Short: Real-Time Bladder Monitoring by Bio-impedance Analysis to Aid Urinary Incontinence

The inability to sense the need to urinate is a persistent trouble for many individuals who have urinary incontinence. There are no products for people with urinary incontinence and their caregivers that target the elimination of involuntary urination instead of containing leaks or warning about leaks. Thus, fulfilling the real-time monitoring of the bladder to notify urination time holds a life-changing innovation potential for people with urinary incontinence. In this paper, we propose a low-cost wearable device in the form factor of an unobtrusive belt that uses bio-impedance measurement across the bladder region to notify the patients or caregivers about the need to urinate. We also present results from human subject tests using the proposed wearable prototype with a custom machine-learning algorithm to evaluate the accuracy of the system. Results from the human-subject tests showed an accuracy of over 90% on the binary task of full versus empty bladder states based on the change in bio-impedance values.


INTRODUCTION
Urinary incontinence (UI) is de ned as involuntary urination or loss of bladder control that leads to social and hygienic problems [1].
According to the statistics of "The National Association of Incontinence" over 200 million people suffer from UI worldwide due to different reasons [6].Risk factors for UI include age, menopause, parity, obesity, vaginal delivery, and past hysterectomy [7].In addition, diseases and disabilities such as Alzheimer, Dementia, Parkinson and cerebral palsy, or spinal cord injury negatively affect awareness of bladder control.Currently, products available in the market for people with UI include absorbent pads, adult diapers, catheters, and urinals to contain urine and wet-detecting alarms or sensors embedded in sheets, underwear, or pads to warn about leaked urine.There are no alternative products for people with UI and their caregivers that target the elimination of involuntary urination instead of containing leaks or warning about leaks.The user problems associated with available UI products call for product innovation.
Wearable technologies, which refer to electronic and computer technologies embedded into textiles, clothing and accessories, worn comfortably on the body [8], can present novel solutions for UI since wearables have intimate proximity to the human body.Such wearable technology can limit UI patients' dependence on absorbent products and enhance psycho-social health by intimate monitoring and warning.Traditionally, bladder fullness measurement is done by ultrasound technology, which is regarded as the golden standard for measurement of the bladder volume [11].However, ultrasound technology relies on relatively big and heavy devices which makes it difficult to embed in a wearable form factor for real-time monitoring.Also, ultrasound devices always need the application of gel on the skin to eliminate the gap between the electrodes and skin to conduct the ultrasound effectively.With the development of wearable technology capable of real-time monitoring [13], it is also possible to fulfill bladder monitoring noninvasively.A current study showcased how personalized medicine and remote monitoring of electrocardiogram(ECG) can be brought to the next level by enabling the patient and doctors with time-sensitive data via a realtime monitoring system [2].Similarly, real-time bladder monitoring can offer time-sensitive data given how critical the timing of a leak warning is for a UI patient.
Among the recent literature on bladder monitoring, using bioimpedance analysis has demonstrated reliable outcomes.Bio-impedance is a parameter that presents how well people's bodies impede electric current.Different compositions of the body (e.g, fat, blood, and  2020) used a high-performance Field Programmable Gate Array (FPGA) and presented good correlation between abdominal admittivity and bladder volume [12].However, none of the studies explored wearable and real-time monitoring.
In this paper, we propose a novel real-time bladder monitoring wearable technology in the form factor of an unobtrusive belt worn an undergarment around the bladder region.The belt houses 7 electrodes arranged as shown in Figure 1 and a snap-on control unit that can wirelessly communicate information to external displays.After the optimization of the hardware design, its validation and integration, we built a custom machine-learning algorithm based on random forest model.Finally, a preliminary longitudinal human subject test (n = 2) was conducted for determining the accuracy of the proposed wearable technology and the algorithm.We measured the bio-impedance using the belt on 2 participants for 10 days before and after urination.We used the first nine days' data to predict the last day data.We employed a computer to receive the bio-impedance data wirelessly and in real-time and to train the machine learning model.The results indicated over 90% accuracy on the binary task of assessing bladder fullness.The overall system shows great potential to help people mitigate UI and bring in more benefits via the internet of things (IoT) technology for advanced healthcare management.The multiplexers had eight channels individually, seven of them were connected to the electrodes on the belt and the last one was connected to a high-resolution reference resistor.
The power management of the circuits was handled by a power management chip (TP5400 NanJing Top Power ASIC Corp.) to fulfill two functions: 1) manage the charging current to the Lithium-ion battery.2) provide a steady power supply to the SoC, Multiplexers, AFE, and LED instruction lights.We tested the power consumption of the PCB board as approximately 15mA.This included the BLE transmission power under 3.7V power supplier (a 500 mWh Lithiumion battery), indicating over one and half days duration for using the wearables continually.PCB board accounted for the main cost of the whole system.The total cost for the PCB is about $42 (shown in Table 1).The bioimpedance measurement array included 7 electrodes which was optimized to achieve the best cost and performance ratio.Thus, our proposed wearable has strong potential to be a low-cost commercial alternative to aid UI compared to the present commercial device which costs about $400 [10].
Finally, we integrated conventional Ag/AgCl Solid Gel (Ambu,Inc) electrode pads which are already used in the bio-impedance measuring systems to the belt prototype.As discussed under the future work section of the paper, we are working on replacing the conventional electrodes to electronic textile-based electrodes, which can be reused and washed, to improve the user experience.The data collected from human subjects at this stage of the study using conventional electrodes would be a good reference to compare the performance of our electronic textile electrodes.[3].In order to automatically discriminate the urination status from the bio-impedance data, we developed a machine learning pipeline.Our bio-impedance array consisted of seven electrodes and each pair of electrodes produced a bio-impedance value.Thus, we collected 21 potential values from the bio-impedance array as the raw data.Then, the raw data were normalized to lessen the dissimilarity between the data from different days as the human subjects' status (such as activity level and realignment mismatch) may affect the measurements and lead to bias.The 21 bio-impedance values served as the input feature of the machine-learning model.We examined random forest, support vector machine and deep neural network as machine-learning algorithm and based on the result of initial experiments described under the results section, random forest was employed in this study.Once the connection was established, MCU communicated with AD5933 to acquire the bio-impedance value while controlling the multiplexers until all 21 values are collected as one point.Then, MCU checked whether the BLE was still connected.After that, MCU sent the data to the computer.Finally, the gain of AD5933 was calibrated through a high-precision 1k resistor.
For the Python script, the main functions were 1) to receive the data from the embedded systems with the targeted number of points.2) to generate the CSV file to save the data automatically with the label and the timestamp.When running the Python script, the script asked for information from the human subjects.Then data collection script waited for the connection of BLE and received the data from the wearable system and accounted for the number of points until enough points have been collected.The computer played a beep sound immediately after receiving a data point to notify the human subjects to move according to the experimental protocol described in section 2.2.Finally, the script saved all data and human subjects' information to a CSV file.

Subject Test
Two human subjects, adult male (subject one) and female (subject two), were recruited in this preliminary testing.Both participants were healthy and did not have any urinary system-related disease.The duration of testing was for 10 days period for each participant.
The experimental protocol consisted of two phases where participants wore a 7-electrode bio-impedance array around their waist leveled with the bladder region in a belt form.Phase one was the acquisition process before urinating and phase two was the acquisition process after urinating.During phase one, a brief introduction was delivered to the participants about how and when to gather data during the 10 days as well as how to don/doff the belt precisely.Once the participant had the desire to urinate, the 7-electrode bioimpedance array, which was evenly distributed across the waistline circumference on a belt form, was placed on the participant.The first wearing was done with the guidance of the researchers to ensure the accurate placement of the belt every time the participant collected data.Once the participant felt comfortable with the belt, the PCB board was turned on and the data collection script was run on the computer.Then, the Bluetooth connection was established automatically between the PCB board and the computer.Data points were collected every 6 seconds while the participants wore the electrode array.Between the intervals of two-point acquisition, the participants were asked to walk about 5 steps steadily to simulate the daily movements to avoid over-fitting of the machine learning model.The total data acquisition for the first phase was 10 minutes every time a participant had the need to urinate during 10 days period of time.
After phase one, the participants emptied their bladders while the belt is on.In phase two, the participants repeated the data acquisition steps from Phase 1 for another 10 minutes every time after urination for 10 days.After the participants collected data for 10 days (two phases per day), CSV files were generated by the data collection script and went through further analysis.The testing protocol generated 200 data points during the two phases as 21 bio-impedance values for a single point per participant.

Statistical Analysis
After data collection, we first conducted statistical analysis to interpret the correlation between urination status and potentials from the electrodes.Figure 4 shows the correlation heat map of subject one.Each square represents the Pearson correlation coefficient between the two attributes and the last row shows the relationship between label (urinate or not) and different features.

Figure 4: Heat map of subject one
In addition, a violin plot was generated to display information about the dispersion of the set of data.It was mainly used to reflect the characteristics of the original data distribution, and could also be used to compare the distribution characteristics of multiple groups of data.As seen in Figure 5, the X-Axis displayed any possible two electrode pairs' impedance in the array while the Y-Axis displayed the bio-impedance value.We employed the two sides of the violin to represent two status, before urination when the bladder is full and after urination when the bladder is empty.From the plot, it is possible to see an apparent difference in bio-impedance data distribution between the two status (e.g., electrode pair 65 in Figure 4).From the statistical analysis, we found that the 21 electrode pairs contained different information and contributed to the discrimination of classification differently.

Machine Learning Analysis
We deployed a machine-learning pipeline to evaluate the performance and the accuracy of the whole system.Machine learning analysis was conducted on two participants separately.Even though the data from different days were independent and identically distributed, we used the data from a period of time as the training set and the day after that period as the testing set instead of randomly shuffling and splitting the order of the data.In this way, we aimed to mimic the process in the real-life scenario that the model was trained at the beginning and used in the follow-up days.
To achieve the optimized settings, we evaluated using different days of data as training set to find the best alternative.The accuracy curve is shown in Figure 6.It is found that on day 9, the accuracy reached the highest and then stayed stable which indicated that the model converged and became steady afterward.Thus, we used the 9th days' data as the training set.As indicated in Table 2, we achieved 90.5% and 88.3% accuracy respectively for each participant in predicting the urination needs using their bio-impedance data to train the machine learning algorithm.As indicated by the results, we could see that as more data was included in the training set, the accuracy increased accordingly.
The accuracy was saturated after 9 days reaching about 90 % accuracy finally.The amount of needed data was large to achieve this accuracy level, note that 9 days of data contained 1800 points that were fed to the machine learning system.We also discovered that bio-impedance values could be influenced by the realignment of the electrode positions every time participants put on and put off the belt during the 10 days which may bring more noise to the machine learning system.Thus, the larger amount of the data collected, the more robust the prediction system gets.According to our results, 9 days of data collection was good enough to make the prediction system robust.
During the data collection period, we also discovered that participants' physical condition (sweat and activity level) could affect the data.The result showed that our present machine-learning model cannot have a good binary-prediction result from a participant who just did vigorous exercise (e.g., playing pickleball or jogging).We assume that sweat may have a large influence on the bio-impedance measurements by changing either electrode-skin contact or shunting the current that goes through the body.Other sources [8] have reported a comparable occurrence where the use of electrical impedance tomography (EIT) to measure lung volume changes is ineffective when a patient perspires.

CONCLUSION
In conclusion, a novel real-time bladder monitoring wearable, which used the bio-impedance data and machine learning algorithm, was proposed and successfully implemented.In particular, we fulfilled accurate prediction results which reached 90% in the preliminary human subject tests.Our findings indicate a strong potential to become a commercial, low-cost wearable device that fulfills the real-time monitoring to aid the urinary incontinence patients.

FUTURE WORK
In future work, expanding the number of participants in experiments is our immediate goal.In this preliminary work only two participants were involved since the protocol was time-consuming (1 hour per day and 10 days in total).As more participants are recruited, we can 1) have more reliable results 2) compare the crossuser accuracy (use one participants to train the model and test on the other participants).In this study, we developed a custom fabric belt to house the 7 conventional Ag/AgCl gel electrodes.We are in the process of developing electronic textile electrodes embroidered directly onto the belt using Silver coated, polyamide conductive thread which will allow reusability and washability [5].The final design and system we continue developing and testing is shown in Figure 7.We aim to develop a fully electronic textile-based intelligent wearable belt for real-time bladder monitoring, to measure the bio-impedance across the bladder region of the torso where the raw data can be processed by the smart-phone and function as a warning system to notify the wearer or a caregiver about the time to urinate.Recent literature also revealed [1] that higher resolution (more electrodes) could result in the ability to predict the volume of the bladder.We will add more electrodes to increase the sensitivity of the wearable system, muscle) have different bio-impedance values.Dorothea et al. (2018) tested a commercial bio-impedance device with 16 electrodes on 4 participants to monitor bladder functions [9].Bruno et al. (

Figure 1 :
Figure 1: Illustration of the real-time bladder monitoring testing system

Figure 2 :
Figure 2: The PCB board:(a) The front of PCB board and battery.(B) The back of PCB board

2. 1 . 3
Data Collection Software.The script for MCU was written in C program and another script was written for the computer, which functioned as the external unit to store bio-impedance data wirelessly in real-time in Python.For MCU, the main functions were 1) control the Multiplexers to connect the AD5933's input and output pins to the corresponding two electrodes.2) Bio-impedance data acquisition from AD5933 via I2C serial communication bus. 3) send the sampled data to the computer via Bluetooth low energy (BLE) wireless communication protocol.When the PCB board was on, the embedded system established BLE connection with the computer.

Figure 3 :
Figure 3: Data Collection Software Flow

Figure 5 :
Figure 5: Violin plot of subject one

Figure 6 :
Figure 6: Accuracy curve with different training set coverage

Figure 7 :
Figure 7: Proposed future work for the wearable system

Table 1 : Cost summary
[4]hine Learning Algorithm.Machine learning algorithms are an important tool for making sense of the vast amounts of data collected from wearable health devices[4]