Identification of Peristomal Skin Alterations Using Convolutional Artificial Neural Networks.

Introduction: Peristomal skin complications (PSCs) are very common among ostomy patients and significantly affect their quality of life and healthcare costs. Although convolutional neural networks (CNNs) offer possibilities for automated diagnosis, specific AI applications for the treatment of PSCs are underdeveloped. Aims: To develop and validate preliminary models based on CNNs for the binary classification of peristomal skin, enabling the distinction between healthy tissue and the presence of skin lesions, thereby laying the foundations for automated diagnostic systems. Design: Prospective study. Methods: The data and images were collected by 24 stoma nurses from 17 hospitals participating in the study. We addressed the classification of peristomal skin images using state-of-the-art pretrained CNNs. The classification models were evaluated using the measures accuracy, F1-score, and the area under the ROC curve. Finally, the Grad-Cam explainability algorithm is applied to the best model. Results: With 1165 images collected, several models were tested. The data were split using standard 10-fold cross-validation. A dual experiment was conducted. First, a standard data split was employed, yielding an accuracy of 0.889, an F1-score of 0.890, and an area under the ROC curve of 0.924 for the best model. Second, the data were split so that images from the same patient would not be distributed across the training and test subsets, thereby preventing data leakage. The best results for this experiment were 0.778, 0.868, and 0.653, respectively. Conclusions: By processing peristomal skin images with artificial intelligence, we developed robust, reliable, preliminary models for detecting peristomal skin alterations. The models allow the automatic detection of any peristomal skin involvement. The automatic detection of peristomal skin changes using a photograph enables remote care and speeds up treatment. Clinical relevance: The model developed using convolutional artificial neural networks is robust and reliable for detecting alterations in the peristomal skin, representing a significant advance in peristomal skin care for all ostomates, with early detection, prevention of complications and cost savings in treatment.