Development and Validation of an Image-Based Deep Learning Tool for Identification of Xylazine-Associated Wounds

Accurate identification of xylazine-associated wounds (XAWs) is critical to providing timely and optimal management; however, discerning the etiology of wounds by appearance alone poses a clinical challenge. This study sought to develop an accessible and accurate approach for XAW diagnosis using a deep learning tool applied to wound photographs.

Publicly accessible wound photographs were curated from academic publications, Reddit, and news media to develop, train, and test the deep learning tool. XAWs were defined by provided clinical confirmation or self-reported descriptions associated with each image. The data set included images of 114 xylazine-associated and 1710 nonxylazine wounds from 17 distinct pathologies. Four deep learning models (DenseNet121, EfficientNetB0, ResNet34, and SENet154) were trained on 1185 images (65%) and 163 for validation (9%) to predict xylazine exposure and tested using 476 unseen wound images (26%).

All 4 deep learning models achieved consistent diagnostic performance on 476 unseen wound images (accuracy: 97.5%-98.5%; AUROC: 96.8%-99.7%; weighted F1 score: 97.2%-98.5%). High specificity, reaching 100.0%, was observed across the 4 models. Sensitivity ranged from 60.0% to 80.0% across the 4 models, with SENet154 demonstrating robust performance across all metrics. Qualitative assessment demonstrated accurate identification of XAWs with high-confidence exclusion of xylazine exposure in wounds attributed to trauma, surgery, pressure, or venous ulcers.

This novel deep learning tool can enable accurate identification of XAW. With further validation, this tool may offer an accessible and automated approach to guide wound care, augment bedside clinical medicine assessments, and equip public health efforts to monitor xylazine's geographic distribution.