Evaluation of Decoding Algorithms for Estimating Bladder Pressure from Dorsal Root Ganglia Neural Recordings

A closed-loop device for bladder control may offer greater clinical benefit compared to current open-loop stimulation devices. Previous studies have demonstrated the feasibility of using single-unit recordings from sacral-level dorsal root ganglia (DRG) for decoding bladder pressure. Automatic online sorting, to differentiate single units, can be computationally heavy and unreliable, in contrast to simple multi-unit thresholded activity. In this study, the feasibility of using DRG multi-unit recordings to decode bladder pressure was examined. A broad range of feature selection methods and three algorithms (multivariate linear regression, basic Kalman filter, and a nonlinear autoregressive moving average model) were used to create training models and provide validation fits to bladder pressure for data collected in seven anesthetized feline experiments. A non-linear autoregressive moving average (NARMA) model with regularization provided the most accurate bladder pressure estimate, based on normalized root-mean-squared error, NRMSE, (17 ± 7%). A basic Kalman filter yielded the highest similarity to the bladder pressure with an average correlation coefficient, CC, of 0.81 ± 0.13. The best algorithm set (based on NRMSE) was further evaluated on data obtained from a chronic feline experiment. Testing results yielded a NRMSE and CC of 10.7% and 0.61, respectively from a model that was trained on data recorded 2 weeks prior. From offline analysis, implementation of NARMA in a closed-loop scheme for detecting bladder contractions would provide a robust control signal. Ultimate integration of closed-loop algorithms in bladder neuroprostheses will require evaluations of parameter and signal stability over time.