Treatment of Nocturnal Enuresis Using Miniaturised Smart Mechatronics With Artificial Intelligence

Abstract

Our study was designed to develop a customisable, wearable, and comfortable medical device - the text so-called "MyPAD" that monitors the fullness of the bladder, triggering an alarm indicating the need to void, in order to prevent badwetting - i.e., treating Nocturnal Enuresis (NE) at the text pre-void stage using miniaturised mechatronics with Artificial Intelligence (AI). The developed features include: multiple bespoke ultrasound (US) probes for sensing, a bespoke electronic device housing custom US electronics for signal processing, a bedside alarm box for processing the echoed pulses and generating alarms, and a phantom to mimic the human body. The validation of the system is conducted on the text tissue-mimicking phantom and volunteers using Bidirectional Long Short-Term Memory Recurrent Neural Networks (Bi-LSTM-RNN) and Reinforcement Learning (RL). A Se value of 99% and a Sp value of 99.5% with an overall accuracy rate of 99.3% are observed. The obtained results demonstrate successful empirical evidence for the viability of the device, both in monitoring bladder expansion to determine voiding need and in reinforcing the continuous learning and customisation of the device for bladder control through consecutive uses. Clinical impact: MyPAD will treat the NE better and efficiently against other techniques currently used (e.g., post-void alarms) and will i) replace those techniques quickly considering sufferers' condition while being treated by other approaches, and ii) enable children to gain control of incontinence over time and consistently have dry nights. Category: Early/Pre-Clinical Research.

Keywords: Nocturnal enuresis; incontinence; long short-term memory recurrent neural networks (LSTM-RNN); reinforcement learning (RL); wearable medical devices.

FIGURE 1.

Main components and techniques of the modular decision-making methodology.

FIGURE 2.

Components of the phantom: a) outer perspex cylinder; b) TMM 1 (top) (2.5 cm thick); c) 8 inner perspex cylinders in different heights (1–8 cm) corresponding to the expansion of the bladder from empty to full; d) TMM 2 (bottom) (3 cm thick); e) TMM preservation fluid; f) whole phantom design: container with TMM 1 (top), TMM 2 (bottom), inner perspex cylinder and TMM preservation fluid between these two layers; g) real phantom.

FIGURE 3.

Cluster design of miniaturised probes with 5 mm apart.

FIGURE 4.

Pulse acquisition interface: Pulses with 4 receivers in their dedicated channels and their composite outcome at the bottom. The first amplitudes in the frames correspond to the echoed pulse coming from the anterior wall whereas the second amplitudes correspond to the pulses from the posterior wall of the bladder. The red circles indicate the places where no echoed pulses are detected from the anterior wall because of the reflection and refraction angles of the emitted US beams from the transmitter. The amplitudes in the yellow ovals indicate the noise caused by the high signal-to- noise ratio.

FIGURE 5.

Components of the electronics and their connection.

FIGURE 6.

Data collection from the phantom (full bladder) (left) and the volunteer (3/4 full bladder) (right).

FIGURE 7.

Phases of data processing.

FIGURE 8.

Training phase of Bi-LSTM-RNN: The top subplot presents the progress of the training accuracy and the bottom subplot shows the cross-entropy loss on each mini-batch - i.e., the reducing error down to zero if the training progresses successfully.

FIGURE 9.

Design of the MyPAD with sensors, electronics, and battery and its use with the undergarment on a manikin.

FIGURE 10.

Early miniaturisation of the device.

FIGURE 11.

First prototype assembly of the device.

Algorithm 1

Reinforcement Learning: C. Alarm triggering & Bladder control module (Fig. 1C)