Enhanced detection of bladder cancer using machine learning analysis of routine urinary sediment data

Bladder cancer (BC) has a high recurrence rate, requiring invasive procedures such as cystoscopy and histopathological examination. Existing non-invasive methods, including urine cytology, fluorescence in situ hybridization (FISH), and tumor markers like nuclear matrix protein 22 (NMP22), have limited sensitivity and specificity. This study assessed the potential of low-cost, non-invasive urine sediment data for BC detection using a machine learning model.

Urine sediment data from 1,341 patients (577 with BC and 764 with benign conditions) were obtained using a fully automated urinary particle analyzer (UF-5000). One preoperative sample per patient was included. A soft-voting ensemble model combining random forest and logistic regression was developed. Class imbalance was addressed using the Synthetic Minority Oversampling Technique, and feature selection was guided by the variance inflation factor. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), balanced accuracy, sensitivity, and specificity, with 95 % confidence intervals estimated via bootstrapping. Model stability was evaluated using fivefold cross-validation.

The ensemble model achieved an AUC of 0.89, with a sensitivity of 0.80 and specificity of 0.88. Balanced accuracy was 0.84, and cross-validation demonstrated stable performance (accuracy 0.80). Feature importance analysis identified atypical cells and red blood cells as the most influential predictors, particularly of non-muscle-invasive BC.

This urine sediment-based, non-invasive approach shows promise for the early detection and monitoring of BC. Its integration into routine clinical workflows may enhance screening efficiency without additional costs.