Research on real-time detection and staging technology for pressure injuries in critically ill patients based on YOLOv8

Research on real-time detection and staging technology for pressure injuries in critically ill patients based on YOLOv8 | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Research on real-time detection and staging technology for pressure injuries in critically ill patients based on YOLOv8 Ningbo Cao, Fengbo Guo, Jingyi Nie, Wenwen Guo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8449503/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background To develop and validate a real-time detection and staging system for PIs using the YOLOv8 deep learning model. Methods A total of 507 PI images from intensive care unit patients (Jan 2023-Jun 2025) were randomly divided into training (414) and test (93) sets. Images were classified into six stages per NPUAP guidelines [ 1 ] . Five YOLOv8 versions were developed using transfer learning, with AdamW optimizer and dynamic learning rate adjustment. The best model was evaluated on precision, accuracy, and inference speed. Results This model effectively enhanced the objectivity and accuracy of pressure injury (PI) staging identification. In testing with 93 PI images, YOLOv8l achieved the optimal balance with 0.854 precision and 0.35 fps/img inference speed, outperforming other versions. Additionally, the model demonstrated high prediction accuracy across all six PI stages: all Stage 2, Stage 4, and Unstageable images were correctly predicted; one image each in Stage 1, Stage 3, and Deep Tissue Injury was misclassified. Conclusions For PI staging identification, the PI assessment system built on the YOLOv8l deep learning model demonstrates high accuracy and efficiency, providing reliable support for clinical decision-making, thereby delivering more personalized care to critically ill patients and significantly reducing pressure injury-related healthcare costs. Biological sciences/Computational biology and bioinformatics Physical sciences/Engineering Health sciences/Health care Physical sciences/Mathematics and computing Health sciences/Medical research Pressure injury YOLO Deep learning Image processing Object detection Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Pressure injury (PI), also known as pressure ulcer [ 2 ] , refers to localized tissue damage caused by intense and/or prolonged pressure, or the combined effects of pressure and shear force, affecting the skin and underlying soft tissues. These injuries typically occur over bony prominences, presenting clinically as damage beneath intact skin or open ulcers. They may also involve mucosal tissues (such as the mucosal surface, submucosa, or submucosal layer) or be associated with medical device use. The injury may be painful or painless, and its development is influenced by multiple factors including the microenvironment, nutritional status, and complications. The intensive care unit (ICU) is a specialized setting for the treatment of critically ill patients. Due to the severity of their conditions, ICU patients often require multiple supportive medical devices, prolonged bed rest, and frequently exhibit sensory impairment, reduced communication and mobility abilities, and altered consciousness. These factors can lead to prolonged pressure on the skin and subcutaneous tissues, ultimately resulting in PI [ 3 ] . Research statistics [ 4 ] indicate that compared to the 5%-15% PI incidence rate among general inpatients, the rate reaches 15%-25% during ICU stays. This increases patients' susceptibility to infections, prolongs hospital stays, and incurs treatment costs that exceed expenditures allocated by individuals and society for health prevention through medical insurance. As frontline caregivers for critically ill patients, ICU nurses play a crucial role in reducing PI incidence through accurate assessment and appropriate nursing interventions [ 5 ] . However, traditional assessment methods are influenced by multiple factors, including individual knowledge gaps, inadequate training, subjective judgment, and environmental variables like lighting and skin discoloration [ 6 ] . These limitations not only increase workload but may also cause patient discomfort during measurement, with varying reliability of outcomes [ 7 ] . Precise diagnosis and assessment of PI enable healthcare providers to objectively evaluate PI staging and select treatment modalities, thereby delivering more personalized care to critically ill patients and significantly reducing PI-related healthcare costs [ 8 ] . In recent years, the continuous advancement of artificial intelligence through deep learning has provided favorable conditions for researchers in the field of intelligent healthcare [ 9 ] . Convolutional neural networks (CNNs), as a machine learning algorithm within deep learning, are widely employed for tasks like wound segmentation and classification due to their efficient image processing capabilities [ 10 , 11 ] . Faster region-based convolutional neural networks (R-CNNs) demonstrate greater accuracy in detecting wound boundaries and performing image detection and classification, though they demand higher computational resources and longer training times [ 12 , 13 ] . In contrast, YOLO (You Only Look Once), introduced in 2015 and refined through multiple iterations, has emerged as a cutting-edge real-time object detection AI solution [ 14 ] . Unlike the two-stage detection strategy (candidate region generation + precise localization) of traditional models like R-CNN, YOLO transforms object detection into a regression problem, completing detection with a single image scan. This approach achieves faster execution and reduced computational resource requirements [ 15 ] . The 2023 YOLOv8 introduced an anchor-free mode, enabling direct prediction of object center, size, and shape without predefined anchor boxes [ 16 ] . This approach streamlines the detection process, making it faster and more efficient, particularly for real-time medical image analysis [ 17 ] . However, domestic research on utilizing deep learning for automated PI staging remains scarce. Therefore, addressing the current lack of precise and unified standards for clinical nurses' subjective assessment of PI staging features, this study develops a novel technology for real-time detection and staging of (PIs) based on the YOLOv8 neural network model. It aims to provide clinicians with a reliable PI staging tool, improve patient treatment outcomes, and alleviate the burden on healthcare professionals. Materials and Methods 1.1 Study Population This study selected 1,024 images of PI from patients admitted to our hospital's intensive care unit between January 2023 and June 2025. After screening, 507 images met the inclusion criteria for analysis. All data were stored in the hospital's dedicated electronic PI information management system. 1.2 Data Collection Nursing staff capture patients' PI images in clinical settings using medical PDAs or personal mobile phones under natural lighting. Images are taken at a distance of 50 cm while maintaining stability during capture. Completed images are subsequently uploaded to the electronic PI information management system. 1.3 Image Annotation This study obtained a total of 1,024 images. After excluding images with poor resolution, overexposure or underexposure, inadequate lighting, and cluttered backgrounds, 507 high-resolution images with appropriate focal length, sufficient lighting, and minimal noise were selected. These were randomly divided into a training set (414 images) and a test set (93 images) at an 8:2 ratio. Image annotation was performed collaboratively by three experienced Wound, Ostomy, and Continence (WOC) nurses from the hospital. Following the 2019 Pressure Ulcer Prevention and Treatment Guidelines published by the National Pressure Ulcer Advisory Panel (NPUAP) [ 1 ] , images were classified into six stages. In cases of disagreement, staging was determined through consensus reached via group discussion. Image annotation was performed using the MakeSense online annotation tool, with annotation files exported in YOLO data format. These verified annotated images were subsequently used to train the deep learning model. Reference criteria for each stage are shown in Table 1 [ 18 ] , the distribution of images across stages is illustrated in Fig. 1, and the study workflow is depicted in Fig. 2. 1.4 Preprocessing To enhance the model's detection accuracy and generalization capability, preprocessing and augmentation operations were applied to the image data. Various modifications were implemented, including random vertical and horizontal flipping, rotation between 0 and 90 degrees, and brightness reduction. Furthermore, the images were resized to 224×224 pixels, and their pixel intensity values were normalized to the range [0, 1] to improve training stability. 1.5 Training Configuration This study employed a transfer learning strategy [ 19 ] , utilizing five versions of the YOLOv8 model pre-trained on the large-scale COCO dataset: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. These models differ in parameter count and inference speed, making them suitable for different scenarios. We retained the shallow layers of the pre-trained models and fine-tuned the deeper layers. During training, the AdamW optimizer was selected, and the learning rate was adjusted according to the configuration file. The learning rate was halved every 5 training epochs to stabilize training and achieve fine-tuning. The training was set for 100 epochs, incorporating the Early Stopping technique to terminate training early if no improvement was observed. The batch size was set to 20 and adjusted based on GPU memory availability to ensure stable training. Additionally, YOLOv8's default compound loss function was used to balance localization and classification losses, and automated mixed-precision training on the GPU was enabled to increase training speed and reduce memory usage. 1.6 Evaluation Metrics In this study, we employed several metrics to evaluate model performance, including precision, recall, F1-score, mAP@50, and mAP@50–95. Precision refers to the proportion of samples predicted as positive by the model that are actually true positives, reflecting the model's accuracy. Recall refers to the proportion of actual true positive samples that are correctly predicted as true positives by the model, measuring the model's comprehensiveness, i.e., its ability to identify actual true positive samples. The F1-score is the harmonic mean of precision and recall, providing a comprehensive evaluation of the model's performance. mAP is a common metric for evaluating models, measuring the overall performance across different categories. mAP@50 refers to the mean average precision when the IoU threshold is set to 0.5, while mAP@50–95 represents the average precision of the model across IoU thresholds ranging from 0.5 to 0.95. Additionally, the confusion matrix was used to visualize prediction results across various categories, aiding in the analysis of the model's classification effectiveness for different stages. Furthermore, evaluation based on frame rate was conducted to reflect the system's real-time image processing capability, ensuring the requirement for immediate feedback in clinical applications. 1.7 Experimental Platform This study was conducted on a high-performance computing platform equipped with an NVIDIA RTX 3080 Ti GPU (12GB VRAM), an Intel Xeon Platinum 8369B processor, and 1TB of storage. Model development, training, and optimization were performed using the PyTorch 2.0 deep learning framework (with CUDA 11.8 acceleration library). Data processing and statistical analysis primarily relied on the Pandas 1.3.5 and NumPy 1.21.3 libraries. Result visualization tasks were implemented using Seaborn 0.12.2 and Plotly 5.12.0. Results 2.1 Model Training Figure 3 illustrates the evolution of loss functions during YOLOv8 training. Bounding box loss represents the discrepancy between the coordinates of the predicted and ground truth bounding boxes. Class loss indicates the probability of each detected object belonging to a specific category. The Distribution Focal Loss, built upon focal loss, enhances the model by distributing the focal loss across multiple scales and classes. The figure demonstrates a declining trend in all loss values as training progresses, indicating that the model is converging toward optimization. Concurrently, Fig. 4 shows an increasing trend in the model's precision, recall, and mAP with the progression of training epochs. This suggests that the model's capability to detect PI at different stages in images continuously improves as the number of training iterations increases. 2.2 Model Performance Evaluation 2.2.1 Bounding Box Regression Prediction Performance of Different YOLOv8 Model Versions Based on Table 2, this study systematically evaluated the bounding box regression prediction performance of the YOLOv8 series models. Results indicate that detection accuracy metrics (e.g., mAP@50, mAP@50–95) significantly improve as model size increases. Among these, YOLOv8l achieves optimal mAP@50 (0.847) and mAP@50–95 (0.577), substantially outperforming other variants. However, inference speed decreases with increasing model size, as detailed in Table 2. 2.2.2 Performance of the YOLOv8l Model at Different Confidence Thresholds As shown in Figure 5, the model's performance across different confidence thresholds was systematically evaluated using the Precision-Recall curve, Precision-Confidence curve, Recall-Confidence curve, and F1 Score-Confidence curve. The P-R curve illustrates the dynamic trade-off between precision and recall, providing critical guidance for model threshold optimization. The Precision-Confidence and Recall-Confidence curves depict the trends of precision and recall at varying confidence levels, respectively. They indicate that the model effectively enhances lesion screening sensitivity at lower thresholds while significantly improving detection specificity at higher thresholds. In the F1-Confidence curve, the F1 score initially increases and then decreases with rising confidence, reaching a peak value of 0.78 for all classes at a confidence of 0.564, achieving the optimal balance between precision and recall. Combined with subsequent validation by the PI nursing team, this multi-dimensional confidence analysis establishes the model's reliability and practicality in real-world applications. Overall, the results from the YOLOv8l model demonstrate its accurate detection and classification of (PIs) within the training and validation datasets. 2.2.3 Predictive Performance of the YOLOv8l Model for PI Staging The YOLOv8l model was used to predict 93 PI images at different stages from the test set, and a confusion matrix was constructed by comparing the predictions with the ground truth labels (Fig. 6). As shown in the figure, all Stage 2, Stage 4, and Unstageable images were correctly predicted, achieving 100% accuracy. However, one image each from Stage 1, Stage 3, and Deep Tissue Injury was misclassified. This indicates that the network model performs well in identifying PI stages, particularly for Stage 2, Stage 4, and Unstageable images. Additionally, Fig. 7 displays representative image analysis results detected by the YOLOv8l model, including the stage, confidence score, and visual annotations. The visualization of prediction outcomes enhances their credibility. Discussion PI is a common and serious complication in patients with prolonged bed rest or limited mobility, such as critically ill patients, and its incidence is closely associated with mortality. The Global Burden of Disease study report [ 20 ] indicates that China has become a high-incidence country for PI due to the dual effects of an accelerating aging population and imbalanced allocation of medical resources. PI not only causes severe pain and increases the risk of infection in patients, but the prolonged and difficult healing process also significantly extends hospital stays and consumes medical resources [ 21 ] . Clinical prevention and management face multiple challenges. Traditional assessment relies on manual observation, which is subjective, time-consuming, and labor-intensive, making it difficult to detect early-stage damage promptly, especially in intensive care settings [ 22 ] . Staging diagnosis requires extensive professional expertise, and delays in judgment can easily lead to deterioration of the condition. Although multidisciplinary collaboration and technologies such as advanced dressings are continuously advancing, there remains an urgent need for real-time, accurate, and automated detection and staging methods. Lei et al. [ 23 ] investigated the capability of four different CNN models for classifying PI images. Their results showed that DenseNet121 achieved the highest classification accuracy of 93.71%, demonstrating strong performance in PI image classification. Ge et al. [ 24 ] compared a CNN-based AI system with traditional assessment methods for PI evaluation. The results indicated that the AI system achieved an accuracy of 90%, superior to the 81.2% recognition accuracy of traditional methods, significantly reducing assessment time and improving the overall efficiency of PI evaluation. Chen et al. [ 25 ] established an intelligent telemedicine diagnosis system based on YOLOv7 and a large language model to assist users in staging and classifying PI, providing real-time and accurate diagnostic and treatment suggestions. In contrast, this study developed a real-time detection and staging model for (PIs) in critically ill patients based on the YOLOv8 neural network. Validation on a clinical dataset demonstrated its capability for rapid identification and accurate staging of PI. This holds urgent practical significance and clinical value for reducing medical costs and improving patient outcomes, representing an important exploration direction for intelligent nursing and precision medicine. The following discussion centers on the research findings, limitations, and future directions. 3.1 Research Findings This study achieved significant results by applying the YOLOv8l model for PI detection. The model demonstrated exceptional performance with an overall accuracy of 96.8%, particularly excelling in the detection of Stage 2, Stage 4, and Unstageable PIs. This validates its efficient object detection capability and the precise capture of injury characteristics through multi-scale feature fusion technology. The model exhibited outstanding real-time performance (processing speed: 0.35 fps/img), meeting the stringent requirements for rapid response in critical care scenarios. Although minor errors occurred in Stage 1, Stage 3, and DTI, primarily due to limited sample size and staging complexity, the model successfully identified a high proportion of (PIs) overall. Compared to traditional manual assessment methods, the technical advantages of YOLOv8l lie in its real-time automated processing, objective and standardized staging, and scalability, effectively avoiding subjective variability and providing a critical time window for early warning.Its clinical value is reflected in: optimizing personalized care decisions (such as pressure relief measures and debridement timing) through accurate staging results, thereby reducing the risk of injury progression; simultaneously, it alleviates healthcare professionals from repetitive tasks, enabling a greater focus on complex clinical judgments. The model's scalability gives it the potential for integration into existing monitoring systems, laying the foundation for intelligent and precise development in the ICU. This represents a deep integration of technological innovation and clinical practice, providing an efficient and reliable AI solution for PI management. 3.2 Study Limitations Dataset Bias: The constructed dataset has an insufficient sample size and an uneven distribution of samples across stages. Furthermore, as the samples were primarily sourced from a single center, the model may lack generalizability validated by multi-center data. Identification of Deep Tissue Injury: The detection accuracy for complex samples requires improvement. Moreover, since changes in tissue pain and temperature often precede skin color changes, future work could incorporate synthetic data augmentation or integrate multi-modal information (e.g., infrared thermal imaging) to enhance performance. Depth of Clinical Validation: The current study only validated the model's technical performance. Prospective controlled clinical trials have not yet been conducted to assess its actual impact on patient outcomes. Integrated Platform Development: Although this study has completed the research on the functional components of a high-performance PI image processing and recognition system, a fully integrated and deployable central detection platform has not been realized. 3.3 Future Work Construct a Multi-center Dataset: Expand the PI dataset in multiple aspects by incorporating data from multiple medical centers to enhance the model's robustness and generalization capability. Multi-modal Fusion: Improve comprehensive decision-making by integrating clinical data (e.g., pressure and temperature sensors, vital signs) with image information. Conduct Randomized Controlled Trials: Quantify the model's effect on improving nursing efficiency and patient outcomes through rigorous randomized controlled trials (RCTs). Edge Computing Optimization: Further compress the model parameters to enable real-time inference on low-power devices, meeting the requirements for bedside monitoring. Personalized Intervention Strategies: Develop an AI-driven recommendation system for personalized nursing care plans based on injury characteristics and staging results. Conclusion This study developed a methodological framework for real-time detection and staging of (PIs) in critically ill patients based on YOLOv8l, validating its technical feasibility and potential value in clinical scenarios. The model's high efficiency and accuracy provide a powerful new tool for the early intervention of (PIs), promising to become an important means of supporting clinical decision-making in wound assessment and helping to reduce the risk of complications and the healthcare burden for critically ill patients. With the future enrichment of PI image databases and advances in cutting-edge technologies, multi-center validation, technical optimization, and clinical empirical studies will continue to promote the in-depth application of AI in critical care nursing. Declarations Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Acknowledgments We sincerely thank all participants for their contributions to this study. Authors Contributions N.C. :Conceptualization, Project administration,Formal analysis, Resources. F.G. : Software,Visualization,Writing original draft,Writing review & editing. J.N. : Data curation, Methodology, Supervision,Visualization. W.G. : Data curation, Supervision,Writing review & editing. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Ethics approval and consent to participate This study was approved by the Ethics Committee of Shanxi Provincial People's Hospital. All participants provided informed consent before being invited to conduct this study and ensured the confidentiality of their data through anonymization and security processing procedures. Consent for publication All participants' consent for publication was sought before this study was published. Competing interests The authors declare no competing interests. Additional information Correspondence and requests for materials should be addressed to F.G. The author confirms that all methods were performed in accordance with the relevant guidelines and regulations. References Kottner, J. et al. Prevention and treatment of pressure ulcers/injuries: The protocol for the second update of the international Clinical Practice Guideline 2019[J]. J. tissue viability . 28 (2), 51–58 (2019). Kottner, J. et al. Pressure ulcer/injury classification today: An international perspective[J]. J. Tissue Viability . 29 (3), 197–203 (2020). Jacq, G. et al. Prevalence of pressure injuries among critically ill patients and factors associated with their occurrence in the intensive care unit: The PRESSURE study[J]. Australian Crit. Care . 34 (5), 411–418 (2021). Cortés, O. L. & Vásquez, S. M. Patient repositioning during hospitalization and prevention of pressure ulcers: A narrative review[J]. Investigación y educación en enfermería , 42 (1). (2024). Kaçmaz, H. Y. et al. Nurses' knowledge and practice in preventing pressure injuries in intensive care units[J]. J. Wound Care . 32 (Sup4), S22–S28 (2023). Lo, H. I. I., Hollywood, E. & Derwin, R. Bridging the gap: ICU nurses’ experiences in detecting pressure injuries across diverse skin tones[J]. J. Tissue Viability . 34 (3), 100891 (2025). Liu, H. et al. The ability of critical care nurses to identify pressure injury and incontinence-associated dermatitis: a multicentre cross‐sectional survey[J]. Nurs. open. 10 (3), 1556–1564 (2023). Stefanelli, A. et al. Develo** an AI-powered wound assessment tool: a methodological approach to data collection and model optimization[J]. BMC Med. Inf. Decis. Mak. 25 (1), 297 (2025). Ganesh, S., Gomathi, R. & Kannadhasan, S. Brain tumor segmentation and detection in MRI using convolutional neural networks and VGG16[J]. Cancer Biomarkers . 42 (3), 18758592241311184 (2025). Cui, C. et al. Diabetic wound segmentation using convolutional neural networks[C]//2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, : 1002–1005. (2019). Talu, M. H. et al. From Image to Diagnosis: Convolutional Neural Networks in Tongue Lesions[J]. J. Imaging Inf. Med. , : 1–11. (2025). Privalov, M. et al. Software-based method for automated segmentation and measurement of wounds on photographs using mask r-cnn: a validation study[J]. J. Digit. Imaging . 34 (4), 788–797 (2021). Zhang, H. & Zhao, F. Deep Learning-based carotid plaque ultrasound image detection and classification study[J]. Rev. Cardiovasc. Med. 25 (12), 454 (2024). Diwan, T., Anirudh, G. & Tembhurne, J. V. Object detection using YOLO: challenges, architectural successors, datasets and applications[J]. multimedia Tools Appl. 82 (6), 9243–9275 (2023). Lv, B. et al. Traditional Chinese medicine recognition based on target detection[J]. Evidence-Based Complement. Altern. Med. 2022 (1), 9220443 (2022). Lee, K. C. et al. Deep-learning-based automated rotator cuff tear screening in three planes of shoulder MRI[J]. Diagnostics 13 (20), 3254 (2023). Huang, K. Y., Chung, C. L. & Xu, J. L. Deep learning object detection-based early detection of lung cancer[J]. Front. Med. 12 , 1567119 (2025). Wanli, C. & Daifeng, H. Interpretation of pressure injury’s definition and staging system of National Pressure Ulcer Advisory Panel in 2016[J]. Chin. J. Injury Repair. Wound Heal . 13 (1), 64–68 (2018). Shin, H. C. et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning[J]. IEEE Trans. Med. Imaging . 35 (5), 1285–1298 (2016). Lan, X. et al. Global, Regional, and National Burden of Pressure Ulcers From 1990 to 2021 and Projections Over the Next Decade: Results From the 2021 GBD Study[J]. Wound Repair. Regeneration . 33 (4), e70064 (2025). Huang, L. et al. Summary of best evidence for prevention and control of pressure ulcer on support surfaces[J]. Int. Wound J. 20 (6), 2276–2285 (2023). Chairat, S. et al. AI-assisted assessment of wound tissue with automatic color and measurement calibration on images taken with a smartphone[C]//Healthcare. MDPI 11 (2), 273 (2023). Lei, C. et al. Convolutional Neural Network Models for Visual Classification of Pressure Ulcer Stages: Cross-Sectional Study[J]. JMIR Med. Inf. 13 , e62774 (2025). Ge, X. et al. Development and application of an intelligent pressure injury assessment system using AI image recognition[J]. Technol. Health Care . 33 (3), 1358–1366 (2025). Chen, C. C. et al. Applying Object Detection and Large Language Model to Establish a Smart Telemedicine Diagnosis System with Chatbot: A Case Study of Pressure Injuries Diagnosis System[J] Vol. 30, e1705–e1712 (Telemedicine and e-Health, 2024). 6. Tables Tables 1 and 2 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files table.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8449503","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":614793567,"identity":"4a8a7872-77f8-400f-a39b-58f2f00c2c03","order_by":0,"name":"Ningbo Cao","email":"","orcid":"","institution":"Shanxi Provincial Peoples Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ningbo","middleName":"","lastName":"Cao","suffix":""},{"id":614793569,"identity":"db4f393e-5747-4627-8d04-0eff65430b3f","order_by":1,"name":"Fengbo Guo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIiWNgGAWjYJCCD0CcwMDe2PggoUJCTp4IHYwzwFp4Dh82eHDGwtiwgWgtEmlpkg/bKhIZDhBQLx+RfLDh447aPH6GHAOJxHkSCYwNzA8f3cCjxfBGWmLjzDPHiyUbzhgYJG6TyGNnYDM2zsGnZUaO+WPetmOJGw72GCQAtRQzNvCwSRPQYtj8F6TlMI/BgcQ5EokNBwhokZcAamFsq0nccIwtsSGxgQgtBjzPEht72w4US/YwH2ZIOCZhbNhMwC/y7cAQ+9lWl8cv/7D954+aOjl59uaHj/HacgBMHUYSYsajHGxLA5iqI6BsFIyCUTAKRjQAAJ3JVbkBaPWsAAAAAElFTkSuQmCC","orcid":"","institution":"Shanxi Provincial Peoples Hospital","correspondingAuthor":true,"prefix":"","firstName":"Fengbo","middleName":"","lastName":"Guo","suffix":""},{"id":614793570,"identity":"751c6104-3fc6-4740-9c56-76561853e27e","order_by":2,"name":"Jingyi Nie","email":"","orcid":"","institution":"Shanxi Provincial Peoples Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jingyi","middleName":"","lastName":"Nie","suffix":""},{"id":614793573,"identity":"79d59ef1-db0a-4b5d-8747-07fc23989fd4","order_by":3,"name":"Wenwen Guo","email":"","orcid":"","institution":"Shanxi Provincial Peoples Hospital","correspondingAuthor":false,"prefix":"","firstName":"Wenwen","middleName":"","lastName":"Guo","suffix":""}],"badges":[],"createdAt":"2025-12-25 14:08:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8449503/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8449503/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105984543,"identity":"c1eb2f88-dcc6-4213-b559-44ef506c1fdd","added_by":"auto","created_at":"2026-04-02 07:14:52","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":15616,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/e8ebe366e34192e0763d7931.png"},{"id":106094381,"identity":"c198d074-8c3f-430e-bd44-688e9c3a6fee","added_by":"auto","created_at":"2026-04-03 11:42:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":70982,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/e7021e41a23de1ea3a12ff3a.png"},{"id":105984551,"identity":"3282575b-3844-454d-9288-3b74239d1415","added_by":"auto","created_at":"2026-04-02 07:14:52","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":25962,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/2625bbc5414c39fbf5a4b838.png"},{"id":106093625,"identity":"09555b46-64ba-4b13-b165-61dfa903688f","added_by":"auto","created_at":"2026-04-03 11:38:19","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":43590,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/964e97f168bf18525844d879.png"},{"id":105984550,"identity":"0d51c849-7b61-4fdf-9e48-212d67729ed7","added_by":"auto","created_at":"2026-04-02 07:14:52","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":95157,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/346627f84afe6b90cb0010bd.png"},{"id":105984552,"identity":"398250b9-aa75-4b15-8288-12a35277a8e7","added_by":"auto","created_at":"2026-04-02 07:14:52","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":59444,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/4339104c5040f67d56dfcefc.png"},{"id":105984553,"identity":"f3264484-cabc-4e8a-9521-fa6e5e183bc3","added_by":"auto","created_at":"2026-04-02 07:14:52","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":370341,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/178dd1b1e1b5363f951e2486.png"},{"id":109178037,"identity":"4572f914-f15f-4730-901c-deebcb41432b","added_by":"auto","created_at":"2026-05-13 09:46:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1021268,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/9c21aed6-9d60-409c-9ccb-e625adb16329.pdf"},{"id":106094387,"identity":"0b2f5364-f5ab-4883-b05b-d93a928c0231","added_by":"auto","created_at":"2026-04-03 11:42:22","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":104895,"visible":true,"origin":"","legend":"","description":"","filename":"table.docx","url":"https://assets-eu.researchsquare.com/files/rs-8449503/v1/0a7f137dd47d9897044c56d7.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Research on real-time detection and staging technology for pressure injuries in critically ill patients based on YOLOv8","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePressure injury (PI), also known as pressure ulcer\u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e, refers to localized tissue damage caused by intense and/or prolonged pressure, or the combined effects of pressure and shear force, affecting the skin and underlying soft tissues. These injuries typically occur over bony prominences, presenting clinically as damage beneath intact skin or open ulcers. They may also involve mucosal tissues (such as the mucosal surface, submucosa, or submucosal layer) or be associated with medical device use. The injury may be painful or painless, and its development is influenced by multiple factors including the microenvironment, nutritional status, and complications. The intensive care unit (ICU) is a specialized setting for the treatment of critically ill patients. Due to the severity of their conditions, ICU patients often require multiple supportive medical devices, prolonged bed rest, and frequently exhibit sensory impairment, reduced communication and mobility abilities, and altered consciousness. These factors can lead to prolonged pressure on the skin and subcutaneous tissues, ultimately resulting in PI \u003csup\u003e[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/sup\u003e. Research statistics \u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003eindicate that compared to the 5%-15% PI incidence rate among general inpatients, the rate reaches 15%-25% during ICU stays. This increases patients' susceptibility to infections, prolongs hospital stays, and incurs treatment costs that exceed expenditures allocated by individuals and society for health prevention through medical insurance. As frontline caregivers for critically ill patients, ICU nurses play a crucial role in reducing PI incidence through accurate assessment and appropriate nursing interventions \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. However, traditional assessment methods are influenced by multiple factors, including individual knowledge gaps, inadequate training, subjective judgment, and environmental variables like lighting and skin discoloration \u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e. These limitations not only increase workload but may also cause patient discomfort during measurement, with varying reliability of outcomes \u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e. Precise diagnosis and assessment of PI enable healthcare providers to objectively evaluate PI staging and select treatment modalities, thereby delivering more personalized care to critically ill patients and significantly reducing PI-related healthcare costs \u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn recent years, the continuous advancement of artificial intelligence through deep learning has provided favorable conditions for researchers in the field of intelligent healthcare \u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e. Convolutional neural networks (CNNs), as a machine learning algorithm within deep learning, are widely employed for tasks like wound segmentation and classification due to their efficient image processing capabilities \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e. Faster region-based convolutional neural networks (R-CNNs) demonstrate greater accuracy in detecting wound boundaries and performing image detection and classification, though they demand higher computational resources and longer training times \u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. In contrast, YOLO (You Only Look Once), introduced in 2015 and refined through multiple iterations, has emerged as a cutting-edge real-time object detection AI solution \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e. Unlike the two-stage detection strategy (candidate region generation\u0026thinsp;+\u0026thinsp;precise localization) of traditional models like R-CNN, YOLO transforms object detection into a regression problem, completing detection with a single image scan. This approach achieves faster execution and reduced computational resource requirements \u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e. The 2023 YOLOv8 introduced an anchor-free mode, enabling direct prediction of object center, size, and shape without predefined anchor boxes \u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003e. This approach streamlines the detection process, making it faster and more efficient, particularly for real-time medical image analysis \u003csup\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. However, domestic research on utilizing deep learning for automated PI staging remains scarce. Therefore, addressing the current lack of precise and unified standards for clinical nurses' subjective assessment of PI staging features, this study develops a novel technology for real-time detection and staging of (PIs) based on the YOLOv8 neural network model. It aims to provide clinicians with a reliable PI staging tool, improve patient treatment outcomes, and alleviate the burden on healthcare professionals.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.1 Study Population\u003c/h2\u003e \u003cp\u003eThis study selected 1,024 images of PI from patients admitted to our hospital's intensive care unit between January 2023 and June 2025. After screening, 507 images met the inclusion criteria for analysis. All data were stored in the hospital's dedicated electronic PI information management system.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.2 Data Collection\u003c/h2\u003e \u003cp\u003eNursing staff capture patients' PI images in clinical settings using medical PDAs or personal mobile phones under natural lighting. Images are taken at a distance of 50 cm while maintaining stability during capture. Completed images are subsequently uploaded to the electronic PI information management system.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e1.3 Image Annotation\u003c/h2\u003e \u003cp\u003eThis study obtained a total of 1,024 images. After excluding images with poor resolution, overexposure or underexposure, inadequate lighting, and cluttered backgrounds, 507 high-resolution images with appropriate focal length, sufficient lighting, and minimal noise were selected. These were randomly divided into a training set (414 images) and a test set (93 images) at an 8:2 ratio. Image annotation was performed collaboratively by three experienced Wound, Ostomy, and Continence (WOC) nurses from the hospital. Following the 2019 Pressure Ulcer Prevention and Treatment Guidelines published by the National Pressure Ulcer Advisory Panel (NPUAP) \u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e, images were classified into six stages. In cases of disagreement, staging was determined through consensus reached via group discussion. Image annotation was performed using the MakeSense online annotation tool, with annotation files exported in YOLO data format. These verified annotated images were subsequently used to train the deep learning model. Reference criteria for each stage are shown in Table\u0026nbsp;1 \u003csup\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e, the distribution of images across stages is illustrated in Fig.\u0026nbsp;1, and the study workflow is depicted in Fig.\u0026nbsp;2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e1.4 Preprocessing\u003c/h2\u003e \u003cp\u003eTo enhance the model's detection accuracy and generalization capability, preprocessing and augmentation operations were applied to the image data. Various modifications were implemented, including random vertical and horizontal flipping, rotation between 0 and 90 degrees, and brightness reduction. Furthermore, the images were resized to 224\u0026times;224 pixels, and their pixel intensity values were normalized to the range [0, 1] to improve training stability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e1.5 Training Configuration\u003c/h2\u003e \u003cp\u003eThis study employed a transfer learning strategy \u003csup\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/sup\u003e, utilizing five versions of the YOLOv8 model pre-trained on the large-scale COCO dataset: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. These models differ in parameter count and inference speed, making them suitable for different scenarios. We retained the shallow layers of the pre-trained models and fine-tuned the deeper layers. During training, the AdamW optimizer was selected, and the learning rate was adjusted according to the configuration file. The learning rate was halved every 5 training epochs to stabilize training and achieve fine-tuning. The training was set for 100 epochs, incorporating the Early Stopping technique to terminate training early if no improvement was observed. The batch size was set to 20 and adjusted based on GPU memory availability to ensure stable training. Additionally, YOLOv8's default compound loss function was used to balance localization and classification losses, and automated mixed-precision training on the GPU was enabled to increase training speed and reduce memory usage.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e1.6 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eIn this study, we employed several metrics to evaluate model performance, including precision, recall, F1-score, mAP@50, and mAP@50\u0026ndash;95. Precision refers to the proportion of samples predicted as positive by the model that are actually true positives, reflecting the model's accuracy. Recall refers to the proportion of actual true positive samples that are correctly predicted as true positives by the model, measuring the model's comprehensiveness, i.e., its ability to identify actual true positive samples. The F1-score is the harmonic mean of precision and recall, providing a comprehensive evaluation of the model's performance. mAP is a common metric for evaluating models, measuring the overall performance across different categories. mAP@50 refers to the mean average precision when the IoU threshold is set to 0.5, while mAP@50\u0026ndash;95 represents the average precision of the model across IoU thresholds ranging from 0.5 to 0.95. Additionally, the confusion matrix was used to visualize prediction results across various categories, aiding in the analysis of the model's classification effectiveness for different stages. Furthermore, evaluation based on frame rate was conducted to reflect the system's real-time image processing capability, ensuring the requirement for immediate feedback in clinical applications.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e1.7 Experimental Platform\u003c/h2\u003e \u003cp\u003eThis study was conducted on a high-performance computing platform equipped with an NVIDIA RTX 3080 Ti GPU (12GB VRAM), an Intel Xeon Platinum 8369B processor, and 1TB of storage. Model development, training, and optimization were performed using the PyTorch 2.0 deep learning framework (with CUDA 11.8 acceleration library). Data processing and statistical analysis primarily relied on the Pandas 1.3.5 and NumPy 1.21.3 libraries. Result visualization tasks were implemented using Seaborn 0.12.2 and Plotly 5.12.0.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Model Training\u003c/h2\u003e \u003cp\u003eFigure 3 illustrates the evolution of loss functions during YOLOv8 training. Bounding box loss represents the discrepancy between the coordinates of the predicted and ground truth bounding boxes. Class loss indicates the probability of each detected object belonging to a specific category. The Distribution Focal Loss, built upon focal loss, enhances the model by distributing the focal loss across multiple scales and classes. The figure demonstrates a declining trend in all loss values as training progresses, indicating that the model is converging toward optimization. Concurrently, Fig.\u0026nbsp;4 shows an increasing trend in the model's precision, recall, and mAP with the progression of training epochs. This suggests that the model's capability to detect PI at different stages in images continuously improves as the number of training iterations increases.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Model Performance Evaluation\u003c/h2\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Bounding Box Regression Prediction Performance of Different YOLOv8 Model Versions\u003c/h2\u003e \u003cp\u003eBased on Table\u0026nbsp;2, this study systematically evaluated the bounding box regression prediction performance of the YOLOv8 series models. Results indicate that detection accuracy metrics (e.g., mAP@50, mAP@50\u0026ndash;95) significantly improve as model size increases. Among these, YOLOv8l achieves optimal mAP@50 (0.847) and mAP@50\u0026ndash;95 (0.577), substantially outperforming other variants. However, inference speed decreases with increasing model size, as detailed in Table\u0026nbsp;2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Performance of the YOLOv8l Model at Different Confidence Thresholds\u003c/h2\u003e \u003cp\u003eAs shown in Figure 5, the model's performance across different confidence thresholds was systematically evaluated using the Precision-Recall curve, Precision-Confidence curve, Recall-Confidence curve, and F1 Score-Confidence curve. The P-R curve illustrates the dynamic trade-off between precision and recall, providing critical guidance for model threshold optimization. The Precision-Confidence and Recall-Confidence curves depict the trends of precision and recall at varying confidence levels, respectively. They indicate that the model effectively enhances lesion screening sensitivity at lower thresholds while significantly improving detection specificity at higher thresholds. In the F1-Confidence curve, the F1 score initially increases and then decreases with rising confidence, reaching a peak value of 0.78 for all classes at a confidence of 0.564, achieving the optimal balance between precision and recall. Combined with subsequent validation by the PI nursing team, this multi-dimensional confidence analysis establishes the model's reliability and practicality in real-world applications. Overall, the results from the YOLOv8l model demonstrate its accurate detection and classification of (PIs) within the training and validation datasets.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3 Predictive Performance of the YOLOv8l Model for PI Staging\u003c/h2\u003e \u003cp\u003eThe YOLOv8l model was used to predict 93 PI images at different stages from the test set, and a confusion matrix was constructed by comparing the predictions with the ground truth labels (Fig.\u0026nbsp;6). As shown in the figure, all Stage 2, Stage 4, and Unstageable images were correctly predicted, achieving 100% accuracy. However, one image each from Stage 1, Stage 3, and Deep Tissue Injury was misclassified. This indicates that the network model performs well in identifying PI stages, particularly for Stage 2, Stage 4, and Unstageable images. Additionally, Fig.\u0026nbsp;7 displays representative image analysis results detected by the YOLOv8l model, including the stage, confidence score, and visual annotations. The visualization of prediction outcomes enhances their credibility.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003ePI is a common and serious complication in patients with prolonged bed rest or limited mobility, such as critically ill patients, and its incidence is closely associated with mortality. The Global Burden of Disease study report \u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/sup\u003e indicates that China has become a high-incidence country for PI due to the dual effects of an accelerating aging population and imbalanced allocation of medical resources. PI not only causes severe pain and increases the risk of infection in patients, but the prolonged and difficult healing process also significantly extends hospital stays and consumes medical resources \u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e. Clinical prevention and management face multiple challenges. Traditional assessment relies on manual observation, which is subjective, time-consuming, and labor-intensive, making it difficult to detect early-stage damage promptly, especially in intensive care settings \u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e. Staging diagnosis requires extensive professional expertise, and delays in judgment can easily lead to deterioration of the condition. Although multidisciplinary collaboration and technologies such as advanced dressings are continuously advancing, there remains an urgent need for real-time, accurate, and automated detection and staging methods. Lei et al. \u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e investigated the capability of four different CNN models for classifying PI images. Their results showed that DenseNet121 achieved the highest classification accuracy of 93.71%, demonstrating strong performance in PI image classification. Ge et al. \u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e compared a CNN-based AI system with traditional assessment methods for PI evaluation. The results indicated that the AI system achieved an accuracy of 90%, superior to the 81.2% recognition accuracy of traditional methods, significantly reducing assessment time and improving the overall efficiency of PI evaluation. Chen et al. \u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e established an intelligent telemedicine diagnosis system based on YOLOv7 and a large language model to assist users in staging and classifying PI, providing real-time and accurate diagnostic and treatment suggestions. In contrast, this study developed a real-time detection and staging model for (PIs) in critically ill patients based on the YOLOv8 neural network. Validation on a clinical dataset demonstrated its capability for rapid identification and accurate staging of PI. This holds urgent practical significance and clinical value for reducing medical costs and improving patient outcomes, representing an important exploration direction for intelligent nursing and precision medicine. The following discussion centers on the research findings, limitations, and future directions.\u003c/p\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n\u003ch2\u003e3.1 Research Findings\u003c/h2\u003e\n\u003cp\u003eThis study achieved significant results by applying the YOLOv8l model for PI detection. The model demonstrated exceptional performance with an overall accuracy of 96.8%, particularly excelling in the detection of Stage 2, Stage 4, and Unstageable PIs. This validates its efficient object detection capability and the precise capture of injury characteristics through multi-scale feature fusion technology. The model exhibited outstanding real-time performance (processing speed: 0.35 fps/img), meeting the stringent requirements for rapid response in critical care scenarios. Although minor errors occurred in Stage 1, Stage 3, and DTI, primarily due to limited sample size and staging complexity, the model successfully identified a high proportion of (PIs) overall. Compared to traditional manual assessment methods, the technical advantages of YOLOv8l lie in its real-time automated processing, objective and standardized staging, and scalability, effectively avoiding subjective variability and providing a critical time window for early warning.Its clinical value is reflected in: optimizing personalized care decisions (such as pressure relief measures and debridement timing) through accurate staging results, thereby reducing the risk of injury progression; simultaneously, it alleviates healthcare professionals from repetitive tasks, enabling a greater focus on complex clinical judgments. The model's scalability gives it the potential for integration into existing monitoring systems, laying the foundation for intelligent and precise development in the ICU. This represents a deep integration of technological innovation and clinical practice, providing an efficient and reliable AI solution for PI management.\u003c/p\u003e\n\u003cstrong\u003e3.2 Study Limitations\u003c/strong\u003e\u003cbr /\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003eDataset Bias: The constructed dataset has an insufficient sample size and an uneven distribution of samples across stages. Furthermore, as the samples were primarily sourced from a single center, the model may lack generalizability validated by multi-center data.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIdentification of Deep Tissue Injury: The detection accuracy for complex samples requires improvement. Moreover, since changes in tissue pain and temperature often precede skin color changes, future work could incorporate synthetic data augmentation or integrate multi-modal information (e.g., infrared thermal imaging) to enhance performance.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eDepth of Clinical Validation: The current study only validated the model's technical performance. Prospective controlled clinical trials have not yet been conducted to assess its actual impact on patient outcomes.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIntegrated Platform Development: Although this study has completed the research on the functional components of a high-performance PI image processing and recognition system, a fully integrated and deployable central detection platform has not been realized.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003cstrong\u003e3.3 Future Work\u003c/strong\u003e\u003cbr /\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003eConstruct a Multi-center Dataset: Expand the PI dataset in multiple aspects by incorporating data from multiple medical centers to enhance the model's robustness and generalization capability.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eMulti-modal Fusion: Improve comprehensive decision-making by integrating clinical data (e.g., pressure and temperature sensors, vital signs) with image information.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eConduct Randomized Controlled Trials: Quantify the model's effect on improving nursing efficiency and patient outcomes through rigorous randomized controlled trials (RCTs).\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eEdge Computing Optimization: Further compress the model parameters to enable real-time inference on low-power devices, meeting the requirements for bedside monitoring.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003ePersonalized Intervention Strategies: Develop an AI-driven recommendation system for personalized nursing care plans based on injury characteristics and staging results.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study developed a methodological framework for real-time detection and staging of (PIs) in critically ill patients based on YOLOv8l, validating its technical feasibility and potential value in clinical scenarios. The model's high efficiency and accuracy provide a powerful new tool for the early intervention of (PIs), promising to become an important means of supporting clinical decision-making in wound assessment and helping to reduce the risk of complications and the healthcare burden for critically ill patients. With the future enrichment of PI image databases and advances in cutting-edge technologies, multi-center validation, technical optimization, and clinical empirical studies will continue to promote the in-depth application of AI in critical care nursing.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe sincerely thank all participants for their contributions to this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eN.C. :Conceptualization, Project administration,Formal analysis, Resources.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eF.G. : Software,Visualization,Writing original draft,Writing review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003eJ.N. : Data curation, Methodology, Supervision,Visualization. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eW.G. : Data curation, Supervision,Writing review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of Shanxi Provincial People\u0026apos;s Hospital. All participants provided informed consent before being invited to conduct this study and ensured the confidentiality of their data through anonymization and security processing procedures.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll participants\u0026apos; consent for publication was sought before this study was published.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCorrespondence and requests for materials should be addressed to F.G.\u003c/p\u003e\n\u003cp\u003eThe author confirms that all methods were performed in accordance with the relevant guidelines and regulations.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKottner, J. et al. Prevention and treatment of pressure ulcers/injuries: The protocol for the second update of the international Clinical Practice Guideline 2019[J]. \u003cem\u003eJ. tissue viability\u003c/em\u003e. \u003cb\u003e28\u003c/b\u003e (2), 51\u0026ndash;58 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKottner, J. et al. Pressure ulcer/injury classification today: An international perspective[J]. \u003cem\u003eJ. Tissue Viability\u003c/em\u003e. \u003cb\u003e29\u003c/b\u003e (3), 197\u0026ndash;203 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJacq, G. et al. Prevalence of pressure injuries among critically ill patients and factors associated with their occurrence in the intensive care unit: The PRESSURE study[J]. \u003cem\u003eAustralian Crit. Care\u003c/em\u003e. \u003cb\u003e34\u003c/b\u003e (5), 411\u0026ndash;418 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCort\u0026eacute;s, O. L. \u0026amp; V\u0026aacute;squez, S. M. Patient repositioning during hospitalization and prevention of pressure ulcers: A narrative review[J]. \u003cem\u003eInvestigaci\u0026oacute;n y educaci\u0026oacute;n en enfermer\u0026iacute;a\u003c/em\u003e, \u003cb\u003e42\u003c/b\u003e(1). (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKa\u0026ccedil;maz, H. Y. et al. Nurses' knowledge and practice in preventing pressure injuries in intensive care units[J]. \u003cem\u003eJ. Wound Care\u003c/em\u003e. \u003cb\u003e32\u003c/b\u003e (Sup4), S22\u0026ndash;S28 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLo, H. I. I., Hollywood, E. \u0026amp; Derwin, R. Bridging the gap: ICU nurses\u0026rsquo; experiences in detecting pressure injuries across diverse skin tones[J]. \u003cem\u003eJ. Tissue Viability\u003c/em\u003e. \u003cb\u003e34\u003c/b\u003e (3), 100891 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, H. et al. The ability of critical care nurses to identify pressure injury and incontinence-associated dermatitis: a multicentre cross‐sectional survey[J]. \u003cem\u003eNurs. open.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (3), 1556\u0026ndash;1564 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStefanelli, A. et al. Develo** an AI-powered wound assessment tool: a methodological approach to data collection and model optimization[J]. \u003cem\u003eBMC Med. Inf. Decis. Mak.\u003c/em\u003e \u003cb\u003e25\u003c/b\u003e (1), 297 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGanesh, S., Gomathi, R. \u0026amp; Kannadhasan, S. Brain tumor segmentation and detection in MRI using convolutional neural networks and VGG16[J]. \u003cem\u003eCancer Biomarkers\u003c/em\u003e. \u003cb\u003e42\u003c/b\u003e (3), 18758592241311184 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCui, C. et al. Diabetic wound segmentation using convolutional neural networks[C]//2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, : 1002\u0026ndash;1005. (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTalu, M. H. et al. From Image to Diagnosis: Convolutional Neural Networks in Tongue Lesions[J]. \u003cem\u003eJ. Imaging Inf. Med.\u003c/em\u003e, : 1\u0026ndash;11. (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrivalov, M. et al. Software-based method for automated segmentation and measurement of wounds on photographs using mask r-cnn: a validation study[J]. \u003cem\u003eJ. Digit. Imaging\u003c/em\u003e. \u003cb\u003e34\u003c/b\u003e (4), 788\u0026ndash;797 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, H. \u0026amp; Zhao, F. Deep Learning-based carotid plaque ultrasound image detection and classification study[J]. \u003cem\u003eRev. Cardiovasc. Med.\u003c/em\u003e \u003cb\u003e25\u003c/b\u003e (12), 454 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDiwan, T., Anirudh, G. \u0026amp; Tembhurne, J. V. Object detection using YOLO: challenges, architectural successors, datasets and applications[J]. \u003cem\u003emultimedia Tools Appl.\u003c/em\u003e \u003cb\u003e82\u003c/b\u003e (6), 9243\u0026ndash;9275 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLv, B. et al. Traditional Chinese medicine recognition based on target detection[J]. \u003cem\u003eEvidence-Based Complement. Altern. Med.\u003c/em\u003e \u003cb\u003e2022\u003c/b\u003e (1), 9220443 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, K. C. et al. Deep-learning-based automated rotator cuff tear screening in three planes of shoulder MRI[J]. \u003cem\u003eDiagnostics\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (20), 3254 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, K. Y., Chung, C. L. \u0026amp; Xu, J. L. Deep learning object detection-based early detection of lung cancer[J]. \u003cem\u003eFront. Med.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 1567119 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWanli, C. \u0026amp; Daifeng, H. Interpretation of pressure injury\u0026rsquo;s definition and staging system of National Pressure Ulcer Advisory Panel in 2016[J]. \u003cem\u003eChin. J. Injury Repair. Wound Heal\u003c/em\u003e. \u003cb\u003e13\u003c/b\u003e (1), 64\u0026ndash;68 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShin, H. C. et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning[J]. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e (5), 1285\u0026ndash;1298 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLan, X. et al. Global, Regional, and National Burden of Pressure Ulcers From 1990 to 2021 and Projections Over the Next Decade: Results From the 2021 GBD Study[J]. \u003cem\u003eWound Repair. Regeneration\u003c/em\u003e. \u003cb\u003e33\u003c/b\u003e (4), e70064 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, L. et al. Summary of best evidence for prevention and control of pressure ulcer on support surfaces[J]. \u003cem\u003eInt. Wound J.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e (6), 2276\u0026ndash;2285 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChairat, S. et al. AI-assisted assessment of wound tissue with automatic color and measurement calibration on images taken with a smartphone[C]//Healthcare. \u003cem\u003eMDPI\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e (2), 273 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLei, C. et al. Convolutional Neural Network Models for Visual Classification of Pressure Ulcer Stages: Cross-Sectional Study[J]. \u003cem\u003eJMIR Med. Inf.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, e62774 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGe, X. et al. Development and application of an intelligent pressure injury assessment system using AI image recognition[J]. \u003cem\u003eTechnol. Health Care\u003c/em\u003e. \u003cb\u003e33\u003c/b\u003e (3), 1358\u0026ndash;1366 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, C. C. et al. \u003cem\u003eApplying Object Detection and Large Language Model to Establish a Smart Telemedicine Diagnosis System with Chatbot: A Case Study of Pressure Injuries Diagnosis System[J]\u003c/em\u003e Vol. 30, e1705\u0026ndash;e1712 (Telemedicine and e-Health, 2024). 6.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\n\u003cp\u003eTables 1 and 2 are available in the Supplementary Files section.\u003c/p\u003e\n"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Pressure injury, YOLO, Deep learning, Image processing, Object detection","lastPublishedDoi":"10.21203/rs.3.rs-8449503/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8449503/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eTo develop and validate a real-time detection and staging system for PIs using the YOLOv8 deep learning model.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA total of 507 PI images from intensive care unit patients (Jan 2023-Jun 2025) were randomly divided into training (414) and test (93) sets. Images were classified into six stages per NPUAP guidelines\u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e. Five YOLOv8 versions were developed using transfer learning, with AdamW optimizer and dynamic learning rate adjustment. The best model was evaluated on precision, accuracy, and inference speed.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThis model effectively enhanced the objectivity and accuracy of pressure injury (PI) staging identification. In testing with 93 PI images, YOLOv8l achieved the optimal balance with 0.854 precision and 0.35 fps/img inference speed, outperforming other versions. Additionally, the model demonstrated high prediction accuracy across all six PI stages: all Stage 2, Stage 4, and Unstageable images were correctly predicted; one image each in Stage 1, Stage 3, and Deep Tissue Injury was misclassified.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003e For PI staging identification, the PI assessment system built on the YOLOv8l deep learning model demonstrates high accuracy and efficiency, providing reliable support for clinical decision-making, thereby delivering more personalized care to critically ill patients and significantly reducing pressure injury-related healthcare costs.\u003c/p\u003e","manuscriptTitle":"Research on real-time detection and staging technology for pressure injuries in critically ill patients based on YOLOv8","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-02 07:14:44","doi":"10.21203/rs.3.rs-8449503/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"1ffe6ee4-6fee-4da8-9452-0b30f4bd1347","owner":[],"postedDate":"April 2nd, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Withdrawn","date":"2026-05-13T09:26:03+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65418378,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":65418379,"name":"Physical sciences/Engineering"},{"id":65418380,"name":"Health sciences/Health care"},{"id":65418381,"name":"Physical sciences/Mathematics and computing"},{"id":65418382,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-05-13T09:42:22+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-02 07:14:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8449503","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8449503","identity":"rs-8449503","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}