SenseX: Design and Deployment of a Pervasive Wellness Monitoring Platform for Workplaces

Abstract

With the increasing number of desk jobs, workplaces have become the epicentre of several health risks. In this paper, we design and develop a pervasive wellness monitoring platform, SenseX, that uses a variety of devices and sensors to track physical activity level of employees in an organization. SenseX platform offers APIs which can be used by 3rd party applications to create services and applications, which can focus on specific interventions (e.g. to reduce prolonged sitting). We performed a real-world evaluation of the platform by deploying it in an IT organization for 6 weeks and observed longitudinal variations. We believe that SenseX platform helps in realizing the vision of “wellness as a service” in modern workplaces, enabling multitudes of different wellness services, which will be a key for sustained adoption of wellness programs.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

With the proliferation of computers and information technology in the last two decades, the number of desk jobs have grown at a phenomenal rate. In the United States (US), less than $20\,\%$ of private sector jobs have moderate levels of physical activity, decreasing by nearly $30\,\%$ compared to the early 60s [9]. Similarly, nearly 4 out of 5 people have desk jobs in the United Kingdom. A survey done in US reports shows that a typical worker spends 7.5 h sitting at work, 8 h sleeping, 4.5 h watching television or at home computer including leisure time, 1 h eating and only 3 h physically active or standing, i.e. sedentary for 21 h out of 24 h everyday [7]. Many studies have identified prolonged sitting as a high risk factor for severe health problems such as diabetes, cancer, heart attack, and stroke. For instance, [8] reports that adults who have moderate-to-high amounts of sitting time (four hours or more) have significantly higher cardio-metabolic risks compared to those who have lesser sitting time (less than three hours). Also, it is also found that the production of enzymes that burn fat declines by as much as 90 % after one hour of continuous sitting [5]. Additionally, it has been found out that excessive sitting results in depression, lower life expectancy, larger waist circumference and slower metabolism and over a term the harmful effects of sitting keep increasing. Surprisingly, researchers have found that regular exercise and balanced diet do not negate the adverse effects of prolonged sitting [4].

It is clearly evident that most of the sedentary behavior is found in workplace environments. Unarguably, they have become the epicenter of serious health risks and as a result, organizations have started investing in wellness programs. Workplace wellness is a $\$6$ billion industry in the United States alone where majority of organizations spending at least $\$521$ per employee per year [1]. Organizations are investing in these wellness programs to improve social, mental, and physical health of their employees as well as to reduce their healthcare payback costs. Most of these wellness programs incorporate costly wearable devices such as Fitbit, Nike Fuel Band, and Jawbone UP for activity tracking with physiological attributes. However, these wearables do not result in sustained adoption due to their obtrusiveness. For the users, it is more of an overhead to carry, wear and maintain an extra device. A study in US shows that more than $50\,\%$ of consumers who owned an activity tracker stopped using them after 6 months [6]. Similar to activity trackers, standing workstations have also failed to have sustained impact and most users stop using those desks after one month of use [2]. Additionally, there are many playstore apps such as Motion24x7, Moves, Google Fit performs activity recognition using sensors available on the smartphone itself. However, these apps too fail to get a sustained adoption due to factors such as limited coverage i.e. users tend to leave their mobile phones stranded on desks while at work, and high battery consumption, which is due to nonstop sensing. Hence, there is a need for an unobtrusive, frugal, and pervasive platform for wellness monitoring and interventions in workplace environments.

In this paper, we design a pervasive wellness monitoring platform SenseX which leverages an employee’s everyday devices and existing infrastructure (i.e. interconnected desktop/laptop, enterprise WiFi) for activity tracking and physiological measurements (i.e. heart rate). SenseX builds services to interface with different workstation devices (i.e. keyboard, mouse, webcam), workplace infrastructure (i.e. enterprise WiFi, calendar, BLE), and mobile phone sensors to sense activities and context of an employee. SenseX employs intelligent sensing approaches such as triggered-sensing i.e. offloading sensing load to infrastructure sensors whenever possible and opportunistic-sensing i.e. switching a sensor ON only when there is an opportunity to sense. Such approaches help in minimizing energy-consumption and increase sensing coverage. SenseX processes the device-specific measurements to infer fine-grained activities of a user as well as to extract high-order contextual information. A cloud-based SenseX service is used for the fusion of multiple device-specific activity profiles into a single profile, which is used for pushing appropriate interventions and notifications. Further, SenseX platform offers APIs which can be used by 3rd party applications to create services and applications, which can focus on specific interventions (e.g. to reduce prolonged sitting) to induce systematic behavior changes. We believe that SenseX platform helps in realizing the vision of “wellness as a service” in modern workplaces, enabling multitudes of different wellness services, which will be a key for sustained adoption of wellness programs. Specifically, this paper makes the following contributions:

• 1. First of its kind pervasive wellness monitoring platform which intelligently combine mobile, workstation, and infrastructure sensors to perform activity monitoring in workplaces.

• 2. Intelligent sensing approaches i.e. triggered-sensing to increase sensing coverage and reduce battery consumption. SenseX provided multitude of services and APIs to enable quick development of workplace-specific intervention and challenges.

• 3. A real-world deployment with 30 participants and detailed evaluation of SenseX platform with StandUp wellness challenge.

2 Related Work

The enhancement in the sensing capabilities of computerized devices along with the advancement in the area of ubiquitous and pervasive computing has resulted into the evolution of many fitness and healthcare applications/systems. We can broadly divide the previous work related to ours in three categories namely, Mobile-enabled sensing, Workstation-based sensing and Wellness platforms and services.

Mobile-Enabled Sensing. Modern-day smartphones are computationally powerful and at the same time are equipped with array of sensors, and thus have been widely used for day-to day sensing, activity recognition and health monitoring systems. Kwapisz et al. [10] proposed and developed a system for android devices, which could recognize, log and maintain simple physical activities like walking, jogging, sitting, standing etc. using phone based accelerometer. Similarly smartphones have been used widely for activity recognition [11, 13, 14]. Physiological parameter monitoring is yet another focus area in which researchers have exploited the capabilities of the smartphone. Aishwarya et al. [21] proposed smartphone based methodology to estimate the range of human blood pressure (BP) using Photoplethysmography. Similar to this [22] discuss a system which enables tracking the heart rate with user’s finger tip placed on smartphone’s camera. Sumida et al. [19] correlate the accelerometer, walking speed readings with the heartrate and estimate its variation while walking. There is no dearth of mobile-based systems designed with physiological parameter monitoring capability using integrated external sensors [25] and only internal sensors like camera etc. [23, 24].

Workstation-Based Sensing. Although workstations generally do not have any dedicated sensors, but many sensing hints can be obtained from inherent components like mouse, keyboard, touchpad, webcam etc. Thus researchers have exploited these sensing capabilities of workstation for measuring a number of health-related parameters like heart-rate, stress levels, emotional state etc. Sun et al. [15] suggest and demonstrate the use of mouse events to detect stress. Epp et al. [26] show that emotional states can be recognized using keystroke features. On the other hand [16–18] discuss and exhibit the ability of webcam equipped workstations to predict heart-rate.

Wellness Platforms and Services. The growing awareness and realization regarding the effectiveness of preventive healthcare has motivated the research community to design and develop several dedicated and ubiquitous healthcare platforms/systems. [28] discuss a service-oriented architecture platform for integration of health-data from various personal health devices. Li et al. designed and implemented a cloud-based platform for personal health sensor data management in a collaborative manner. Similarly [30] discuss the implementation of a scalable, robust cloud-based platform which can manage the semistructured, unstructured, and heterogeneous physiological signal effectively and can satisfy high concurrent requests from ubiquitous healthcare services.

SenseX builds upon the research work done in mobile and workstation-based sensing outlined above. However, there is lack of a platform that provides pervasive sensing capabilities such as SenseX and flexibility to create new interventions/challenges especially in workplaces.

Fig. 1.

System design of SenseX platform with different layers.

3 System Design

Modern workplace environment consists of existing infrastructure of inter-connected desktops (i.e. laptop/PCs) equipped with web cameras, and employees carry their mobile phones with them. SenseX platform is designed to perform multi-modal sensing that encomapasses existing workplace infrastructure as sensor hints along with mobile-based sensors to accurately track activity levels as well as wellness of an employee. SenseX platform follows a layered architecture involving three different layers designed for sensing raw data, and subsequently extracting high-level attributes to enable appropriate interventions. Figure 1 presents a snapshot of proposed three layered design with each layer consisting of a set of components. These components are part of the various services that run on end-user devices (i.e. mobile, workstation) and a cloud instance. SenseX follows a decentralized model where most of the sensing and processing of data is performed on end-user devices and aggregation is performed using a cloud instance. Further, the rationale of having layered design is that it enables flexibility of adding or removing layers as well as components by third party applications and services. These services can use activity tracking and context inference capabilities of SenseX and build their own wellness management and intervention solutions.

3.1 Sensing and Actuation Layer

This layer is responsible for interfacing with hardware devices and sensors to collect sensory measurements. Some of the primary components of this layer are device-specific integrations, sensor controller and scheduler, data collector, and a rule engine. SenseX using following device-specific integrations.

• 1. Mobile Phone: Mobile phone is equipped with a variety of different sensors such as location, light, proximity, accelerometer, gyroscope, microphone, etc. SenseX has a dedicated service running on the mobile phone, which can start/stop sampling any of these sensors based on the application requirements. All these sensors provide raw data values, which are later processed in higher layers to infer user-specific activity and contextual attributes. For example, accelerometer sensor data capture provides x, y, and z coordinate values corresponding to the motion in three axes/dimensions of a mobile phone while location sensor can provide the geo-coordinates or WiFi/GPS/Bluetooth fingerprint of user’s current location.

• 2. Workstation Devices: Many people use desktop/laptop in their workplaces. SenseX runs a workstation service to capture different sensing cues from the workstation devices including activity on mouse/keyboard, name of foreground application (s), and webcam video feed. Based on the application requirements, SenseX service subscribe to one or more of these cues from the OS to collect raw sensor data. For example, an OS-based interrupt is generated whenever the keyboard or mouse is used. SenseX listens for such interrupts and records them with their respective time stamps. This data is processed in subsequent layers to infer high-order activities i.e. to find if a person was sitting or not in a given time interval. The main assumption is that if there is a steady stream of keyboard and mouse interrupts, the user is sitting continuously. Similarly, webcam-based video feed can be recorded as part of sensing, which could be used for simply detecting the presence of a person at workstation or to do more complex ones such as extracting physiological parameters.

• 3. Infrastructure-Based Devices/Services: Most of the organizations use different hardware sensors and software services as part of their day-to-day functioning. Some of these services include RFID card for access control, enterprise WiFi for ubiquitous network access, Bluetooth low energy (BLE) devices for tracking intra-organization movements, organization-wide calendar services for maintaining schedules, and messenger services for communication. SenseX provides customized services to gather sensor cues from available organization-wide sensors and services. For example, cues from calendar service can indicate the schedule of employees whereas enterprise WiFi or BLE can be used for localization of employees in the workplace environment.

Sensing Service Controller and Scheduler. Some of the above sensors are battery-constrained (i.e. mobile sensors) whereas some of them have near continuous power supply (i.e. workstation devices). One of the main responsibilities of this layer is to control different sensors based on the application requirements, posed by higher layers. This module also identifies opportunities where battery-constrained sensors can be complemented with the data from infrastructure or workstation sensors. In a nutshell, it employs a novel triggered-sensing based approach to dynamically start/stop a device-specific service whenever there is an opportunity to use infrastructure-based sensors. The main idea behind triggered-sensing is to reduce sensing data redundancy as much as possible, thus improving battery life.

We present a use-case to demonstrate applicability and effectiveness of triggered-sensing based approach. For example, SenseX need to monitor a user Alice’s sitting and standing patterns to provide an intervention for excessive sitting. Lets assume that Alice goes to her office by foot. While walking, SenseX mobile service will remain ON to keep track of Alice’s activities. After, Alice reaches her workplace and starts using her workstation devices (i.e. keyboard, mouse), a trigger is generated by SenseX workstation service to the cloud instance and it pushes a notification to mobile service for switching off mobile sensors as shown by the first trigger in Fig. 2. Meanwhile, workstation service keeps reporting the current activity (i.e. sitting) to the cloud instance. Whenever, there is an activity change detected by the SenseX workstation service i.e. there is no activity for a certain time, it starts sampling the webcam feed as represented by second trigger of Fig. 2. If it concludes that Alice is not present on workstation, then cloud instance send a trigger to mobile service to resume activity tracking.

Another advantage of triggered-sensing based approach comes in the form of increased sensing coverage where different sensing schemes can complement each other in terms of data collection. Consider the above scenario, Alice may place her phone on the table while working on the workstation and in such a case, a system using only mobile sensors will not be able to sense current state/activity. However, SenseX can sense the activity of Alice based on the usage of workstation devices. Further, the applicability of triggered-sensing based approach is not restricted to above described use-case only and it can be easily extended to a variety of other scenarios. For example, SenseX may use enterprise WiFi-based tracking for finding current location of a user rather than actively scanning mobile-based WiFi.

Fig. 2.

An instance of triggered-sensing based approach employed in SenseX. Red color markers shows different triggers, which initiates other sensors (Color figure online)

Rule engine component translates higher order activity inference requirements into low-level sensor mappings. For instance, if an application wants to track sitting-standing pattern of an employee, it enables mobile service and workstation-based sensors to track activity patterns. One of primary challenges faced by SenseX platform is in terms of heterogeneity that includes different devices and lack of a standardized representation in sensing data. Sensing and actuation layer handles device heterogeneity by implementing different sensing interfaces for different devices and provide ability to control as well as collect sensory data from them. However, data heterogeneity still remains a challenge and higher layers as presented in Fig. 1 provide a solution for the same.

3.2 Activity Inference and Fusion Layer

The primary responsibility of this layer is to infer various high-level activities (i.e. sitting, standing, walking) as well as contextual attributes (i.e. inMeeting, Indoor/Outdoor, DeviceOnTable) from the raw data collected by sensing and actuation layer. After sensor/device specific inference, the different activities are fused together to create a wholistic activity profile for every user. Additionally, this layer hosts components to estimate physiological parameters from the sensing data.

Activity and Context Inference. This component aims to recognize high order activities from raw sensor measurements. Some of the activity inferences are straightforward i.e. if a person is typing on a keyboard/mouse or her presence is detected using webcam, her activity can be classified as ‘sitting’. However, some of the activity inferences need processing and in most of the cases require three different steps i.e. pre-processing of sensor data, extracting features, and applying a classification scheme to recognize the activity. Due to significant heterogeneity among sensed data values in SenseX, these processing steps need to be implemented distinctively for each device/sensor category. For example, accelerometer sensors provide X, Y, Z values representing motion in three axes/dimensions and it can be used to infer user activities such as sitting, standing, walking, climbing stairs, etc. The pre-processing steps involves dividing accelerometer data stream into segments (say x seconds) based on application and accuracy requirements. For each segment, it calculate set of statistical (mean, standard deviation, skewness, RMS), time-domain (integral, auto-correlation) and frequency-domain (spectral energy, spectral entropy) features. These feature values are then fed into a classification framework such as decision tree or SVM to recognize users’ activity.

Contextual information is an integral part of a wellness platform and it enriches the user-specific activity profile as well as notification delivery/intervention mechanism. For example, notifications could be smarter and will not be delivered when a person is in meeting. Some of the contextual information is easy to sense, for example, context information of a person being in a meeting or not, can be easily sensed from the calendar information. However, some of the contextual attributes such as detecting mobile phones’ presence in pocket require processing and fusion of multiple sensors data. For example, to detect whether the phone is in pocket or not, SenseX uses accelerometer sensor readings to find angle of inclination which is fused with proximity sensor information to classify whether a phone is in pocket or lying flat while facing up or lying flat while facing down, etc. SenseX also interfaces with enterprise WiFi network and BLE technology to infer current location of the user [20].

Activity Fusion. Aforementioned components infer activity and context from device-specific sensor measurements and convert data representation from low-level to high-level activity or contextual constructs. SenseX utilizes different kinds of devices/sensors for activity and context tracking and most of higher order inferences (activity recognition) are performed on the device itself, while periodically reporting to SenseX cloud instance. This module perform fusion of activity constructs coming from two or more different devices/sensors. For example, the SenseX workstation service may detect that a person was sitting from 09:00 to 09:44 AM where as mobile service infers that the person was walking from 9:45 to 10:15 AM. This information is fused to make a single activity profile. In some of the cases, time information across two different activities may overlap due to errors in inferencing. In such cases, this module assigns higher priority to the infrastructure and workstation based sensors.

Physiological Analysis. Considerable research has been done to extract physiological parameters such as heart rate from activity data [19] as well as webcam-based video feed [18]. Further, there have been efforts to use webcam-based video feed to detect respiratory problems, emotions, etc. The focus of SenseX platform is to use existing research work as a pervasive mechanism to measure vital signs in day-to-day life. Heart rate measurement is an important vital parameter to diagnose various ailments such as anxiety, stress, cardio-vascular disease, etc.

Opportunistic Sensing. This component works in-conjunction with sensing service controller component to enable sensor tracking only when there is an opportunity detected. For example, physiological analysis does not work when there is motion w.r.t. subject’s face in the video. In case of SenseX, there could be two kinds of opportunity-based sensing. Based on the detection of an opportunity, this component sends trigger to sensor controller module to enable tracking.

• Context-directed/Detection: In this case, detection of contextual attributes decide whether sensing should be enabled or not. For example, an application may require sensing of heart rate just after lunch time. To enable such sensing, required context is continuously tracked and a notification is issued to the user when there is an opportunity.

• Sensing-directed/Prediction: This is a completely automated and non-obtrusive approach where sensing is dependent on end-result i.e. automatically figuring out the instances where there is likely to be little motion so that heart rate tracking is possible using webcam. SenseX achieves this by continuously tracking system usage logs (i.e. mouse activity, keyboard activity, open applications) and using this information to predict the right moments with the help of a supervised classifier.

3.3 Analytics and Notifications

This layer performs high-level analytics on user-specific activity profiles and contextual attributes. High-level analytics is performed to detect broad patterns across an organization or a group of people as well as personalized behavior specific to a user. Based on these patterns, appropriate notifications can be generated and issued to users at opportune moments. This layer also contains a user profile component which tracks longitudinal data of users’ response to the interventions/notifications, contextual information, privacy policies and the end result i.e. whether a user complied with the notification or not. The context tracking along with notification history captures specific behavioral attributes such as user X does not like to get disturbed while coding. This data in conjunction with behavior modelling component is used to employ the right persuasion strategies to motivate the user to comply with notifications. For example, an application dealing with prolonged sitting of employees can sense the sitting time using SenseX and accordingly, send notification in time to alert user to take break periodically. Based on the contextual attributes and longitudinal history, the notification may look like following:

“You have a meeting in “10” min and working from last 35 min. Please take a break now and walk for 2 min”.

4 Implementation Details

We developed SenseX platform with most of components described in Fig. 1. To demonstrate the efficacy of the platform for performing pervasive sensing and appropriate interventions, we developed an application StandUp to monitor activity levels of employees in an organization.

Mobile Service: It is a native OS service developed for Android devices due to its popularity and large market-share, especially in developing markets. SenseX mobile service is able to perform low-level sensory measurement i.e. accelerometer, proximity, microphone, etc. and contains activity inference algorithms to extract high-order activities and contextual attributes. StandUp is implemented as a native mobile application that accesses SenseX service to sense various activities (i.e. sitting, standing, walking, unknown, etc.). An activity sensed from mobile service is categorized as ‘unknown’ if the phone is kept on a flat surface (e.g. table). StandUp visualizes sensed activities in a day-based activity profile along with timeline as shown in Fig. 3a. Apart from day-based activity profile, users can see their weekly performance and an organization leaderboard where they can compare their performance with peers as shown in Fig. 3c. StandUp also uses SenseX intervention mechanism to provide engaging notifications to the users which provide “just-in-time” alerts for daily goals, compliance level etc.

Fig. 3.

a,b,c. Snapshots of StandUp Mobile Application d. Snapshot of workstation service ticker and notification

Workstation Service: SenseX workstation service is implemented using Microsoft’s .NET Framework 3.5 and supports all desktops/laptops running Windows OS. The service is configured to run whenever system is rebooted or awakens after sleep mode. It supports functionality to track keyboard activity, mouse activity, webcam-based feed, and front ground application usage. StandUp uses the workstation service to track activity of an employee and builds a simple windows form based application to visualize such activity levels. StandUp application places a ticker on the users’ desktop to show the current station of user activity as shown in Fig. 3d. Similar to mobile service, StandUp application uses workstation service to provide timely notifications in case of prolonged sitting as shown in Fig. 3d.

Cloud-Based Instance: The server is hosted on a VM (virtual machine) provided by a public cloud service provider. The end-device services such as workstation service and mobile service communicate with the cloud-based instance using HTTP Get/Post requests. SenseX uses Google Cloud Messaging (GCM) to push updates on mobile devices. StandUp application uses APIs to provide a web-based dashboard which can be accessed by organization administrator to track installation, status of running services, application usage, organization leaderboard, and broad patterns.

5 Evaluation

SenseX is a platform designed for pervasive wellness monitoring platform in workplaces; consists of different services running across devices. As described in Sect. 4, StandUp application uses these services to provide activity monitoring and interventions for prolonged sitting in workplaces. The system evaluation of a platform such as SenseX is important to establish the effectiveness of different features (i.e. triggered-sensing) where as a typical wellness monitoring system will be questioned on how much can it contribute towards increasing activity levels, especially in a workplace setting. Specifically, for SenseX and StandUp, we conduct a mix of both system-evaluation (i.e. measuring impact of triggered-sensing, sensing coverage) and a user-study evaluation to capture the personal characteristic of users. We frame following research questions to guide both of these evaluation mechanisms.

• R1: How much sensing coverage is provided by different services of SenseX? How does sensing coverage vary across different users of the system?

• R2: What is the impact of triggered-sensing on the SenseX system? How much is the trigger-delay among different services? Does it help in saving energy for battery-constrained devices?

• R3: What are the activity patterns (i.e. continuous sitting time, daily steps) of employees working in a workplace environment? How do these patterns fair when put along organization wide leaders and slackers?

• R4: What is the impact of nudging an employee to take a stroll in case of prolonged sitting? How long does she take to respond to these notifications?

• R5: What is the impact of gamification and incentivizing the top performing employees? What is the observed difference between leaders and slackers in the system?

Deployment Details: We did a pilot deployment of proposed platform in an IT organization with a total of 30 participants (25 males and 5 females). The ages of participants ranged from 21 to 45 years, with a mean age of 29 years (male = 28, female = 31). All the participants used StandUp mobile and desktop application which were built upon SenseX services. The participants were given flexibility to specify the time intervals where activity tracking can be performed on desktop and mobile applications; default interval was set from 7AM to 7PM. To make the challenge engaging and competitive, we announced awards for top 3 weekly leaders chosen based on their performance over two metrics, i.e. average step count and average notification compliance score. Initially, we planned to run the deployment for 4 weeks but extended it for another 4 weeks keeping in mind engagement and interest of the participants. During the pilot, SenseX logged several parameters related to participants’ performance and interaction with the system. We analyzed the collected data to answer aforementioned research questions.

5.1 Sensing Coverage

The sensing coverage in the context of SenseX is defined as the fraction of time where activity monitoring is possible in a participant’s day-to-day life. SenseX uses mobile sensors along with contextual and environmental information, sensed by the workstation service. Using logs, we characterized activity monitoring w.r.t. sources i.e. mobile, workstation, or unknown. An activity is said to be ‘unknown’ if mobile device happened to be placed on a flat surface and there was no activity sensed by the workstation service. Across all the participants in our deployment study, we observed that mobile service sensed nearly $50\,\%$ of the total time, nearly $25\,\%$ of the time was monitored by the workstation service and rest of the time was categorized as unknown as can be inferred from Fig. 4c. This brings an interesting observation that mobile-based wellness applications such as Moves and Google Fit are not able to track activities for more than $50\,\%$ of time. Further, Fig. 4c shows the sensing coverage pattern for the subset of participants who had both mobile and desktop services actively running throughout the deployment study. We observe that for some users (ID: 5 & 16), more information is obtained from workstation service as compared to mobile sensors. It could be due to limited monitoring because some of the users enabled activity tracking during workplace timings only. Our analysis show that workstation-based service increases the sensing coverage to a large extent especially during workplace timings.

Fig. 4.

a. Measured trigger-delay between mobile service to workstation service b. Battery decay comparison for both mobile-only activity tracking approach and SenseX based triggered-sensing with baseline c. Comparison of sensing coverage for mobile and workstation based services across different participants

5.2 Triggered-Sensing and Battery Consumption

Section 3 presents triggered-sensing approach to offload sensing responsibilities to workstation service when there is an opportunity. In such cases, workstation service with the help of cloud instance need to send a trigger to mobile service indicating that it should stop sensing. We use push-based notification system to send trigger to a mobile service instead of a pull-model where it has to continuously look for a trigger by sending repeated request to the cloud instance. It is important to characterize the time to send this trigger from workstation service till it reaches to mobile service, represented as trigger delay. We emulated trigger-delay using SenseX platform where a total of 1200 requests where sent from workstation service to mobile service for a duration of 8 h. We present the observed trigger-delay in Fig. 4a, most of the time the delay was less than 10 s. Similarly, we compared energy consumption of SenseX activity tracking services w.r.t. only mobile-based tracker and baseline as shown in Fig. 4b. In mobile-based tracking approach, a continuous activity tracker that uses accelerometer ran on a MotoE phone till the battery reached to a low level where as baseline shows the battery decay without any external service. We found that SenseX using its triggered-sensing approach could increase the battery life time by nearly $25\,\%$ of the time. From our experimental results, we conclude that triggered-sensing works in near real-time and helps in increasing sensing coverage as well as result in significant energy-savings.

Fig. 5.

a. CDF of day-wise steps across different participants b. CDF of sitting time sessions across different weeks c. CDF of the response time observed for all the notifications during the deployment d. Average number of leaderboard hits per user each week during the deployment

5.3 Workplace Activity Patterns

StandUp challenge was designed to increase the overall activity level of employees at a workplace. Figure 5a presents the CDF of the number of steps for each week across all participants. We see a clear shift among the first and sixth week, indicating that more people have increased their number of steps over the course of the pilot. Figure 5b shows the CDF of the sitting time of users, i.e., the time for which a user sits continuously at a stretch. We see that the duration of sitting segment of the users have marginally decreased from week 1 to week 6, which indicates that the users started adhering to the notifications and had been avoiding sitting continuously for longer durations. We conclude that StandUp challenge was effective in increasing the overall activity level among employees. Considering that, there was no effective way to capture baseline (i.e. participants’ performance before intervention), we have only provided performance comparison on week-to-week basis.

5.4 Effect of Notifications

One of the important aspects of StandUp challenge was to continuously monitor the sitting-standing pattern of the people and send notifications if they sit for more than a threshold time t. In case of StandUp, t was set to 40 min based on the prior medical studies and repetitive notifications were pushed every 5 min. The participants also had an option to stop the notification temporarily if they were busy. We compute a compliance score which is an indicator of percentage of time a participant has acted on notification or remain active at least 2 min in an hour. Figure 5c shows the CDF of response time for both workstation and mobile notifications. Response time is the time between when a notification is pushed and the user moves out of his sitting position (and proceeds to stand or walk). We see that for nearly half of the notifications, participants have acted in less than 10 min. We also observe that there is a slight preference of mobile-based notifications over workstation ones. Also, long tail of larger response times are observed because many time participants did not carry their mobile phone when they took a break.

5.5 Effect of Incentives and Gamification

With the help of SenseX analytics capabilities, StandUp maintained an organization-wide leaderboard which could be accessed by the participants. The leaderboard was established based on two metrics i.e. average step count and compliance score. At the end of each week, cash prizes worth 15 USD each were announced for the top 3 leaders. We did not collect any qualitative data to measure the impact of leaderboard or incentives. However, we collected quantitative data on how many times leaderboard page was accessed by different participants during the deployment and we used that as a proxy to capture the interest of participants. Figure 5d presents the average number of leaderboard hits per user each week, (i.e., the number of times the leaderboard was visited by a user). We see that initially, it was visited two times per day on average, and the frequency gradually reduced as the initial excitement of the pilot settled in. However, the leaderboard visits have remained stable over the last 3 weeks, converging at around one visit every day by each user on an average.

Next, we analyzed the performance attributes of leaders and slackers during the deployment study. The participants above the 95th percentile of step count are taken as the leaders and those below the 5th percentile of step count are taken as the slackers. It may be noted that the set of leaders and slackers may vary in each week. We see a clear correlation between the leaders/slackers and the average sitting time, response time, and leaderboard hits, i.e., the step count-based leaders also feature in the above-average category for leader board hits, and below average (i.e., shorter duration) for sitting time and response time, and vice versa for the slackers.

6 Discussion

Modern day workplace forces a sedentary lifestyle that involves reduced physical activity and prolonged sitting, which poses a risk of severe ailments including diabetes, cancer, heart attack and stroke. Current wellness solutions are costly, dependent on wearable devices, which do not have a sustainable impact. We designed and developed a comprehensive wellness monitoring platform i.e. SenseX, which intelligently uses combination of mobile, workstation, and infrastructure sensors to do pervasive activity and physiological monitoring in workplaces. With multitude of different services, SenseX enables easy creation of different wellness interventions and challenges that can drive sustained behavioral changes among employees. As an example, we created a wellness challenge i.e. StandUp to increase the activity levels of employees in an organization. We ran the challenge for nearly 6 weeks and observed several key insights i.e. increased daily step count across weeks, reduced sitting sessions, positive effect of gamification/incentives, etc. From a systems’ evaluation perspective, SenseX uses triggered-sensing approach to increase the sensing coverage by nearly $25\,\%$ and reduced the battery consumption considerably.

Through its pervasive sensing capabilities, we believe that SenseX helps in realizing the vision of “wellness as a service in modern workplaces, enabling plethora of different wellness services and interventions”. We provide a sample of some of the other challenges/interventions that can be created using SenseX.

• 1. Vitamin D Challenge: IT workers suffer from lack of vitamin D, which is obtained by sun-exposure. A service can be created using SenseX activity monitoring and contextual inference capabilities that infer the presence of a person in outdoor and uses weather data to estimate the vitamin D exposure.

• 2. Healthy Meetings: A significant amount of sitting time in an IT workplace happens during the meetings. SenseX can be used to create a challenge to have meetings where people opt to stand, which can be compared across different departments in a workplace.

In essence, SenseX can help in creating many such challenges, which will drive sustained adoption of wellness programs. The layered architecture of SenseX can be used to enhance specific components, for example, accurate behavior modelling will help in enhancing the response time of the platform notifications. Similarly, pervasive sensing capabilities of SenseX could be enhanced to measure conditions such as sensing stress using keyboard and mouse activity.