Artificial Intelligence for the prediction of postoperative/postablative renal function in elderly and/or comorbid patients
Nuclear Medicine, Radiotherapy, Radiobiology
Final Report Abstract
The project aimed to predict acute kidney injuries (AKI) as well as chronic kidney disease (CKD) following surgical or ablative interventions for renal tumors. Initially, comprehensive clinical and operative parameters, as well as pre-therapeutic CT data, were collected for 172 patients. Manual segmentations of the kidneys and renal tumors were performed to create radiomics. Subsequently, predictive algorithms were developed: clinical data and CT-based radiomic features were analyzed using machine learning methods. A logistic regression model with 48 variables achieved an AUC of 0.649 (95 % CI: 0.55–0.749) for predicting AKI, with a sensitivity of 53.3 % and a specificity of 72.5 %, while better values were achieved for CKD (AUC 0.796, 95 % CI: 0.686–0.907, sensitivity 87.1 %, specificity 73.6 %). Additionally, a deep learning approach using a pre-trained convolutional neural network (MobileNet v3 small) was evaluated; however, it fell short of expectations for CKD, whereas qualitatively better predictions were obtained for AKI. Due to capacity constraints on the part of the project partners, a collaboration with the German Cancer Research Center (DKFZ) was established. For 98 patients, random forest models were created based on 42 clinical variables and radiomic features. The models differentiated between exclusive use of radiomics (training AUC 0.759, test AUC 0.646), only clinical parameters (training AUC 0.857, test AUC 0.769), and a combination of both data sources, with the latter proving to be the most powerful method (training AUC 0.891, test AUC 0.707). Furthermore, deep learning analyses provided insights into the relevance of individual parameters for CKD prediction. Currently, two publications are in preparation and a conference abstract has been submitted. In summary, the project demonstrates that the combination of clinical data and radiomics using machine learning methods represents a promising approach for predicting AKI and CKD. The results, particularly for CKD, indicate a clinically relevant improvement in risk assessment, which is of foremost importance for future research.
