Cross-Dataset Reliability of Deep Learning Models for Pressure Injury Staging: a Multiarchitecture Evaluation and Prediction-Error Analysis.

Deep learning models have been proposed for automated pressure injury staging, but reliability across external wound image datasets remains uncertain. This study assessed cross-dataset reliability across multiple architectures. After dataset curation, including duplicate-image screening, 1081 images from the Pressure Injury Image Dataset (PIID), 1844 images from an institutional Hallym University Medical Center (HUMC) dataset, and 141 Kaggle images were analyzed. Six convolutional neural network and Transformer architectures were trained on PIID or HUMC and evaluated on internal and external datasets without additional fine-tuning or threshold adjustment. The primary outcome was the internal-to-external generalization gap in macro-averaged F1 score. Models trained on PIID achieved a mean internal macro-averaged F1 score of 0.75, decreasing to 0.36 on HUMC and 0.56 on Kaggle. Models trained on HUMC achieved a mean internal macro-averaged F1 score of 0.66, decreasing to 0.53 on both external datasets. No architecture consistently outperformed others in external validation. Prediction-error analysis showed dataset-dependent differences in error direction. Understaging and overstaging errors were defined as predictions below and above the true stage. Models trained on PIID predominantly produced understaging errors on HUMC, and models trained on HUMC predominantly produced overstaging errors on PIID. Exploratory analysis of frozen feature vectors showed limited grouping of images in feature space, with dataset source more closely associated with feature structure than pressure injury stage. These findings support the importance of cross-dataset validation in evaluating automated pressure ulcer staging models. The observed performance degradation and dataset-dependent error patterns suggest sensitivity to dataset-level heterogeneity. Consequently, this study supports two important points. First, in the field of pressure ulcers where image capture protocols are not fixed, external validation using diverse datasets is necessary to evaluate generalization performance. Second, a unified protocol must be included in the data acquisition process to develop reproducible models.