Machine Learning for Automated Bladder Event Classification from Single-Channel Vesical Pressure Recordings

Analyzing urodynamic study (UDS) tracings can be prone to error in the presence of artifacts and subjective due to lack of standardization in clinical UDS interpretation. As such, the diagnosis of patients undergoing UDS would greatly benefit from a standardized, automated method to assist clinicians in interpreting UDS tracings. In this work, we evaluated a machine learning framework for automatically classifying bladder events from single-channel vesical pressure recordings
$(P_{VES}) (N=60)$
into 4 possible classes: abdominal event (i.e., cough or Valsalva), voiding contraction, detrusor overactivity (DO) and no event. Wavelet multiresolution analysis of
$P_{VES}$
was used to extract time-frequency localized wavelet coefficient vectors which were segmented into 0.8 second segments with 55 statistical features per segment. Feature selection was subsequently applied for three classifier architectures: a k-nearest classifier (KNN), an artificial neural network classifier (ANN) and a support vector machine classifier (SVM). Each classifier was trained and evaluated using five-fold cross validation, from which we derived the sensitivity, specificity, F1 score and AUC for all four classes and the overall classification accuracy for each classifier. The KNN, ANN and SVM classifiers labeled 7,861 0.8 second
$P_{VES}$
segments with 91.5%, 90.8% and 82.4% accuracy, respectively. We have thus proposed the first framework for automated multi-event bladder classification using single-channel UDS data.