Machinfunction based on pre-implantation kidney imagese learning model to predict delayed graft

• Grant number: 23/13415-6
• Support Opportunities: Regular Research Grants
• Start date: April 01, 2024
• End date: March 31, 2026
• Field of knowledge: Health Sciences - Medicine - Medical Clinics
• Principal Investigator: Luís Gustavo Modelli de Andrade
• Grantee: Luís Gustavo Modelli de Andrade
• Host Institution: Hospital das Clínicas da Faculdade de Medicina de Botucatu. Universidade Estadual Paulista (UNESP). Campus de Botucatu. Botucatu , SP, Brazil
• Associated scholarship(s): 24/06622-8 - Machine function based on pre-implantation kidney images learning model to predict delayed graft, BP.TT

Abstract

Artificial intelligence (AI) is currently being used to learn patterns from images and relate them to clinical outcomes. Recent increases in computational power and data volume have led to substantial developments in machine learning techniques for addressing significant healthcare issues. In the case of kidney transplantation, one of the primary challenges is delayed graft function (DGF), which is difficult to predict. Objective: To create models for predicting DGF based on pre-implantation images of deceased donor organs, combining them with clinical characteristics. The first model uses only pre-implantation images of the graft, while the second model combines images with clinical data from the donor/recipient. Materials and Methods: An estimated total of 1500 photos of pre-implantation grafts from different recipients, taken with various smartphones from a minimum distance of 30 cm, and captured by different members (delegates and trained staff) of transplant teams at various transplant centers will be used. Electronic medical record information will also be collected. A convolutional neural network (CNN)-based model will be used to extract features from the images so that important information can be obtained for regression adjustment. Five CNN architectures will be evaluated for image analysis in this study: VGG-face, VGG16, ResNet50, DenseNet121, and MobileNet. For the classification task, five different classification algorithms will be employed: SVM, K-Nearest Neighbor (KNN), Linear Regression (LR), Decision Tree (DT), and Linear Discriminant Analysis (LDA). During training, the data will be split into two sets: 70% for training and 30% for testing. Additionally, using a 20% split, each classifier will be trained using k-fold cross-validation, with k=5. Finally, for regression prediction tasks, the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), R-squared value, and Explained Variance Score (EVS) will be calculated. Accuracy will be evaluated using the area under the Receiver Operating Characteristic curve (AUC-ROC). Conclusion: Despite numerous factors that have been identified as affecting renal function after transplantation, there is still no reliable model to predict functional outcomes of the allograft. AI could be the game-changer in this paradigm and thus assist healthcare professionals in making clinical decisions in the best interest of patients. (AU)