IntelliTrace: Intelligent Contact Tracing Method Based on Transmission Characteristics of Infectious Disease

Abstract

The COVID-19 pandemic has underscored the necessity for rapid contact tracing as a means to effectively suppress the spread of infectious diseases. Existing contact tracing methods leverage location-based or distance-based detection to identify contact with a confirmed patient. Existing contact tracing methods have encountered challenges in practical applications, stemming from the tendency to classify even casual contacts, which carry a low risk of infection, as close contacts. This issue arises because the transmission characteristics of the virus have not been fully considered. This study addresses the above problem by proposing IntelliTrace, an intelligent method that introduces methodological innovations prioritizing shared environmental context over physical proximity. This approach more accurately assesses potential transmission events by considering the transmission characteristics of the virus, with a special focus on COVID-19. In this study, we present space-based indoor Wi-Fi contact tracing using machine learning for indoor environments and trajectory-based outdoor GPS contact tracing for outdoor environments. For an indoor environment, a contact is detected based on whether users are in the same space with the confirmed case. For an outdoor environment, we detect contact through judgments based on the companion statuses of people, such as the same movements in their trajectories. The datasets obtained from 28 participants who installed the smartphone application during a one-month experiment in a campus space were utilized to train and validate the performance of the proposed exposure-detection method. As a result of the experiment, IntelliTrace exhibited an F1 score performance of 86.84% in indoor environments and 94.94% in outdoor environments.

1. Introduction

The coronavirus disease 2019 (COVID-19) pandemic caused public health crises globally, which revealed the need for highly efficient contact tracing for the effective containment of infectious diseases. In a pandemic situation, an electronic support system capable of preventing the spread of secondary infection through contact tracing is required [1]. The study presented in [2] reported that a smartphone contact tracing app can be useful in terms of automating the challenging task of tracing the contacts of people infected with COVID-19. The University of Oxford published research that showed a significant decrease in the proportion of COVID-19-confirmed patients following the use of the contact tracing app [3]. In South Korea, an electronic support system known as the Epidemic Investigation Support System (EISS) was quickly introduced. By February 2022, EISS had been utilized for 90% of confirmed cases, thereby demonstrating its effectiveness [1,4]. Contact tracing methods for infectious diseases can be categorized into location-based and distance-based methods for the detection of proximity contact. Methods that use GPS data [5] or a Wi-Fi fingerprint [6] are representative location-based methods that analyze the location information of a confirmed case to determine the contact status. However, since location-based contact detection methods require the use of personal information, privacy issues have been raised. Therefore, it is difficult to apply these methods in many countries. To address this problem, methods using Bluetooth for distance-based contract detection instead of location-based information have been developed [7]. The Google Apple Exposure Notification (GAEN) [8,9], which was jointly developed by Google and Apple, is a distance measurement-based contact tracing method that uses Bluetooth signals. After exchanging tokens between users of GAEN, when a confirmed case occurs, the users who exchanged tokens are notified of their contact status. GAEN can protect personal information and inform contacts in real time.

However, the GAEN method faces the following challenges for large-scale adoption and application. The first problem is related to the radio frequency (RF) signal strength accuracy of Bluetooth, which is used by GAEN. Leith and Farrell [10,11] conducted a GAEN API measurement-based evaluation on commuter trams and buses in Dublin, and their results indicated that there was a low correlation between the Bluetooth-received signal strength and the physical distance among the devices installed with the app. In addition, in an experiment related to detecting proximity contact using the GAEN API-based COVID-19 contact tracing app, approximately half of the device pairs were not judged to exposure despite being less than 2 m apart. These previous studies show that the accuracy of GAEN API-based contact tracing is not high for practical applications because of the accuracy limitations of the Bluetooth radio signal. The second problem is that GAEN simply determines the contact based on the distance between people without considering the transmission characteristics of the SARS-CoV-2 virus. Therefore, more contacts are traced than necessary, which generates overly frequent contact cases. For example, in the case of people waiting for a bus at an outdoor bus station without close contact, a large number of unnecessary contacts are detected when determining the contact status by GAEN. These two problems limit the global application of GAEN. The Ministry of Health and Welfare of South Korea also developed a GAEN-based app, but it was not used for practical applications because of concerns that the occurrence of unnecessary contacts may lead to an excessive load in the process of infectious disease management.

Exposure status should be determined according to the transmission characteristics of each type of virus, instead of simply detecting the distance between people. The differences in the transmission patterns of the SARS-CoV-2 virus between indoor and outdoor environments were studied in [12]. This study reports that the transmission of COVID-19 is approximately 10 times higher in indoor environments than in outdoor environments. In addition, the US Centers for Disease Control and Prevention (CDC) announced that the spread of COVID-19 from a single patient to several customers in a restaurant in Guangzhou in January 2020 was through aerosol transmission via air conditioners [13]. These studies clearly demonstrate that the transmission patterns of the COVID-19 virus differ significantly between indoor and outdoor environments. Specifically, when individuals are in the same indoor space, there is a high probability of virus transmission, even if people maintain a considerable distance. On the other hand, when people are close to each other outdoors, the virus does not propagate easily. Hence, it is crucial to take into account the differences in transmission characteristics between indoor and outdoor environments.

Health authorities have been hesitant to adopt approaches like GAEN, which primarily considers the physical distance between individuals without adequately addressing key transmission risk factors, such as co-location. Our study aims to advance the field of contact tracing by prioritizing the identification of whether individuals share the same environmental context, such as being in the same room or trajectory, which is a more relevant measure for assessing potential transmission events than mere physical distance. This approach is particularly pertinent given the heterogeneity of COVID-19 transmission rates reported in different settings. For indoor spaces, we propose space-based indoor Wi-Fi contact tracing, which uses the similarity of Wi-Fi fingerprints of access points in spaces and recognizes the co-location statuses of individuals in the same indoor space; it detects contacts based on this information. Unlike existing Wi-Fi-based contact tracing methods that rely on measuring physical distance or user location using received signal strength indicators ($RSSI$) [6], space-based indoor Wi-Fi contact tracing focuses on similarities in access point signals to identify co-located individuals. This is achieved through the implementation of a machine learning model.

On the other hand, for outdoor spaces, we suggest trajectory-based outdoor GPS contact tracing. This method utilizes GPS data to track the movement patterns of individuals and detects potential exposure by identifying whether people were on the same trajectory and in close proximity. By analyzing the trajectories and distances, this approach enables the identification of potential contacts and exposure risks in outdoor environments.

By implementing two distinct contact tracing methods—indoor space (space-based indoor Wi-Fi contact tracing) and outdoor space (trajectory-based outdoor GPS contact tracing)—we can effectively address the challenges of each environment and improve the overall contact tracing performance.

We conducted experiments to validate the IntelliTrace. We obtained data from 28 experiment participants, consisting of groups in which regular contact may occur at the university. We installed an app that periodically collected GPS, Wi-Fi, and Bluetooth low energy (BLE) beacon information on the smartphones of participants for one month, with their agreement (the experiment was conducted in compliance with research ethics procedures with prior consent for the collection of personal information of participants). To prevent the excessive collection of personal information, we set up a geofence around the university campus so that personal information was collected only when the participants were on campus. The experiment was approved by the Institutional Review Board (IRB). We followed the guidelines of the IRB to protect personal information and personally identifiable information of the research participants.

The contributions of this study are as follows: We propose a novel contact tracing method using exposure detection based on machine learning, considering that a virus has different transmission characteristics, depending on indoor and outdoor conditions. For an indoor space, we propose space-based indoor Wi-Fi contact tracing to detect exposure by determining the co-location status of people in the same space. For an outdoor space, trajectory-based outdoor GPS contact tracing is used to detect exposure by determining the companion status of people, such as the same movement in their trajectory. We evaluated the performance of the proposed method using real data collected from 28 people for a month. This study can also be used as guidance for future research works in the development of effective contact tracing methods for COVID-19 and other infectious diseases by considering the transmission characteristics.

The structure of this paper is organized as follows. Section 2 reviews research works related to contact tracing. Section 3 proposes indoor and outdoor contact tracing methods of intelligent contact tracing based on transmission characteristics of COVID-19 (IntelliTrace). Section 4 introduces the experimental design. Section 5 discusses the analysis of exposure detection based on data obtained from actual experimental results. Section 6 presents the conclusions of this study.

2. Related Work

People infected with COVID-19 transmit the virus even when they are in the asymptomatic stage. Therefore, there has been increasing emphasis on the importance of digital contact tracing to prevent infectious disease [14]. Many countries have shown considerable interest in smartphone contact tracing apps, which are useful for automatic contact tracing, and several studies with various approaches have been actively conducted. For COVID-19, contact tracing is performed based on various types of data, such as GPS, Bluetooth, and Wi-Fi.

2.1. GPS-Based Contact Tracing

In GPS-based methods, contact tracing is performed based on the user’s location information obtained through GPS. Berke et al. [15] proposed a contact tracing method that utilized recent GPS location histories and a private set intersection protocol. Wang et al. [5] conducted contact tracing through a spatial transmission trajectory generated using GPS data from contact tracing app users. However, accuracy and privacy issues have been raised for location-based contact tracing methods. Moreover, GPS does not generally operate properly in indoor spaces, and its performance is poor in outdoor spaces, such as central business districts with high-rise buildings [2]. Recognizing the limitations of traditional GPS-based contact tracing that prioritizes distance, our research emphasizes the importance of shared movement paths. Our approach aims to provide a more comprehensive reflection of the transmission characteristics of the viruses, enhancing the predictive capability of contact tracing.

2.2. Bluetooth-Based Contact Tracing

Alternative methods of contact tracing have emerged to address the problems of the GPS-based method described above, where contact is detected by exchanging signals based on Bluetooth. The Bluetooth-based contact tracing method collects only anonymized information from devices that are in proximity, utilizing Bluetooth signals, and has the advantage of protecting personal information [16]. Private automated contact tracing (PACT) [17] detects exposure to an infected person based on estimating the distance between users by utilizing Bluetooth. GAEN, which is compatible with Android and iOS, is a representative Bluetooth-based contact detection framework that determines the status of exposure to COVID-19 [8,9]. GAEN exchanges tokens by interacting with other smartphones through Bluetooth. GAEN measures the signal strength between devices using Bluetooth, and the distance between two smartphones is estimated based on the signal strength. If it is determined that the distance between the smartphones is shorter than a specified distance for a set duration (e.g., 10 min), the beacon information from nearby smartphones starts to save. GAEN conducts exposure detection through the distance-based token exchange but does not reflect the transmission characteristics of the COVID-19 virus. Using this approach, GAEN makes judgments simply based on contact distance, without considering the transmission characteristics of indoor or outdoor environments. Therefore, numerous contact alerts, including casual contacts with a very low transmission probability, may occur in indoor and outdoor environments [10,11,18]. However, to capture the nuanced risk of virus transmission, IntelliTrace moves beyond mere distance estimation to a spatial-based contact detection method that considers the environmental characteristics conducive to virus spread.

2.3. Wi-Fi and Cellular Network-Based Contact Tracing

For contact tracing based on Wi-Fi signals, the location visited by users is recognized through Wi-Fi signals. WiFiMon [19] recognizes the visited locations based on the AP connection log messages of user devices. WiFiTrace [6] proposes a network-centric contact tracing approach that relies on passive Wi-Fi sensing without client-side involvement. WiFiTrace monitors and tracks the movement of an infected person and contacts through Wi-Fi-based indoor positioning based on Wi-Fi APs.

For cellular network-based contact tracing, various methods for location-based contact detection have been developed, such as the tracking of contacts through a location–recognition method based on the base stations of mobile carriers led by the Korea Disease Control and Prevention Agency (KDCA) [1]. However, issues regarding privacy, such as the invasion of personal information, have been raised [20].

Traditional contact tracing methods often rely solely on location data and do not account for co-location. Particularly, space-based analyses have not been extensively applied within indoor environments, leading to a challenge in differentiating between casual and close contact, especially outdoors. To overcome these limitations, our study introduces a contact tracing method that comprehensively considers the transmission dynamics of the virus. By prioritizing the shared occupancy of spaces over mere physical proximity, we aim to enhance the precision and applicability of contact tracing, ensuring it more accurately reflects the complex nature of pathogen spread.

3. Intelligent Contact Tracing Based on Transmission Characteristics

Existing contact tracing methods do not consider the transmission characteristics of the virus. This leads to difficulties in distinguishing between casual contacts and close contacts. As a result, in some cases, an excessive number of contacts are detected, while in other cases, the contact tracing method fails to detect cases in which exposure to close contact occurred. To address these limitations, we propose a novel method called “IntelliTrace”, which provides intelligent contact tracing based on the transmission characteristics of COVID-19.

3.1. Overall Process of IntelliTrace

Algorithm 1 represents the exposure assessment procedure in the IntelliTrace system, which determines “close contact” between a confirmed case (u) and a potential contact (v). Here, $t_{confirm}$ represents the time when u was confirmed, and $t_{infectious}$ indicates the infectious period, which is the time during which an infected individual can transmit the virus to others. The starting time (t) of the exposure assessment algorithm is initialized as the time obtained by subtracting $t_{infectious}$ from $t_{confirm}$. The IntelliTrace classifies the location of the confirmed patient as either indoors or outdoors based on satellite information received from the mobile phone GPS sensor, which is generally used to distinguish between indoors and outdoors in smartphones, in order to reflect different transmission characteristics of COVID-19 in indoor and outdoor spaces ($isIndoor(u,t)$ in Algorithm 1). A strong GPS signal strength detected by the phone’s GPS sensor indicates a ’GPS provider’ and $isIndoor$ is classified as outdoors. If the signal is weak, suggesting reliance on the ‘Network provider’, then $isIndoor$ is classified as indoors [21]. For indoor spaces, it uses space-based indoor Wi-Fi contact tracing to confirm the co-location status for at least $t_{minExposurePeriod}$ in the same space ($IsExposureIndoor(u,v,t,tminExposurePeriod)$ in Algorithm 1). The implementation details of $IsExposureIndoor$ are expanded in Section 3.2. For outdoor spaces, it employs trajectory-based outdoor GPS contact tracing to determine whether v has moved along the same path as the confirmed case for at least $t_{minExposurePeriod}$ ($IsExposureOutdoor(u,v,t,tminExposurePeriod)$ in Algorithm 1). The implementation details of $IsExposureOutdoor$ are expanded in Section 3.3. Here, $t_{minExposurePeriod}$ refers to the minimum exposure time required to classify it as a “close contact” when someone who is susceptible to infection has been exposed to an infectious case. The CDC defines “close contact or exposure” as someone who spends at least 15 min within six feet of a confirmed coronavirus case. During the loop, if exposure is not detected, the algorithm increments ‘t’ by 1, progressing in 1-minute intervals. If exposure is not found after completing the loop, the algorithm returns false, indicating that there was no close contact between u and v.

3.2. Space-Based Indoor Wi-Fi Contact Tracing

Unlike conventional Wi-Fi-based contact tracing methods that require comprehensive mapping of Wi-Fi fingerprints, our system employs a machine learning model that leverages spatial similarity derived from AP similarity metrics to determine co-location [6]. For space-based indoor Wi-Fi contact tracing, Wi-Fi fingerprints of a confirmed case (u) and a potential contact (v) are used to calculate similarity values, as shown in Figure 1. The status of exposed cases is determined through a machine learning model using calculated similarity values. The steps that we followed in the space-based indoor Wi-Fi contact tracing method are described in the following sub-sections.

Algorithm 1: Exposure Assessment

3.2.1. Step 1: Calculate Similarity

To enhance the precision of indoor contact tracing, we initially applied a path-loss model to Wi-Fi data to calculate a distance vector. This allows us to compute Euclidean, Cosine, and Jaccard similarities, which serve as robust metrics to assess the proximity between a confirmed patient and other individuals.

Consider A, a set of basic service set identifiers (BSSIDs) of Wi-Fi APs collected through the IntelliTrace system. A is represented as follows:

$$\begin{array}{ccc}\hfill A& =& \bigcup _{i=1}^{N_{pe}}{A}_i\hfill \\ & =& \{a_1,a_2,\dots ,a_j,\dots ,a_{N_{ap}}\}\hfill \end{array}$$

where $N_{pe}$ is the number of individuals whose data are in the system, and $A_i$ is the set of BSSIDs of Wi-Fi APs received by the individual (i). $N_{ap}$ denotes the total number of APs in the IntelliTrace system.

The received signal strength indicator ($RSSI$), which is computed in decibels (dB), indicates the strength of the received radio signals. A greater $RSSI$ value represents a stronger signal. $r_{ij}$ means the $RSSI$ for AP BSSID $a_j$ detected by the individual (i). When determining the contact between the confirmed patient and potential contact, an estimated distance using $RSSI$ is utilized to calculate the similarity. Owing to the characteristics of Wi-Fi signals, the distance between an individual and an AP increases exponentially as $RSSI$ decreases. We obtain $\overrightarrow{d_i}$, a distance vector for set A of the individual (i), by applying the path-loss model to $r_{ij}$ to convert the $RSSI$ signal into the distance, $d_{ij}$, between the individual (i) and AP ($a_j$).

$$\overrightarrow{d_i}=\left[\begin{array}{cccccc}{d}_{i1}& d_{i2}& \cdots & d_{ij}& \cdots & d_{N_{pe}{N}_{ap}}\end{array}\right]$$

$$d_{ij}=\left\{\begin{array}{cc}{10}^{\frac{(r_0-r_{ij})}{10\delta }},\hfill & \mathrm{if}{r}_{ij}>RSS{I}_{min}.\hfill \\ d_{max},\hfill & \mathrm{otherwise},\mathrm{or}\mathrm{is}\mathrm{NaN}.\hfill \end{array}\right.$$

(1)

$r_0$ indicates the $RSSI$ at a unit distance ($d_0$). $RSS{I}_{min}$ denotes a minimum $RSSI$ value to ensure distance accuracy and $\delta$ is a path-loss coefficient according to the Wi-Fi signal environment. In general, the value of $\delta$ in an indoor environment is set to 3–5 [22]. Additionally, the Wi-Fi signal reception environment is indoors with many obstacles; therefore, the path-loss coefficient ($\delta$) is set to 4. If $r_{ij}$ is less than $RSS{I}_{min}$ or is NaN, it is set as a constant $d_{max}$, which means the maximum radius between an individual and a Wi-Fi AP. The IntelliTrace uses the BSSI set ($A_i$) and the distance vector ($\overrightarrow{d_i}$) for Wi-Fi signals detected by the individual (i) to calculate similarity.

In this study, three similarity measures were chosen—Euclidean similarity, Cosine similarity, and Jaccard similarity—to compute the similarities, as described below in

Table 1. Similarities utilized for space-based indoor Wi-Fi contact tracing.

	Euclidean Similarity ($\mathit{Sim}{(\overrightarrow{{\mathit{d}}_{\mathit{u}}},\overrightarrow{{\mathit{d}}_{\mathit{v}}})}^{\mathit{EUC}}$)	Cosine Similarity ($\mathit{Sim}{(\overrightarrow{{\mathit{d}}_{\mathit{u}}},\overrightarrow{{\mathit{d}}_{\mathit{v}}})}^{\mathit{COS}}$)	Jaccard Similarity ($\mathit{Sim}{(\overrightarrow{{\mathit{d}}_{\mathit{u}}},\overrightarrow{{\mathit{d}}_{\mathit{v}}})}^{\mathit{JACC}}$)
Description	The similarity that measures the Euclidean distance between vectors	The similarity between vectors obtained using the cosine angle in an N-dimensional space	The similarity between sets of data to see which members are shared and distinct
Target feature	The Wi-Fi fingerprint	The direction of movement	Shared Wi-Fi APs
Equation	$1-\frac{\left|\right|\overrightarrow{d_u}-\overrightarrow{d_v}\left|\right|}{2d_{max}}$	$\frac{\overrightarrow{d_u}·\overrightarrow{d_v}}{\left|\right|\overrightarrow{d_u}\left|\right|\left|\right|\overrightarrow{d_v}\left|\right|}$	$\frac{|A_u\cap A_v|}{|A_u\cup A_v|}$

. Euclidean distance was employed for simple distance calculations, Cosine similarity was utilized to assess the similarity of AP signal vectors, and Jaccard similarity was used to evaluate the similarity of overlapping AP patterns.

3.2.2. Step 2: Determine the Close Contact with Machine Learning

The space-based indoor Wi-Fi contact tracing determines the exposed (close contact) case based on the time duration of co-location in the same space after detecting whether the potential contact and confirmed patient were in the same space. We propose the application of a machine learning model for this task due to its ability to discern complex patterns and relationships within the data.

First, the system determines whether the confirmed patient and potential contact person were co-located in the same space. When it is identified that an individual has been in the presence of a confirmed COVID-19 patient at a specific moment (t), it is classified as a casual contact ($C_t=1$).

We employed $I_{threshold}$ as an adjustment factor to address the inherent uncertainties present in our dataset. Subsequently, if it is determined that the total number of casual contacts exceeds the threshold value for indoor ($I_{threshold}$) during the period $t_{minExposurePeriod}$, which indicates a potential risk of infection, the system designates it as a close contact event between the two individuals. This value was determined through an iterative experimental process, where we systematically fine-tuned it to align with observed trends and patterns in the data. Machine learning models were trained and used for the prediction of co-location in the same space based on the three similarity features. For training the machine learning models, we used BLE-based co-location data obtained through the BLE beacons installed in each indoor space as the ground truth. The feature data X are set as a three-dimensional feature vector that includes $Sim{(\overrightarrow{d_u},\overrightarrow{d_v})}^{EUC}$, $Sim{(\overrightarrow{d_u},\overrightarrow{d_v})}^{COS}$, and $Sim{(\overrightarrow{d_u},\overrightarrow{d_v})}^{JACC}$. The data label Y (0 or 1) indicates whether or not they are co-located in the same space, 0 denotes non-contact, and 1 denotes casual contact (co-location). The casual contact status is determined by machine learning models trained using similarity features and BLE data.

As the contact prediction models, we use the decision tree (DT), support vector machine (SVM), k-nearest neighbor (kNN), and artificial neural network (ANN) models, which are widely utilized as prediction models. To obtain accurate labels, the machine learning model was executed with datasets obtained through the pre-arranged scenario. Figure 2 shows how the space-based indoor Wi-Fi contact tracing method detects close contact. As a time window of $C_t$ with a size of $t_{minExposurePeriod}$ slides within $t_{infectious}$, if the sum of $C_t$ during $t_{minExposurePeriod}$ exceeds the threshold ($I_{threshold}$), it is determined as an exposed case (close contact).

3.3. Trajectory-Based Outdoor GPS Contact Tracing

Considering that transmission of the virus is related to human activity patterns [23], we focus on the trajectory using GPS data for outdoor contact tracing. We clarify the distinction between casual contact and close contact using both distance and the similarity of the trajectory vector between the confirmed case and potential contact, which factors in accompanying movement. Unlike recent GPS-based contact tracing methods that primarily track current location history, our trajectory-based approach leverages vectors from specific time periods alongside confirmed case data. This reduces the rate of false detections by accounting for temporal dynamics [15].

Figure 3 shows the distance and similarity between individuals used for the trajectory-based outdoor GPS contact tracing. The trajectories of each individual are denoted by dotted lines, and the points $G_u$, $G_{v_1}$, and $G_{v_2}$ represent the positions of the confirmed patient and the potential contacts, respectively, at time t. $D(u,v_1)$ means the distance between the patient and the first potential contact. $S_{current}(u,v_1)$ signifies their similarity at the current time. We minimize GPS errors by applying the extended Kalman filter (EKF) [24] to the GPS data ($G_t$) of confirmed cases and potential contacts, considering the error range, which is the propagation characteristic of GPS. Next, the estimated distance between the confirmed person and potential contact is calculated using the filtered GPS data. We define a casual contact when the calculated distance is within the maximum infection range ($R_{max}$) for COVID-19, and when the cosine similarity, calculated using the trajectory vector ($\overrightarrow{T_t}$) of the two individuals, is greater than the threshold value ($S_{max}$). If the number of casual contacts during $t_{minExposurePeriod}$ exceeds a threshold value for outdoors ($O_{threshold}$), it is judged as an exposed case.

3.3.1. Step 1: Calculate Contact

To bolster the accuracy of outdoor contact detection, we first refine the GPS data with the EKF. Subsequently, we compute the cosine similarity of trajectory vectors.

When GPS data, $G_t$, at a specific time t, are represented as a pair of latitude ($la{t}_t$) and longitude ($lo{n}_t$), the GPS data after applying the EKF are expressed as $G_t^f$. $G_{t,u}$ means the confirmed person’s GPS data, and $G_{t,v}$ refers to the GPS data of a potential contact, v, at time t. To calculate the distance using the filtered location information of a confirmed person and potential contacts, we utilize the Haversine formula [25], which is used to calculate the distance between points on a sphere. The distance ($D_t(u,v)$) is calculated using $G_{t,u}^f$ and $G_{t,v}^f$.

If $D_t(u,v)$ is greater than the maximum infection distance ($R_{max}$), it is considered non-contact at that time t. If $D_t(u,v)$ is within $R_{max}$, we calculate the similarity of the movement between the confirmed person and potential contact, rather than simply determining contact based on distance information. The similarity of the movement path ($S_t(u,v)$) is obtained through the trajectory vector ($\overrightarrow{T_t}$) expressed as Equation (2), containing the movement tendency information of a person.

$$\overrightarrow{T_t}=\overrightarrow{G_t^f{G}_{t+1}^f}$$

(2)

We set the moving tendency vector of the confirmed person to $\overrightarrow{T_{t,u}}$ and the vector of potential contacts to $\overrightarrow{T_{t,v}}$. $S_t(u,v)$ can be obtained with Equation (3).

$$S_t(u,v)=\frac{\overrightarrow{T_{t,u}}·\overrightarrow{T_{t,v}}}{\left|\right|\overrightarrow{T_{t,u}}\left|\right|\left|\right|\overrightarrow{T_{t,v}}\left|\right|}$$

(3)

If $S_t(u,v)$ exceeds the threshold value ($S_{max}$), it is called casual contact. The determination of the casual contact status ($C_t(u,v)$) between u and v at a specific time t is obtained by Equation (4).

$$C_t(u,v)=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}{D}_t(u,v)<R_{max}\mathrm{and}{S}_t(u,v)>S_{max}.\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(4)

3.3.2. Step 2: Determine Exposed Case

To reinforce the reliability of detecting close contacts outdoors, our system extends beyond the analysis of recent GPS data and incorporates historical trajectory information. The process is as follows.

We consider a time window of $C_t(u,v)$ with a size of $t_{minExposurePeriod}$ to determine whether a casual contact is a close contact or not. As the time window slides within $t_{infectious}$, if the sum of $C_t(u,v)$ during $t_{minExposurePeriod}$ exceeds the threshold ($O_{threshold}$), it is determined to be an exposed case (close contact). Figure 4 shows how exposed cases are determined during $t_{infectious}$.

4. Experimental Design

4.1. Data Collection

• Data gathering

We developed a smartphone application for data collection and utilized it for data gathering. The boundary of the Sejong University campus was established using geofence technology, which creates a virtual perimeter based on location information, since participants were involved in the experiment. To prevent the excessive collection of personal information, data were gathered only when the users’ real-time locations were detected within the geofence, and data collection was discontinued when the users moved out of the geofence. The detailed items of the data collected through the data-gathering application are as follows: LOG TIME, SSID, BSSID, and $RSSI$ for Wi-Fi; LOG TIME, DEVICE NAME, and $RSSI$ for BLE Beacon; LOG TIME, altitude, latitude, longitude, and provider for GPS. Figure 5 shows the main screen of the data-gathering application used in the experiment. Participant recruitment within the university institution was conducted via an online portal. We set the number of participants to 28, taking into account ethical issues and the need for smooth and accurate management of the participants. The cohort consisted of 22 students and 6 staff members, with a balanced gender distribution. Comprehensive explanations of the experimental procedures were provided, and all participants gave their informed consent prior to the commencement of the study. For data collection, participants were equipped with mobile phones and were instructed to use pseudonyms in place of their real names to safeguard their privacy and ensure the anonymity of the collected data. Participants were required to consistently carry their mobile phones to maintain data integrity. Regarding research ethics, the experiment was conducted in strict accordance with the ethical guidelines set forth by our institution’s Research Ethics Board. Additionally, the confidentiality of participant data was prioritized, and measures were implemented to prevent unauthorized access to sensitive information. During the course of conducting this experiment, the scenario was not conducted with pre-confirmed patients designated in advance. Following the scenario, ’confirmed patient’ were assigned to data randomly to maintain the integrity of the study’s realism. Furthermore, data received from the app utilized pseudonyms and unique user identifiers (UIDs) rather than real names. Because of this design, participants were unaware of each other’s behavior and did not modify their own behavior. We provided participants with a small stipend to offset any expenses or inconveniences incurred during the experiment.

• Training dataset

We created training datasets according to the scenario to increase the accuracy of training through precise labeling. The training data were collected from 10 individuals during a scenario in which their movements and exposure statuses were pre-determined. During the experiment, each participant (10) took turns acting as the confirmed patient, while the remaining 9 participants were considered potential contacts, thereby augmenting the dataset. This design enabled the collection of a diverse dataset, simulating various contact scenarios and providing a foundation for training our contact tracing model. Whenever the participants visited these spaces, the BLE data of the beacon installed in places were simultaneously collected in the data-gathering application on participants’ smartphones. The collected data were used to confirm the accuracy of the collected datasets and train the machine learning model for the proposed contact tracing method. Using the data-gathering application, a total of 300,645 fingerprints of Wi-Fi APs and 200,000 GPS data points were collected for the training dataset.

• Validation dataset

For the experiment, we obtained consent from 28 participants, consisting of Sejong University students and faculty members. Data were gathered while the participants used the data-gathering application installed on their smartphones. Before starting the experiment, we installed BLE beacons in enclosed spaces that participants regularly visited, such as laboratories, seminar rooms, and campus cafes. These locations were areas where frequent contact was likely to occur, serving as the ground truth for our study. We collected data at the Sejong University campus in South Korea at one-minute intervals for approximately a month, from September to October 2021. While participants in the training dataset followed designated times and locations set by the researchers, participants in the validation dataset performed routine activities in a variety of environments and situations. We do not distinguish between outdoor and indoor scenarios for the validation dataset, since the validation dataset is collected from participants’ daily lives. Through the data-gathering application, a total of 27,517,386 fingerprints of Wi-Fi APs and 486,329 GPS data points were collected for the validation dataset.

4.2. Data Preparation and Processing

For contact tracing of IntelliTrace, $t_{infectious}$ was set to 14 days. When there was exposure to the SARS-CoV-2 virus for 10–20 min on average, the exposure led to a confirmed case [26]. Based on this information, $t_{minExposurePeriod}$ was set to 10 min. When applying the path-loss model to convert $RSSI$ to distance, the unit distance $d_0$ was set to 1 m. Depending on the frequency values, $r_o$ was set to −26 dBm at 2.4 GHz and −30 dBm at 5 GHz. For $RSSI$, since small variations caused large errors, an $RSSI$ value with a sensitivity that satisfied the condition of $RSSI$ ≈ PER being 8% or less was set as $RSS{I}_{min}$, which is a reliable range based on the specification of the Wi-Fi receiver module mainly used in smartphones, in practice [27]. Then, −98.4 dBm (2.4 GHz) and −94.5 dBm (5 GHz) were set as $RSS{I}_{min}$ according to the frequency.

5. Results and Discussion

5.1. Indoor Experimental Results

• Similarity of Wi-Fi fingerprints

Figure 6 shows some of the calculated results of the Euclidean, Cosine, and Jaccard similarities between two individuals. The irregularities observed in the Euclidean similarity during the first 2 h of the trace are influenced by the specifics of our experimental setup. The varying proximity of rooms A–C, some adjacent and others more distant, contributes to the diverse data patterns. The beacon data served as the ground truth for determining the casual contact. If one or more network addresses of the same beacon were shared between a confirmed patient and a potential contact person, it could be determined that they were co-located in the same space at the time (marked as the filled circle in Figure 6). The ground truth was set to true, indicating contact in this case, while without a shared network address of beacons, it was set to false, indicating non-contact (marked as the empty circle). As shown in Figure 6, the accuracy of the beacon signal-based ground truth is approximately 70%, leaving about 30% of the data with inaccuracies. This discrepancy arises despite participants being physically present in the same space in the scenario. It is crucial to recognize and address these challenges in order to provide a comprehensive understanding of the experimental conditions. The accuracy of the beacon signal-based ground truth has an impact on the value of $I_{threshold}$.

• Machine learning-based co-location recognition

For the refinement of training, we trained machine learning models that predict a contact status using the training dataset, which is obtained in a scenario where we know a certain route and have already set an exposure status. Our machine learning model was trained using the Wi-Fi similarity features and beacon data as the ground truth and returned the casual contact status as the output. We used a decision tree (DT), k-nearest neighbor (kNN with $k=3$), and support vector machine (SVM) learning models. Additionally, we utilized a shallow neural network model with three hidden layers. The number of hidden neurons for each hidden layer was 30, 12, and 8. For the implementation of these models, we utilized the scikit-learn toolkit, a widely used machine learning library in Python that provides a variety of tools for classification, regression, clustering, and more. We analyzed each model by comparing the co-location recognition performance of predicted contacts to validate whether the trained machine learning models correctly recognize the space and judge the casual contact through accurate co-location recognition. Precision, recall, accuracy, and the F1 score were used as performance metrics.

The co-location recognition accuracy of each model is shown in

Table 2. Instantaneous co-location recognition performance of machine learning models trained to find casual contacts.

Model	Accuracy	Precision	Recall	F1 Score
ANN	83.06	69.00	83.06	75.38
DT	84.61	84.52	84.61	84.56
SVM	85.22	83.91	85.22	84.34
kNN	85.01	84.44	85.01	84.69

.

All of the trained machine learning models exhibited co-location recognition accuracy of at least 83%. For the space-based indoor Wi-Fi contact tracing methods, DT was selected for the detection of casual contacts because it exhibited the best performance. We set the threshold value ($I_{threshold}$) for determining close contact as 50% of the time window ($t_{minExposurePeriod}$), based on the accuracy of the beacon signal-based ground truth. For the machine learning model that predicts contact for the input similarity features, if the model predicts casual contacts for more than the $I_{threshold}$ within the set time duration ($t_{minExposurePeriod}$), the case is determined as an exposure.

We analyzed the performance of the space-based indoor Wi-Fi contact tracing method using the validation dataset obtained from the month-long experiment. The performance of the exposure detection algorithm (according to the $t_{minExposurePeriod}$ of the DT trained on the similarity features calculated from the fingerprint data of Wi-Fi APs and the beacon data) is summarized in

Table 3. Performance of space-based indoor Wi-Fi contact tracing.

Precision	Recall	Accuracy	F1 Score
96.68	78.82	88.41	86.84

.

Through these indicators, it can be observed from Table 3 that 97% of exposures predicted by the algorithm were real exposures, and 79% were correctly predicted as exposed cases among the real exposures. We demonstrated the effectiveness of the proposed space-based indoor Wi-Fi contact tracing method in a real-life environment. To address the challenge of manually labeling the participant’s movement in each space, we employed a BLE beacon for automatic labeling during the collection of datasets. However, since the reception rate of the BLE beacon signals was low, there was a limitation in improving the recall performance.

5.2. Outdoor Experiment Results

Since the contact detection performance of the trajectory-based outdoor GPS contact tracing varies, depending on $R_{max}$ and $S_{max}$, the experiment was conducted by changing $R_{max}$ and $S_{max}$. The performance of casual contact detection is shown in the graph in Figure 7.

To evaluate exposure detection performance, we set the combination with the best accuracy (94%), $R_{max}$ at 30 m, and $S_{max}$ at 0.5, considering the trade-off between $R_{max}$ and $S_{max}$. The performance of the exposure detection is shown in the graph of Figure 8.

We need to obtain the appropriate threshold $O_{threshold}$ to determine the close contact (see Figure 4). If the precision is low, excessive contact is detected, and if the recall is low, the actual contact may not be detected. Considering this, we set $O_{threshold}$ to 7 to achieve a balance between the two metrics. With this setting, precision showed 96% and recall showed 96%, as shown in

Table 4. Performance of trajectory-based outdoor GPS contact tracing.

Precision	Recall	Accuracy	F1 Score
96.31	96.60	97.50	94.94

.

5.3. Discussion

Direct experimental comparisons are complicated by the fact that GAEN is focused on confirmed case interactions, whereas IntelliTrace emphasizes co-location within the same space. Due to the proprietary nature of GAEN, which typically restricts development to a single entity per nation, direct experimental comparisons with GAEN are inherently limited. Therefore, access to GAEN’s API for developmental purposes is not universally available, limiting the capacity for empirical comparative studies. Furthermore, GAEN focuses on confirmed case interactions, whereas our approach emphasizes co-location within the same space.

Consequently, we resorted to a literature-based performance assessment.

Table 5. Comparative performance metrics of our study against contact tracing using GAEN [28].

	GAEN [28]	Our Study
Precision	70.00	Indoor	Outdoor	Overall
		96.68	96.31	96.49
Recall	47.00	78.82	96.60	87.71
Accuracy	79.00	88.41	97.50	92.95
F1 Score	56.24	86.84	94.94	90.89

presents our findings and the development of Germany’s official COVID-19 exposure notification app, the “Corona-Warn-App”, as detailed in [28].

In the landscape of digital epidemiology, South Korea’s use of QR code-based electronic entry logs (KI-Pass) has been instrumental in facilitating spatially informed responses to the pandemic [29]. Despite its utility, this system has presented challenges, including the inconvenience of manual QR code scanning by individuals, and a dearth of practical alternatives. Our IntelliTrace system offers an innovative and seamless alternative, circumventing the need for manual efforts, as evidenced by our performance comparison with the GAEN system, as outlined in Table 5.

Wi-Fi adoption continues to grow globally, with installations in high-traffic public areas becoming increasingly common [30]. Nonetheless, the assumption of ubiquitous Wi-Fi availability in our system could restrict its use in areas without such infrastructure. Furthermore, the proposed method’s efficacy is also contingent on the reception quality of BLE beacon signals, which are essential for the labeling of spatial information. There is a need for further research to refine the labeling process, ensuring that the data used to train and validate our system is as precise as possible.

6. Conclusions

As the COVID-19 pandemic continues, there has been a surge in research focused on contact tracing. We introduced IntelliTrace, a novel contact tracing method that incorporates co-location—being in the same space or same trajectory—with the consideration of the transmission characteristics of the virus. We employed a space-based contact tracing method using Wi-Fi for indoor spaces and a trajectory-based contact tracing method using GPS for outdoor spaces. The indoor Wi-Fi contact tracing method utilized values of Euclidean, Cosine, and Jaccard similarities derived from the fingerprint data of Wi-Fi access points as features. Furthermore, machine learning models were employed to detect exposure by making decisions regarding the co-location of confirmed cases and potential contacts in the same space. The outdoor GPS contact tracing method relied on calculating distances between individuals and similarities between trajectory vectors using GPS data. Next, we evaluated the performance of our proposed method using a validation dataset, which was gathered over a one-month experiment involving 28 participants. IntelliTrace demonstrated an F1 score performance of 86.84% in indoor environments and 94.94% in outdoor environments. Through comparative evaluations, we have substantiated the efficacy of our approach, demonstrating its superior performance in relation to established methods. The proposed algorithm is expected to contribute to digital means of containing the spread of infectious diseases in the future, especially when combined with other methods such as distributed data storage and anonymization of personal information.