Design and Implementation of Health and Risk Level Assessment for Socially Disadvantaged Patients

Abstract

The platform developed in this study provides users with mental and physical health status monitoring services. Recently, it has been difficult for people with various respiratory diseases and mental health conditions to receive regular examinations/observation from specialized medical institutions. Therefore, in this study, because the need for the continuous health status measurement/monitoring of subjects was recognized, a device for constantly monitoring trends and changes in blood oxygen saturation was developed to ensure safety. Oxygen saturation below 90 indicates an emergency. In addition, we conducted a hypothesis test for analyzing the relationship between sensor data and depression and anxiety. The sensor data used in the research were location, gyroscope metrics, touch, data usage, and sleep mode from a smartphone. Correlation analysis and multiple regression analysis were conducted to analyze the relationship between the collected sensor data and depression and anxiety. As a result of the research, it was found that among the general characteristics, income, location, sleep duration, and gyroscope were found to have a negative effect on depression and anxiety, whereas touch was shown to have a positive (+) effect. It was confirmed that sensor data measured over runtime can be used as information that can explain users’ depression and anxiety.

1. Introduction

Over the past 50 years, humanity has undergone rapid changes throughout society due to industrialization, westernization, and urbanization. There has been a decrease in the availability of jobs for the elderly and an increase in the unemployment rate due to population aging. Recently, COVID-19 has led to an increase in mental health problems such as mental illness, depression, and suicide due to the social isolation and the social disconnection of vulnerable individuals. This study developed a healthcare analysis technology to assess individual mental and physical health conditions by utilizing various sensor information acquired from a user’s smartphone, including oxygen saturation and accelerometer, gyroscope, touch screen, and sleep mode information. We continuously monitored data, visualized the information from each sensor, and synchronously accumulated the data on a cloud computing platform. We analyzed the user’s health condition/risk level by analyzing the data continuously collected from the user. In addition, the system alarmed cooperative organizations in the event of an emergency. The service developed in this study is expected to be in high demand among vulnerable users who have difficulty accessing help in their environment, such as elderly people living alone or self-employed individuals during times when many respiratory infectious diseases occur. We developed a biometric information monitoring system that enables continuous monitoring by transmitting the information measured via a sensor to a server. The IoT device developed in this study is equipped with computing power and is closely linked with the cloud server for raw data input from various sensors due to the need for algorithm execution/version management/emergency response.

Notably, Republic of Korea ranks first in the rate of increase in suicide among OECD member countries, and the rate of suicide attempts is high in the high-risk group for mental health. In the past, the burden of mental illness and mental health problems was shared among the family or community, but recently, it has become difficult to share this burden due to changes in the social structure such as nuclear families and low marriage and high divorce rates. This leaves people easily exposed to problems. According to the “2021 COVID-19 National Mental Health Survey” conducted by the Ministry of Health and Welfare, in March 2020, the early stage of the COVID-19 outbreak, the proportion of people thinking about suicide had increased by more than 40%, and more than 20% of respondents were at risk of depression. In addition, according to the study conducted in [1], the rate of respondents experiencing depression or anxiety symptoms reached 53% during the COVID-19 pandemic. According to the WHO, about 10–15% of those who have experienced depression eventually attempt suicide or die by suicide. Depression in old age is a serious mental illness and is known to affect physical diseases such as anxiety disorders, sleep disorders, and dementia. Individuals belonging to the clinical high-risk group experience changes in thoughts, behaviors, and their bodies such as depression and anxiety, insomnia, poor concentration, poor memory, mood instability, social isolation, and physical symptoms. These mental health problems are mainly caused by stress, and in some cases, they will heal on their own over time. However, due to the possibility of developing a psychotic disorder, the subject is deemed to belong to a high-risk group and is regarded as being in an “at-risk state”, i.e., a period of heightened vulnerability. If an individual manages or receives treatment via appropriate methods during this “dangerous state” period, the pain caused by mental illness can be reduced, deterioration of health can be prevented, and good mental health can be restored.

In the field of mental health, researchers have been trying to find a correlation between smartphone use and mental health problems for a long time. According to Agnew’s (1992) general tension theory, stress was explained as one of the main causes of deviant behavior. It was explained that when humans are placed in a stressful situation, emotional problems such as depression and anxiety symptoms occur due to stress, and deviant behavior can be performed in an attempt to cope with the stressful situation and emotional problems. One of these deviant behaviors is smartphone addiction, including social network use, video viewing, and gaming, and studies are being conducted on the relationship between smartphone overdependence and depression. According to [2], it was confirmed that the higher the smartphone addiction, the greater the lack of sleep that occurs, and this lack of sleep is known to have a significant relationship with depression. According to a study carried out by [3], the GPS sensor information of a smartphone is related to the severity of depression. As such, there is a trend towards actively utilizing smart device sensor information to solve mental health problems. However, studies that predict or diagnose mental health problems using smartphone sensor data often rely on machine learning, which is insufficient to explain the relationship between smart device sensor information and mental health problems.

Accordingly, in this study, a platform was built that is capable of detecting depression and emergency situations through time series data measured using smartphone sensor information. As mental health problems intensify, various studies are being conducted to detect and improve such problems at an early stage. Mental health problem prediction and management systems using sensor information from smart devices are one of these approaches. Smartphones or wearable smart devices have become popular enough for most people to carry at least one. In the case of a smartphone, since it is carried by a user almost always, it can be used to track the user’s behavior in real time. Therefore, smartphone sensor information is being utilized in various fields. We analyzed the correlation with mental health for a total of five types of sensor data collected from smartphones: oxygen saturation, accelerometer, gyroscope, touch screen, and sleep mode information. This research proposed a smart method for predicting and diagnosing mental health issues by correlating depression and anxiety with data collected from smartphone sensors.

2. Related Work

Depression is a chronic, recurrent disease that generally exhibits symptoms such as depressed mood, loss of interest and pleasure, and increased fatigue. Depression is accompanied by sadness, loss of interest or pleasure, guilt, low self-esteem, disturbed sleep or appetite, fatigue, and difficulty in concentrating. Since depression includes cases with mild symptoms, a more accurate expression is the major depression disorder, as defined by the American Psychiatric Association (APA) [4]. In this study, the “Patient Health Questionnaire-9”, a tool developed by [5] and standardized by [6], was used to measure depression [4,5,6].

Anxiety is an emotional state that many people feel very commonly in modern society, where social relationships have become more complex and demands have increased. Therefore, it is a research topic in which psychologists studying psychological maladjustment and psychopathology have consistently shown great interest [7]. In the DSM-IV, anxiety disorders exhibit a complex form of various sub-symptoms and disorders and are subdivided according to the pattern in which pathological anxiety appears or the object and situation in which anxiety is felt [4]. In this study, “Generalized Anxiety Disorder 7”, developed by [8], was used to measure anxiety.

In Ref. [3], a study was conducted to build a watch-type device and server system to monitor and manage the health information of dementia patients in real time. We built a server system that collects physical activity information through a watch-type device and checks the patient’s health information based on it. It was found that a relationship exists between smartphone sensor information and depression; it was confirmed through a study where GPS information collected from smartphones was related to the severity of depression. The researchers built a smart jewelry system that can be worn like an accessory to collect body information and built a system to manage depression based on this. The researchers collected heart rate, exercise time, exercise consumption, sleep time, and illumination information through the smart jewelry, and based on the collected information, a system was built to monitor depression. In addition, studies on depression relatedness based on sensor information such as smartphones and smart watches have been conducted, but research limitations remain due to the irregularity of sensor data, lack of collected variables, and difficulty in real-time analysis [9].

According to the research carried out in [2], community models for mental health promotion have been established. Suwon city established an approach by developing programs for each target group and operating intensive case management services through the ACT program. Suwon city’s community model is characterized by its ability to be accessed by class through facilities such as the Suwon City Welfare Center. Hwaseong city carries out a support project for long-term hospitalized mental patients to return to and settle down in the community and supports various educational programs. The community model of Hwaseong city is characterized by the strengthening of public medical service infrastructure and policies centered on suicide prevention. Jeonju city operates a weekly rehabilitation program and provides counseling and treatment services to citizens who find it difficult to take care of their mental health due to social and personal conditions. The community model used in Jeonju city is characterized by promoting detailed projects centered on suicide prevention projects and children and youth projects. In [10], a machine learning model was utilized to detect depression in elderly people based on data extracted from personal wearable devices. The researchers built a model based on supervised learning, and three-axis accelerometer, temperature, BVP (blood volume pulse), HR (heart rate), and EDA (electrodermal activity) data were used as learning qualities. As a result of the experiment, a maximum F1 score of 0.971 was measured. Ref. [11] analyzed the correlation between depressive symptoms and biometric information collected through a smart band, and a study was conducted on a model that could predict the degree of depression. Depression scale (PHQ-9), sleep time, sleep start time, and wake-up time were used as the learning qualities, and a model was built using SVM (Support Vector Machine) and linear regression algorithms [12,13]. As a result of the experiment, a model was constructed in which the depression score was high due to irregular and short sleep times. Ref. [14] conducted research into a depression prediction model using the naive Bayes model. A total of 23 types of information, including sex, age, income, education, number of family members, marital status, stress perception, and education, were used as the learning qualities, and the model construction result showed an F-Measure of 0.974. Ref. [15] investigated a model that analyzes users’ emotions based on images posted on Instagram. After analyzing the images using CNN-based models to extract emotions, a model was built to predict depression based on the extracted emotions. In [16], research was conducted using a method for predicting depression scores on the PHQ-9 scale using survey data provided by the National Health and Nutrition Examination Survey (NHANES) [7,8]. The authors built a multi-class classification model based on the DNN algorithm, demonstrating an F1 score of 0.6, and in binary class classification, the F1 score was measured as 0.71 using a decision tree model. VAE is an algorithm that operates with the goal of extracting features that can sufficiently explain the input value X as one of the autoencoders and generating completely new data similar to X from the extracted features. Since VAE is based on a probabilistic model, flexible calculation is possible, but since it does not directly implement density, its performance may be lower than that of a model that does [17]. DAGMM (Deep Autoencoding Gaussian Mixture Model) is an algorithm studied based on GMM (Gaussian Mixture Model) and is a model in which Gaussian distribution is mixed. DAGMM uses an autoencoder to maintain the characteristics of the reduced dimension and restoration error, so that important information regarding the input value can be maintained even in the low dimension. Also, as a structure suitable for end-to-end learning, DAGMM does not fall into local optima [18]. COPOD is an unsupervised learning-based algorithm suitable for detecting outliers. COPOD analyzes the dependent structure between random variables and judges it as an outlier depending on whether it exceeds the threshold of the predicted outlier score. Therefore, it is an algorithm suitable for research that detects special situation changes [19]. LGBM is a tree-based learning algorithm with a gradient-boosting framework. In general algorithms, the tree expands horizontally, whereas in LGBM, the tree expands vertically, that is, LGBM is a leaf-wise form. To expand a tree, a leaf with maximum delta loss is selected, and when extending the same leaf, the leaf-wise algorithm is characterized by being able to reduce more loss than the level-wise algorithm. There are many studies that have analyzed time series data through LGBM; it has been applied in the prediction of the amount of wind power generated and in the prediction of missing values of air pollution information data.

3. Platform Design and Implementation

3.1. Server–Client Architecture for Patient-Specific Risk Assessment

In this research, data were collected for about two months to build a risk assessment model. We collected data from at-risk and non-risk subjects. For data collection, we developed and used a smartphone sensor collection application. In addition, we collected and labeled the diagnosis results by conducting expert consultations with application users. We collected the distance traveled, the number of steps, the number of screen touches, data usage, sleep time, and oxygen saturation data through a device developed separately from the application used in this study. This application utilizes smartphone sensor data to collect information such as the physical activity, social activity, and emotional state of the user as shown in Figure 1.

The collected data are largely divided into user information, sensor data, diagnosis information, and depression diagnosis results. User information includes member account number, age, sex, name, region, occupation, income, and household member information. Sensor data are information collected from a smartphone every 15 min, and includes the date and time of data creation, distance traveled (km), number of steps, number of screen touches (number of times), data usage (Mb), and sleep time (time when mobile phone is not used). Diagnosis information includes the anxiety scale (Generalized Anxiety Disorder 7) and the depression screening tool (Patient Health Questionnaire-9) as psychological test results. Sensor data were used as the learning data, and the values of “depressed” and “normal” were used for the label based on the psychological test results and the expert’s opinion.

Sensor data were collected at 15 min intervals. The reason for this is that frequent data taken less than 15 min apart are more than necessary to determine depression, and difficulties arise in terms of system resource management. On the other hand, it was judged that it would be difficult to cope with emergency situations if data were collected over a longer period of more than 15 min. If the communication state is unstable or the application is deactivated, it is handled by a strategy designed to deal with missing data. In this research, we adopted a supplementary method instead of a method of removing missing values in order to use temporal information meaningfully. Missing values were obtained by substituting the mean (μ) and standard deviation (σ) of values measured at the same time on different days so that time information could be learned meaningfully. When comparing the method of supplementing missing values used in this study and the method using the average of values immediately before/after the missing values, the method proposed in this research was able to compensate for the missing values more precisely. In the formula, s used a random real value between 0 and 0.01.

Figure 2 shows the smartphone application storyboard developed in our research. In designing the information architecture, we focused on explicitly categorizing the menus and avoiding excessive depth. Our app consists of about 20 pages with various functions such as an intro page, login page, measurement page, main page, ID/PW search page, information sharing page, and analysis result inquiry page.

In addition, the characteristics of time were created and used as variables to learn the characteristics of the time series data collected in real time. Time information was divided into hours and minutes. To secure independence among the variables, we created numerical variables using OneHotEncoding as a categorical variable and StandardScaler, a standardized method that is relatively less affected by outliers. In the case of VAE, we used 70% of the data as training data, 20% as validation data, and the remaining 10% as test data. VAE’s training and validation data only used the data from general subjects. For other models, we divided the ratio of training data, validation data, and test data as 7:2:1, respectively. The ratio of the data from depressed subjects and normal subjects was equalized.

In terms of back-end, we exploited the Springboot-based server technique to perform dynamic web development and data management, adopting a model-view-data design pattern. The Springboot is a Spring sub-project that avoids complicated XML configuration, enabling a simpler environment configuration. Based on the Spring framework, our server architecture provides various core services such as inversion of control (IoC), dependency injection (DI), aspect-oriented programming (AOP), and java object persistence API (JPA).

Figure 3 shows the overall system architecture of our platform, consisting of four components. The first component is a user-level application, which utilizes the oxygen saturation, accelerometer, gyroscope, touch screen, and sleep mode information from the smartphone to acquire sensor information time series data and transmit them to the server using HTTP communication. The second component is the server. Sensitive data such as the user’s personal information and password must be secured quite thoroughly, so most applications create a web application server (WAS) to store and process data. The WAS is a server that provides dynamic content according to client requests, such as various algorithms, business logic, and DB inquiries, and sends a response for data analysis and visualization to the smartphone along with a response code. The third component is the database management system (DBMS), which is a software (MySQL version 8.0.28) that manages and operates databases. The fourth component is the Google FCM service. Firebase cloud messaging (FCM) is a cross-platform messaging solution that reliably delivers messages for free and allows you to distribute messages to your app in three ways: single devices, groups of devices, or devices subscribed to a topic. If you create a notice in the admin page, you can receive push notifications in the user-side app.

Figure 4 shows the navigation graph, which shows all navigation-related information. This graph shows all the flows that the user can proceed with in the application, and the fragments in the application can be depicted briefly. Movement in the navigation graph is performed by the navigation controller. The navigation controller is an object that the navigation host uses to manage application navigation. The navigation controller is responsible for steering the transition of target content on the navigation host as the user navigates through the app. To use the navigation controller in fragments, we imported it through the findNavController() function.

The default navigation host implementation, NavHostFragment, navigates fragments as defined in the navigation graph. We specified NavHostFragment in the android:name item of XML. With this, we specified that a particular fragment is a NavHostFragment. The NavHostFragment is a container, and a fragment can be regarded as being put into a container according to the navigation graph defined by the developer.

The execution of the application largely takes place according to a two-layer architecture, as depicted in Figure 5. The first layer is the foreground layer including identity, and stores sensor data by classifying users. From then on, the data collection phase continues in the background layer with the appropriate user authority. JWT, a token-based authentication system for authenticating and identifying users, is introduced. Using JWT, user data can be exchanged in a stateless environment such as RESTful. When a user attempts to log in, the server performs user authentication from the DB. If authentication is successful, an Access Token and Refresh Token are issued to the user. The user stores the issued token in an Android storage location called shared preferences. The application developed in this study allows sensor data to be continuously collected even if the user executes another application in the foreground. Therefore, even if the user does not interact with the app, it operates as a background service using the Android service to periodically transmit sensor data. Nevertheless, data are temporarily stored in the internal storage in order to cache the sensor data in case transmission failure occurs. Afterwards, when the smartphone becomes ready for transmission again, the temporarily stored sensor data are transmitted at once.

In this research, user authentication was implemented using the Spring Security function of Springboot. In particular, JSON Web Token (JWT) was utilized. Based on the ID/password user information, a check is made to confirm whether the user is registered, and if authentication is successful, the user’s principal and credential information is stored in Authentication, which is stored in SecurityContext. When authentication is complete, it is stored in HttpSession, so it can be used throughout the application afterwards.

3.2. Models and Parameters

In this research, variational autoencoders (VAE), deep autoencoding Gaussian mixture model (DAGMM), and light gradient boosting machine (LGBM) algorithms were used to construct a depression detection model. Variational autoencoders (VAE) are a type of autoencoder whose encoding distribution is regularized during the training in order to ensure that its latent space has good properties, allowing us to generate some new data. Moreover, the term “variational” comes from the close relation there is between the regularization and the variational inference method in statistics. The deep autoencoding Gaussian mixture model (DAGMM) consists of two major components: a compression network and an estimation network. DAGMM works as follows: (1) the compression network performs dimensionality reduction for input samples using a deep autoencoder, prepares their low-dimensional representations from both the reduced space and the reconstruction error features, and feeds the representations to the subsequent estimation network; and (2) the estimation network takes the feed, and predicts the likelihood/energy in the framework of the Gaussian Mixture Model (GMM). Light gradient boosting machine (LGBM) is a gradient boosting framework that uses tree-based learning algorithms. It is designed to be distributed and efficient with the following advantages: faster training speed and higher efficiency; lower memory usage; better accuracy; supports parallel, distributed GPU learning; and capable of handling large-scale data.

In order to properly configure the parameters of these classification models, hyperparameters were applied to each algorithm and are shown in

Table 1. Parameters for VAE.

Parameters	Configuration
hidden_neurons	4~512 (by power of 2)
hidden_activation	tanh, sigmoid
output_activation	X
loss_function	X
learning_rate	0.01~0.1
optimizer	SGD
epochs	30~200
batch_size	8~1024 (by power of 2)
dropout_rate	0.2~0.5
validation_size	0.1~0.2
contamination	Depending on data

,

Table 2. Parameters for DAGMM.

Parameters	Configuration
hidden_neurons	4~512 (by power of 2)
n_gmm	1~21
dropout_rate	0.2~0.5
loss_function	X
epochs	30~200
batch_size	8~1024 (by power of 2)
validation_size	0.1~0.2
learning_rate	0.0001~0.1

and

Table 3. Parameters for ECOD and COPOD.

Parameters	Configuration
n_estimators	100
num_leaves	31
min_child_samples	20
learning_rate	0.1
colsample_bytree	1
reg_alpha	0
reg_lambda	0

.

In this research, we trained a depression detection model using depressed subject data to build an emergency detection model. As for the detection of emergency situations, the outliers among the depression detection results were judged as emergency situations. Interquartile Range (IQR), Isolation Forest, and Density-Based Spatial Clustering are the methods used for judging outliers. In this study, an IQR-based anomaly score was calculated to judge outliers based on the normality of the data. For the labeling of emergency situations, a model with an F1-Score of 0.8 or higher was used for the classification of depression, and if the anomaly score was in the 0.01% category, it was labeled as an emergency situation. An anomaly score is a value that judges abnormality or normality, and the higher the number, the more critical the situation.

3.3. Personal Information Privacy and Research Ethics

The services are application components for background tasks. They do not provide a UI and continue to run in the background even when the user switches to another application. Users without a UI have problems recognizing the type of service running and the resource being used, so it is displayed as a notification in the status bar. The service process can read data, such as oxygen saturation, accelerometer, gyroscope, touch screen, and sleep mode information from the smartphone. For this purpose, we registered the SensorEventListener to the SensorManager in order to receive information. Through this, we obtained three values for the X, Y, and Z axes from the corresponding sensor. In this research, we collected the user location, distance traveled, sleep, and Internet usage information of the users. To utilize such personal information, we needed to request access rights from the users in advance. Starting with Android API 23, permissions classified as dangerous permissions were changed so that permission was granted every time the application was executed, not at the time of installation. For example, functions such as location, camera, microphone, contacts, phone calls, text messages, schedules, and sensors are classified as risk rights.

Since this is personal information belonging to the research participants, it needs to be managed/protected very carefully from the viewpoint of research ethics. To this end, we conducted the research after obtaining Institutional Review Board (IRB) approval with control number 2022-1001-003, as shown in Figure 6 and also in Appendix A. The IRB is an independent consensus-setting organization that deliberates on the ethical and scientific aspects of human subject research for the rights and safety of subjects participating in clinical research. In accordance with the Bioethics and Safety Act, the IRB not only reviews research plans, but also investigates and supervises research progress and results, and conducts research ethics education for researchers and related persons.

Obtaining permission to access location information allows apps to share or receive location information only once or for a prespecified period. In particular, the location information access authority is more specifically divided into the authority to access location information in the foreground and the authority to access background location information. That is, even when the application is not running in the foreground, a function allows the user and location information to be received. In this study, we requested and utilized both foreground and background location information access rights so that the app can utilize user location information even in the background.

3.4. Development of IoT Computing Device That Performs Oxygen Saturation (SpO₂) Measurement and Acquisition

For measuring oxygen saturation (SpO₂), the input signal is differentially amplified, passed through a band pass filter, and adjusted to an appropriate range to be transmitted to the IoT cloud server through a thermal printer or communication network. As shown in Figure 7, the input signal is differentially amplified and passed through a band pass filter. We adjusted the signal to an appropriate range and transmitted the data to our server in the cloud through a communication network. We calculated the SpO₂ value by measuring the amount of light passing through a finger and the amount of light reflected from a finger. Using the characteristic that the amount of current varies according to the amount of light irradiated to the photo transistor or photo diode, the change in blood flow is measured.

We calculated the Steinhart–Hart model coefficients A, B, and C based on the curve of the measured temperature resistance pair. We obtained a constant value as the resistance value corresponding to the NTC sensor and calculated the temperature value using the formula below. Finally, the calculation that was performed is as follows:

$$\frac{1}{T}=A+B\ln \left(R\right)+C{\left[\ln \left(R\right)\right]}^3$$

A method of measuring the amount of light passing through a finger and a method of measuring the amount of light reflected from a finger were used. Using the characteristic that the amount of current varies according to the amount of light irradiated to the photo transistor or photo diode, the change in blood flow was measured. Figure 8 shows our implementation of SpO₂ measurement board.

3.5. Development of Smartphone Application

The Android background service collects oxygen saturation, accelerometer, gyroscope, touch screen, and sleep mode data from the device, stores them in a list collection, and sends them to the server periodically. To classify the sensor data, the field_name attribute is declared in the user_activity entity. Also, the type is declared as float to express the field_value attribute as a real number. Each table in the database is implemented as a JPA Entity.

Table 4. Libraries and modules for system implementation.

	Modules	Version
Package Management	gradle	7.4.1
Compiler	OpenJDK	11.0.13
Server Framework	Springboot	2.7.0
JWT Library	jjwt-api, jjwt-jacson	0.11.5
HTTP API Library	volley	1.2.1
DBMS	Mysql Server	8.0.28
Pub/Sub Framework	Firebase-core, firebase-messaging	21.1.0
JSON Processing	Json-simple	1.1.1
Kafka Client	kafka.js	v1.15.0

shows the versions of the modules used in this research. The runtime in which the server runs is composed of Node.js based on the JavaScript language, and a package suitable for Node.js is configured so that the server can work well.

This system consists of an MVC structure, and the application created with Android Studio is in charge of the view, and the server is in charge of the model and controller. A model represents the elements handled by an application. The ER model of the database corresponding to the model in this project is as follows (Figure 9).

-

Region Entity: To designate the scope of access to notice for each user, the user must belong to a region when registering as a member. One region belongs to several users.

-

Notice Entity: Manages notices that users can see. The business logic separated by the service layer is used to designate whether to send a notice to a separate user or to multiple users belonging to a region.

-

user_account entity: Separately manages the ID and password for each user to access the system. The user transmits the sensor data as well as the JWT token.

-

user_activity entity: Manages various sensor data for each user.

In this research, we designed and implemented an application storyboard. In designing the information architecture, we focused on the explicit categorization of menus and avoidance of excessive depth. Our application consists of 20 pages and various functions such as splash page, login pages, signing-up pages, measurement and analysis pages, monitoring report sharing pages, dashboard pages, information sharing pages, and psychometric feedback pages, as shown in

Table 5. UI layouts and screenshots.

	UI Screenshots
Login pages
Signing-up pages
Measurement and analysis pages
Monitoring report sharing pages
Dashboard pages
Psychometric feedback pages

below.

We continuously measure sensor data and draw graphs every 15 min. The inspection results are provided as a bar graph, and the data collected are represented as a line graph. On the signing-up pages, we make use of where users agree to the terms and conditions. On the measurement and analysis page, there are screens that briefly provide the results measured by the sensor. We can check announcements or notification messages provided from the system through these pages. We are also able to check the health status of users at a distance, and the user can use the information sharing function in the application. In other words, “share my information” can be chosen from the menu provided by the application and the person can be selected in advance to share information with. On the monitoring report sharing page, the analysis results are provided, measured, and analyzed by the sensor. The psychological measurement page allows the subject to conveniently conduct a simple psychological test with a smartphone without the need to visit a hospital.

This research aimed to verify whether the information measured using digital biomarkers is related to the depression and anxiety of the subjects. The general characteristics of data using this research are as shown in

Table 6. General characteristics of research subjects.

Factor	AUC	SE	95% Confidence Interval		p
			Lower	Upper
co_rev (GPS)	0.981	0.024	0.935	1	<0.001
af_av (Touch)	0.944	0.048	0.851	1	0.001
ex_rev (Gyro)	0.963	0.041	0.883	1	<0.001
mo_av (Data)	0.537	0.132	0.278	0.796	0.776
sl_av (Sleep)	0.231	0.11	0.017	0.446	0.039
de_av (PHQ-9)	0.44	0.13	0.186	0.694	0.644

. Prior to the experiment, the research questions that were tested are as follows: (1) the relationship between digital biomarkers and depression, and (2) the relationship between digital biomarkers and anxiety. In this research, the depression and anxiety scores were measured using a depression screening tool and an anxiety scale test for 300 adults living in South Korea. Data from 278 people (85 males and 193 females) who responded faithfully were used for the analysis (

Table 7. General characteristics of research subjects.

Property	Value	Frequency (N)	Ratio (%)
Sex	Man	85	30.6
	Woman	193	69.4
Occupation	Unemployed	175	62.9
	Employed	103	37.1
Age	20–29	15	5.4
	30–39	18	6.5
	40–49	24	8.6
	50–59	75	27.0
	60–69	71	25.5
	70–79	48	17.3
Income	More than 80	27	9.7
	Less than 1000	188	67.6
	Between 1000 and 2000	67	24.1
	Between 2000 and 3000	9	3.2
	Between 3000 and 4000	3	1.1
	More than 4000	11	4.0
Socially disadvantaged or not	Not disadvantaged	117	42.1
	Socially disadvantaged	161	57.9

). In addition, the research subjects installed a sensor measurement application on their smartphone, and collected oxygen saturation, accelerometer, gyroscope, touch screen, and sleep mode information at regular intervals through the application for a period of 2 weeks. This study was conducted after receiving approval from the Public Institutional Bioethics Committee.

As for the sex of the research subjects, there were 85 males (30.6%) and 193 females (69.4%), with about twice as many females as males. As for whether or not they had a job, 175 (62.9%) were unemployed and 103 (37.1%) had a job, indicating that there were more unemployed participants than those with a job. The age of the study subjects was evenly distributed from the 20s to 80s, and among them, 75 (27.0%) were in their 50s, which was the highest proportion. In terms of monthly income, 188 people (67.6%) were the lowest earners with less than KRW 1 million. Finally, 117 people (42.1%) were in the second category, and 161 people (57.9%) were recipients of slightly more income than average.

When the subjects participated in this research, a survey was conducted using the anxiety scale and the depression scale. Based on the results of the survey and counseling, only those who were judged to be eligible to participate in the study were selected as study subjects after confirming once again through guidance and counseling. Subjects who agreed to participate in the study installed an application named “Dr. Forest”, a digital biomarker collection application, on their smartphones to provide location, gyroscope, touch, data usage, and sleep mode information. Statistical analysis was performed by refining the normally collected data without any losses, collected by the smartphone sensors. Digital sensor data were collected through the “Dr. Forest” application, which collects sensor data in real time using information from the smartphone. A total of five digital biomarkers were collected, i.e., location, gyroscope, touch, data usage, and sleep mode. Location is the distance traveled based on the GPS sensor, and gyroscope is the number of steps. Data usage is the network data usage, touch is the number of smartphone touches, and sleep mode relates to sleep time. The sensor data were collected by the application for each subject. The location and gyroscope data were used as information explaining physical activity, touch and data usage were used for smartphone addiction, and sleep mode was used as information explaining users’ sleep state.

3.6. Measurement Tools

Using the conventionally popular screening tools, we were able to judge the users as moderate, weak, severe, and dangerous, if they fell within the score ranges listed below. Learning data were collected to build a depression detection model.

(1)

PHQ-9: Patient Health Questionnaire-9 as a depression screening tool

This is a depression screening tool, standardized and developed by Spitzer et al. (1999). The depression screening tool consists of a total of nine items, and each item is composed of a four-point scale ranging from 0 to 3 points. A total score of 4 or less is none-minimal, 5 to 9 is mild, 10 to 14 or less is moderate, 15 to 19 is moderately severe, and 20 or more is severe. We also used the “Patient Health Questionnaire-9” as the depression screening tool consisting of a total of nine items, and each item is composed of a four-point scale ranging from 0 to 3 points. A total score of 4 or less is none-minimal, 5 to 9 is weak, 10 to 14 is moderate, 15 to 19 is severe, and 20 or more is dangerous.

(2)

GAD-7: Generalized Anxiety Disorder 7 as an Anxiety Scaling tool

GAD-7 is a brief measure for assessing generalized anxiety disorder. An anxiety scale diagnosis was used as an adaptation of the Generalized Anxiety Disorder Assessment (GAD-7) tool developed by Spitzer (2006). It is an easy-to-use, self-administered patient questionnaire that uses seven items to measure or assess the severity of generalized anxiety disorder (GAD). Each item asks the individual to rate the severity of his or her symptoms over the past two weeks. Each item is composed of a four-point scale ranging from 0 to 3 points. A total score of 4 or less is normal, 5 to 9 is mild, 10 to 14 is moderate, and 15 or more is dangerous. The GAD-7 score is calculated by assigning a score of 0, 1, 2, or 3 to the response categories of “not at all”, “several days”, “more than half the days”, and “nearly every day”, respectively, and adding together the scores for the seven questions. Scores of 5, 10, and 15 are taken as the cut-off points for mild, moderate, and severe anxiety, respectively. When used as a screening tool, further evaluation is recommended when the score is 10 or greater.

4. Performance Analysis

This system continuously measures health status every 15 min. For this purpose, we developed an application that is installed on smartphones for target users to measure the sensor data and a device for constantly monitoring trends and changes in blood oxygen saturation to ensure the safety of target users. The measurement and analysis method is described in detail below.

4.1. ROC Analysis

We conducted ROC analysis for the data correlation analysis of the measured data. An ROC curve is a graph showing the performance of a classification model at all classification thresholds. Before going into detail, we set the cut-off value, above which the participants were predicted as positive (they have the disease) and below which were predicted as negative (they do not have the disease). We set the thresholds at 5 and 15, respectively, as shown in Figure 10 and Figure 11.

Figure 10 and Figure 11 are diagrams showing the AUROC. ROC is one of the most important evaluation metrics for checking any classification model’s performance. The area under the curve has a meaningful interpretation for the classification of monitoring data, determining the optimal cut-off value. It illustrates how much the model is capable of distinguishing between classes. Our model predicts whether a patient has a disease or not and is reasonably effective in distinguishing between patients with disease and without.

4.2. SpO₂ Measurement and Notification Accuracy

For oxygen saturation (SpO₂) monitoring to detect cases where the SpO₂ value drops below 0.90% (by the user removing their hand from the measuring device), an experiment was conducted to confirm that the server side receives the notification normally, as shown in Figure 12.

Alerts = Oxygen saturation (SpO₂) < 0.90%

SpO₂ analysis/notification accuracy = whether the alarm was received normally/number of experiments.

Under normal circumstances, blood oxygen saturation (SpO₂) levels do not usually fall below 0.95%. Therefore, to artificially test the dangerous situation notification function, we tested whether the device operated normally by inducing a dangerous situation by removing the user’s hand from the measuring device.

We monitored the SpO₂ value measured by the SpO₂ measuring device by continuously sending it to a server in the cloud. If an emergency situation (a situation where the measured SpO₂ value falls below 0.90) occurs, the system generates an alarm to the user and related personnel.

The user caused an emergency situation by removing his hand from the device, and then checked whether the alarm was activated normally on the smartphone. By repeating these experiments 20 times or more, we confirmed that the number of times the alarm was generated accurately when the experimenter removed his hand from the monitor was more than 95% of the total.

4.3. Sensor Data Measurement, Analysis, and Consideration

In this research, to determine the relationship between digital sensor data and mental health, five types of information (location, gyroscope, touch, data usage, and sleep mode) were selected as independent variables, and anxiety and depression scales were selected as dependent variables. Cronbach’s alpha was calculated to verify the reliability of all variables, and all statistics were verified under the significance level of 0.05. Correlation analysis and multiple linear regression analysis were conducted to analyze the association between the independent variables and the dependent variables set in this study, and all data analysis was performed using SPSS 28.0.

As a result of this research, the descriptive statistics results shown in

Table 8. General characteristics of research subjects.

	Min	Max	Mean	Standard Deviation
age	19	92	59.92	15.455
depression	0	27	7.42	7.334
anxiety	0	21	6.28	6.032
location	0.002	11.886	9.016	2.748
gyroscope	19	6878	5824.5	1241
data usage	4.152	473.54	93.192	86.82
touch	0	1103	254.15	207.52
sleep mode	3.575	13.205	9.601	2.087

could be measured. The age of the study subjects ranged from a minimum of 19 years to a maximum of 92 years, and the average age of the research subjects was measured as 59.92 years. The depression scale scores of the research subjects ranged from a minimum of 0 to a maximum of 27, with an average score of 7.42. The anxiety scale measurement score for the research participants ranged from a minimum of 0 to a maximum of 21, with an average score of 6.28. The location moving distance of the study subjects ranged from a minimum of 0.002 to a maximum of 11.886, with an average of 9.016. The number of steps ranged from a minimum of 19 to a maximum of 6878, with an average of 5824.53. Data usage ranged from a minimum of 4.152 to a maximum of 473.54, and the average was 93.192. Touch ranged from a minimum of 0 to a maximum of 1103, with an average of 254.15. The sleep mode time ranged from a minimum of 3.575 to a maximum of 13.205, with an average of 9.601.

As a result of comparing the averages of depression, anxiety, and measured sensor information by sex (

Table 9. Depression/anxiety level and digital sensor data trend according to sex.

	Sex	Size	Mean	Standard Deviation
depression	man	85	9.6	7.024
	woman	193	6.46	7.279
anxiety	man	85	6.81	5.45
	woman	193	6.04	6.271
location	man	85	8.597	2.769
	woman	193	9.201	2.725
gyroscope	man	85	5361	1499.82
	woman	193	6028	1049.5
data usage	man	85	95.149	89.816
	woman	193	92.33	85.703
touch	man	85	254	206.843
	woman	193	253	208.356
sleep mode	man	85	9.611	2.24
	woman	193	9.597	2.022

), it was found that depression was higher in men (M = 9.60) than in women (M = 6.46). Anxiety was also found to be higher in men (M = 6.81) than in women (M = 6.04). This is interpreted as the fact that men’s mental health was worse than women’s among the study participants. The location data showed that the women’s values (M = 9.201) were higher than the men’s (M = 8.597), and the gyroscope data also showed that the women’s values (M = 6028) were higher than the men’s (M = 5361). This indicates higher levels of physical activity in the women than in the men. The data usage showed that males (M = 95.149) used more data than females (M = 92.330), and the touch data showed that males (M = 254) had higher values than females (M = 253), but the difference was small. This can be interpreted to mean that men use smartphones more frequently than women. The sleep mode values were higher in males (M = 9.611) than females (M = 9.597), but the difference was small. This indicates that the men had slightly better sleeping conditions than the women.

As a result of comparing the averages of depression, anxiety, and our application’s inspection results according to occupation/job status (

Table 10. Depression/anxiety level according to occupational status: average digital sensor data value trend.

	Occupation	Size	Mean	Standard Deviation
depression	Yes	175	9.69	7.553
	No	103	3.56	4.988
anxiety	Yes	175	7.96	6.466
	No	103	3.42	3.792
location	Yes	175	8.61	2.89
	No	103	9.707	2.343
gyroscope	Yes	175	5642	1292.4
	No	103	6134	1085.9
data usage	Yes	175	105.67	93.081
	No	103	71.99	70.52
touch	Yes	175	302	227.66
	No	103	172	133.82
sleep mode	Yes	175	9.114	2.282
	No	103	10.429	1.359

), we found that depression was higher in the unemployed (M = 9.69) than in the employed participants (M = 3.56). Anxiety was also higher in the unemployed (M = 7.96) than the employed participants (M = 3.42). This can be interpreted to mean that the mental health of the unemployed participants was worse than that of the employed participants. In terms of the location, the values for employed participants (M = 9.707) was higher than those of the unemployed participants (M = 8.61), and for the gyroscope, the values were also higher for the employed (M = 6134) than the unemployed participants (M = 5642). This is because employed individuals are more physically active due to social activities such as commuting. The data usage was higher among the unemployed participants (M = 105.66) than for those with jobs (M = 71.99), and touch was also higher for unemployed participants (M = 302) than for those with jobs (M = 172). This is thought to be due to the long time spent using smartphones in the case of unemployed people. The sleep mode was higher for the unemployed (M = 9.114) than for the employed participants (M = 10.429).

As a result of comparing the averages of depression, anxiety, and digital biomarker information according to whether the participant was a recipient of the lowest income bracket or not (

Table 11. Depression and anxiety and average digital biomarkers according to recipients of the lowest income.

	Socially Disadvantaged	Size	Mean	Standard Deviation
depression	No	117	4.37	5.851
	Yes	161	9.64	7.519
anxiety	No	117	4.01	4.421
	Yes	161	7.93	6.506
location	No	117	9.454	2.571
	Yes	161	8.698	2.835
gyroscope	No	117	6056	1153.4
	Yes	161	5655	1278.2
data usage	No	117	72.415	72.13
	Yes	161	108.291	93.45
touch	No	117	193	169.47
	Yes	161	298	221.63
sleep mode	No	117	10.179	1.574
	Yes	161	9.181	2.307

), depression was higher in these recipients (M = 9.64) than in those from other incomes (M = 4.37), and anxiety was also higher in these recipients (M = 4.37) 7.93) than those in receipt of other incomes (M = 4.01). This can be interpreted as the fact that recipients of lower incomes have poorer mental health than the average person. For the location, the values for participants with average income (M = 9.454) were higher than for recipients of lower income (M = 8.698), and for the gyroscope, average income recipients (M = 6056) showed higher values than those on lower incomes (M = 5655). This can be interpreted to mean that the recipients of lower incomes engage in less physical activity than the average person. For data usage, recipients of lower incomes (M = 108.291) were higher than those on higher incomes (M = 72.415), and for the touch, recipients of lower incomes (M = 298) showed higher values than those on other incomes (M = 193). This can be interpreted to mean that there is a higher frequency of smartphone usage among those in lower income brackets than others. The sleep mode was higher in participants with a higher income (M = 10.179) than in those who were more socially disadvantaged (M = 9.18).

Prior to identifying the research question of this study, we analyzed whether the general characteristics of the research subjects were related to their mental health (

Table 12. Correlation between general characteristics and mental health.

	PHQ-9 Depress	GAD-7 Anxiety	Age	Income
PHQ-9 Depress	1	0.811 **	0.168 **	−0.201 **
GAD-7 Anxiety	0.811 **	1	0.146 *	−0.180 **
Age	0.168 **	0.146 *	1	−0.435 **
Income	−0.201 **	−0.180 **	−0.435 **	1

* p < 0.05, ** p < 0.01.

). The result of the correlation analysis between general characteristics and mental health showed that age and income had a statistically significant relationship. For age, the relationship between depression and positive (+) was statistically significant (r = 0.168, p < 0.01), and the relationship between anxiety and positive (+) was statistically significant (r = 0.146, p < 0.05). This can be interpreted to mean that mental health deteriorates with age. For income, the relationship between depression and negative (−) was statistically significant (r = −0.201, p < 0.01), and the relationship between anxiety and negative (−) was statistically significant (r = −0.180, p < 0.01). This can be interpreted to mean that a lower income is associated with a poorer mental health.

The result of the multiple regression analysis following correlation analysis found that only income information had a statistically significant negative (−) effect on mental health. For income, depression and negative (−) effects were statistically significant (t = −2.622, p < 0.01), and anxiety and negative (−) effects were statistically significant (t = −3.65, p < 0.001). This indicates that there is a negative effect on depression and anxiety as income decreases.

To confirm research question 1 of this study, correlation analysis and multiple regression analysis between our application and depression were conducted (

Table 13. Correlation between depression and anxiety and digital sensor data.

	Location	Gyroscope	Data Usage	Touch	Sleep Mode
PHQ-9	−0.682 **	−0.681 **	0.505 **	0.654 **	−0.645 **
GAD-7	−0.549 **	−0.550 **	0.669 **	0.839 **	−0.786 **

** p < 0.01.

and

Table 14. Results of multiple regression analysis between digital biomarkers and depression and anxiety.

	Var.	β	t	p	R	${\mathit{R}}^2$	F	Durbin–Watson
depression	Location	−0.3	−8.518 ***	0	0.888	0.788	82.138 **	1.957
	Gyroscope	−0.254	−6.807 ***	0
	Data usage	0.004	0.116	0.908
	Touch	0.207	4.712 ***	0
	Sleep mode	−0.153	−3.714 ***	0
anxiety	Location	−0.09	−3.164 **	0.002	0.928	0.86	135.969 **	1.837
	Gyroscope	−0.098	−3.227 **	0.001
	Data usage	0.146	4.833 ***	0
	Touch	0.455	12.728 ***	0
	Sleep mode	−0.283	−8.465 ***	0

** p < 0.01, *** p < 0.001.

). The location (11.88 ± 9.01) showed a negative (−) correlation (r = −0.682, p < 0.01) with depression (7.42 ± 7.33), and gyroscope (5824.53 ± 1241.05) and depression showed a negative (−) correlation (r = −0.681, p < 0.01). The data usage (93.19 ± 86.82) showed a positive (+) correlation with depression (r = 0.505, p < 0.01), and touch (254.15 ± 207.52) showed a positive (+) correlation with depression (r = 0.654, p < 0.01). Also, sleep mode (9.01 ± 2.74) showed a negative (−) correlation (r= −0.645, p < 0.01) with depression. It was confirmed that all five items of sensor information in our application correlated with depression, and showed significant correlations in the order of location, gyroscope, touch, data usage, and sleep mode. Based on this, we conducted an additional multiple regression analysis to closely observe the relationship between the variables, as shown in Table 13.

The result of the multiple regression analysis between the digital biomarkers and depression, as shown in Table 14, showed that the explanatory power of the regression model was 78.8%, and the regression equation was found to be statistically significant for location, gyroscope, touch, and sleep mode, but not for data usage (F = 82.138, p < 0.001). In terms of the independent variables, location was found to have a statistically significant negative (−) effect on depression (t = −8.518, p < 0.001), and gyroscope also appeared to have a negative (−) effect (t = −6.807, p < 0.001). Data usage showed a statistically significant negative (−) correlation in the correlation analysis with depression but was found to have no statistically significant effect in the multiple regression analysis. Touch appeared to have a statistically significant positive (+) effect on depression (t = 4.712, p < 0.001). Finally, sleep mode was found to have a statistically significant negative (−) effect on depression (t = −3.714, p < 0.001).

To test the second hypothesis of this research, correlation analysis and multiple regression analysis were conducted using the sensor data. Location showed a negative (−) correlation (r = −0.549, p < 0.01) with anxiety (6.28 ± 6.03), and a negative (−) correlation (r = −0.550, p < 0.01) was found between gyroscope and anxiety. Data usage showed a positive (+) correlation with anxiety (r = 0.669, p < 0.01), and touch showed a positive (+) correlation with anxiety (r = 0.839, p < 0.01). Also, sleep mode showed a negative (−) correlation with anxiety (r = −0.786, p < 0.01). It was confirmed that all five items of sensor information from the digital biomarkers correlated with anxiety, and showed significant correlations in the order of location, gyroscope, touch, data usage, and sleep mode. Based on this, an additional multiple regression analysis was conducted to closely observe the relationship between the variables. The result of the multiple regression analysis between the digital biomarkers and depression showed that the explanatory power of the regression model was 86.0%, and our application’s measurement of location, gyroscope, touch, data usage, and sleep mode were found to be statistically significant in the anxiety and regression equations (F = 135.969, p < 0.001). In terms of the independent variables, location was found to have a statistically significant negative (−) effect on anxiety (t = −3.164, p < 0.01), and gyroscope also had a statistically significant negative (−) effect on anxiety (t = −3.227, p < 0.001). Data usage was found to have a statistically significant positive (+) effect on anxiety (t = 4.833, p < 0.001), and the touch was also found to have a statistically significant positive (+) effect on anxiety (t = 12.728, p < 0.001). Finally, sleep mode was shown to have a statistically significant negative (−) effect on anxiety (t = −8.465, p < 0.001).

This research was conducted to determine whether it is possible to measure users’ mental health problems using smartphones, as mental illness and mental health problems are increasing. First of all, the mental health relevance according to the general characteristics of the research subjects was analyzed. It was found that depression was more likely to occur in women than in men, but in this research, it was found that men had poorer mental health than women. This is thought to be influenced by the fact that most of the study subjects were unemployed or recipients of low income. Marital status and family relationships had a significant negative effect on mental health. In addition, there was a difference in mental health depending on whether the participant was employed and whether or not there was a pension recipient; the result of multiple regression analysis between income and mental health showed that income had a negative (−) effect on mental health. As is widely known, economic activity and income level have an effect on mental health, which is consistent with the results of this research. Therefore, it is necessary to address any issues with economic activity and income level in order to manage mental health, and as the aging population increases, job creation and economic activity participation policies for the aging population are necessary.

In this research, the correlation and influence of digital biomarkers and users’ depression and anxiety, measurable using a smartphone, were analyzed using correlation analysis and multiple regression analysis. The application used in this research used data relating to location, which measures the distance traveled; gyroscope, which measures the number of steps and other physical activity information; data usage; touch, which measures the number of smartphone touches; and sleep mode, which measures sleep time.

In this research, the location and gyroscope information were measured and used as variables to explain the physical activity of the participants. The results showed that location was found to have a statistically significant positive (−) effect on depression (p < 0.001), and gyroscope was shown to have a statistically significant negative (−) effect on depression (p < 0.001). In addition, location was found to have a statistically significant negative (−) effect on anxiety (p < 0.001), and gyroscope was shown to have a statistically significant negative (−) effect on anxiety (p < 0.001). It can be interpreted that a higher physical activity reduces depression and a lower physical activity increases depression, and it has been confirmed that physical activity has a greater effect on anxiety than depression. As is widely known, participation in physical activity in the elderly population has a positive effect on mental health. In this research, it was also confirmed that the higher the level of physical activity, the lower the depression and anxiety. It was found that only regular light walking exercise can reduce stress and depression; therefore, it is necessary to establish a systematic system that can help encourage steady and regular physical activity.

In this research, smartphone usage was measured using data usage and touch information, and the degree of smartphone addiction was used as a decision variable. The results of the research showed that data usage did not have a statistically significant effect on depression, but touch did have a statistically significant positive (+) effect (p < 0.001). Also, data usage had a statistically significant positive (+) effect on anxiety (p < 0.001), and touch also had a statistically significant positive (+) effect on anxiety (p < 0.001). This can be interpreted to mean that the greater the smartphone usage and the more touches made on the smartphone, the more negative the effect on mental health. As is widely known, the higher the level of smartphone addiction, the poorer the mental health status. The effect of data usage and touch measured in this research on depression and anxiety coincided. In particular, it was confirmed that the frequency of smartphone use had a greater impact on anxiety than depression. Anxiety can lead to smartphone dependence, and smartphone overdependence can increase depression and anxiety. To break the negative link of smartphone overdependence and mental health deterioration, appropriate prevention education programs and treatment support programs are needed.

In this research, sleep mode was used as a variable to explain the sleep state as the information obtained by measuring the sleep time. The result showed that sleep mode had a statistically significant negative (−) effect on depression (p < 0.001). Also, sleep mode was found to have a statistically significant negative (−) effect on anxiety (p < 0.001). This can be interpreted to mean that the greater the lack of sleep, the more negative the effects on depression and anxiety. In general, there is a statistically significant negative (−) correlation between sleep quality and mental health, and when comparing people who achieve an adequate amount of sleep with those who do not, the more insufficient or excessive sleep is, the more likely it is that depression will develop. Therefore, the negative (−) effect of sleep mode on depression and anxiety found in this study was consistent with the results of previous studies. In addition, to improve the quality of sleep, it is known that appropriate physical activity and leisure activities reduce stress and improve sleep quality, and that the temperature during sleep also affects the quality of sleep. Based on these preceding studies, measures related to physical activity are needed for positive changes in mental health.

This study investigated whether there is a relationship between mental health and digital biomarkers measurable on smartphones. Based on previous studies, the digital biomarker information was classified into three major categories. Location and gyroscope were classified as physical activity, data usage and touch as smartphone addiction, and sleep mode as sleep state. The results of the research confirmed that all measurements made using our application had a statistically significant effect on depression and anxiety. This can be interpreted to mean that the information collected from these digital biomarkers can explain the users’ mental health condition. In addition, based on the data that our application measure, we can determine that the users’ mental health status can be predicted.

5. Conclusions

Nowadays, it is difficult for people with various respiratory diseases and mental health conditions to receive regular examinations/observations from specialized medical institutions. Therefore, as the need for the continuous health status measurement/monitoring of subjects was recognized, a device for constantly monitoring the trends and changes in blood oxygen saturation was developed to ensure safety (oxygen saturation measurement was conducted through a separate device). In addition, we conducted hypothesis testing for analyzing the relationship between sensor data and depression and anxiety. The sensor data used in this research were location, gyroscope, touch, data usage, and sleep mode, taken from a smartphone. Correlation analysis and multiple regression analysis were conducted to analyze the relationship between the collected sensor data and depression and anxiety. As a result of the research, it was found that among the general characteristics, income, location, sleep duration, and gyroscope were found to have a negative effect on depression and anxiety, whereas touch was shown to have a positive (+) effect. It was confirmed that sensor data measured at runtime can be used as information that can explain users’ depression and anxiety.