This cross-sectional comparative study evaluated the performance of 3 participant groups on a standardized 25-question wound care certification examination: (1) MLLMs (n=3), (2) unimodal large language model (n=1), and (3) human clinical experts (n=4). The study was conducted at Etimesgut Şehit Sait Ertürk State Hospital, Ankara, Turkey, between January and March 2025. This study was designed and reported in accordance with the Standards for Reporting of Diagnostic Accuracy Studies—Artificial Intelligence extension and the COSMIN (Consensus-Based Standards for the Selection of Health Measurement Instruments) guideline [9,10].

Twenty-five multiple-choice questions were systematically compiled from established wound care certification examination resources and clinical practice guidelines. Primary sources included the National Alliance of Wound Care and Ostomy WCC Practice Examination Bank (2024); the American Board of Wound Medicine Certified Wound Specialist Examination Preparation Guide (2023‐2024); the Wound, Ostomy, and Continence Nursing Certification Board Certification Content Outline (2024); the National Pressure Injury Advisory Panel Clinical Practice Guideline (2019); the Wound Healing Society Diabetic Foot Guidelines (2023); and the International Wound Infection Institute Consensus Document (2022). Question distribution across clinical domains is presented in Table 1 [11-16].

Four clinical experts representing diverse wound care backgrounds participated in this study: a board-certified general surgeon (GS) with 10 years of wound care experience specializing in debridement, negative pressure wound therapy, grafting, and flap reconstruction; a certified wound care nurse (WCN) with 8 years of experience in assessment, dressing selection, and compression therapy; and 2 internal medicine physicians with palliative care experience (IM-1: 5 y; IM-2: 3 y) focusing on wound management in chronic disease and comfort-oriented care settings. All 4 experts completed the examination simultaneously in the same room under standardized proctored conditions on February 15, 2025 (14:00-15:30) at Conference Room B-204, Etimesgut Şehit Sait Ertürk State Hospital, with desks arranged at 2-m intervals with privacy dividers, standardized lighting (500 lux), and temperature control (22 °C). Two independent proctors supervised throughout: Proctor 1 (front) managed timing and instructions, while Proctor 2 (rear) monitored for communication; duties included identity verification, electronic device collection, simultaneous sealed packet distribution, continuous monitoring, time announcements (15-min and 5-min warnings), and sealed answer sheet collection at 90 minutes. Each participant received identical sealed packets containing instruction sheets, blinded answer sheets (codes WC-A through WC-D), question booklets (25 multiple-choice questions), high-resolution wound photographs (10×10 cm, 300 DPI, color calibrated), and clinical history sheets, with written instructions prohibiting reference materials, electronic devices, and interparticipant communication. Completion times were 67 minutes (GS), 72 minutes (WCN), 81 minutes (IM-1), and 85 minutes (IM-2).

Group differences in overall accuracy were assessed using one-way ANOVA with Tukey honestly significant difference post hoc tests for pairwise comparisons. Effect sizes were calculated using eta-squared (η²) for ANOVA and Cohen d for pairwise comparisons. Domain-specific analyses employed the Kruskal-Wallis H test. Pairwise comparisons used Fisher exact test with Bonferroni correction. Performance correlations with clinical experience and specialization were assessed using Pearson correlation coefficient (r). All statistical analyses were performed using R version 4.3.2 (R Foundation for Statistical Computing), with statistical significance set at P<.05.

This study was approved by the Ankara Provincial Directorate of Health Non-Interventional Ethics Committee (decision number 2025-10-3; October 24, 2025). As this study used anonymized examination data and involved no direct patient intervention, individual informed consent was not required per applicable institutional guidelines.

Three distinct participant groups were evaluated on a standardized 25-question wound care certification examination designed to assess competency across 4 clinical domains: Diagnosis, Treatment, Complication Management, and Wound Subtype Knowledge. The participant groups included MLLMs (n=3 models: Med-PaLM 2, LLaVA-Med, and BioGPT), a unimodal large language model (ChatGPT 5.0; n=1 model), and human clinical experts (n=4 participants: GS, WCN, IM-1, and IM-2). Substantial and statistically significant performance differences were observed across groups, with a clear hierarchical pattern emerging that reflects the fundamental importance of visual processing capabilities for clinical wound assessment (Figure 2).

Human clinical experts achieved the highest aggregate performance with a mean accuracy of 86% (SD 9.1%), corresponding to 21.5 out of 25 questions answered correctly on average. Individual human expert scores ranged from 76% to 96%, with a 95% CI for the group mean of 71.5% to 100.5%. The coefficient of variation within the human expert group was 10.6%, indicating moderate within-group heterogeneity attributable to differences in specialized training and clinical experience. All 4 human experts (100%) exceeded the 70% certification passing threshold, demonstrating consistent competency across the group despite varying levels of wound care specialization.

MLLMs achieved the second-highest aggregate performance with a mean accuracy of 78.7% (SD 12.2%), corresponding to 19.7 out of 25 questions answered correctly on average. Individual MLLM scores demonstrated substantial variability, ranging from 68% (BioGPT) to 92% (Med-PaLM 2), with a 95% CI for the group mean of 48.4% to 109%. The coefficient of variation within the MLLM group was 15.5%, higher than that observed for human experts, reflecting considerable heterogeneity in model architecture, training data, and multimodal fusion strategies. Two of the 3 MLLMs (66.7%) exceeded the 70% certification threshold: Med-PaLM 2 at 92% and LLaVA-Med at 76%, while BioGPT narrowly failed at 68%.

The unimodal ChatGPT 5.0 achieved the lowest performance at 64% accuracy, correctly answering 16 of 25 questions. As a single-model comparator, no within-group variability statistics are applicable; however, the model’s performance fell 6 percentage points below the 70% certification threshold, representing a clinically meaningful failure to demonstrate wound care competency. Despite representing the state of the art in text-based AI with enhanced medical knowledge encoding and sophisticated reasoning capabilities, ChatGPT 5.0’s text-only architecture proved insufficient for certification-level performance in this visually dependent clinical domain.

One-way ANOVA revealed statistically significant overall differences among the 3 groups (F_2,5=8.42, P=.018). The effect size was large (η²=0.74), indicating that group membership—reflecting the presence or absence of multimodal processing capabilities and human expertise—explained approximately 74% of the total variance in examination performance. This finding demonstrates that architectural differences in information processing have profound implications for clinical decision-making accuracy in wound care assessment, with visual processing capabilities representing a critical determinant of performance.

Post hoc pairwise comparisons using Tukey Honestly Significant Difference test revealed the specific nature of between-group differences. Human clinical experts significantly outperformed ChatGPT 5.0 with a mean difference of 22 percentage points (95% CI 8.4%‐35.6%; P=.006), representing the largest between-group difference observed and a very large effect size (Cohen d=2.12). This 22-percentage-point advantage for human experts over the most advanced unimodal AI system underscores the continued importance of human clinical judgment for tasks requiring the integration of visual assessment with experiential pattern recognition.

MLLMs significantly outperformed ChatGPT 5.0 with a mean difference of 14.7 percentage points (95% CI 2.3%‐27.1%; P=.03) and a large effect size (Cohen d=1.38). This finding provides direct evidence that multimodal processing capabilities—specifically, the architectural ability to directly analyze clinical images rather than relying on verbal descriptions—confer substantial and statistically significant advantages for wound care decision-making. The multimodal advantage translates to approximately 3.7 additional correct answers per 25-question examination, representing a clinically meaningful improvement in diagnostic accuracy.

Human clinical experts demonstrated a trend toward higher accuracy compared to MLLMs, with a mean difference of 7.3 percentage points (95% CI −2.1% to 16.7%; P=.09) and a medium effect size (Cohen d=0.68). Although this difference did not achieve statistical significance at the conventional α=.05 threshold, the substantially underpowered nature of this comparison (post hoc power=0.52) and the wide CIs spanning zero preclude definitive conclusions regarding either superiority or equivalence. Thus, while the direction of the effect favors human experts, adequately powered studies are necessary to confirm this trend.

Examination completion time differed markedly between human experts and AI models, with implications for clinical workflow integration. Human experts required a mean of 76.3 (SD 8.1) minutes to complete the 25-question examination (range 67‐85 min, median 76.5 min). Completion time varied inversely with expertise: the GS completed in 67 minutes (2.68 min/question), the WCN in 72 minutes (2.88 min/question), IM-1 in 81 minutes (3.24 min/question), and IM-2 in 85 minutes (3.40 min/question). The correlation between completion time and accuracy was strongly negative (r=−0.89, P=.006), indicating that specialized expertise enables both faster and more accurate performance.

All AI models generated responses essentially instantaneously, with total examination times under 1 minute regardless of model type. Med-PaLM 2 averaged 2.3 (SD 0.4) seconds per question (total 57.5 s), LLaVA-Med averaged 1.8 (SD 0.3) seconds (total 45 s), BioGPT averaged 1.1 (SD 0.2) seconds (total 27.5 s), and ChatGPT 5.0 averaged 0.9 (SD 0.2) seconds (total 22.5 s). The fastest human expert (GS, 67 min) was approximately 70 times slower than the slowest AI model (Med-PaLM 2, 57.5 s), representing a substantial efficiency advantage for AI systems. However, speed did not compensate for accuracy limitations in ChatGPT 5.0, which was both fastest and least accurate.

This study provides the first systematic comparison of MLLMs, unimodal large language models, and human clinical experts on a standardized wound care certification examination. Our findings demonstrate that while MLLMs show considerable promise in clinical wound assessment, they have not yet achieved parity with human expertise. Human clinical experts achieved the highest overall accuracy (86, SD 9.1%), followed by MLLMs (78.7, SD 12.2%), while the text-only ChatGPT 5.0 performed significantly below the certification threshold (64%). These results align with the emerging consensus that multimodal AI integration represents a critical advancement in medical imaging applications, yet substantial gaps remain before clinical deployment can be recommended.

The most striking finding was the pronounced multimodal advantage in visually dependent domains. MLLMs outperformed ChatGPT 5.0 by 38 percentage points in diagnosis (P=.006) and 22 points in complication management (P=.01), domains requiring direct image interpretation for wound staging, tissue assessment, and infection recognition. Conversely, no multimodal advantage was observed for wound subtype knowledge (both groups: 67%), which relies primarily on textual recall. This domain-specific pattern supports the theoretical framework proposed by Jung et al [17], who emphasized that MLLMs achieve their greatest utility when visual and textual information must be simultaneously integrated for clinical reasoning. Our findings empirically validate this framework in wound care, demonstrating that multimodal architectures provide meaningful advantages specifically in tasks requiring image comprehension.

Our results contribute to the growing body of evidence examining AI performance against human clinicians. Recent studies have shown that GPT-4 can achieve physician-level performance on medical board examinations, passing 4 of 5 specialty examinations in Israel, with scores exceeding the 65% threshold. Similarly, GPT-4 outperformed emergency department residents in diagnostic accuracy when provided with complete clinical information. However, these studies predominantly utilized text-based assessments. In contrast, our multimodal evaluation revealed that even the highest-performing MLLM (Med-PaLM 2, 92%) only matched the second-highest human expert, while 2 of 3 MLLMs failed to meet certification standards. This performance gap is consistent with Jin et al [18], who found that GPT-4V frequently presents flawed rationales despite achieving correct final answers, particularly in image comprehension tasks.

The superiority of human experts over ChatGPT 5.0 is statistically robust (P=.006, Cohen d=2.12). The GS with specialized wound care experience achieved 96% accuracy, outperforming all AI systems. However, for the MLLM-human comparison, the 7.3-percentage-point difference (P=.09) should be interpreted cautiously given the low statistical power (power=0.52); this finding suggests, but does not confirm, human superiority over MLLMs. Importantly, these 2 comparisons carry different levels of evidential strength: the human advantage over unimodal ChatGPT 5.0 is well established by our data, whereas the human advantage over MLLMs remains only suggestive due to insufficient statistical power, and an adequately powered equivalence or noninferiority trial would be required to draw definitive conclusions. This finding aligns with Kücking et al, who demonstrated that factors related to expertise—such as formal qualifications, deliberate practice, and diagnostic confidence—play a significant role in clinical judgment during wound care assessments [19]. Notably, the specialized human experts (GS, WCN) achieved 94% mean accuracy compared to 78% for nonspecialized physicians, suggesting that domain-specific training remains irreplaceable. The recent systematic review by Reifs Jiménez et al [20] similarly concluded that while AI systems excel at standardized pattern recognition, human clinicians demonstrate superior contextual reasoning and integration of atypical presentations.

From a clinical implementation perspective, our findings suggest that MLLMs may serve as valuable decision-support tools rather than autonomous diagnostic systems. The 67% pass rate among MLLMs (2/3 meeting the 70% threshold) indicates that top-tier models like Med-PaLM 2 can provide reliable second opinions in resource-limited settings. Grunhut and Nagarsheth [21] recently proposed that AI-powered wound assessment tools are best positioned for triage, preliminary documentation, and alerting clinicians to potential complications—functions that complement rather than replace expert judgment. Similarly, Barakat-Johnson et al [22] demonstrated that AI-assisted wound imaging improved standardization of assessments during the COVID-19 pandemic while maintaining physician oversight. Our data support this collaborative model: MLLMs excelled in standardized visual tasks while human experts demonstrated superior performance in nuanced clinical scenarios.

Recent advances in AI-powered wound assessment tools further contextualize our findings. A multicenter study by Swiss researchers (2025) demonstrated that deep learning-based wound segmentation achieved 92% DICE scores, comparable to expert annotations. However, tissue classification accuracy varied considerably across wound types, particularly for fibrin and necrosis—findings that parallel our observation of variable MLLM performance across clinical domains [23]. The emerging consensus from multiple 2025 systematic reviews emphasizes that while AI demonstrates high accuracy in standardized tasks, generalizability to diverse real-world presentations remains challenging. Russ et al [24] identified key barriers, including data quality heterogeneity, limited interpretability, and the need for standardized evaluation frameworks before widespread clinical adoption.

The efficiency differential between AI and human assessors merits attention. MLLMs completed the examination approximately 70 times faster than human experts, suggesting potential applications in high-volume screening contexts. Mohammed et al [25] reported similar findings, demonstrating that AI-powered wound measurement reduced assessment time by 75% compared to manual methods. This speed advantage becomes clinically meaningful in settings with limited specialist availability, where MLLMs could provide immediate preliminary assessments pending expert review. However, the inverse correlation between human completion time and accuracy (r=−0.89; P=.006) suggests that experienced clinicians process wound images more efficiently through pattern recognition developed over years of practice—a capability that current AI systems have not fully replicated.

Several limitations should be acknowledged. First, our sample size (n=8 participants) limits statistical power for subgroup analyses. Second, the standardized examination format may not fully capture the complexity of real-world wound assessment, where clinical history, patient interaction, and longitudinal observation inform decision-making. Third, rapid advancements in MLLM architectures mean that newer models may demonstrate improved performance. Fourth, the observed unimodal performance deficit likely represents a composite effect of (1) inherent text-only processing limitations and (2) potential information loss or observer-dependent bias during the human-mediated translation from clinical images to standardized verbal descriptions; the verbal description protocol was not independently validated, and the 12-item template (detailed in Multimedia Appendix 2) may not have captured all visually relevant features. Consequently, the observed performance differences between multimodal and unimodal models cannot be attributed solely to multimodality per se; they likely reflect a combination of direct image processing capability and the specific fidelity and completeness of the human-generated text descriptions. Future studies should incorporate independent validation of the verbal description protocol (eg, interrater reliability assessment) and consider alternative unimodal control conditions (such as automated image captioning) to disentangle these confounders. The recent development of region-grounded MLLMs with segmentation-aware spatial tokens, as described by Stefanelli et al [23], represents a promising direction for enhancing wound-specific AI capabilities. Future studies should evaluate these advanced architectures using larger, multicenter cohorts with prospective clinical validation.

MLLMs demonstrate significant performance advantages over the state-of-the-art unimodal ChatGPT 5.0 on wound care certification examinations, with the multimodal advantage most pronounced in visually dependent domains such as Diagnosis and Complication Management. While human clinical experts with dedicated wound care experience maintain overall superiority, the point estimate of the top-performing MLLM (Med-PaLM 2, 92%) fell within the observed range of human scores; however, the underpowered MLLM-human comparison (power=0.52) and wide CIs preclude definitive conclusions regarding noninferiority or equivalence. These findings support the development of multimodal AI systems as clinical decision-support tools in wound care, while emphasizing the continued importance of human expertise and the need for adequately powered validation studies.