Hypothesis / aims of study
Non-invasive diagnosis of detrusor underactivity (DU) is in high demand, however the similarity of low flow rate with bladder outlet obstruction (BOO) hampers the possible non-invasive diagnostic development. There are a few articles proposing uroflowmetry parameters for discriminating DU from BOO non-invasively, but the diagnosing power is still limited (1). A recent study proposed to combine non-invasive parameters to differentiate DU from BOO by employing a linear statistical model, and yield optimised sensitivity of 73.1% and specificity of 84.6% (2). in this study, the aim is to test the possibility of using feedforward artificial neural network (ANN) method, which has the key feature of nonlinearity and capable of learning and adaptability, to non-invasively differentiate DU from BOO by employing non-invasive urine flow parameters and flow shape.
Study design, materials and methods
158 BOO and 135 DU free flow data from male patients who underwent pressure flow studies (PFS), studied in a single specialist centre between 2012 and 2018 were selected for analysis in this study. Free uroflowmetry was performed before each PFS which was carried out according to ICS guidelines, with sampling rate of 10Hz. All urine flow rate data have been pre-processed for start and end points with a threshold value of 0.5ml/s. The analysis was conducted in the MATLAB version 2018b.
The non-invasive parameters employed for generating ANN model are derived by analysing urine flow rate in both time and frequency domains, which are the same as in previous analysis of a multivariate analysis of variance (MANOVA) model and have been proved to have significant statistical difference between the two groups (2). In order to test the diagnostic utility on pure flow shape with ANN, each set of urine flow data is normalised into 1000 re-sample points and a maximum amplitude of 1ml/s (3).
All data are split randomly with the percentage of 70, 15, 15 for training, validation and testing respectively. The ANN models for non-invasive parameters and flow shape are built of one hidden layer with 53 and 815 hidden neurons separately which are chosen by five-fold cross validation on overall accuracy.
Interpretation of results
The result shows that feedforward ANN model holds promise to discriminate DU from BOO by employing non-invasive uroflowmetry parameters, with 79.5% overall accuracy, 75.6% sensitivity and 82.9% specificity. The designed ANN model on urine flow shape has a lower overall diagnosing accuracy of 70.6%, with 54.3% sensitivity and 82.7% specificity. Both models are fairly robust as the overall accuracies are similar to training accuracies. The data used in this study, however, were preselected to be patients with either DU or BOO urodynamically proven. A full clinical trial will need to use a more mixed population.