Deep learning-based pressure injury staging: a multicentre study involving 59 hospitals

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Mask R-CNN with Swin Transformer achieved the highest performance: mAP = 0.894, AP50 = 0.900, AP = 0.757, AR100 = 0.792

• Primary Outcome: Model performance for pressure injury detection and staging using mean average precision (mAP) across three deep learning architectures

• Nnt Or Sensitivity: Sensitivity and specificity not reported; per-stage diagnostic accuracy not provided. mAP of 0.894 (AP50 threshold) is an object detection metric, not directly equivalent to clinical sensitivity/specificity

• Confidence Interval: Not reported for any metric

Clinical Application

Smartphone-based deployment is conceptually feasible and low-cost. However, the model requires prospective clinical validation, standardised image acquisition protocols, regulatory approval, and integration with existing electronic medical record systems before clinical deployment. Real-world performance in variable lighting, diverse skin tones, and wound dressings in situ remains untested. In Australia, pressure injury prevention and management is governed by the Australian Commission on Safety and Quality in Health Care (ACSQHC) National Safety and Quality Health Service (NSQHS) Standards, specifically Standard 8 (Recognising and Responding to Acute Deterioration). The NPUAP/EPUAP/PPPIA staging system is used nationally. Any AI-based wound staging tool would require TGA regulatory approval as a Class IIa or higher medical device software (SaMD) under the TGA's Software as a Medical Device framework before clinical use. The RACGP and AWMA (Australian Wound Management Association) would need to endorse clinical integration pathways. PBS implications are indirect but relevant if the tool influences wound care product selection or specialist referral thresholds. The study's Chinese hospital context limits direct generalisability to Australian patient demographics, skin tone diversity, and care workflows. Hospitalised patients with pressure injuries requiring staging assessment, particularly in settings where wound care expertise may be limited or where rapid bedside assessment tools are needed

Abstract

BACKGROUND: Accurate assessment of pressure injury staging is essential for guiding appropriate care, reducing patient suffering, alleviating the healthcare burden, and improving quality of life. We aimed to develop a deep learning-based model for pressure injury recognition and to translate the best-performing model into a preliminary smartphone application. METHODS: We conducted a multicentre retrospective study using pressure injury images collected from 59 hospitals. We applied to the data set three artificial intelligence-based image analysis models, Mask R-convolutional neural network (CNN) with ResNet-18, Mask R-CNN with Swin Transformer, and Segmenting Objects by Locations, version 2. We assessed model performance using standard evaluation metrics, including precision-recall curves, average recall (AR), average precision (AP), and mean average precision (mAP). The model achieving the best overall performance was selected for subsequent translation into a smartphone-based tool. RESULTS: We included a total of 1903 pressure injury images, with 1713 used for model training and 190 for validation. Among the models evaluated, Mask R-CNN with Swin Transformer demonstrated the highest overall performance (mAP = 0.894, AP50 = 0.900, AP = 0.757, AR100 = 0.792), surpassing the corresponding metrics of the other two models. The model, based on Mask R-CNN with a Swin Transformer, was integrated into a smartphone for preliminary translation. CONCLUSIONS: The proposed deep learning-based system showed promising performance in pressure injury staging and may provide meaningful support for clinical decision-making. Future studies should focus on further validation and refinement using larger and more diverse data sets to enhance their applicability in routine clinical practice.

References

1. 1.Zhou, L., Zhang, Z., Wang, J., Zhang, H., Zhong, C., Zong, T., Sun, L., Luo, M., Qiao, L., Hu, H., Zhang, W., Wang, C., Yang, R., Zhou, Y., & Wang, L. (2026). Deep learning-based pressure injury staging: a multicentre study involving 59 hospitals. Journal of Global Health, 16, 04175. https://doi.org/10.7189/jogh.16.04175