Real-Time Wavelet Processing and Classifier Algorithms Enabling Single-Channel Diagnosis of Lower Urinary Tract Dysfunction

Cystometry measures the behavior of the bladder and is commonly used to evaluate how the lower urinary tract functions during urine storage and voiding phases. Cystometry measures the internal bladder vesical pressure (PVES) while abdominal pressure (PABD) is superimposed. Traditionally, cystometry uses two catheters, one in the bladder and one in the vagina or rectum, so that PVES and PABD are independently measured. The pressure produced by the detrusor (bladder) muscle PDET is then derived by subtraction of simultaneous signals. In this work we demonstrate two strategies for extracting the PDET signal from a single-channel measurement of PVES, which can enable single-catheter cystometry. In addition, it is also needed for emerging catheter-free cystometry sensors. This chapter demonstrates two approaches for real-time signal processing of PVES pressure data. Because cystometry is performed with a live view of recorded pressures, both examples were designed for minimal delay and low computational complexity. First, we demonstrated a real-time, wavelet-based signal reconstruction algorithm for extracting the low-frequency PDET signal from PVES data. Due to spectral overlap, simple filtering was insufficient for reconstruction, and a nonlinear event-driven reconstruction weighting approach was adopted. Four statistical features were derived to detect cough and Valsalva artifacts, and reconstruction weights were determined via least mean square fitting. This approach demonstrated detection accuracy for cough and Valsalva events of 99.5% and 84.3%, respectively. After event-based reconstruction, this approach extracted the PDET signal with overall root mean square error of 10.7 ± 2.1 cm H2O and R = 0.88 ± 0.6 (N = 20). Second, we evaluated the performance of three machine learning classifiers for event detection, operating on extremely short data segments. To minimize delay, classifiers were operated only on 0.8-second-long data segments, representing only eight pressure samples with the traditional cystometry sample rate of 10 Hz. Despite the short segment length, 55 features were extracted for classifier training. This feature set included discrete wavelet transform (DWT) coefficients up to five levels, DWT cross-correlations, and time-domain statistical features independently calculated for each 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 fivefold cross validation, from which we derived the sensitivity, specificity, F1 score, and receiver operating characteristic for all four classes. The KNN, ANN, and SVM classifiers labeled 16,000, 0.8 second PVES segments with 90.8%, 88.8%, and 82.5% accuracy, respectively. The ANN classifier presented the best balance between accuracy and computational efficiency for real-time use. This framework demonstrated that automated multi-event bladder classification is feasible using single-channel UDS data.