Improving the Accuracy of a Wearable Uroflowmeter for Incontinence Monitoring Under Dynamic Conditions: Leveraging Machine Learning Methods.

Urinary incontinence affects many women, yet there are no monitoring devices capable of accurately capturing flow dynamics during everyday activities. Building on our initial development of a wearable personal uroflowmeter, this study enhances the device's performance under realistic, dynamic conditions similar to those encountered in daily living. We integrated an optimized eight-vane Etoile flow conditioner with a 0.2D opening into the device. Both computational fluid dynamics simulations and experimental tests demonstrated that this flow conditioner significantly reduced turbulence intensity by 82% and stabilized the axial velocity profile by 67%, increasing the R of flow rate measurements from 0.44 to 0.92. Furthermore, our machine learning framework-utilizing a support vector machine (SVM) and an extreme gradient boosting (XGBoost) model with principal component analysis (PCA)-accurately predicted the true flow rate with high correlations, robust performance, and minimal overfitting. For the test dataset, the SVM achieved a correlation of 0.86, an R of 0.74, and an MAE of 2.8, whereas the XGBoost-PCA model exhibited slightly stronger performance, with a correlation of 0.88, an R of 0.76, and an MAE of 2.6. These advances established a solid foundation for developing a reliable, wearable uroflowmeter capable of effectively monitoring urinary incontinence in real-world settings.