# Construction and validation of a multi-function artificial intelligence–assisted system for pressure injury recognition

## TL;DR

This paper introduces an AI system for detecting and measuring pressure injuries with high accuracy and speed, improving clinical assessment and remote management.

## Contribution

A novel AI system (PI-3DAS) integrating automatic detection, staging, and wound size measurement using YOLOv11 and ArUco markers is developed and validated.

## Key findings

- The YOLOv11s model achieved high localization performance with precision of 0.854 and mAP50 of 0.842.
- Staging classification reached 92.64% accuracy, with deep tissue injury showing the highest accuracy at 96.45%.
- Wound size measurement was highly consistent with reference standards, with ICCs of 0.996 for length and 0.994 for width.

## Abstract

With the acceleration of population aging, the incidence of pressure injury (PI) continues to rise, making early identification and accurate staging essential for preventing disease progression and improving prognosis. Conventional manual assessment relies heavily on clinical experience and subjective judgment, limiting real-time, objective, and quantitative evaluation.

This study aimed to develop and validate an artificial intelligence model based on the YOLOv11 neural network that integrates automatic PI detection, intelligent staging, and wound size measurement, thereby enhancing the timeliness, accuracy, and objectivity of PI assessment.

A total of 1,815 PI images collected from the electronic PI management systems of two medical centers between January 2021 and June 2025 were included. According to the 2019 National Pressure Ulcer Advisory Panel (NPUAP) guidelines, images were classified into six categories: Stage I, Stage II, Stage III, Stage IV, unstageable, and deep tissue injury. Transfer learning was applied to train YOLOv11 models of different scales (v11n/s/m/l/x). Lesion localization and staging performance were compared to identify the optimal model. Automatic wound size measurement was achieved by integrating ArUco marker recognition with pixel-to-centimeter conversion.

For bounding box localization, the YOLOv11s model demonstrated superior performance, with a precision of 0.854, recall of 0.766, mAP50 of 0.842, mAP50–95 of 0.629, and an inference speed of 4.8 ms per image. On the test set, overall staging classification accuracy reached 92.64%, with a sensitivity of 79.79%, specificity of 95.56%, and a false-positive rate of 4.44%. The highest accuracy was observed for deep tissue injury (96.45%), while Stage III showed the lowest accuracy (85.04%). In wound size measurement, PI-3DAS demonstrated high agreement with the reference standard, with a length mean absolute error (MAE) of 0.155 cm and intraclass correlation coefficient (ICC) of 0.996, and a width MAE of 0.137 cm and ICC of 0.994. The mean time for AI-based measurement was 0.691 s, representing a 36.8-fold reduction compared with manual measurement (25.414 s; P < 0.001).

The YOLOv11-based PI-3DAS system enables automated PI detection, staging, and non-contact wound size quantification with high accuracy and consistency, while substantially improving measurement efficiency. This system provides a portable and practical tool to support clinical nursing assessment, therapeutic follow-up, and remote PI management.