In this position paper we present various aspects of comparing human physical activities using time series data within the selfBACK project. The goal of the project is to develop a decision support system for patients su ering from non-speci c low back pain. The system will give users advice in form of a self-management plan that is based on self-reported physical and psychological symptoms as well as activity stream data collected by a wristband. Here, we discuss the activity stream representation di culties as well as various challenges that arise when comparing the resulting activity streams.
Introduction
selfBACK [
Two of the many challenges during matching and similarity assessment when using the CBR methodology, is developing a suitable abstraction of the wristband activity stream and an adequate comparison algorithm for such streams. The abstractions and the comparison algorithm must catch the important aspects and compare these in a meaningful way relevant in a non-speci c low back pain domain.
We will in the next section shortly present the di erent aspects and challenges that need to be considered in order to develop such a comparison. Speci cally, we will focus on the challenges arising when comparing the activity streams of patients. 2
The main goal of the selfBACK project is to give non-speci c low back pain patients personalized advice based on their reported symptoms and activity patterns collected by a wristband. Three major challenges arise when comparing the activity streams: 1. The activity stream needs to contain the knowledge needed and used for the treatment of non-speci c low back pain. 2. The comparison of such streams must be made in a way that is meaningful when comparing patients with non-speci c low back pain. 3. Comparing activity streams with missing data. The wristbands have limited battery time and are often not water proof and needs to be taken o during showering or swimming activities. Users might nd it uncomfortable to wear the wristbands continuously and try to loosen the strap or take breaks in wearing them, which might introduce gaps in recordings.
Because these problems are very novel, we have very little knowledge on what
to include and how to compare such activity streams. There is some evidence
that sleeping patterns and prolonged periods of inactivity are correlated with an
increase in non-speci c low back pain [
The activity streams that we have experimented with can contain activity data from periods of up to several weeks. This is because to see the e ect of a change in behaviour, by applying some advice, one often needs to wait for about 4-6 weeks. Including all kinds of information into an activity stream (type of activity, intensity, skin temperature, pulse, moisture of the skin, etc.), comes with an increase in the computational cost when comparing the resulting activity streams against each other. See gure 1 for examples of abstractions that can be included.
We have also experimented with di erent resolutions on our activity streams. We are able to capture and transform data into 1 second as well as 1 hour windows to name a few. Decreasing the resolution, lowers the computational comparison cost, but important data may be lost during the transformation. Di erent resolutions can be combined throughout the activity stream to prevent loss of information. For example a bigger window can be used during prolonged periods of inactivity or sleep. On the other hand, during periods of activity where the intensity or activity types changes frequently, a smaller window can be used to capture and conserve that information.
Other ideas we have been experimenting with include comparing the activity streams based on the activity intensity rather than activity type. Clinicians with experience in the eld of non-speci c low back pain have suggested, instead of comparing the streams directly, to compare sleeping patterns and periods of prolonged inactivity. This way of comparing activity streams produces very di erent results than when comparing activity streams containing activity types. This is shown in gure 2.
As for the comparison algorithms themselves, we are currently experimenting
with a string based approach. We convert the activity stream into a string of
characters. When comparing two strings, we compute the percentage of each
character per string and the number of consecutive sequences of those characters.
We then also compute the longest common sub-sequence, the sequence distance,
the number of similar k-mers and the number of unique similar k-mers. The
sequence distance is the Levenshtein distance which is the least number of single
character insertions, deletions or substitutions that is required to transform one
string into the other [
We have presented and discussed various challenges that arise when comparing activity streams in the selfBACK project.
Our current work focuses on the representation and comparison of the activity stream data that is meaningful in the non-speci c low back pain domain. We are investigating which time resolution is bene cial and which information should be included in the activity streams. We are also exploring whether and how domain knowledge from non-speci c low back pain experts can be incorporated. As shown in gure 2, the representation of the activity streams can di er a lot depending on which information is most relevant (alteration of activities in S1 and S2 vs. the reduction of pro-longed inactivtiy in S3). Therefore we are working on how the representation a ects the similarity comparison of activity streams collected from real patients.
Our next steps include the investigation whether activity stream rearrangement could yield towards a more precise comparison. Therefore we investigate if parts of the activity stream are more important than others (for example sleep patterns). When the data collection from the planned randomized control trial is complete, it will also be possible to mine the collected activity streams for behavioral patterns that contribute to an increase or a decrease in non-speci c low back pain. For example we could discover that a certain amount of time standing followed by a certain amount of time sitting is increasing non-speci c low back pain etc. Finally we will conduct experiments on how much weight the activity stream comparison should have compared to the self-reported reported questionnaire data from the patients.
Acknowledgement The work has been conducted as part of the selfBACK project, which has received funding from the European Unions Horizon 2020 research and innovation programme under grant agreement No 689043.