This review was based on the recommendations of the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines [13] and used the methodological framework developed by Peters et al [14]. The completed PRISMA-ScR checklist is provided in Checklist 1. This study was registered with PROSPERO on December 12, 2025 (registration number CRD420251251573).

The inclusion criteria were defined using the population, concept, and context framework. Detailed inclusion and exclusion criteria can be found in Multimedia Appendix 1.

To identify potentially relevant studies, the following bibliographic databases were searched up to December 10, 2025: PubMed, Cochrane Library, IEEE Xplore, and Web of Science. We limited the search to English-language publications and included specific article types (journal articles and conference proceedings), explicitly excluding preprints. The search strategies were drafted by an experienced librarian and further refined through team discussion. We supplemented our search with forward and backward citation chasing to minimize the risk of missing key studies.

This review was conducted by combining the terms “deep learning,” “multi-modal,” “artificial intelligence,” “convolutional neural networks,” and “pressure injury.” The search was conducted on April 15, 2026 (see Multimedia Appendix 2 for the complete database search strategies).

Two researchers conducted the literature search using the above-mentioned search terms. The final search results were exported into EndNote (Clarivate Analytics), and duplicates were removed by a researcher. Two independent investigators screened the retrieved studies based on their titles and abstracts against the eligibility criteria. Subsequently, the same 2 researchers conducted full-text screening.

We extracted data using the framework proposed by Peters et al [14] and populated data extraction forms. The 2 reviewers independently charted the data, resolving discrepancies through discussion, and continuously updated the data charting form in an iterative process.

The following data were extracted from each included study: article identifiers (author, year, and country), dataset size and source, modality type, model type, task, validation type, and primary metrics.

According to the methodological framework outlined in the PRISMA-ScR guidelines, this study did not undertake a formal critical appraisal or risk-of-bias assessment of individual studies. The primary aim of a scoping review is to map the breadth and nature of the existing evidence rather than synthesize effect sizes or assess the effectiveness of interventions. Therefore, the inclusion of studies of varying methodological quality, including those of lower quality, was intended to contribute to a comprehensive overview of the current state of research in this field.

We synthesized the studies according to their dataset size and source, modality type, model type, task, validation type, and primary metrics.

This scoping review used only previously published secondary data and did not involve direct human participant interaction.

The scoping review process is detailed in the flow diagram shown in Figure 1, presenting the stepwise filtering and the number of studies included at each step.

Table 1 presents an overview of the 15 included studies. Regarding study design, all studies were retrospective. Geographically, 33.3% (n=5) of the studies were conducted in China; 20% (n=3) were from South Korea; 13.3% (n=2) were from Turkey; and 6.7% (n=1) each were from the United Kingdom, the United States, Germany, Japan, and Saudi Arabia.

In total, 40% (6/15) of the studies applied multimodal integration, whereas 60% (9/15) applied unimodal imaging. This review involved 26 models, including AlexNet, VGGNet16, ResNet18, DenseNet121, SE-Swin Transformer, Cascade R-CNN, vision transformer, ConvNeXtV2, EfficientNetV2, Meta Former, TinyViT, CCM, binary classification model, ResNext+wFPN, SE-Inception, Mask-R-CNN, SE-ResNext101, Faster R-CNN, ResNet50, ResNet152, DenseNet201, EfficientNet-B4, YOLOv5, Inception-ResNet-v2, InceptionV3, and MobilNetV2. In total, 46.7% (7/15) of the studies [17,21,23-25,27,28] applied a single model for training, comparing the accuracy of intelligent PI recognition achieved through model application against that of subjective assessments conducted by health care professionals to evaluate model training effectiveness. A total of 40% (6/15) of the studies [15,16,18,19,26,29] applied 2 to 6 models for training, comparing the training outcomes of different models to determine the optimal model capable of achieving the highest accuracy in intelligent PI recognition. In total, 13.3% (2/15) of the studies [20,22] applied multiple models in combination to assess their accuracy in intelligent PI identification. Among all models mentioned, VGGNet16 was applied most frequently, featuring in 20% (3/15) of the studies [15,26,29]. DenseNet121 was applied in 13.3% (2/15) of the studies [15,29], ResNet50 was applied in 13.3% (2/15) of the studies [26,29], and ResNet152 was applied in 13.3% (2/15) of the studies [26,29]. The remaining models were each applied once. CNNs included algorithms such as AlexNet, VGG, ResNet, DenseNet, and EfficientNet. AlexNet, proposed by Krizhevsky et al [30], used a substantial number of parameters, resulting in extended training times and significant computational demands. VGG [31] used small convolution kernels and greater depth but had a large number of parameters and high computational cost. ResNet [32] introduced residual connections, allowing the network to learn identity mappings. DenseNet introduced dense connections, where each layer connects directly to all preceding layers [33]. EfficientNet [34] involved the coordinated scaling of the network’s depth, width, and resolution rather than scaling a single dimension.

Among the studies included in this review, the accuracy rates for intelligent PI staging identification varied across different models, fluctuating between 54.8% and 93.7%. We present performance figures for illustrative purposes only and do not compare the models’ accuracy. One study indicated that the DenseNet121 model achieved an identification accuracy of 93.7% (not reported [NR]) [15], whereas another study reported an accuracy rate of 67.2% (NR) for the same model [29]. VGGNet16 demonstrated relatively stable accuracy across 20% (3/15) of the studies [15,26,29], achieving values of 83.7%, 82.4%, and 71.8% (NR), respectively. ResNet152 recorded accuracy rates of 84.1% and 77.4% (NR) in 13.3% (2/15) of the studies [26,29]. ResNet50 achieved accuracy rates of 82.9% and 74.6% (NR) in 13.3% (2/15) of the studies [26,29]. The classification performance of AlexNet, VGGNet16, ResNet18, and DenseNet121 yielded overall accuracies of 87.7%, 82.4%, 92.4%, and 93.7% (NR), respectively, with corresponding F₁-scores of 94.7%, 88.2%, 98.2%, and 97.0% (NR) [15]. The SE-Swin Transformer model proposed in the study by Huang et al [16] achieved an accuracy of 87.1% (NR), surpassing other feature extraction models by at least 0.5 percentage points, as well as a mean absolute precision of 80.9% (NR). The results of the study by Cho and Yoo [17] indicated that the PUC-ViT model achieved an accuracy rate of 95.6% (NR) and F₁-score of 95.4% (NR). Among the 4 models applied in the study by Brehmer et al [18]—ConvNeXtV2, EfficientNetV2, MetaFormer, and TinyViT—the transformer-based TinyViT model demonstrated the most effective performance in the binary classification of pressure ulcers, achieving an F₁-score of 93.2% (NR). Among the 2 models applied in the study by Kimura et al [19], the CCM achieved an accuracy of 81.4% (NR), a precision of 81.6% (NR), and an F₁-score of 80.8% (NR). The binary classification model attained an accuracy of 87.1% (NR), a precision of 84.1% (NR), and an F₁-score of 85% (NR). The study by Wang et al [20] increased the accuracy rate for stage 3 PI diagnosis from 60.3% to 76.2% (NR) using ResNext+wFPN. The F₁-scores for stage 1 to 4 PI diagnosis were 81.5%, 80.8%, 81.6%, and 81.1% (NR), respectively. In the study by Gui et al [21], the SE-Inception model achieved an accuracy of 93% (NR) and an area under the curve of 94% (NR). In the study by Zalluhoğlu et al [22], a semiduplex CNN achieved an accuracy of 93% (NR) and F₁-score of 93% (NR). In the study by Swerdlow et al [23], the Mask-R-CNN model achieved an overall classification accuracy of 92.6% (NR) and a segmentation accuracy of 93.0% (NR). For stages 1 to 4, the model obtained F₁-scores of 84.2%, 94.7%, 90.7%, and 94.4% (NR) and Dice coefficients of 92%, 85%, 93%, and 91% (NR), respectively. In the study by Kim et al [24], the model achieved top-1 accuracy scores of 79.3% and 71.7% (NR) and F₁-scores of 79.1% and 71.5% (NR) on the internal and external test sets, respectively. The study by Fergus et al [25] used a confidence score threshold of 75% (NR), yielding an average precision of 67.9% and F₁-score of 67.8% (NR). In the study by Seo et al [26], for EfficientNet-B4, the macro–F₁-score was calculated to be 89.4% (NR), and the average performance of 2 experienced nurses was reported as 87.8% (NR). EfficientNet-B4 achieved an accuracy of 91.4% (NR). In the study by Aldughayfiq et al [27], the YOLOv5 model achieved an overall mean average precision 50 of 76.9% (NR) on the validation set, enabling it to accurately detect and classify PIs with a high degree of confidence. In the study by Liu et al [28], the Inception-ResNet-v2 model achieved an accuracy of 98.5% and an F₁-score of 94.1% (NR). In the classification task of necrotic tissue, the model attained an accuracy of 97% and an F₁-score of 93.5% (NR). Finally, in the study by Ay et al [29], ResNet152, ResNet50, VGGNet16, DenseNet121, InceptionV3, and MobileNetV2 achieved overall accuracies of 77.2%, 74.6%, 71.8%, 67.2%, 63.1%, and 54.8% (NR), respectively, with corresponding mean area under the curve values of 93%, 92%, 91%, 91%, 85%, and 85% (NR), respectively.

The validation Sankey diagram (Figure 2 [15-29]) delineated a critical attrition curve, revealing that while all 15 studies had retrospective designs, the flow of evidence was sharply constricted at the validation checkpoint, with only 26.7% (4/15) incorporating external test sets. For the purpose of this review, “external validation” was broadly defined to include any testing on data not used during model training, “multicenter” was defined as testing on data from a physically different institution from those of the training data, and “public dataset” was defined as testing on a publicly available benchmark dataset distinct from the authors’ private training data.

This bottleneck was further elucidated by the grouped bubble chart (Figure 3), which mapped the intersection of modality and clinical stage. While most of the research (11/15, 73.3%) remained concentrated at stage 1 (development), a notable finding was that 54.5% (6/11) of the studies in stage 1 were multimodal. This suggested that while the field was dominated by retrospective single-modality development, multimodal integration appeared to be a distinguishing characteristic of the studies that will progress toward clinical validation.

This scoping review mapped the contemporary landscape of deep learning in PI identification, revealing a field characterized by a profound modality chasm between computational sophistication and clinical utility. While the evolution of unimodal imaging exhibits diminishing returns, with marginal performance gains despite increasing model complexity, the integration of multimodal data remains nascent and fragmented.

The notable scarcity of true multimodal studies identified in this review underscores the next frontier for research. Our findings suggest that meaningful transition requires the integration of heterogeneous data sources. Currently, existing multimodal attempts are limited to incorporating thermal or basic clinical data. However, to bridge the translational gap, future systems must evolve from superficial sensor fusion to deep multimodal integration. This requires synergizing imaging data with longitudinal electronic health record, nutritional biomarker, and mobility data to capture a holistic view of patient health [43].

A limitation of this scoping review is the absence of a formal methodological quality appraisal of the included studies. According to standard scoping review methodology, the focus was on mapping the landscape of available evidence rather than assessing the risk of bias or the certainty of evidence.

The current landscape of deep learning in PI identification is heavily skewed toward unimodal imaging, creating a modality gap that hinders clinical translation. While unimodal models demonstrate high accuracy in controlled settings, they lack the physiological context required for robust wound assessment. Future research must prioritize bridging the gap from unimodal imaging to true multimodal integration, shifting the focus from algorithmic architecture to the fusion of heterogeneous data sources to achieve clinical-grade reliability.