Effects of a transformer-based AI-based application to support incontinence-associated dermatitis and pressure injury assessment, nursing care and documentation: Controlled pilot intervention study

TL;DR

This study tested an AI app to help nurses assess and manage wounds like pressure injuries and incontinence-related skin issues, finding it improved guideline adherence but not workload.

Contribution

The first controlled pilot study of a transformer-based AI app for nursing wound assessment and documentation in pressure injury and incontinence-associated dermatitis care.

Findings

The AI app increased guideline adherence and documentation time but did not reduce nursing task load.

AI predictions showed high accuracy in wound type identification but lower accuracy in wound category classification.

Nurse qualification and wound severity were significant predictors of care duration.

Abstract

Artificial intelligence (AI) is playing an increasingly important role in nursing care, including wound management. Differentiating pressure injuries from incontinence-associated dermatitis is clinically challenging, often leading to misclassification. Although AI-based wound assessment is advancing, few models specifically address incontinence-associated dermatitis, and clinical evidence remains limited. The KIADEKU project developed and piloted a transformer-based AI app to support care for these wounds.

The aim of this pilot intervention study was to assess the impact of the AI-based app on duration of wound assessment, dressing changes, documentation, nursing staff task load, and guideline adherence. Secondary aims included evaluating the AI’s accuracy and app usability compared to standard systems.

This monocentric, non-randomized controlled study was conducted in two sequential phases: a control phase with conventional wound management, followed by an intervention phase utilizing the AI-based app.

The study included 88 voluntary nurses caring for pressure injuries and incontinence-associated dermatitis in adult patients on seven participating wards of LMU Hospital.

Wound care was systematically observed, and nurses completed questionnaires on task load, usability, and covariates. Outcomes were measured using standardized protocols, validated tools (NASA Task Load Index (NASA-TLX), Usability Metric for User Experience (UMUX-LITE)) and expert-defined indicators. Statistical analyses included descriptive statistics, group comparisons (t-test, Mann-Whitney U test), and multivariate linear regression adjusting for covariates. An independent wound assessment validated AI-generated predictions, with accuracy evaluated using F1-scores.

A total of 88 wound care sessions were analysed. The intervention group had a statistically significantly longer mean duration of care and documentation (12.84 vs. 9.20 min; p = 0.002; 95 % CI: –5.59; –1.41 min) and higher guideline adherence (mean rank = 50.91 vs. 38.38; p = 0.017). Nurse task load showed no statistically significant group differences. Regression analysis identified AI app use, nurse qualification, and wound severity as statistically significant predictors of care duration, while AI use did not predict task load or guideline adherence. Usability ratings were similar to standard systems. Model performance showed high accuracy in identifying wound types, but lower accuracy in classifying their categories.

This pilot study is the first to evaluate an AI-based app supporting nursing wound management for pressure injuries and incontinence-associated dermatitis. While the app did not reduce care duration or nurse workload, it may have potential to improve guideline adherence. Limitations included limited user experience and sample bias. Future multicentre studies with larger samples and randomized trials are needed to validate findings and support clinical integration.

www.drks.de DRKS00031355. Registered 05/04/2023, first recruitment 31/05/2023.

Figures