EP01.15: Exploring the role of machine learning models to improve prediction of prognosis in unilateral congenital uropathies

To assess the prognostic value of radiomics and explore the role of subjective texture assessment of renal parenchyma in fetuses with unilateral obstructive uropathy. A retrospective study was conducted between 2013-2023 and fetal ultrasound images of fetuses with suspected unilateral obstructive uropathies were retrospectively collected in order to extract radiomics shape features (ureteral dilatation, renal diameters, renal volume, pyelum measurement and bladder aspect). Moreover, renal texture features, including kidney hyperechogenicity, corticomedullary differentiation and calyceal dilatation were subjectively evaluated. Composite adverse outcome was defined as abnormal kidney function on nuclear scan and/or need for postnatal surgery. Sensitivity, specificity and area under the curve were calculated. The study enrolled 86 fetuses with a confirmed postnatal diagnosis of unilateral obstructive uropathy. A machine learning model, only based on radiomics shape features showed an AUC of 0.79 (SD 0.10, sensitivity 0.76, specificity 0.67). The subjective evaluation of texture features, including kidney hyperechogenicity, calyceal dilatation and loss of corticomedullary differentiation was significantly associated with unfavourable postnatal outcomes (OR 1.34, OR 1.66, OR 1.20). However, the addition of these variables to the model led to a comparable accuracy (AUC 0.78 vs 0.79, sensitivity 0.79 vs 0.76, specificity 0.65 vs 0.67). A machine-learning model only based on radiomic shape features showed promising results in predicting prognosis of unilateral congenital obstructive uropathies. Although subjective renal texture features were significantly correlated with unfavourable outcomes, their addition to the machine learning model did not lead to a significant improvement in accuracy. These results highlight the potential of radiomics for evaluating renal fetal aspect and suggest the need to further explore texture features by using radiomics or deep learning algorithms.