Machine Learning Accurately Predicts Muscle Invasion of Bladder Cancer Based on Three <scp>miRNAs</scp>

The aim of this study was to validate the diagnostic potential of four previously identified miRNAs in two independent cohorts and to develop accurate classification models to predict invasiveness of bladder cancer. Furthermore, molecular subtypes were investigated. The miRNAs were isolated from pTa low-grade (lg) (n = 113), pT1 high-grade (hg) (n = 133) and muscle-invasive bladder cancer (MIBC) (n = 136) tumour tissue samples (FFPE) after either transurethral resection of a bladder tumour (TURB) or cystectomy (CYS). In both cohorts, the expression of miR-138-5p and miR-200a-3p was significantly lower, and the expression of miR-146b-5p and miR-155-5p was significantly higher in MIBC compared to pTa lg. A k-nearest neighbours (KNN) classifier trained to distinguish pTa lg from MIBC based on three miRNAs achieved an accuracy of 0.94. The accuracy remained at 0.91 when the classifier was applied exclusively to the TURB samples. To guarantee reliable predictions, a conformal prediction approach was applied to the KNN model, which eliminated all misclassifications on the test cohort. pT1 hg samples were classified as MIBC in 32% of cases using the KNN model. miR-146b-5p, miR-155-5p and miR-200a-3p expressions are significantly associated with particular molecular subtypes. In conclusion, we confirmed that the four miRNAs significantly distinguish MIBC from NMIBC. A classification model based on three miRNAs was able to accurately classify the phenotype of invasive tumors. This could potentially support the histopathological diagnosis in bladder cancer and therefore, the clinical decision between performing a radical cystectomy and pursuing bladder-conserving strategies, especially in pT1 hg tumors.