Development and validation of a multimodal artificial intelligence model for diagnosing peristomal skin complications: A multicenter retrospective study
Xu Zhang, Hui Ren, Subinuer Aihemaitiniyazi, Xiaomeng Wang, Zheng Xu, Xueping Jiao, Hongyan Li, Zhiwen WangBackground
Peristomal skin complications (PSCs) are common after stoma surgery and are frequently underrecognized outside specialist care. Existing automated approaches mainly rely on images and rarely incorporate symptom information.
Objective
To develop and internally validate a multimodal artificial intelligence model integrating peristomal skin images and standardized symptom descriptors for PSC recognition.
Methods
This multicenter retrospective diagnostic study used specialist-acquired peristomal skin photographs and clinical documentation from five tertiary hospitals in China. Images with substantial blur, insufficient lighting, or lesion occlusion were excluded. Symptom text was standardized using a diagnosis-indexed terminology table and should be interpreted as diagnosis-derived standardized descriptors rather than independently collected patient-reported symptoms. StomaBiomedCLIP was developed using a pretrained BiomedCLIP backbone and cross-attention fusion. Performance was evaluated on a held-out test set and by stratified 5-fold cross-validation.
Results
The dataset included 1,128 image–text pairs. On the held-out test set, StomaBiomedCLIP achieved an accuracy of 0.967 (95% CI, 0.950–0.985), sensitivity of 0.956 (95% CI, 0.917–0.987), specificity of 0.994 (95% CI, 0.991–0.998), and AUC of 0.993 (95% CI, 0.986–0.999). Under the null-text condition, performance decreased to an accuracy of 0.611 and a macro-F1 score of 0.539. Image-ablated analysis retaining the full standardized descriptors achieved an accuracy of 0.950 and a macro-F1 score of 0.849, suggesting substantial category-discriminative information in the standardized descriptors.
Conclusions
StomaBiomedCLIP showed strong internal performance under specialist-curated image conditions with diagnosis-derived standardized symptom descriptors. External prospective validation using independently collected symptom descriptions, automated image-quality control, and clinical safety safeguards is required before real-world implementation.