessments of wound tissue types. A study by Ramachandram et al. found that clinicians demonstrated poor inter‐rater agreement for epithelialisation and slough, and moderate agreement for granulation and eschar. All assessments were conducted under uncontrolled lighting conditions and variable viewing angles. Moreover, the study utilised a browser‐based image annotation tool, which is more efficient at labelling tissue types; however, the authors indicated that such tools are not accessible in routine clinical practice, suggesting that real‐world assessments may exhibit even greater inter‐rater variability than what was observed in their study [7].
While granulation, slough, and eschar exhibit distinguishing visual features, factors such as ambient lighting, wound position on the body, and the experience of the clinician can all influence the accuracy of assessments [20, 28]. Traditional methods of wound assessment have tended to lean rather heavily upon clinicians' experience and subjective interpretation, thus creating variability in wound characterisations that may cause inconsistency in treatment approaches [14, 29, 30].
Despite that, our study recorded an overall high inter‐rater reliability across different clinician groups for these tissue types. This contrasts with previous research, which has linked higher clinical expertise, usually acquired through repeated exposure to complex cases, with enhanced wound assessment accuracy through critical reflection and pattern recognition [26, 29]. This may be because the wounds viewed in this study were AI‐captured images. The use of AI‐captured images might have played a critical role in facilitating more accurate differentiation of tissues, regardless of the clinician's experience. Specifically, the technology can reject images that are out of focus or have poor lighting, before applying colour correction to the captured image to account for variable lighting conditions. Research suggests that AI‐driven digital planimetry provides high‐resolution imaging with increased contrast and colour vibrancy, facilitating tissue differentiation, particularly in the case of subtly distinct or overlapping tissue types, irregular borders or complex presentations [31, 32]. A study conducted by Kabir et al. evaluated several AI‐enhanced imaging models, all of which proved to optimise the wound colour contrast. This enhancement facilitated an accurate segmentation and delineation between granulation from necrotic tissues, suggesting that these models are well‐calibrated for various tissue types, even in complex presentation [33]. Our study's findings align with this body of research, as we observed a high inter‐rater reliability and a strong positive correlation with AI‐generated scores. This suggests that the AI‐captured images, even without segmentation technology, may reduce experience‐related variability if used for assessment; using the segmentation technology would further enhance standardised assessments across clinicians, ensuring consistent assessments.
In contrast, epithelialisation presented a significant challenge in this study, with an average identification rate of 60% and poor agreement (ICC: 0.392–0.553), highlighting a high variability between clinicians across the groups for this tissue type. Notably, wound care specialists showed the highest ICC for epithelialisation at 0.553, yet this score remains relatively low, emphasising the challenge in assessing this tissue type. This finding aligns with other studies where clinicians also reported inconsistency with respect to the epithelialisation area [7, 31].
The challenge in detecting epithelialisation, even among the wound care specialists, stems from the inherent biological and visual characteristics of epithelial tissue [34]. Unlike slough and granulation, which feature distinct colours and textures, epithelial tissue often presents as a thin, nearly translucent layer that merges with the surrounding skin, complicating differentiation [35, 36]. Clinicians had to subjectively determine the wound boundaries, which may have introduced variability in identifying epithelialisation, a thin and subtle tissue type that typically forms at the wound edge. Unlike more prominent wound bed tissues, epithelialisation can be easily overlooked, contributing to lower ICC values and reduced agreement between clinicians. This inherent challenge in visualising epithelial tissue likely played a role in the observed discrepancies. Wound care specialists likely reached a higher ICC for epithelialisation due to their expertise and knowledge in wound care, enabling them to recognise subtle changes better than non‐specialists. However, their agreement was not as unanimous still compared to other tissue types, possibly because epithelialisation lacks strong visual markers [37] that would support more standardised assessments.
Moreover, clinicians' average ratings compared to the AI‐generated scores for epithelialisation showed weaker relationships with moderate correlation, representing greater variability among clinicians in all groups with the AI for the assessment of this tissue type. As clinicians exhibited high variability in their assessments and weak correlation with AI, this suggests that AI has a more standardised approach to detecting epithelial tissue compared to human observers. This could indicate that AI is better at identifying epithelialisation, likely because it applies the same predefined imaging processing algorithms and learned patterns without subjective interpretation.
Since clinicians gave rather inconsistent ratings on epithelialisation, regardless of experience level, the use of AI may therefore be used as a supplementary tool to facilitate segmentation and accuracy. By introducing AI‐assisted segmentation into clinical workflows, clinicians can perform their evaluations more uniformly while also reducing inter‐rater variability. This would be particularly helpful for epithelialisation, where subtle visual cues make it difficult for perceptual classifiers. Rather than replacing clinicians' judgement, the AI could act as a supportive tool to strengthen objectivity and assist clinicians in reliable tissue identification, particularly in difficult types of wound tissue.
Supporting this potential, a study by Bulten et al. developed a deep learning approach designed for the automatic segmentation of epithelial tissue in digitised prostatectomy slides. The researchers' method achieved an impressive score on an external validation set [38]. The notable level of recorded accuracy underscores the potential of AI to accurately identify epithelial regions, thereby promoting more consistent and objective evaluations of epithelialisation [38]. Similarly, Kumar et al. explored the application of combined optical coherence tomography and deep learning techniques for the effective segmentation of skin layers, including epithelium, in wound healing. This study proved the model could accurately delineate epithelial layers and provide enhanced accuracy in recording the progression of healing of a wound [39].
Collectively, these studies suggest that the AI‐driven tools could improve the accuracy and consistency of epithelial tissue segmentation, which may lead to great improvement in the evaluation of epithelialisation within the clinical setting.
4.1. Implications for Clinical Practice
At the patient‐clinician level, AI‐driven wound assessment tools have the potential to enhance efficiency in decision‐making and patient outcomes through standardised segmentation for objective and consistent evaluations. Standardised wound assessments facilitate the identification of a deteriorating wound earlier for the patient and thereby facilitate the opportunity to avert complications—such as infections and delayed healing [40, 41]. Improvement in consistency allows the visual and accurate tracking of wound healing, thereby improving clinician‐patient communication and fostering patient engagement in their care [42].
Beyond clinical accuracy, the use of AI‐driven assessment tools for evaluation also decreases operational costs, enabling the accomplishment of previously time‐consuming manual wound assessment faster. Research indicates that AI‐assisted wound assessment is 79% faster than the traditional manual approaches [20]. By leveraging AI, clinicians can use a more consistent and reliable method to evaluate wounds, reducing the risk of human error and ensuring an overall assessment quality [43]. These efficiencies can help prevent extremely costly clinical complications, like progressing stages of pressure injury, whose treatment, according to the Agency for Healthcare Research and Quality (AHRQ), could reach up to $151 700 US per patient [44]. Therefore, the incorporation of AI‐based tools into clinical practice has the potential to increase detection and earlier treatment of patients, thereby reducing the financial burden that wound complications can inflict [45].
At the organisational (meso) level, AI‐powered wound assessment tools can improve care coordination and assist organisations in tracking wound healing trends, adjusting treatment protocols, and evaluating departmental performance. By detecting early signs of wound deterioration, AI can help prevent longer stays at hospitals and associated expenditures [45], nurturing substantially lower expenses and better patient outcomes. Machine learning models for wound care, rating wound severity and prognostics, were compared to clinical tools (e.g., BWAT and PUSH) and performed better when using AI‐enabled tissue typing than clinician documentation alone [46].
The financial impact of preventing just one hospital admission for a wound‐related complication can save up to $30 000 per patient episode [44]. Additionally, with the help of AI, assessment and documentation accuracy are improved, lowering errors concerning billing and expediting the reimbursement process [14]. Reports indicate that nearly 20% of all claims for billing are denied due to insufficient documentation and approximately 60% of returned claims are never resubmitted. The cost to rework or appeal denials is $25 per claim on average for practices and $181 per claim for hospitals [47, 48]. Efficient assessments and documentation standardisation could lower denial rates of reimbursement, translating into substantial retained revenue for health facilities.
4.2. Limitation
While this study provides insightful information, several limitations must be considered. The study employed a non‐randomised purposive sampling approach, using mainly recruitment channels through a single conference and professional networks that might have limited generalisability of the findings.
Further, the survey used AI‐captured images; hence, it did not allow assessment of real‐time clinical patient assessments, which may have influenced the clinicians' evaluations. AI‐imaging standardises wound visualisation by holding the contrast and differentiation of colour into consideration but does not replicate real‐life circumstances where factors such as light exposure, position of the wound, and other surrounding aspects could affect the accuracy of the assessment. The controlled imaging environment may have contributed to higher inter‐rater reliability than what might be expected in an uncontrolled clinical setting.
Finally, the self‐reported clinician confidence levels could create a possible response bias. It is possible that more experienced clinicians reported higher confidence in identifying tissue types even when their assessments did not significantly differ from those offered by less experienced clinicians.
4.3. Conclusion and Future Directions
This study highlights the variability in clinician assessments of wound tissue, particularly epithelialisation, and demonstrates the potential of AI‐driven tools to standardise wound evaluation and improve consistency. By providing more reliable, consistent assessments, AI could lead to earlier detection of healing progress, more tailored treatment plans and ultimately, improved clinical decision‐making.
Expansion of studies with larger, more diverse clinician populations, real‐time patient assessments, and complex tissue presentation will also aid in generalisability. Future studies should validate the AI‐generated assessment against expert consensus or, where possible, enhanced clinical reference standards to further support its clinical reliability. Finally, future studies could investigate how the integration of an AI segmentation tool within routine clinical workflows affects patient outcomes, clinician efficiency, and health costs.
Ethics Statement
Our study maintained full compliance with ethical confidentiality standards, as no identifiable patient or clinician information was collected. A de‐identified database from participating wound care facilities using AI‐enabled digital wound imaging was used to select random de‐identified wound images; no patient‐related metadata or identifiable features were included or extracted for this study. As a quality improvement initiative, this study was exempt from ethical review, as it assessed clinician agreement in wound assessment rather than human subject research. This study was considered a quality improvement study and had been granted an exemption from ethics review by Pearl IRB LLC, an independent institutional review board (ID: 2023‐0100).
Consent
Clinician participation was voluntary and anonymous, with implied consent obtained through survey instructions outlining the study's objectives and methods.
Conflicts of Interest
H.T.M., R.D.J.F., S.W., Z.L., J.A. and A.C. are all current employees of Swift Medical Inc. The other authors declare no conflicts of interest.
Supporting information
Data S1. Supporting Information.
Funding: The authors received no specific funding for this work.
Data Availability Statement
The authors have nothing to report.
References
- 1. Swann G., “The Skin Is the Body's Largest Organ,” Journal of Visual Communication in Medicine 33, no. 4 (2010): 148–149, 10.3109/17453054.2010.525439. [DOI] [PubMed] [Google Scholar]
- 2. Baranoski S. and Ayello E. A., Wound Care Essentials, 3rd ed. (Wolters Kluwer, 2015). [Google Scholar]
- 3. Alhajj M. and Goyal A., “Physiology, Granulation Tissue,” in StatPearls [Internet] (StatPearls Publishing, 2025). [PubMed] [Google Scholar]
- 4. Pastar I., Stojadinovic O., Yin N. C., et al., “Epithelialization in Wound Healing: A Comprehensive Review,” Advances in Wound Care (New Rochelle) 3, no. 7 (2014): 445–464. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Flanagan M., “The Characteristics and Formation of Granulation Tissue,” Journal of Wound Care 7, no. 10 (1998): 508–510, 10.12968/jowc.1998.7.10.508. [DOI] [PubMed] [Google Scholar]
- 6. Stremitzer S., Wild T., and Hoelzenbein T., “How Precise Is the Evaluation of Chronic Wounds by Health Care Professionals?,” International Wound Journal 4, no. 2 (2007): 156–161, 10.1111/j.1742-481X.2007.00334.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Ramachandram D., Ramirez‐GarciaLuna J. L., Fraser R. D. J., Martínez‐Jiménez M. A., Arriaga‐Caballero J. E., and Allport J., “Fully Automated Wound Tissue Segmentation Using Deep Learning on Mobile Devices: Cohort Study,” JMIR mHealth and uHealth 10, no. 4 (2022): 36977, 10.2196/36977. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Wendelken M. E., Berg W. T., Lichtenstein P., Markowitz L., Comfort C., and Alvarez O. M., “Wounds Measured From Digital Photographs Using Photo‐Digital Planimetry Software: Validation and Rater Reliability,” Wounds 23, no. 9 (2011): 267–275. [PubMed] [Google Scholar]
- 9. Vermeulen H., Ubbink D. T., Schreuder S. M., and Lubbers M. J., “Inter‐ and Intra‐Observer (Dis)agreement Among Nurses and Doctors to Classify Colour and Exudation of Open Surgical Wounds According to the Red‐Yellow‐Black Scheme,” Journal of Clinical Nursing 16, no. 7 (2007): 1270–1277, 10.1111/j.1365-2702.2007.01789.x. [DOI] [PubMed] [Google Scholar]
- 10. Frykberg R. G. and Banks J., “Challenges in the Treatment of Chronic Wounds,” Advances in Wound Care 4, no. 9 (2015): 560–582, 10.1089/wound.2015.0635. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Vains A. S. and Finlayson K., “How Generalist Nurses Inform Their Clinical Decision‐Making in Wound Management: A Scoping Review,” Wound Practice and Research 30, no. 2 (2022): 112–118, 10.33235/wpr.30.2.112-118. [DOI] [Google Scholar]
- 12. Lindholm C. and Searle R., “Wound Management for the 21st Century: Combining Effectiveness and Efficiency,” International Wound Journal 13, no. Suppl 2 (2016): S5–S15, 10.1111/iwj.12623. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Chen M. Y., Cao M. Q., and Xu T. Y., “Progress in the Application of Artificial Intelligence in Skin Wound Assessment and Prediction of Healing Time,” American Journal of Translational Research 16, no. 7 (2024): 2765–2776, 10.62347/MYHE3488. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Alonso M. C., Mohammed H. T., Fraser R. D., Ramirez Garcia Luna J. L., and Mannion D., “Comparison of Wound Surface Area Measurements Obtained Using Clinically Validated Artificial Intelligence‐Based Technology Versus Manual Methods and the Effect of Measurement Method on Debridement Code Reimbursement Cost,” Wounds 35, no. 10 (2023): E330–E338, 10.25270/wnds/23031. [DOI] [PubMed] [Google Scholar]
- 15. Wang S. C., Anderson J. A. E., Evans R., et al., “Point‐of‐Care Wound Visioning Technology: Reproducibility and Accuracy of a Wound Measurement App,” PLoS One 12, no. 8 (2017): e0183139. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Barakat‐Johnson M., Jones A., Burger M., et al., “Reshaping Wound Care: Evaluation of an Artificial Intelligence App to Improve Wound Assessment and Management Amid the COVID‐19 Pandemic,” International Wound Journal 19, no. 6 (2022): 1561–1577, 10.1111/iwj.13755. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Cassidy B., Hoon Yap M., Pappachan J. M., et al., “Artificial Intelligence for Automated Detection of Diabetic Foot Ulcers: A Real‐World Proof‐of‐Concept Clinical Evaluation,” Diabetes Research and Clinical Practice 205 (2023): 110951, 10.1016/j.diabres.2023.110951. [DOI] [PubMed] [Google Scholar]
- 18. Queen D., Botros M., and Harding K., “International Opinion—The True Cost of Wounds for Canadians,” International Wound Journal 21, no. 1 (2023): 14522, 10.1111/iwj.14522. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Al‐Gharibi K. A., Sharstha S., and Al‐Faras M. A., “Cost‐Effectiveness of Wound Care,” Sultan Qaboos University Medical Journal 18, no. 4 (2018): e433–e439, 10.18295/squmj.2018.18.04.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Mohammed H. T., Bartlett R. L., Babb D., Fraser R. D. J., and Mannion D., “A Time Motion Study of Manual Versus Artificial Intelligence Methods for Wound Assessment,” PLoS One 17, no. 7 (2022): e0271742, 10.1371/journal.pone.0271742. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Koo T. K. and Li M. Y., “A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research,” Journal of Chiropractic Medicine 15, no. 2 (2016): 155–163. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Akoglu H., “User's Guide to Correlation Coefficients,” Turkish Journal of Emergency Medicine 18, no. 3 (2018): 91–93. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Barnard A. and Allison K., “The Classification and Principles of Management of Wounds in Trauma,” Trauma 11, no. 3 (2009): 163–176, 10.1177/1460408609336188. [DOI] [Google Scholar]
- 24. Mark Allen Group , “Developing Education,” Journal of Wound Care 4, no. 9 (1995): 389, 10.12968/jowc.1995.4.9.389. [DOI] [PubMed] [Google Scholar]
- 25. Rogers L. C., Bevilacqua N. J., Armstrong D. G., and Andros G., “Digital Planimetry Results in More Accurate Wound Measurements: A Comparison to Standard Ruler Measurements,” Journal of Diabetes Science and Technology 4, no. 4 (2010): 799–802, 10.1177/193229681000400405. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. McCluskey P. and McCarthy G., “Nurses' Knowledge and Competence in Wound Management,” Wounds 8, no. 2 (2012): 37–47. [Google Scholar]
- 27. Bulman C., Lathlean J., and Gobbi M., “The Concept of Reflection in Nursing: Qualitative Findings on Student and Teacher Perspectives,” Nurse Education Today 32, no. 5 (2012): e8–e13, 10.1016/j.nedt.2011.10.007. [DOI] [PubMed] [Google Scholar]
- 28. Bowling F. L., Paterson J., and Ndip A., “Applying 21st‐Century Imaging Technology to Wound Healing: An Avant‐Gardist Approach,” Journal of Diabetes Science and Technology 7, no. 5 (2013): 1190–1194, 10.1177/193229681300700518. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Flanagan M., “Improving Accuracy of Wound Measurement in Clinical Practice,” Ostomy/Wound Management 49, no. 10 (2003): 28–40. [PubMed] [Google Scholar]
- 30. Mekkes J. R. and Westerhof W., “Image Processing in the Study of Wound Healing,” Clinics in Dermatology 13 (1995): 401–407. [DOI] [PubMed] [Google Scholar]
- 31. Chairat S., Chaichulee S., Dissaneewate T., Wangkulangkul P., and Kongpanichakul L., “AI‐Assisted Assessment of Wound Tissue With Automatic Color and Measurement Calibration on Images Taken With a Smartphone,” Healthcare (Basel) 11, no. 2 (2023): 273, 10.3390/healthcare11020273. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Jiménez D., Casanova‐Lozano L., Grau‐Carrión S., and Reig‐Bolaño R., “Artificial Intelligence Methods for Diagnostic and Decision‐Making Assistance in Chronic Wounds: A Systematic Review,” Journal of Medical Systems 49 (2025): 29, 10.1007/s10916-025-02153-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Kabir M., Rahman M. M., Smith A., and Brown J., “Deep Learning‐Based Wound Tissue Segmentation: A Benchmark Study,” arXiv Preprint arXiv:2502.10652, https://arxiv.org/abs/2502.10652.
- 34. Labib A. and Winters R., “Complex Wound Management,” in StatPearls [Internet] (StatPearls Publishing, 2023), https://www.ncbi.nlm.nih.gov/books/NBK576385/. [PubMed] [Google Scholar]
- 35. Winnipeg Regional Health Authority , “Tissue Types in the Wound Bed,” 2024, https://professionals.wrha.mb.ca/files/Tissue‐Types‐in‐Wound‐Bed‐2024.pdf.
- 36. The WoundSource Editors , “Chronic Wound Management: Types of Wound Tissue,” WoundSource, April 19, 2018, https://www.woundsource.com/blog/chronic‐wound‐management‐types‐wound‐tissue.
- 37. Rennekampff H. O., Fimmers R., Metelmann H. R., Schumann H., and Tenenhaus M., “Reliability of Photographic Analysis of Wound Epithelialization Assessed in Human Skin Graft Donor Sites and Epidermolysis Bullosa Wounds,” Trials 16 (2015): 235, 10.1186/s13063-015-0742-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Bulten W., Kartasalo K., Pinckaers H., et al., “Epithelium Segmentation Using Deep Learning in H&E‐Stained Prostate Specimens With Immunohistochemistry as Reference Standard,” arXiv, Published August 17, 2018, 10.48550/arXiv.1808.05883. [DOI] [PMC free article] [PubMed]
- 39. Kumar P., Dhara S., Gope A., Chatterjee J., and Mandal S., “Deep Learning‐Based Skin‐Layer Segmentation for Characterizing Cutaneous Wounds From Optical Coherence Tomography Images,” arXiv, Published June 2, 2023, 10.48550/arXiv.2306.01252. [DOI] [PubMed]
- 40. Ahn C. and Salcido R., “Advances in Wound Photography and Assessment Methods,” Advances in Skin & Wound Care 21, no. 2 (2008): 85, 10.1097/01.asw.0000305411.58350.7d. [DOI] [PubMed] [Google Scholar]
- 41. Gould L. J. and Li W. W., “Defining Complete Wound Closure: Closing the Gap in Clinical Trials and Practice,” Wound Repair and Regeneration 27, no. 3 (2019): 201–224, 10.1111/wrr.12707. [DOI] [PubMed] [Google Scholar]
- 42. Healthy.io , “Digitising Wound Management: A Standardised, Data‐Driven Approach to Healing,” Published 2022, accessed March 4,2025, https://lp.healthy.io/wp‐content/uploads/2022/03/Digitising‐Wound‐Management‐White‐Paper.pdf.
- 43. Zoppo G., Marrone F., Pittarello M., et al., “AI Technology for Remote Clinical Assessment and Monitoring,” Journal of Wound Care 29, no. 12 (2020): 692–706, 10.12968/jowc.2020.29.12.692. [DOI] [PubMed] [Google Scholar]
- 44. Agency for Healthcare Research and Quality , “Preventing Pressure Ulcers in Hospitals: A Toolkit for Improving Quality of Care,” U.S. Department of Health and Human Services, accessed March 4, 2025, https://www.ahrq.gov/patient‐safety/settings/hospital/resource/pressureulcer/tool/pu1.html.
- 45. Padula W. V. and Delarmente B. A., “The National Cost of Hospital‐Acquired Pressure Injuries in the United States,” International Wound Journal 16, no. 3 (2019): 634–640, 10.1111/iwj.13071. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46. Gupta R., Goldstone L., Eisen S., et al., “Towards an AI‐Based Objective Prognostic Model for Quantifying Wound Healing,” TechRxiv, 2022, 10.36227/techrxiv.21067261.v1. [DOI] [PubMed]
- 47. Poland L. and Harihara S., “Claims Denials: A Step‐by‐Step Approach to Resolution,” Journal of AHIMA, Published 2022, accessed March 4, 2025, https://journal.ahima.org/page/claims‐denials‐a‐step‐by‐step‐approach‐to‐resolution.
- 48. Mello M. M. and Rose S., “Denial—Artificial Intelligence Tools and Health Insurance Coverage Decisions,” JAMA Health Forum 5, no. 3 (2024): e240622, 10.1001/jamahealthforum.2024.0622. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data S1. Supporting Information.
Data Availability Statement
The authors have nothing to report.