1. Introduction

The number of degenerative/metabolic diseases has been increasing owing to aging, urbanization, and lifestyle changes. This increase has resulted in growing attention to healthmanagement in society and individuals. Accordingly, the smart healthcare industry, which specializes in providing continuous healthcare in everyday life, has attracted public interest. The healthcare big data linkage platform currently provides users with valid information through the collection and analysis of series of data related to users’ health by means ofambient sensors. Smart health services provide diagnosis/prevention-based medical servicesin the field of disease treatment-based medicine [1]. Recently, precision medicine focusing on personal diagnosis/disease prevention has attracted public interest. Meanwhile, the health care industry is developing personalized health management and lifestyle improvement models and services with the aim of providing universal precision medicine [2]. Biosensors, networks, and knowledge bases were investigated for collecting and analyzing data, such as users’ status, surrounding environment, and surrounding circumstances [3]. A technology featuringintegrated data collection is being developed in the field of biosensors to allow the transition from a single-modal system capable of using one sensor to collect one object data to a multi-modal system capable of collecting diverse data through ambient sensor networks [4]. In the network for the multi-modal system, the data simultaneously collected with an ambientsensor in the same environment are used to learn the associated relation among multi-modal systems. Based on this process, a deep learning network capable of extracting therepresentation shared among modes from the multi-modal input is constructed. Deeplearning-based artificial intelligence (AI) has limitations, such as a lack of information insingle-modal systems, lack of effective convergence among multi-modal input information, and the knowledge bottleneck phenomenon. These problems can be resolved by devising a multi-modal machine learning method where shared representation and artificial intelligence AI factors extracted based on modal input information are converged [5]. In the field of networking, a common data model was investigated that uses multi-modal-based knowledge expression, acquisition, and reasoning in a mesh-form hybrid peer-to-peer (P2P) networksystem based on the host device [6]. This is a dynamic structure that is advantageous in terms of sharing and distributed processing of data based on diverse connectivities and is used tointegrate data collected from wearable devices. The structure consists of users, servers, and gate ways and is capable of efficient data management and collection through diverse routes [7].

Diverse technologies capable of collecting, supplying, processing, and analyzinglarge-scale data received from multi-modal sensors, mobile devices, RSS, and XML are being developed in the field of knowledge bases. In the current healthcare industry, data quantified through connections between electronic medical records (EMRs), personal health devices (PHDs), and personal health records (PHRs) are being integrated and preprocessed [8-10]. The big data in healthcare generated from mobile-based wearable devices are collected by different means according to the characteristics of the structured and non-structured data. Adeep learning method capable of producing knowledge learned based on machine learning is being used to acquire logic-based knowledge and expand the ontology and logic knowledge bases. This method allows the extraction and expansion of significant knowledge through theacquisition of knowledge from a knowledge base and then the refinement of this knowledge. Accordingly, a technology must be developed that is capable of integrating the rapidly increasing healthcare big data and lifelogs and integrating and processing heterogeneous big data that are closely related to health, such as nutrition, environment, and meteorological data. In addition, given the rapid increase in the number of patients per medical worker due to overpopulation resulting from urbanization, it is necessary to universalize medical servicesthrough the supply of smart health. Currently, the universal wearable mobile-based devices are the smartwatch and health band, and the market size of these devices is expanding continuously. However, the integration of the standards or specifications for managing collected data remains to be achieved. In addition, as the mobile-based wearable devices lackan understanding of the users’ tendency, purpose, use, and behavioral changes, they provideonly standardized health management services. This issue can be resolved through machinelearning that mimics the high-dimensional cognitive skills of humans, active machine learning based on knowledge of the real world, multi-modal-based knowledge expression, and acquisition and reasoning technology.

This paper is structured as follows. Section 2 discusses mobile healthcare provided by means of wearable sensors. Section 3 presents the proposed prediction model of user physicalactivity that implements a data characteristics-based long short-term memory (DC-LSTM) recurrent neural network (RNN). Section 4 describes the performance evaluation and results. Section 5 concludes the paper.

2. Mobile Healthcare Provided by Means of Wearable Sensors

The sensors of health bands, smart caps, and ECG-measuring smartwear are used to collect context information, without the user’s awareness, on a mobile health platform regardless of time and the user’s location. Wearable sensors allow the collection, recording, and storage of lifelog data, such as time, place, location, movement, biosignals, and calories.

The sensors of previously developed health bands that are worn on the wrist and frequently used by individuals while exercising [1] collect the user’s biosignals and provide the user with his/her context-based health status. In addition, the smart cap, which is worn in everyday life[8], collects location-based context information and provides five types of health-weatherindices and eight types of life-weather indices for every geographical region through a sensorattached to it. Weather indices can be used to provide services according to a certain location and to allow the user to actively prepare his or her physical body for climate changes. Moreover, phased context awareness-based cautions according to the current location of the user, detected through a GPS, can be provided. The level of the influence of meteorologicalelements on health or the level of the possibility of an adverse weather occurrence, displayed as an index that refers to the probability of such an occurrence under specific conditions, isquantified using a prediction model developed based on meteorological data. The public dataportal [11], which contains actualized index information per context, provides an open API based on weather index information that is applicable to everyday life and healthcare.

Fig. 1. Smartwear with a wearable sensor

Wearable sensors are capable of measuring body temperature, humidity, illumination intensity, temperature, and UV and, using Bluetooth communication technology, can transferthese data from mobile devices featuring GPS reception. These sensors use 2.4 GHz ZigBee, Atmega 128L, on TinyOS 2.X to share Bluetooth communication with mobile devicesfeaturing GPS reception for the transfer of data packets. In the research conducted by Kim et al.[7] and Chung et al. [1], a wearable sensor was used for context recognition and health careservice provision. An ontology-based context information model can be created using the OWL reasoning process based on the Jena API [12]. A knowledge base can be constructed based on service reasoning, health reasoning, and context reasoning rules, and the context peruser can be recognized to establish an evolutionary reasoning rule. The data packet consists of packet values starting with 7E and ending with 7E for serial communication with wirelesssensor networks and can be expressed in the form of hexadecimal numbers based on the consecutive numerical values indicating body temperature, humidity, illumination intensity, and UV. The packets received from a serial port can be expressed as “7E 46 04 FF FF 03 0A 0000 08 00 03 1B 00 38 00 05 00 70 00 72 D4 7E.” In addition, a wearable sensor can be manufactured that can be attached to and detached from body regions, thus causing nohindrance to the user’s movements, which resolves the structural problems of complicated smartwear. The really simple syndication (RSS) based weather data [13] from the Korea Meteorological Administration and the data received through IEEE 802.15.4 Standard Wireless Transfer can be analyzed according to the location in a region to provide a health-weather index. Given that meteorological and context information varies and users & rsquo; health status changes fluidly according to their current location, a GPS receiving module is used to provide adequate services in real-time. RSS-based meteorological data are forms of data established based on XML and JSON, the two switched data standards, and provideupdated information every three hours using the interface standard REST(Get) method. To collect meteorological data, queries are used to save data in the form of an index and a display, and a Document Object Model (DOM) parser is used to extract, exchange, and process datainto an XML format [14]. This method expresses contents and structures as objects and provides a standard interface that can be managed by the user. Fig. 1 shows smartwear with anattached wearable sensor, including a health band [15], smart cap [10], and ECG measuring smart wear [16] previously developed on a mobile health platform.

Chung et al. [1] developed a user-adaptive decision-making simulation for which smartwearto which a sensor was attached and a meteorological WebBot were implemented. In theirresearch study, emotions that change according to meteorological elements were analyzed and applied to decision making. In a previous research study, ECG measuring smartwear was developed that uses cardiac biosignal data to monitor heart rate variability (HRV) in real-time [16]. Because it requires an attached wearable sensor and a circuit for transferring the ECG and heart rate (HR) signal data, it was manufactured in the form of a small attachable pocket. In addition, the design of the smartwear considered the electrode’s position and volume, the wearer & rsquo;s movements, battery, clothing pressure, and clothing type. To provide smartwearcapable of producing stable ECG measurements, an electrode was fixed to a highly elastic band in the form of embroidery, which allowed stable contact of the smartwear and biosensor with the body, and thus, the ECG and HR signal data could be analyzed accurately. This ECG measuring smartwear was worn by a subject and the ECG/HR data were collected in the 1-100Hz frequency band. To examine the ECG waveform according to the user status, the powerspectrum was analyzed using the low frequency/high frequency ratio. The time series analysis according to time was conducted by limiting the size of the window within the ECG data. A peak detection algorithm was used to calculate the R-R value, and Fourier transform was used to analyze low frequencies. Fig. 2 shows a mobile health service in which wearable sensors are implemented. The figure shows the HRV-based stress index service [6], chatbot-based mobile health care [17], and dietary nutrition recommendation service developed on a mobile health platform for the management of obesity among teenagers [18].

In the research study conducted by Yoo et al. [1], HRV was used to analyze the activity of sympathetic and parasympathetic nerves on a mobile health platform to determine users’ stresslevels. According to the input biosignal data, the HRV-based frequency domain can be divided into a positive status and negative status based on the negative feedback provided from the sympathetic and parasympathetic nerves of the user. Using this process, users can monitor their cardiac impulse and ECG analysis in real-time and set their health management andexercise based on this analysis. In addition, this process is valuable when used to monitor the status of patients with heart diseases, as well as for detecting incidences of respiratory disturbance-related diseases. The chatbot-based mobile health service developed by Park et al.[17] is an intelligent chatting interface that provides prompt treatment for emergencies that may occur in everyday life and responds to the changing status of patients with chronic diseases. The diagnosis/treatment program, which can be installed in diverse types of mobile devices, was expanded through the analysis of the interaction among data using naturallanguage processing. The dietary nutrition recommendation system for obesity management developed by Jung et al. [18] collects context information from mobile device health data anduses knowledge base-based cooperative filtering to predict missing values in the {User, Diet}matrix. Using the constructed {User, Diet}-merged matrix, a customized diet is recommended for obesity management according to context. This service allows recipes and diets to be provided through a mobile device regardless of time and the user’s location.

Fig. 2. Mobile health service implementing wearable sensors

3. Prediction Model of User Physical Activity using Data
Characteristics-based Long Short-term Memory Recurrent Neural
Networks

The current activity of a user is predicted using his/her record of past activity according to adate or day of the week. For example, if the average activity of a user on previous Sundays burned 1,000 calories, this method predicts that the activity of the user on the subsequent Sunday will also burn 1,000 calories. However, this prediction method does not consider the user context, and the error of recommending outdoor activity in a context where such activity is impossible occurs if the recorded activity is insufficient, that is, in this case burns less than 1,000 calories. In order to avoid such an error, data that are highly relevant to the user & rsquo;sactivity and can easily be collected from healthcare data are selected and used for predicting his/her future activity. The range of the numeric data in the domain is clear; however, in the case of time series data, an LSTM model is used in the recommendation method in order to analyze the data in time intervals and determine whether such data are normal or abnormal. Given that users’ health data and surrounding context data change in real-time in everyday life, the renewal cycle, collection method, overlapping significance, acquisition convenience, andutility must be considered. The level of difficulty of the data sequence and the collection of mobile health data vary according to the collection methods. In addition, the scope and management of collection vary according to the user’s field of interest or owned devices. Accordingly, to analyze health data efficiently in a mobile environment the characteristics of the data must be considered. Health data are specific in that they are mutually influenced byeach other, either directly or indirectly. For instance, users’ BMI and body fat percentage change because they are influenced by the users’ height and weight, and users’ blood sugarchanges because it is influenced by their intake of sugar. Given the mutual, direct relationship between health data, it is difficult to find a hidden significance through data analysis. The maximum and minimum temperatures are used to calculate the daily temperature difference using an arithmetic operation. However, it is possible that an overlapping significance may be shared by the variables. Between data that do not indicate a clear association, such as dailytemperature difference, activity, temperature, and sleep, it is possible to find a significantrelationship by analyzing their hidden relationship [19-22]. In order to provide efficient health services, these relationships must be constructed based on data that can be conveniently collected by users. In this paper, we propose a user physical activity prediction model that uses the DC-LSTM RNN. The proposed model is a prediction method that implements a neural network, the construction of which is based on the characteristics of health-related datacollectable from mobile host devices. Health data and weather data continuously change overtime because of their time series characteristics. Accordingly, an old state affects a new state. Therefore, in this study an RNN was used to predict user activity. The structure of an RNN issuch that the old state is used in the learning process as additional input. In the case of an RNN, it is likely that the long-range dependence issue that the slope disappears as learning continues will occur, and therefore, only short-term data can be considered. Therefore, an LSTM RNN model, which allows long-term data to be considered, was selected. Fig. 3 shows the configuration of the proposed physical activity prediction model.

Fig. 3. Configuration of the proposed physical activity prediction model

3.1 Selection and Collection of Multi-modal Health Data

The term multi-modal health data refers to all health-related data for the same target that arecollectable from mobile host devices. These data can be divided according to the collection methods and data sources, such as PHRs, Meteorological Information Systems (MISs), and lifelogs. In addition, because they are obtained through diverse collection means, these data are multi-modal. Data representing the same target are collected using an ambient sensornetwork. The collected data have mutually different characteristics related to the same targetand show incompleteness, noise, and inconsistency problems [20,23]. A PHR consists of personal health-related variables, such as gender, age, height, weight, family medical history, allergy, smoking status, drinking status, and health-related medical examination information. Meteorological information consists of weather, temperature, humidity, UV, dailytemperature difference, fine dust, and weather index. The term lifelog refers to all recorded data on the environments surrounding an individual’s everyday life. The data are collected by means of diverse wearable devices, personal health devices, and location-based services. Wearable devices, such as health bands [1], smart caps [10], and ECG measuring smartwear[16], use an ambient sensor to collect health-related lifelogs. These devices can be used tocollect and analyze multi-modal health data. They use short-distance communication means, such as Bluetooth, WiFi direct, and Beacon to achieve real-time interaction with mobile host devices. Ambient sensors can be divided into photoplethysmography (PPG) sensors, pulse wave velocity (PWV) sensors, gyro sensors, and GPS. PPG sensors measure the HR by detecting the blood flow with an LED light. ECG refers to electrocardiography, which measures the electrical activity of the heart. The users’ personal HR, sleep, activity, altitude, atmospheric pressure, and cadence can be measured using these sensors. A personal health device is a small device capable of measuring the user’s bio status and health status and is used to measure the user’s blood pressure, blood sugar, heart rate, and sleep. Location-based service refers to all the services based on the users’ location, such as travel routes, weather, traffic conditions, surrounding facilities, and emergency facilities. Table 1 shows a list of collectable multi-modal health data.

Table 1. Multi-modal health data

In general, more types and a greater volume of data allow a more accurate analysis. Giventhat different users own different devices, universally collectable data must be structured and the scope of the data must be set. In this study, related data were set as variables for predictinguser physical activity. Activity refers to the data collected from smart devices on calories burned during physical activities such as walking and running. In general, activity iscalculated as the activities accumulated from morning to the instant the user falls asleep. To beconsistent with general users’ everyday life patterns, in this study the data on calories burned during the day were collected at 12:00, 18:00, and 24:00. As the definition of availability, which is applicable to actual circumstances, ease of collection and frequency of occurrence were considered in the selection of the variables. Ease of collection confirms whether or not avariable is collectable by general users. For frequency of occurrence, variables indicating anoteworthy decrease in the input value, such as gender, age, and height, were excluded from neural network learning. In this study, the variables were selected by comparing the ease of collection and frequency of occurrence of an activity and the output variable with those of other variables. As a result, the collected variables were as follows: activity, weather, temperature, mean daily temperature, humidity, UV, fine dust, asthma and lung diseaseprobability index, skin disease probability index, cadence, travel distance, mean HR, and sleephours.

3.2 Preprocessing and Classification According to Data Characteristics

Mobile health data are preprocessed using data integration, data cleaning, data conversion, and data reduction. Data integration is a process in which users are designated as identifiers, unnecessary data are removed from diverse data sources, and data are integrated based on users. In this process, data are integrated into a group of consistent units using a map-reducemethod to repeat the mapping and reduction of heterogeneous data. Given that heterogeneous big data are diversely utilized, medical institutions and hospitals are attempting to use diversemethods and programs for collecting and utilizing medical data [24,25]. Lifelogs arecontinuously accumulated over time, and a common data model is used to manage users & rsquo; lifelogs. In general, data access methods consume substantial time and costs owing to therequired repeated collection and analysis processes and context-based knowledge acquisition and refinement. However, a common data model can be used to process data in real-time without any loss and to construct a data network quickly. Data cleaning is a method used tointegrate one sequence into one transaction. Given that overlapping health data can besimultaneously collected via a smartwatch, health band, and smartphone, one eigenvalue is stored and the overlapping data are deleted. Data conversion is a process in which the value of a property is separated from non-structured data containing a number of properties diversely generated and collected according to wearable devices. Data reduction is a method used todisplay diversely expressed health data in an integer format for the purpose of calculation and analysis. For instance, a code can be assigned to data, such as allergy, weather, and familymedical history, so that they can be displayed in an integer format.

To construct a neural network, the variables were classified according to theircharacteristics identified during preprocessing. Characteristics refer to data properties, such as data type and collection method. Codes were assigned to data by means of data classification. In the initial phase, the variables were classified according to their data type. Based on their data type, categorical variables were classified into weather (c01), asthma and lung diseaseprobability index (c02), and skin disease probability index (c03). Variables other than the categorical variables, such as document type variables, coordinate type variables, and compound type variables, were excluded, because they are not suitable for neural networklearning. Integer-type variables were classified according to their collection method. In thesecond phase, the variables were classified according to their collection method. The variablescollected through the Korea Meteorological Administration were classified into temperature (w01), mean daily temperature (w02), humidity (w03), UV (w04), fine dust (w05), andultra-fine dust (w06). Variables collected through users’ devices were classified into cadence (d01), travel distance (d02), mean HR (d03), sleep hours (d04), and activity (d05). Table 2 shows the data prior to processing, where TID denotes the transaction ID and seq. denotes the sequence. For TID, 18060711256 signifies that User u1256 is the 1^st sequence on 7^th June 2018,18060722334 signifies that User u1256 is the 2^nd sequence on 7^th June 2018, 18060712334 signifies that User u2334 is the 1^st sequence on 7^th June 2018, and so on. Data collected from 26 users by means of smart bands and smart phone applications for 110 days from March 2, 2018 to June 19, 2018 were used.

Table 2. Data prior to processing

3.3 Recurrent Neural Network-based Time Series Data

Mobile health data are specific given that that they are consecutively collected in time series astime progresses. In everyday life health data, sequentially earlier data (s-1) frequently influence the subsequent data. Valuable information can be provided to users by predicting the change in mobile health data according to time series. In this study, the RNN described in [19]was used to learn mobile health data and predict users’ context information changes. The RNNuses actual data learning up to a certain sequence to predict the s + 1, s + 2, s + 3, …, s + nsequence. Fig. 4 shows the RNN-based prediction of time series data [26-29].

Fig. 4. Recurrent neural network-based prediction of time series data [19]

In an RNN, data runs from the input layer to the hidden layer, and the obtained results areentered back into the input layer. In an RNN, the previous status influences the subsequentstatus according to the data sequence. An RNN operates in the feedforward neural network (FNN)-based structure, and the data flow operates in the order of input layer-hiddenlayer-output layer. RNN models can be divided into fully connected RNN (FRNN), recurrent multilayer perceptron (RMLP), and simple recurrent network (SRN) models according to their feedback method. For an RNN, the pattern of the arrangement appearing according tosequence is used for calculating the arrangement of the next sequence [26-29]. Fig. 5 shows an RNN module, where x_s represents input and ys and hs represent new outputs according tonode function. f_w represents the node function. h_s-1 represents the recurring old output. Newinput can be calculated by substituting the new state (x_s) and the old state (h_s-1) for node function.

Fig. 5. RNN module [26-29].

3.4 Long Short-term Memory Recurrent Neural Network Modeling for Prediction

In the prediction of mobile health data, the old state influences the new state because of theirtime series characteristics. To consider both states, mobile health data are used in the RNNstructure. General RNNs are likely to cause the long-term dependency problem followed by the vanishing gradient problem. Mobile health data generate a large number of sequences astime progresses and are constructed based on LSTM [30-33], where the long-term dependency problem is ameliorated. An LSTM model uses a gate mechanism to resolve the vanishing gradient problem. It is constructed based on a number of gate-connected cells, which can be used to read/write information. Fig. 6 shows an LSTM neural network module.

Fig. 6. Long short-term memory neural network module [30-33]

In Fig. 6, i represents the input gate, f represents the forget gate, o represents the output gate, g represents the candidate hidden state, c_t represents the unit’s internal memory, and s_trepresents the hidden state. i, f, and o, serving as gates, are converted into a value between 0 and 1 using the sigmoid function to conduct an element-wise calculation for each element of the input vector. The input gate adjusts the information transfer rate to confirm the extent to which the current input value is transferred. The forget gate adjusts whether the input value is long-term or short-term memory. The output gate adjusts the information transfer rate to confirm the extent to which the state information is output. The candidate hidden state (g) displays the candidate output value calculated using the activation function tanh and the pre-existing state. Instead of providing the candidate output value itself as the output, LSTM provides only part of it as the output through gate calculation. c_t is the sum of c_t-1 stored in the previous memory, the value calculated for each element of the hidden gate value, and the value calculated for each element of the candidate state and input gate. It represents the combination of the previous memory and current input. s_t calculates the final output value by calculating the c_t value and each element of the output gate.

Mobile health data have a direct/indirect association established in the collection ordeduction process depending on the variables, which needs to be considered. Associations can be divided into direct relations and indirect relations. A direct relation is a relation with a BMI or daily temperature difference, which are variables calculated from other variables. Anindirect relation is exemplified by the relation between travel distance and weather, which arevariables that change as the user’s location changes. An LSTM network was constructed based on the characteristics of mobile health data. Fig. 7 shows the DC-LSTM RNN model. The proposed model is a complex neural network, where the classified input variables c{c01~c03}, w{w01~w06}, and d{d01~d05} are used to construct the many-to-many LSTM for each classified variable group, the output values are totaled by fully connecting the feed forward fusion layer, and the totaled data are again set as the input to LSTM to obtain the final userphysical activity prediction.

Fig. 7. Data characteristics-based long short-term memory recurrent neural network model

4. Performance Evaluation and Results

The physical activity prediction model uses keras on RStudio for learning. The s of twareenvironment comprises Windows 10 Pro, R 3.5.0 [34], RStudio 1.1.453 [35], Keras 2.1.6 [36 ], and TensorFlow 1.8.0 [37]. The hardware comprises Intel i5-4690 CPU 3.50 GHz (4 CPUs), 16384 MB RAM, and GeForce GTX 970. RStudio is an integrated development environment for R and provides diverse functions through its packages [38]. Keras is a library that can beeasily used based on TensorFlow and is capable of using a sequential model to actualizemulti-layers [36]. In this study, the R packages used were timetk, cowplot, recipes, rsample, yardstick, and keras. The timetk package is a time series processing tool. The cowplot package is a data visualization tool. The recipes package is a preprocessing tool for design matrices. The Rsample package is a general resampling infrastructure tool. The yardstick package is atool for attaining accuracy. The keras package is an R-based neural network API and is amodeling tool that supports both circuit-based and recurrent networks [34,35,38].

The neural network learning using R was operated in the following order: package call, datainput, separation between learning data and test data, data preprocessing, LSTM modeling, model learning, prediction, and model evaluation. In R, the package call was made in a library (keras) format. The data were divided into 80% training data and 20% testing data, and the key values assigned to the two data groups were “train” and “test,” respectively. The sequential data for 22 days following the old state were used as the testing data in order to consider their time series characteristics. The sequential data for the remaining 88 days was used as training data. The LSTM algorithm requires that input data be centrally aligned and scaled and uses therecipes package for data preprocessing. It uses the recipes package’s step_sqrt for dataconversion and singular value reduction. The proposed LSTM model was built in amany-to-one structure [30-33] that is capable of predicting one activity from various inputs. The input value had 14 columns, which were divided into a group of 3, a group of 6, and agroup of 5 according to the data characteristics. Accordingly, three LSTMs were constructed having sequence lengths of 3, 6, and 5, respectively, and their output was modeled as an LSTM with a sequence length of 3. In the LSTM model, tanh was used for the activation function. The constructed LSTM RNN model used the data having the key value “train” for learning. When the learning process was complete, the data having the key value “test” were used to predict and evaluate user physical activity.

User physical activities are data with time series characteristics and the most representative time series prediction methods are statistical regression methods and neural network methods. The statistical regression methods used were the autoregressive (AR) model and autoregressive integrated moving average (ARIMA) model, which are modified versions of the moving average (MA) model [39]. These statistical models use past data for tendency prediction. The neural network methods include multi-layer perceptron (MLP), convolutional neural networks (CNNs), and RNNs. In this study, an evaluation was conducted to measurethe performance of the activity prediction method using the ARIMA model, CNN, RNN, and the proposed DC-LSTM RNN. The activity of a user was predicted using each method and the difference between the actual value and the predicted value was evaluated as an error. Thetesting data were collected from 26 users for 22 days. The activity from March 2 to March 13 was predicted according to the input variables c{c01~c03}, w{w01~w06}, and d{d01~d05} of the user collected on March 2, 2018 and compared with the actual value. A total of 572 errors, that is, 22 errors for each user, were evaluated using each method and the root mean squareerror (RMSE) was calculated as the average value according to the prediction time (+1 day, + 2days, +3 days, …, +11 days). Fig. 8 shows the user physical activity prediction according to the time series prediction method.

Fig. 8. User physical activity prediction based on the time series prediction method

Table 3 shows the RMSE of the user physical activity prediction that was yielded using the ARIMA model, CNN, RNN, and DC-LSTM. A sequence refers to a sequential flow, such astime flow. s represents the current time point and s + n represents the current time point plusthe next n time point. For example, Sequence 0 of DC-LSTM indicates the activity of a user at the current time and Sequence 1 indicates the activity of a user one day after that time, which is predicted at the current time. The user physical activity prediction results achieved using the prediction methods were evaluated using the RMSE. The evaluation results showed that the activity prediction methods incurred greater errors because more sequences were implemented. Overall, the proposed DC-LSTM showed an outstanding RMSE as compared to other methods despite the fact that more sequences were implemented.

Table 3. Root mean square error of user physical activity prediction using ARIMA, CNN, RNN, and DC-LSTM 

5. Conclusion

In this paper, a prediction model of user physical activity was proposed that uses the DC-LSTM RNN. The proposed method is an LSTM RNN constructed by selecting, collecting, preprocessing, and classifying data according to their characteristics. Various methods exist for predicting or calculating the activity level of a user; however, such methods most frequently use step count and traveling distance data. These methods predict the activity for the current day or future days using the data for activity performed in the previous week ormonth. If low activity due to unfavorable weather or the surrounding context is measured, awarning is given to the user. These methods suffer the problem that outdoor activity may berecommended through a comparison with the activity in the past based on the date, even if the current surrounding context is not favorable for such activity. In this study, such errors wereminimized by means of the learning of various variables to obtain an activity prediction thatreflects the surrounding context and personal context. In the case of mobile health data, the old state affects the new state because of the data’s time series characteristics. In order to considerthis, an LSTM model that allows the use of long-term memory was configured. Based on datacharacteristics, data that can be conveniently collected and utilized by universal users wereselected. The selected data were collected from wearable devices, the Korea Meteorological Administration, and meteorological information applications. The collected data wereconstructed into a transaction by preprocessing, including data integration, data cleaning, dataconversion, and data reduction. Data were classified according to their types and collection methods revealed during preprocessing and were set as the variables of a neural network. Data classification was used to construct an LSTM RNN and the learning process was performed. The RMSE of the activity prediction method was evaluated using the ARIMA model, CNN, RNN, as well as the DC-LSTM RNN. The evaluation results showed that the mean RMSEvalue of the proposed DC-LSTM RNN model, 0.616, was the best value.

In the future, we plan to conduct in-depth research on differentiated personalized smarthealth services by expanding the scope of data collection and increasing the number of target prediction variables. In addition, the development of an application is planned for providinguser-based smart health services. This application would overcome the problems of existing standardized health services and provide information that is more valuable to users.