Voiding difficulty due to bladder outlet obstruction (BOO) poses a diagnostic and therapeutic clinical challenge. The cornerstone for diagnosing and assessing this disorder is the urodynamic examination, namely the Pressure/Flow (P/F) study, but its disadvantage lies in the fact that it is invasive and involves bladder catheterization. Our aim was to develop a model based on artificial intelligence and machine learning that will distinguish between obstructive and non-obstructive voiding according to parameters obtained exclusively from uroflowmetry testing.
Data from 179 urodynamic tests of men age ranged 17-83 were included. Parameters from the uroflowmetry (free flow) and from the P/F study were extracted. Tests were categorized into obstructive and non-obstructive patterns according to the Bladder Outlet Obstruction index (BOOI) or the Bladder Contraction Index (BCI). Equivocal results were excluded from the statistical analysis. Using a machine-learning algorithm and the multiple decision trees method (random forest), two artificial intelligence (AI) models were examined, each model separately guesstimated the BOOI and the BCI according to uroflow parameters only. The model building process involved the construction of 100 different decision trees, each of which randomly assigned patients to a training group, whose data was used to construct the decision tree, and a validation group, which was classified according to the obstructed/ non-obstructed decision tree system and used them to estimate the success rate (accuracy) of the tree. The new random recursion and division for each tree prevents bias and allows the weighting of all decision trees for a single model with a predicted percentage of approximation to the P/F test results.
The random forest algorithm, using solely the patients’ uroflowmetry measurements, correctly estimated BCI correctly, compared to the value obtained in the P/F test, 84.2% of the time (sensitivity-92.5%, specificity-27.3%) and BOOI correctly compared to the value obtained in the P/F test for 57.4% of patients (sensitivity – 59%, specificity – 56%).
Our results show that by perfecting the method and selecting the right parameters, the use of artificial intelligence and machine learning can be a new mean of predicting obstructive/non-obstructive voiding dysfunction in men. We believe that in the field of urodynamics, there is great potential for developing models based on artificial intelligence using machine learning that will result in less invasive urodynamic testing in the future. This seminal study lays the path for using big data in urodynamics for both sexes will undoubtedly refine and produce optimal results for such models.