Actitracker - Fordham University Computer and Information Sciences
Contents
1. 5 agree Figure 6 Response to app provides accurate predictions 60 50 40 30 20 A gt m 1 2 3 4 5 balanced under predicts over predicts Figure 7 Response to app predicts proper amount of activity 4 3 4 3 3 9 3 7 3 3 l 1 T T 1 Standing Walking Stairs Jogging Sitting Lying Figure 8 Response to app accurately predicts specific activity for example the timeline perfectly indicates when there is a long walk with virtually no misclassifications during the walking peri od Similarly one may walk up and down between levels in one s home and the short episodes of stair climbing are nonetheless correctly reflected in the timeline chart However there are other times when basic activities are mistaken and these results are glar ing We believe that a major source of such errors is due to chang ing orientations of the phone within the pocket which are not adequately taken into account in our current system Solutions to this problem are discussed in Section 7 Some of the specific comments that we have received also shows that the system is able to properly reflect a mixture of ac tivities For example the activity of snow shoveling appears as a mixture of walking and standing Most users agree that training yields much improved results and that the results are quite poor without2. via the for mation of higher level features The transformed data is then passed to a classification model which determines the activity corresponding to that 10 seconds of accelerometer data Actitracker currently identifies five activities walking jogging stairs up or down standing and sitting lying down Sitting and lying down are combined because it is very difficult to distinguish these two activities from a smartphone positioned in one s pocket If the user has provided sufficient training data through the Actitracker self training process then a personalized activity recognition classification model is applied otherwise an imper sonal universal classification model is applied The results are stored in a database which is subsequently accessed by the web interface to provide activity reports to each user through their personal Actitracker account The details of these system compo nents and the processing are described in the next few sections the smartphone client is described in Section 3 the web interface is described in Section 4 and the data mining process compo nents which implement activity recognition are described in Section 5 The high level system architecture shown in Figure 1 which is consistent with our current implementation involves several im portant design decisions and tradeoffs First virtually all function ality is offloaded from the phone to the server This simplifies the design of the client
3. Strategies for addressing many of the weaknesses are briefly mentioned in this section or are discussed in Section 7 under future work 6 1 Usability Usability relates to the ease of use and learnability of a human made object Actitracker was designed to be easy to use ideally you just set it and forget it Overall our responses indicate that Actitracker is an easy tool to learn This is probably due to the extremely simple client interface and the quality documentation However there were two key issues related to ease of use which involve the positioning of the device on the body and battery us age i e power consumption 6 1 1 Body Location of Phone The original activity recognition research that was done in sup port of Actitracker assumed that the phone was located in the front pants pocket 7 10 and we do recommend that the user carry their phone in that location If they do not the results could be compromised although our training mode will allow other body locations to work if the movement at that location is suffi ciently different to distinguish between the various activities The question of where and how people carry their smartphones has been studied 2 5 and quantitative statistics have been cap tured While many factors can come into play e g gender age nationality our user experiences are quite simple the recom mended phone location is not an issue for men but is a huge issue for women Both
4. about battery drain and provides a space for free form comments The compre hensive survey is meant to be filled out only occasionally and asks about twice as many questions including some higher level questions about the usefulness of the service The survey results help us to describe and quantify user experiences in Section 6 Users are prompted upon registration to fill out a profile that includes their gender age height and weight Some of the profile information is used to determine the user s peer group so that we can compare that user s activity results with those in the same peer groups based on age gender and body mass index in the comparison charts described in Section 4 2 2 In the future user profile information may also be used to improve activity recogni tion performance since this information could be incorporated into the impersonal universal activity recognition models The demographic information collected in the user profile can also be used to advance our research on identifying personal traits like gender height and weight from accelerometer data 19 4 2 Activity Results The primary responsibility of the web interface is to visualize activity information for the user This is controlled through their activity dashboard 4 2 1 Activity Dashboard The central component of the web interface for each user is the activity dashboard The reports area takes up two thirds of the dashboard This includes a disp
5. ated reports only via a web interface Thus the Actitracker app provides no facility for viewing results Users can view the results by logging into their secure web account from their computer or from their smartphone While a mobile web interface is provided and usable the preferred method is to view results from a com puter with a larger screen since this permits the best viewing of some of the more comprehensive reports like our timeline 3 Smartphone Client Application Actitracker currently supports Android smartphones and the Actitracker client app is available for free from the Google Play store Support for the iPhone should arrive in June 2014 The central responsibility of the smartphone client is to poll the accel erometer sensors and transmit the data to the Actitracker server in the future we will include the gyroscope sensor if it exists and the GPS sensor As mentioned earlier and discussed in Section 7 additional functionality may be migrated from the server to the client over time The client interface is extraordinarily simple and has only three major components a button to toggle the service on and off a button for adjusting settings and a button to enter training mode First time users are prompted to create a secure account on the Actitracker server this account is used by the phone to submit data and by the web server to authenticate users attempting to view their data Users may also create accounts from the
6. but increases the load on the server and hence impacts the scalability of the system The primary reason for do ing things this way is that it enables the server to receive and store all raw sensor data This is useful because the raw data can then be shared with other researchers 20 and alternate feature encod ings can be applied at a later date Given that users supply some labeled data via the self training mode we can evaluate such al ternate encodings with minimal effort even long after the data have been collected Nonetheless because these advantages are mainly for researchers we are in the process of migrating the preprocessing and transformation code to the client app we can then enable this optionally via our administrative interface By doing this we may save battery life on the phone because even though the processing demands will increase the amount of data that will need to be transmitted using the phone s radios given our current encoding scheme will decrease to 2 9 of the current size 9 We have also thought about moving additional function ality to the client These alternatives and their pros and cons are discussed in detail in our prior work 18 At the furthest extreme all of the functionality is migrated to the phone including model induction at that point the system is perfectly scalable but re searchers lose access to all data and results Another design decision was to provide the results and associ
7. dashboard will display a con gratulatory message 4 2 4 Achievements The activity dashboard includes a panel of fitness achieve ments for the user Lifetime achievements allow the user to view achievements based on their lifetime usage of Actitracker The se achievements include total calories burned total hours using Actitracker and total hours expended performing each activity The best achievements are geared to the performance over a single day and include maximum time walking jogging and climbing stairs Because sedentary activity can be detrimental to health in the future we will add an achievement for least time sitting We also plan to add medals to a user s account for significant achievements and provide the ability to link these to the user s Facebook account to garner additional social support More creative achievements will also be added such as one for people who go a week without sitting continuously for more than 20 minutes at a time 5 DATA MINING PROCESS The Actitracker server implements and automates the entire da ta mining process for supporting activity recognition Every step from data cleaning and preprocessing to model induction and activity prediction is automatically performed by the server as it receives data from client phones without human intervention 5 1 Data Segmentation and Quality Control Each data connection from a client leaves the server with an ordered list
8. of accelerometer records a timestamp and x y and z acceleration readings collected over a period of several minutes or hours These lists of data are first divided into 10 second non overlapping segments a simple task since the client sends data in multiples of 10 seconds This has several advantages First it allows us to use standard classification techniques which require discrete examples Second this allows us to handle each example independently which facilitates parallelization Finally the use of discrete examples enables clients to connect and submit data peri odically without requiring data from separate connections to be matched up during processing Due to the way android smartphones provide accelerometer da ta to apps such as ours repeat and null values must be expected Usually they are the result of the operating system or hardware being too busy to provide new sensor readings at the requested rate We implemented Algorithm 1 in order to allow a 10 degree of tolerance for bad sensor readings as long as 90 of the ALGORITHM 1 end time r i time duration while r i time lt end time if is null r i amp amp is repeat r i example add r i itt if example r count gt 0 9 rate duration process example readings are good an example is generated We find that in prac tice this algorithm discards very few examples 5 2 Feature Generation The raw accelerometer records in each exampl
9. this by making our models more flexible so that they can adapt to differ ent body locations This can be done by having users provide training data while the phone is in different locations which can enable the models to learn to identify the phone location We will also address this issue by providing support for alternative sen sors such as those located in smartwatches We are currently working on integrating the Pebble smartwatch into Actitracker and hope to release a Pebble app The movement of the phone as a person moves has also been recognized as a problem We plan to address this by either coming up with features that are not so sensitive to the phone s orienta tion or by explicitly compensating for changes in orientation which can be determined by the accelerometer and gyroscope We have also found that battery power is an issue although this was not totally unexpected given the resource limits of mod ern smartphones We believe that we can improve the energy efficiency of our app in many ways such as using an adaptive scheme that does not poll the sensors as frequently when the user i e phone appears to be stationary We are also investigating the impact of moving the data preprocessing and transformation steps to the phone which will greatly reduce the amount of power needed for data transmission We have already planned some additional significant en hancements to the system which should make it more effective Th
10. training We do not have much information on the performance using the universal models in the wild because we recommended that all users utilize the training mode as soon as possible But our study in a controlled environment shows that the personal models greatly outperform the impersonal models 10 One of the metrics that we created was the FitDex which as described in Section 4 3 summarizes a user s total level of activi ty in a single number When asked whether this single number appropriately summarized their total amount of activity the mean response using our normal 5 point scale 1 disagree and 5 agree was 3 5 Overall we have found that the FitDex does tend to cor relate with how active one is each day but the values are poor when one runs Actitracker for only part of the day This is a con sequence of our current assumption which is that a user is active only when the tool is running and confirms that they are active We could address this by setting the FitDex relative to the number of hours it is collecting data but we wish to keep the current scheme to encourage users to run Actitracker all day Thus we plan to modify the service to apply a threshold and only generate the value when that threshold has been met 6 3 Objective Assessment of Performance Our prior research was able to measure the accuracy of our ac tivity recognition models with relative ease because the experi mental setup was very precise subje
11. transmission settings The Actitracker training mode screen enables a user to provide labeled activity data so that the system can then automatically generate a personalized activity recognition model which will override the default impersonal universal model The training phase is highly recommended since our research studies summa rized later show that personal models far outperform the imper sonal ones 10 The training mode guides the user to progress though the set of activities A timer is started after the user selects start at which point the user is expected to perform the activity until the time expires and the phone emits a siren sound The user can ignore the siren and continue to collect additional training data and they can come back at a later time to generate additional training data As long as a minimum amount of data is available for each of the required activities jogging and stairs are not required activities a personal model will automatically be creat ed and enabled For most activities we require only 2 or 3 minutes of data although our research has shown that good results are possible with even less data 10 As described in Section 5 the labeled training data is sampled in a manner that ensures that class imbalances do not unduly bias the induced model 4 Web Interface The web interface provides each user with access to the Actitracker results It allows the user to obtain docu
12. 6 Physical activity all cause mortality and lon gevity of college alumni New England Journal of Medicine 314 605 613 13 Physical inactivity a leading cause of disease and disability warns the World Health Organization http www who int mediacentre news releases release23 en 2002 14 Powell K E Thompson P D Caspersen C J and Kendrick J S 1987 Physical activity and the incidence of coronary heart disease Annual Review of Public Health 8 253 287 15 Reed B Wearable computer shipments seen hitting 64 million in 2017 Yahoo News October 25 2013 http news yahoo com wearable computer shipments seen hitting 64 million 2017 014532724 html 16 Siscovick D S LaPorte R E and Newman J M 1985 The disease specific benefits and risks of physical activity and exercise Public Health Reports 100 180 188 17 United States Public Health Service Office of the Surgeon General et al 1996 Physical Activity and Health A Report of the Surgeon General Government Printing Office 18 Weiss G M and Lockhart J W 2012 A comparison of alternative client server architectures for ubiquitous mobile sensor based applications In Proceedings of the Ubicomp 2012 Ist International Workshop on Ubiquitous Mobile In strumentation Pittsburgh PA 19 Weiss G M and Lockhart J W 2011 Identifying user traits by mining smart phone accelerometer d In Proceedings of the Fifth Internati
13. 780 5 Ichikawa F Chipchase J and Grignani R 2005 Where s the phone A study of mobile phone location in public spac es In Proceedings of Mobility 2005 Conference on Mobile Technology Applications and Systems 797 804 6 Koplan J P Liverman C T and Kraak V I 2005 Pre venting childhood obesity health in balance National Acad emies Press Washington DC 7 Kwapisz J R Weiss G M and Moore S A 2010 Activi ty recognition using cell phone accelerometers ACM SIGKDD Explorations 12 2 74 82 8 Leonard H There Will Soon Be One Smartphone For Every Five People In The World Business Insider February 7 2013 http www businessinsider com 15 billion smartphones in the world 22013 2 9 Lockhart J W Weiss G M Xue J C Gallagher S T Grosner A B and Pulickal T T 2011 Design considera tions for the WISDM smart phone based sensor mining ar chitecture In Proceedings of the KDD Fifth International Workshop on Knowledge Discovery from Sensor Data San Diego CA 25 33 10 Lockhart J W and Weiss G M 2014 The benefits of per sonalized models for smartphone based activity recognition In Proceedings of the 2014 SIAM International Conference on Data Mining 11 Oberman A 1985 Exercise and the primary prevention of cardiovascular disease American Journal of Cardiology 55 10D 20D 12 Paffenbarger R S Jr Hyde R T Wing A L and Hsieh C C 198
14. Actitracker A Smartphone based Activity Recognition System for Improving Health and Well Being Gary M Weiss Jeffrey W Lockhart Tony T Pulickal Paul T McHugh Isaac H Ronan Jessica L Timko Department of Computer amp Information Science Fordham University 441 East Fordham Road Bronx NY 10458 gweiss lockhart pulickal mchugh ronan timko cis fordham edu ABSTRACT This paper describes Actitracker a smartphone based activity monitoring service that helps people ensure that they receive suf ficient activity to maintain proper health Unlike most other such services Actitracker requires only a smartphone no watches bands or clip on devices This free service allows people to set personal activity goals and monitor their progress toward these goals Actitracker uses data mining to generate its activity recog nition models It initially uses a universal impersonal model that is generated from labeled activity data from a panel of users but will automatically generate and deploy much more accurate personal ized models once a user completes a simple training phase De tailed activity reports and statistics are maintained on the Actitracker server and are available to the user via a secure web interface Actitracker has been deployed for several months and currently has over 250 registered users This paper discusses user experiences with the service as well as challenges and tradeoffs associated with building and deploy
15. acker accuracy for the three subjects also identified where most of the errors oc cur Subject 1 had the walking and standing activities occa sionally classified as sitting lying down These errors were spo radic and were generally within long periods of correct classifica tions so a user would most likely ignore or barely notice these errors in the timeline view Subject 2 sometimes had walking mislabeled as stairs which is the most common misclassifica tion for the walking activity 10 but also had a very substantial amount of sitting misclassified as standing Further analysis showed that these errors came from the end of the day when the subject sat in a chair that is different from his normal chair This other chair accounted for 3 hours and 23 minutes of sitting and this activity was mislabeled for 1 hour and 32 minutes of that Table 2 Actitracker Activity for 3 users over a day Subject Accuracy 1 91 2 2 79 5 3 98 1 Ave 89 6 time this one error explains most of subject 2 s errors Subject 3 had a very high accuracy and most of the errors came from mis classifying short periods of walking as lying down sitting This evaluation process was quite informative although it does place a burden on the subject We are expanding our number of test subjects and are trying to simplify the process of keeping the diary We will develop a simple app to
16. assist with this which will let a subject select an activity from a list of high level activities e g dishwashing and basic activities e g walking The app will then append a timestamp automatically which will be far more precise than the current method that only tracks the time at the level of minutes This app would however have to be run on a second phone or other device to avoid interference with the normal activity recognition process 7 CONCLUSION AND FUTURE WORK This paper described a deployed smartphone based activity monitoring application called Actitracker This service provides the user with an accurate assessment of the amount of activity that they are performing and allows them to monitor the impact of any behavioral changes The goal of this service or tool is to improve people s health by combating the many harmful diseases and con ditions associated with inactivity Thus this is a tool that uses data mining for social good The collected data will also be made available to researchers and can also be used for large scale epi demiological studies of activity thus providing additional bene fits for society We faced many challenges while developing and deploying this tool and learned several lessons which we plan to use to improve the tool One key lesson that we learned was that some people mainly women rarely carry their phone in their pocket This impacts the utility of our service We hope to address
17. atement and 5 indicates that they agree with the statement A 3 is interpreted as meaning that the app results are often correct i e more than 50 of the time but do not fully meet expectations Figure 7 shows the re sponses to the assertion Do you think that the app accurately measures the overall amount of physical activity that you do In this case the scale is set so that 3 corresponds to the ideal value balanced whereas 1 corresponds to under predicting the total amount of activity and 5 corresponds to over predicting it Note that one can misclassify some activities while providing an accu rate assessment of the overall amount of physical activity The results in Figure 6 and Figure 7 are quite positive but show that there are errors and that there is room for improvement The results concerning the accuracy of the specific activities are summarized in Figure 8 The assertion was I feel that my _ results were accurate where the blank was filled in with the ap propriate activity The scale was the same used for Figure 6 where 1 corresponded to disagree and 5 to agree Only the averaged values are shown The survey results show that the service does generally provide good results but with room for improvement Specific comments from users indicate that often the results are highly accurate 35 30 25 20 15 10 Ea 0 T T T 7 1 3 4 1 disagree 2
18. cts were told to perform specific activities for specific amounts of time in specific loca tions 7 10 It is extremely difficult and time consuming to gen erate such objective performance metrics in the wild i e in a fully natural setting Nonetheless such objective metrics are important even given the user satisfaction metrics from the sur vey results described in Section 6 2 In this section we discuss the performance results for three Actitracker subjects who kept a detailed diary of their activities for a 12 hour period These sub jects had sufficient training data so that they each had personal Table 1 Partial activity diary entry for subject 1 Duration Activity 18 24 18 34 Standing 18 35 19 27 Sitting 19 27 19 33 Wander about house walk stand Sit down to eat chili Get water to drink thirsty Eat while sitting down Walk to work desk Sit at work desk 19 33 19 40 19 40 19 42 19 42 19 55 19 55 19 58 19 58 21 27 activity recognition models A sample of one of these diaries is provided in Table 1 The times are noted to the nearest minute and the activities are described briefly but are not always mapped precisely to a single basic Actitracker activity For example wander about house really means walking about but also includes time spent chatting while standing In cases like these it is not simple to identify indi vidual activities and any real attempt to do so would requ
19. e are trans formed into a set of 43 summary features that have been empiri cally shown to perform well 7 These features are simple statis tics including the means and standard deviations of acceleration per axis the binned distribution of values and a heuristic measure of wave period These features are computationally easy to gener ate ensuring that the system remains scalable and that it can pro vide results to the end user in near real time We are exploring more advanced features and will incorporate them if they are shown to improve accuracy Additionally the code which performs feature generation is written in a Java library shared by both the client and server This will make it easy for us to move the data transformation step from the server to the clients We are currently exploring this avenue since while it does add extra load to the device processor it may save battery by reducing the amount of data transmitted by the phone 5 3 Model Induction Our prior work has demonstrated that models built using train ing data from one user are excellent at predicting that user s activ ity from unlabeled data with accuracies in the high 90 s 10 However unlike our prior work and that of most other activity recognition groups Actitracker s data is not collected in the lab but in real time and in the wild If the incoming data has an activity class label attached to it i e via the app s training mode it is co
20. ed to be a critical public health threat 6 Inactivity also impacts the ability of the elderly to live independently which has huge quality of life and economic implications given our rapidly aging population If a caregiver could identify dramatic changes in activity for an elderly person this could help identify health related problems before they become serious The good news is that according to a report from the US Sur geon General even moderate amounts of exercise can substantial ly improve one s health 17 Activity recognition technology can address the problems associated with inactivity by providing peo ple with accurate information about their activities along with guidelines concerning the amount of physical activity that is re quired for good health In this paper we describe Actitracker a deployed free smartphone based service that provides activity monitoring and thus supports the social good The fitness market currently offers several activity tracking products such as the Fitbit www fitbit com and Nike Fuelband www nike com fuelband Both of these require the purchase of additional sensor hardware Fitbit requires the purchase of a clip on or wristband device in the 60 130 range while the Nike Fuelband costs 149 Fitbit requires an additional annual 49 fee should the user want to download their data and access supple mental features Smarthwatches such as the Pebble smartwatch getpebble com have rec
21. ently begun to enter the market and will support activity tracking applications the cost of these watches starts at about 100 The Actitracker service offers several advantages over the commercial products just listed The most obvious is that it re quires no additional hardware This results in cost savings but perhaps more importantly it makes activity tracking immediately available to a much broader audience there are currently about 1 5 billion smartphone users in the world 8 whereas the project ed market for wearable fitness sensor devices in 2017 is just 64 million 15 There is also some incremental effort associated with owning and maintaining these wearable accessories e g charg ing taking off and on Although smartphones require similar effort most people carry smartphones for reasons other than activ ity monitoring and thus this effort must be expended anyway Actitracker also provides more granular results than most com mercial products i e our timeline displays activities in 10 second intervals and recognizes some additional activities Actitracker s focus on basic daily activities like walking and climbing stairs is not particularly limiting because for most people including the elderly such activities are the most realistic and sustainable way to maintain healthy levels of activity There are also benefits associated with the fact that Actitracker although positioned as a commercial quality service is de
22. ese include iPhone support which is under development and integration with social media sites such as Facebook to better encourage changes in activity via positive social influences Fu ture work will also assess the tool differently rather than focus ing on accuracy of the models it will focus on how effective the tool is at fostering positive changes in behavior 8 ACKNOWLEDGMENTS This material is based upon work supported by the National Science Foundation Smart Health and Wellbeing program under Grant No 1116124 a Google Faculty Award and a variety of Fordham University faculty and student research awards 9 REFERENCES 1 Blair S N Kohl H W Paffenbarger R S Clark D G Cooper K H and Gibbons L W 1989 Physical fitness and all cause mortality A prospective study of healthy men and women Journal of the American Medical Association 262 2395 2401 2 Cui Y Chipchase J and Ichikawa F 2007 A cross cul ture study on phone carrying and physical personalization In Proceedings of the 2nd international conference on Usability and Internationalization 483 492 3 Fox S M Naughton J P and Haskell W L 1971 Physi cal activity and the prevention of coronary heart disease An nals of Clinical Research 3 404 432 4 Gerhardsson M Norell S E Kiviranta H Pedersen N L and Ahlbom A 1986 Sedentary jobs and colon cancer American Journal of Epidemiology 123 775
23. est model This univer sal model was generated by researchers on high quality data gath ered under lab conditions as described in our prior work 10 This model is known to perform reasonably well on unseen users with an average accuracy of 75 although there is considerable variation around this number from user to user While the per formance of these impersonal models is not nearly as good as that of the personal models which have accuracies generally above 95 they do perform at least as good as the impersonal models generated by other researchers working with similar activities 10 This use of a default impersonal model allows new users of Actitracker to start using the system right away They can later evaluate whether they are satisfied with the performance of the system and train a personal model for increased accuracy Prediction results are stored as a set of probabilities For each 10 second example the probabilities that during that time the user is walking jogging sitting standing or climbing stairs are all stored This allows displays and summaries of activity to take into account the classifier s uncertainty 6 USER EXPERIENCES User experiences have been collected via a formal on line sur vey which also allows free form comments and by more infor mal feedback from our own researchers and early adopters The surveys include a short daily survey and a more comprehensive survey meant to be filled out only once
24. ght and weight values stored in the profile This user is more sedentary than others of similar age gender and activity In the future we can incorporate recommended values since the average values very possibly do not reflect a healthy lifestyle Figure 4 shows an example of the timeline report This report shows the results at the most granular level at 10 second incre ments The user can scroll left or right to see different time peri ods Because of the highly granular nature of the data it may ap pear that a user is performing multiple activities at any time but Zoom 1h 2h 6h 12h All Jogging Stairs 1 vity Walking Standing Sitting Lying Down 30 Jul 02 00 04 00 06 00 Standing Stairs 2 23 Jogging 5 Walking 12 Sitting Lyi ng Down 58 Figure 2 Acitivty breakdown for one day Walking Avg by BMI Josei E Avg by Age ogging sems E Avg by Gender xs Stairs E Your Activity 2 g Standing Sitting Lying Down 0 2 4 6 8 10 Hours Figure 3 Activity comparison in this chart only one activity is identified at each 10 second in terval which is clear upon zooming in to any specific time span In the future we plan to aggregate the results so that the user can roll up to longer intervals of time e g one minute 4 2 3 FitDex and Calories Burned Actitracker provides users with two numerical values that summarize their daily activity a calo
25. h not all users are active each month While the usage of the tool is quite modest the existing user base is sufficient for us to provide a reasonable assessment of the strengths and weaknesses of the tool We expect these usage numbers to increase slowly over the next six months and then plan to promote the tool aggressively starting in the summer of 2014 Current usage statistics are generated dynamically and available from the Actitracker com main page 2 SYSTEM ARCHITECTURE Actitracker utilizes a client server model to perform activity monitoring and to present the results to the user A high level view of the basic Actitracker system architecture is provided in Figure 1 which shows the major system components The Acti Internet Server Client Interaction Cleaning amp Segmentation Feature Generation Raw Data Model Generation AR Models J Results Database Figure 1 Actitracker system architecture Smartphone Classification tracker client runs on the smartphone and transmits sensor data to the Acitracker server for processing The client app collects this data when it is enabled and the phone is unplugged with the screen off the phone is normally expected to be in one s pocket The server performs data cleaning and preprocessing steps and then transforms the time series accelerometer data into examples that each describe 10 second intervals of activity
26. ing the service Categories and Subject Descriptors J 3 Life and Medical Sciences health General Terms Algorithms Measurement Performance Design Human Factors Keywords Activity recognition sensors smartphones data mining health mobile health fitness 1 INTRODUCTION A lack of adequate physical activity is an enormous problem in our society because physical inactivity dramatically increases the health risks for many diseases including cardiovascular disease 3 11 14 colon cancer 4 diabetes mellitus hypertension and osteoporosis 16 According to the World Health Organization inactivity is responsible for approximately two million deaths per year 13 while a healthy amount of physical activity has been shown to significantly reduce the risk of all cause mortali Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page To copy otherwise or republish to post on servers or to redistribute to lists requires prior specific permission and or a fee KDD 14 August 24 27 2014 New York NY USA Copyright 2014 ACM 1 58113 000 0 00 0010 15 00 ty 1 12 Inactivity is especially associated with health related societal problems like childhood obesity which is now generally consider
27. ire tracking activities at the second level rather than minute level Tracking at this fine level of granularity would tend to interfere with the subject thus invalidating the purpose of the evaluation unless a less intrusive mechanism such as video tracking were used But this would be very costly and time consuming e g each motion would also have to be analyzed carefully and might still interfere with the natural behavior of the test subjects Fur thermore some activities are fundamentally a mixture of two or more basic activities and such a system would still not enable one to break them down precisely into their component parts The evaluation process involves querying the Actitracker data base for the activity results for the appropriate user over the speci fied times and comparing those results so the actual activity or activities noted by the user Because Actitracker issues predic tions at 10 second intervals but the users track times at the minute level if the evaluator noticed a small time shift in the transition to a new activity the transitions were essentially aligned to remove any errors due to minor inaccuracies in time In cases like the wander about house activity mentioned earlier where the user notes that it is a mixture of two activities our evaluation process counts either activity as correct The results for the three test sub jects are provided in Table 2 The evaluation process used to measure Actitr
28. lay area for the graphical results with three tabs at the top corresponding to the three reports time line activity comparison and activity breakdown The remaining one third of the dashboard is taken up by other elements includ ing e A graphical date selector to specify the time period over which the various charts and statistics are computed Buttons are provided as shortcuts to view the current day week and month e A visual indicator of the FitDex value a single number that maps total activity to an easy to understand value The user s average FitDex and user supplied goal are also displayed e Calories burned as well as average calories burned per day and user supplied calorie goal e Lifetime achievements computed over the lifetime of one s Actitracker usage and best achievements typically com puted over a single day 4 2 2 Activity Reports Actitracker currently supports three activity reports each of which provides a different view of the user data The relevant timeframe is determined by the date selector Figure 2 provides a breakdown of the activities for a specific user over the period of the day when the user is awake The figure shows that the user spent most of his time sitting or lying down The comparison chart in Figure 3 shows how a user s activity compares to that of other Actitracker users based on three different demographic categories BMI Body Mass Index and is computed from the hei
29. mentation and information about Actitracker provide feedback about the service via several online surveys and provide information to Actitracker via a user profile However the most important function of the web interface is to provide the user with access to the activity recognition results from Actitracker s database through a personal dashboard 4 1 User Documentation Surveys and Profile Actitracker documentation assists the user in learning about the service Under the About menu there is a very short promotion al description of the service a quick start guide that should enable most users to become familiar with the tool within a few minutes and a user manual that provides some additional details A privacy policy and document that explains the terms of use are also provided Although most of the data is not particularly sensi tive privacy issues will become much more important if we de cide to enable GPS sensor collection although we would make this an opt in service Finally we provide a frequently asked question list with detailed answers Given that we wish to improve the Actitracker service and evaluate it effectiveness we provide two online surveys located under the Contact Us menu The short daily survey is intended to capture feedback about the service for a single day of use and most of its questions are concrete and relate to the accuracy of the various activity predictions That survey also asks
30. n reducing the battery life by 20 30 Although the survey does not explicitly ask what impact this reduction in battery life has on their phone usage it is clear based on our per sonal experiences that this additional drain sometimes prevents the phone from lasting a complete day A closer analysis of the data further confirms our own experience which is that the impact of Actitracker usage on battery life seems to fluctuate dramatical ly on different days with no immediately apparent explanation We are investigating this phenomenon 45 40 4 35 30 25 20 4 15 10 0 T T T T None 10 20 30 30 40 40 50 gt 50 Figure 5 Perceived reduction in battery life due to app 6 2 Accuracy and Efficacy In this section we discuss the accuracy and efficacy of the sys tem mainly through the perceptions of the users This could argu ably be more important than actual accuracy since perceptions will drive whether users continue to use the service Our surveys ask about the perceived accuracy of the service both at an overall level and for specific activities The results for the overall accuracy are summarized in Figure 6 and Figure 7 Figure 6 shows the responses to the assertion Overall the app is accurate at measuring my activities The scale for this statement is the same as for most of the accuracy based questions where 1 means that they disagree with the st
31. nsidered training data New training data is combined with all previously received training data from the same user and then is automatically used to generate a personal activity classification model for the user The model is then stored for later use in classi fying unlabeled data from the user Currently Actitracker induces and stores several kinds of models that have been shown in our prior research to perform well in controlled environments 10 The Actitracker system only uses Random Forest models which were shown to perform very well 10 because of their relative speed when being built and when classifying examples Each model is represented as a Java object and they are both cached in RAM for speed of access and serialized to long term storage for persistence More information about this implementation and alternatives is available in our prior work 9 18 Because training data is user submitted it may be the case that the user has not submitted training data for all activities or that they have submitted insufficient training data e g only a few examples for some activities Thus before inducing a personal model Actitracker checks to make sure that it has at least one minute of accelerometer data for each of the walking sitting and standing activities before building a model This way the user s personal model will be able minimally to predict these activities If they later submit training data for additional activities
32. onal Workshop on Knowledge Discovery from Sensor Data San Diego CA 61 69 20 Wireless Sensor Data Mining WISDM Datasets Fordham University WISDM Lab http www cis fordham edu wisdm dataset php
33. rie count and their FitDex The calorie count is based on the activity results and more specif ically the time spent performing each activity The conversion from activity type to calories burned is accomplished via pub lished estimates Actitracker computes calories burned without the manual effort required by many fitness tools Figure 4 Activity timeline The FitDex metric maps a user s daily activity to a value be tween 0 and 1 000 The FitDex is calculated by taking the weighted sum of the activities where the weights are determined based on the relative amount of calories burned for that activity The values are scaled such that an average healthy amount of activity yields a FitDex of 500 A logarithmic scale is employed so that moderate increases in activity will yield significant im provements in FitDex at the low end of the scale This will pro vide motivation to those users who are not particularly active and who may require the most positive feedback As users achieve higher FitDex values more substantial increases in activity are required to move the FitDex higher which ensures that highly active users remain properly motivated Both the FitDex and the calorie count are currently displayed on the activity dashboard with the ability to view the last two weeks of values in a line graph From the dashboard the user can also set a goal value for the FitDex and calories burned and when the user reaches their daily goal the
34. the research and our user experiences show that women carry their phones in a variety of locations namely purse back pocket front pocket hand and jacket pocket Women typi cally do not use the front pants pocket because the phone will not fit But although the back pocket would probably provide an ade quate sensor signature for activity recognition women transfer the phone to their hand or a table prior to sitting down This is not ideal since they may continue to move around although if they do remain seated the phone positioned on a stationary table will for tuitously lead to the proper activity since it is in a similar posi tion as if it was in the user s pocket Nonetheless the issues are significant enough that women are very likely to avoid this service or stop using it This is backed up by our user statistics which show that women make up 20 of our registered users but signif icantly generate only 13 of our activity data As discussed in Section 7 we plan to address this by integrating smart watches into Actitracker but this will remove one key advantage of our system for most women that it only requires a smartphone 6 1 2 Battery Usage The other key issue with usability is the impact that Actitracker has on the smartphone battery life The results of the user survey provided in Figure 5 are consistent with our team s experience Actitracker seems to have a modest but significant impact on bat tery life ofte
35. their model will be rebuilt using both their old and new training data and the new model will be able to predict the additional activities In the future we may implement hybrid models where data for ALGORITHM 2 k lg e input_count output_count for i 0 i lt output count i select ceil 1 e exp i k missing activities is taken from a panel of other users or from users with similar physical characteristics Users have a high degree of control over what training data they submit so it is possible that they will submit training data with a severe class imbalance e g one hour of sitting data and only a few minutes for other activities To prevent this from un duly biasing the predictions of the induced activity recognition models we down sample any activity for which there are more than three times that user s mean number of examples per activity We assume that the most recent examples will be the most useful in classification since a user s gait may change over time due to injury or age However old examples may still contain useful and diverse information about their gait Thus we designed our sub sampling process as an exponential function that selects a few old examples and many recent ones It is defined in Algorithm 2 5 4 Prediction Unlabeled examples are classified using the personal Random Forest model for their user if one is available or else with our default impersonal universal Random For
36. titracker system or the user interface component including the Actitracker activity reports This is also the first paper to describe user experiences with the complete system in a fully naturalistic setting Although Actitracker is being developed in a university re search lab and is free it was designed to mimic a commercial service This was to remedy some complaints with prior work on activity recognition including that they do not address problems in a fully natural setting and are terminated before they gain real users To help ensure that Actitracker does not suffer the fate of some of these prior systems thousands of development hours have been spent on user interface design and providing profes sional quality documentation approximately half a million dollars has been expended on Actitracker research and development over the past four years The basic functionality of Actitracker has been operational since March 2013 at which point the system entered Alpha re lease After a number of enhancements and bug fixes the system entered Beta release in September 2013 and then general availa bility in November 2013 However very little effort has thus far been expended to promote the Actitracker service and most of that effort has focused mainly on the Fordham University Rose Hill campus Nonetheless users have found the service including many international users As of February 2014 there were 250 registered Actitracker users althoug
37. veloped and maintained by a University research lab and is not profit driven Aside from the fact that the service is free anonymized Actitracker data is shared with activity recognition researchers via publically available datasets 20 Furthermore if the service be comes sufficiently popular the data can also be used to enable large scale epidemiological studies of activity and behavior that would otherwise be impractical For example Actitracker activity data can be used to track activity levels over time by season by geographic region and by various demographic factors such as age profession and weight Also unlike most commercial prod ucts Actitracker s design and algorithms are documented in a form available to the public and research community 7 9 10 Actitracker will also support the exporting of low level user data so that external third party services and applications can be used to benefit the user Currently most commercial products want to retain control over their data and only allow the exporting of rela tively high level results There have been several research papers that have described some of the research underlying Actitracker including papers that have focused exclusively on activity recognition 7 10 19 How ever that prior work focused exclusively on highly structured settings where users were directed to perform specific tasks at specific times Additionally no prior work discussed the overall Ac
38. web The preferences screen allows users to change data transmis sion and storage settings Users have control over the circum stances under which the sensor data is uploaded to the server a major factor in battery life Wi Fi Transmission and Mobile Network Transmission checkboxes permit the phone to send the data when connected to a Wi Fi network mobile network or both selecting only Wi Fi tends to conserve power and prevent the app from using up limited data plans but delays results until the de vice is connected to Wi Fi The user can also select the Charge Only option which overrides the other options and ensures that data is only transmitted when the phone is charging which for most people will lead to results being posted only once per day Most smartphones can easily store many days or months worth of data Finally users can specify transmission frequency which applies only when the specified transmission mechanism s are available Frequency options range from once per minute to once per hour although we are planning a real time option to support live demos of the service Most people check their activity reports only once a day or at most a few times a day so most configu rations are more than adequate The data storage and management options do not normally need to be changed but can allow one to not cache unsent data clear cached data or force send all cached data disregarding the current
Download Pdf Manuals
Related Search