Hospitals offer a variety of tools to help clinicians care for patients receiving diuretics, including urination charts. Although urination charts are described in textbooks as one of the most important diagnostic tools for urination [1-3], these blank charts can contain inaccurate data because they are written by the patient, similar to how journal entries are written. Even if patients are well trained, as is the case in clinical studies, many studies have found that the recorded data may be inaccurate [4-6]. Furthermore, because many variables are involved in the interpretation of these records (usually because patients’ abilities to record their activities are different), it is difficult to use urinary charts to obtain accurate measurements of urinary activity. If these problems could be solved, it would be possible to provide more systematic and efficient care for patients receiving diuretics. Under these circumstances, it is increasingly necessary to come up with a plan to apply information technology to a recording system of urinary activity. Accurate detection techniques using urination monitoring capabilities can systematically and efficiently manage patient urination by duplicating the functions of urination logs. To this end, a management system was developed to recognize patients’ urinary activity by collecting and analyzing information on movements detected by smart bands (magnetic and distance information) worn by patients.

Various methods and learning algorithms for motion recognition have been proposed. Most studies have used static algorithms such as artificial neural networks [7-11] and K-means clustering [12-14], or dynamic time warping [15] in combination with algorithms. However, a time series algorithm should be used to predict or classify dynamically changing time series data. Typical time series algorithms include dynamic Bayesian networks [16], hidden Markov models (HMMs) [17], and recurrent neural networks (RNNs) [18]. The HMM and RNN methods have been widely used for time series data. However, HMMs are not appropriate for learning sequential data because each step is only influenced by the previous one.

In this study, we propose an algorithm for recognizing the movements involved in urination by analyzing data from acceleration signals and gyro signals collected from smart bands. To do so, we developed an analytical technique that enables high accuracy, while solving the limitations of existing studies. Thus, we applied an RNN-based long short-term memory (LSTM) [19,20] method for urination recognition. In addition, functions were developed to show and manage patient urinary activity history information to provide diagnostic assistance for t clinicians. This was integrated into the overall system-level functionality to efficiently manage urological patients. A conceptual diagram of the urination management system is shown in Fig. 1.

Participants’ average age was 28.73 years (range, 26–34 years), and the study was limited to patients with no signs of urination disorders. Users completed a 3-day urination log and wore a smart band to monitor the number of urinations. The smart band used in this study recognized urinary activity by combining gyro sensor information from male patients with distance information (i.e., height information relative to the ground). In this study, an improved method of LSTM-based RNN was applied based on changes in patients’ posture to increase the accuracy of urinary research. In conventional studies [21], complex data are classified simultaneously, so this system has been used to solve problems with techniques that are difficult to analyze in multiple dimensions. This method determines the class to which newly entered signal information belongs based on learned boundaries, such as characterization and classification of gyro sensor information in male patients and distance information (i.e., the height from the ground). A paper urinary diary was reviewed by 2 urologists and compared to the data reported by the smart band.

We also developed a system that enables efficient patient management by gathering and managing the history information of urological patients for clinical diagnostic assistance. This system visualizes the results of urinary activity recognition, enabling appropriate feedback for urological patients.

The performance of the urination recognition technology was analyzed statistically using a confusion matrix. A total of 30 patients were selected, and the performance of the algorithm was evaluated based on data collected from smart bands (acceleration, inclination angle) over 60 days. The final accuracy of the algorithm was calculated based on clinical guidelines used by urologists. The proposed experimental method was divided into 2 main steps for recognizing urinary activity, and urinary time was calculated in the group of patients wherein urinary activity was recognized.

Information on acceleration and inclination angles was collected at frequencies between 12 and 13 Hz, and coordinates of the x-, y-, and z-axes were stored on the smartphone as time series data. The data were reformatted in XML (extensible markup language) and divided into a 3:7 ratio of learning and test data. The feature information we tried to extract included 3 steps: a forward motion to prepare for urination, a backward motion to clean oneself after urination, and urination step. The final urination recognition results were provided considering these 3 levels of information on acceleration and the inclination angle. The highest probability distribution value of the sequence data was selected as the output of the network. The output nodes were labeled for each task and learned using back-propagation algorithms, where x, y, and z represented the structure in which the accelerometer and gyro sensor signal data were entered. This comprised the process of calculating node continuity and the state of the current point at time (t) and the next point at time (t+1). Processing the target network involved 2 main steps. From phase I, the final urination phase was determined by comparing the perceived urination phase in the network with the perceived urination phase in the second-stage gyro data through the input of accelerometer data.

After the experiment was carried out, a confusion matrix (Table 1) for the proposed method and the existing method was calculated, and the accuracy was shown through actual differences in terms of seconds (Table 2) and receiver operating characteristic (ROC) curves (Fig. 4). Each number in Table 1 denotes the actual number of tests, with a total of 60 confusion tests. The numbers displayed in Table 2 have units in seconds; this table represents the average absolute values of time differences obtained through testing in 60 cases. The results showed an average accuracy of 95.6% in recognizing urinary activity, confirming the validity of urinary activity recognition. The ROC curve for the system proposed herein is shown in Fig. 4 along with each comparison method.

This paper proposes a method to assist in making an efficient clinical diagnosis through the establishment of an automatic urinary management system for urological patients. To this end, an LSTM-based RNN method was used to recognize patients’ urinary activity, and the results were visualized through a web-based program to assist clinicians’ diagnosis.

Existing classification methods are capable of recognizing the number of urination events, but it is difficult to analyze urination time using these methods due to the loss of location information when converting spatial data. Instead, the RNN method using the LSTM structure is appropriate for this goal because it analyzes the urine collection time and collects data continuously. The calculation of urination time is processed by identifying positional information over time in the forward and backward movements involved in urination. Processing is conducted through the recognition method of the RNN structure. Specifically, the proposed urination recognition technology collects acceleration and inclination angle information from smart bands and applies an LSTM structure-based RNN to derive the final urination frequency and time.

The performance of the proposed system for urinary activity recognition was evaluated using a study population of 30 actual urological patients who collected data for 60 days. The results showed an average accuracy of 95.6% in recognizing urinary activity, confirming the validity of urinary activity recognition. In addition, the urination time can be calculated after urinary activity is recognized, which will ultimately allow a comprehensive evaluation of urination in patients with urinary dysfunction. On this basis, the ultimate goal is to consider the personal characteristics of individual patients when calculating the doses of medications affecting urination. Some inaccurate results in recognizing urinary activity resulted from errors in the calculation of weights for accelerometers and gradient data. For urinary activity time measurements, the recognition rate of LSTM was found to be lower than that of comparable methods. However, the neural network was composed of random initial weights, so the average recognition accuracy was still high. To improve recognition accuracy, we plan to utilize Dempster-Shaper theory [23] or bidirectional LSTM [24], wherein 2 hidden nodes are placed in the forward and backward directions on RNN input and output nodes. In addition, if the amount and quality of data are ensured in the future, a CNN model can be applied to time series data; future research should also consider analyzing big data and developing CNN and RNN fusion deep-learning models.

In conclusion, recognizing urinary activity and providing feedback through a urinary patient management system makes it possible to monitor a patient’s actual urinary patterns. The system described in this study can be a useful option for diagnostic assistance.