MARIKA: A Mobile Assistance System for Supporting

Home Care

Tobias Umblia, Albert Hein

, Ilvio Bruder and Thomas Karopka

University of Rostock, Institute of Computer Science, Rostock, Germany

IT Science Center R

ugen gGmbH, Putbus, Germany

Abstract. Documentation of care activities is a very time-consuming task of

home healthcare, but necessary due to legal requirements. Automating the care

documentation would relieve the nurses from writing it down by hand, resulting

in more time for the patients

. The MARIKA project presented in this paper is

addressing this problem through the development of a system which assists the

home care personnel by automatically recording care activities, and integrates

with other systems involved in the care process. Technical approaches for two

major building blocks of such a system are described: A hybrid, sensor-based ac-

tivity recognition approach, and a mediator-based data and function integration

approach.

1 Introduction

The elderly currently are the fastest growing demographic group in the USA [1], a

trend which is similar to other, for example European regions due to increasing life ex-

pectancy and decreasing birth rate. So the aging human is more and more getting into

the focus of not only demographic research. Graceful aging is a common demand of the

current and future population. Most of the people wish to stay at home as long as possi-

ble, but the home care for elderly or sick people is a serious burden for family members.

One third of admissions to nursing homes are done not because of deterioration of the

persons’ health conditions, but because of the so-called “caregiver burnout”.

An alternative to stationary treatment is the professional ambulant care at home.

Especially this kind of service needs accurate documentation of care activities due to

legal requirements and to allow for correct accounting with the health insurances. The

usual documentation process is to this day still done manually and takes up to 40% of

the working time, is error-prone and mostly inaccurate because it does not happen in

situ but afterwards.

This requires solutions with which documentation tasks are accomplished “along

the way”, yet properly. The personnel thus is given free hand to attend to their actual

role. Additional functions like reminders or giving decision guidance assist in the ev-

eryday work. These are requirements analogous among developments in the medical

and afﬁliated sectors, like the wearable assistant for hospital ward rounds presented in

The terms patient and client will be used interchangeably throughout this text.

Umblia T., Hein A., Bruder I. and Karopka T. (2009).

MARIKA: A Mobile Assistance System for Supporting Home Care.

In Proceedings of the 1st International Workshop on Mobilizing Health Information to Support Healthcare-related Knowledge Work, pages 69-77

DOI: 10.5220/0001815200690077

 SciTePress

[2]. Easily operated, mobile devices and unobtrusive sensors contribute to meeting the

socio-technical conditions also common to these scenarios. That is, to improve user ac-

ceptance of such systems on the one hand, as well as, on the other hand, the acceptance

by the patients who understandably demand the care personnel’s attention, an aspect

emphasized in [2], too.

In this paper we ﬁrst give a short overview of the MARIKA project in section 2,

followed by a description of the problems we are addressing in section 3. In section 4.1,

we present our activity recognition approach along with ﬁrst test results, the data and

function integration concept for the system is described in section 4.2. Concluding with

section 5, we give an outlook on our further research.

2 The MARIKA Project

The demographic situation described above and the emerging challenges turn home

healthcare into one of the fastest growing areas of healthcare provision. The MARIKA

project (Mobile Assistance for Route Information and Electronic Health Record) aims

at providing support for the home care service by automating the obligatory documen-

tation in the care process. Particularly dedicated to accompanying the nurse’s everyday

work ﬂow, different techniques are brought together in the project. Two of these are sub-

ject matter of the following sections: sensor-based activity recognition (see section 4.1)

and data and function integration (see section 4.2). Other research areas involved are

combined indoor and outdoor positioning, data privacy, e-learning, as well as content

and knowledge management [3].

One of the key challenges is the organization and coordination of all stakeholders

that take part in the care process, and the dissemination of information between them.

Several projects have been developed recently aiming at providing a communication

and coordination platform for home healthcare [4–6], showing that this is a ﬁeld with

innovation potential. In this respect, developments within the MARIKA project are also

focusing on a platform that allows all participants of the care process, namely general

practitioner, home help service and home care nurse to exchange information. Although

cost effectiveness studies are rare, it is evident that coordination platforms have a great

potential for enhancing quality of life and reducing costs in the care process.

The collaboration with a professional outpatient care service, working in rural areas

and small towns mostly, allows us to test and optimize the components of the system

with regard to user acceptance and reliability.

3 Use Case and Problem Analysis

To identify the technical problems our project is addressing, a simpliﬁed day-to-day use

case is illustrated as it will look like in the future:

A daily schedule exists for the home care service personnel (the nurses), contain-

ing information about which client will be visited and when. Clients will be given care

mainly at their homes, but also at nursing homes. Usually, the nurses take off their route

by car. At the client’s home, the predeﬁned care activities are performed by the nurse

and documented automatically. When leaving the client, the care documentation process

is completed and the associated information will be transferred to the central care docu-

mentation and planning system of the home healthcare provider. Spontaneous changes

of the schedule can occur at any time, in which cases the required route information,

client data etc. will be updated automatically.

There are several devices involved in this scenario: The car is equipped with a GPS-

based navigation system, which dynamically adapts the route, combined with an elec-

tronic logbook, which records the duration of a tour and the distance covered. The nurse

is using a mobile device, which collects and processes data en-route. One of the now

omnipresent smartphones, PDAs or tablet PCs come into question, preferably dedicated

to applications in the healthcare sector. Devices adhering to the mobile clinical assistant

(MCA) reference architecture introduced by Intel recently [7] will be the target systems

of our ﬁrst implementations, as these already provide serviceable features like RFID

scanner and digital camera. Additionally, wearable sensors will be incorporated into

the system.

Hence a high degree of mobility and autonomy of the involved players and compo-

nents, and a deﬁnite request for ease of use, portability and adaptability, as well as data

privacy are the resultant basic conditions we are confronted with.

The organizational framework of a care documentation system is given by the statu-

tory health insurance. All care activities are grouped into so-called nursing care levels,

their exact duration, extent and cost are ﬁxed and retained in a service catalog, thus

building the basis for accounting. In the MARIKA assistance system for home care

documentation these activities should be detected automatically. Therefore, they have

to be modeled, and the recorded sensor data then has to be matched with these models

adequately. In reality, there often are ambiguities between different activities sharing

the same motions or gestures like carrying a glass of water or carrying a pillbox and

also ambiguities between different activities using the same objects, which is e.g. using

an object or simply carrying it around. These two problems are not clearly distinguish-

able using only a single sensing method. That is why a direct motion measurement with

inertial sensors and the detection of object interactions with RFID are combined in our

activity recognition approach.

Our ﬁrst main objectives are to investigate whether such a combined sensor conﬁg-

uration is technically viable in terms of delivering reliable data without explicit com-

pliance of the test subject, if a static classiﬁer or a hybrid model is able to infer usable

estimates on a continuous time trace and how much each sensor type actually con-

tributes to the ﬁnal classiﬁcation results. The activity recognition approach is described

in section 4.1.

Another main problem besides activity recognition is the integration of all subsys-

tems into an overall architecture. The data and functions necessary for care documenta-

tion are distributed across these components and heterogeneous regarding their structure

and semantics. The different types of data are sensor data streams, representing care ac-

tivities, text and multimedia documents like photos or voice memos complementing

the care documentation, and precise temporal and spatial logging. Altogether they al-

low for accurate accounting by giving answers to questions like: Who did what, when

and where, and how long did it take? Furthermore, there are special functionalities re-

quested during care service, its analysis and planning, e.g. plausibility checks in the

work ﬂow and giving decision guidance in case of ambiguities. The MARIKA system

will ﬁnally be integrated with existing care documentation software (CDS), which is al-

ready in use at home care services, and aiming at a larger-scale coordination platform.

The integration methods required to achieve this are described in 4.2.

4 Technical Approaches

4.1 Activity Recognition

Currently, sensor based activity recognition is widely seen as an essential technology

for providing mobile assistance in ambient assisted living (AAL) scenarios. An ambient

sensor infrastructure as in the iDorm [8] or the PlaceLab [9] is mostly unobtrusive,

but (at least for now) too expensive and complex for home care settings where each

apartment of every care patient would have to be equipped with environmental sensors.

Wearable sensors as an alternative are already showing good results in human activity

recognition, either through direct motion sampling [10, 11] or through capturing object

interactions [12].

We are aiming to combine these two methods of direct motion measurement with

inertial sensors and detection of object interaction with RFID for high level activity

recognition. Our system uses hierarchical sensor fusion on different levels of abstrac-

tion for simultaneously integrating many channels of heterogeneous sensor data. For

inferencing high level activities we use a layered hybrid discriminative and model based

generative approach. This will enable us to integrate prior knowledge into the decision

process in the future to reduce the amount of training while keeping the probabilistic

model simple. First parts of this approach have already been evaluated on the exper-

imental setting of a home care scenario where the hybrid approach reached accuracy

rates of 96% although the RFID object sightings were not yet considered.

For the tests we roughly rebuilt the ﬂoor plan of an apartment consisting of a bed-

room, a bathroom, a living room and a kitchenette in our laboratory. The test runs were

performed by professional care personnel (a geriatric nurse) and a student who helped

out as a patient (Example still frame see Fig. 1 (left)). The test agenda and the scenario

have been developed in close collaboration with a nursing service, which also provided

authentic equipment for the tests. The activities were directly taken from the service ac-

counting catalogue of the health insurances: “general service” (greeting, fetching news-

paper, ...), “big morning toilet” (including washing whole body, brushing teeth), “mic-

turition and defecation”, “administration of medications”, “bandaging”, “preparation of

food” and “documentation”.

For the motion recordings we used three small sensor boards, equipped with a 3-axis

accelerometer, 3-axis gyroscopes and a 2-axis magnetometer. Because of the compact

size (51x41x23mm) the boards could be attached at unobtrusive positions: at the dom-

inant wrist for recording gestures and object motion, at the chest/upper back and at the

hip. RFID is a popular technology for contactless identiﬁcation of objects. A basic sys-

tem consists of a reader module with an antenna and several active or passive tags in the

form of small boxes, stickers or even implants. Especially, passive RFID stickers are a

cheap, battery free solution for reliable object detection. RFID practically does not re-

turn any false positive object sightings. Because currently no wearable RFID modules

Fig. 1. Still frame taken from a surveillance ﬁsheye camera. The test person is equipped with

sensors at the hip and upper back and dominant wrist (left). RFID Wrist Antenna and Inertial

Sensor Board (right).

are available commercially, we used a multi function reader with a custom-made wrist

antenna (see Fig. 1 (right)). Depending on object geometry and material it has a reading

range between 10 and 30cm. All data streams were wirelessly transmitted to a laptop

computer where they were immediately formatted and synchronized.

Based only on the motion sensor data we compared different supervised and un-

supervised learning methods in conjunction with an automatically synthesized Hidden

Markov Model for evaluating the general feasibility of our hybrid approach with very

promising results with recognition accuracies up to 96% (Table 1). For a detailed de-

scription of the approach we have to refer to [13]. As the presented classiﬁcation results

are based on very few experimental training data, they have to be treated carefully in

respect of generalization. Future experiments will have to follow under real world con-

ditions to allow a reliable evaluation of a comprehensive set of care activities and mul-

tiple test subjects. This is expected to provide more realistic results allowing an outlook

on everyday use.

Table 1. Comparison of several supervised and unsupervised discriminative approaches: The

table shows the speciﬁc recognition accuracies in % for both test runs. (The ﬁrst test run has 451,

second has 471 single observations.)

1st test run 2nd test run

Decision Tree 86.9 86.6

Support Vector Machine 85.8 85.6

Naive Bayes 77.2 70.9

k-Means (k = 35) 98.2 96.0

4.2 Data and Function Integration

The recorded motion and RFID sensor data are the basis for the activity recognition

component. In the intended overall architecture it will be combined with several other

technical components to form a mobile assistance system for care documentation. A

proper integration strategy is pursued in order to utilize data and services associated

with these components. This comprises sensor data streams, spatial and temporal data,

context and knowledge information, history and logging data, as well as multimedia

documents and healthcare records. Furthermore, components may offer special func-

tions and services towards applications.

These data and function sources can be characterized as being highly heteroge-

neous regarding their structure and semantics, autonomous, distributed and subject to

frequent changes, therewith describing a common integration problem [14]. Its solution

is expected to run automatically, that is, without the need for user interaction after it has

been implemented, allow for adaptations, and to stay both ﬂexible and consistent across

evolution stages.

This is also demanded from the applications’ point of view, where the integration

approach mainly aims at building an extension to existing care documentation software

(CDS), and moreover, merging the MARIKA subsystems into a communication and

coordination platform for home healthcare.

To achieve this, we choose a mediator based integration approach, utilizing logical

views of the data, which is widely accepted as an applicable solution for such integra-

tion problems [15–17]. Here we follow the Local-as-View concept at ﬁrst: The source

schemata are deﬁned as views of a global schema. Data in the global schema is not

materialized, but made available dynamically on request. Its the mediator’s responsibil-

ity to translate queries from applications and to respond accordingly by means of those

view deﬁnitions. Resulting from this concept, the mediator will perform different tasks,

depending on the perspective, as depicted in Table 2.

Table 2. Tasks of the mediator based integration, from the sources’ perspective in contrast to the

applications’ perspective.

Sources’ Perspective Applications’ Perspective

Deﬁne local source schemata as views (within wrappers) Deﬁne unique, global schema

Provide data containers (tables) and function hooks Provide data and functions

Permit write access Permit read access

Deciding advantages of approaching an LaV model are that sources can be kept

distributed and autonomous, that changes of their schemata as well as adding or re-

moval of sources can be controlled better, thus allowing for easier maintenance and

evolution of the overall system. The global schema, representing the interface towards

applications, remains constant and does not need to be changed frequently. This meets

the requirements of the targeted users, namely healthcare instead of IT professionals, to

spare them the necessary adaptation and update procedures. A complicated query trans-

lation process is seen as the main drawback of the LaV concept, since it leads to poor

performance in case of frequent, complex queries. For now, these are not part of the

MARIKA use cases and can be disregarded. Evaluating the combination of concepts of

Local-as-View and Global-as-View to solve this problem, as proposed in literature [18,

19], is part of our future work.

A mediator based integration always implies the implementation of wrappers. They

wrap the information sources by imposing the local schemata on them, in the form of

views of the global schema. For our scenario different wrappers have to be deﬁned for

the data and functions associated with the components listed above. Thereof, wrapping

sensor data streams is one of the demanding tasks in the integration process. To address

it, database system extensions for stream data processing are evaluated, which utilize

caching of data and allow for dropping irrelevant data based on ﬁlter mechanisms. This

is part of our idea to optimize the performance of the integrated system in the long term:

Data and processes which are not inherent in the ﬁnal care documentation will, archi-

tecturally, be kept at a low level, that is inside of the wrapper. That applies particularly

to motion sensor data streams or GPS tracking data. On the other hand, information

and functions essential for the care documentation will be modeled at a higher level,

that is within the mediator, to ensure proper responsiveness of the applications above

it. That applies to more complex tasks like, for example, analyzing the convalescence

after an injury by comparing photos taken during the home care. Implementation of

the mediator and wrappers will focus on object-relational database techniques, which

have been standardized with SQL:1999 and further extended with SQL:2003 [20, 21].

They facilitate integration of the several subsystems by supporting user-deﬁned data

types and methods in the data model. Furthermore, the foreign table concept deﬁned in

the SQL/MED part of the standard allows for integrating sources residing at distributed

places and regardless of their SQL capabilities (SQL-aware and non-SQL-aware foreign

servers), [22]. Using these standardized techniques is supposed to be future-proof, since

external information systems could be integrated with less effort and tools adhering to

the standards would be available for modeling and implementation.

5 Conclusions and Future Work

Activity recognition and data and function integration are two important research areas

within the MARIKA project towards automatic healthcare documentation. A combina-

tion of sensor data, knowledge and context information, as well as spatial and temporal

data is needed to differentiate between activities in the nurse’s work ﬂow. First tests are

described in this paper and ﬁrst results show the feasibility of the approaches.

Regarding the activity recognition, our next steps will be, in close collaboration

with a home care service, to collect a much larger and more realistic out-of-lab dataset

recorded during the day-to-day work and integrating the RFID sensor data into the

recognition module. There are still open problems in integrating more data in the deci-

sion making process and more analysis is needed for increasing the recognition quality.

Implementation of the mobile assistance system will primarily focus on devices

conforming to Intel’s MCA speciﬁcation. Case studies involving the home care service

personnel will be tackled to prove whether the features are suitable for our scenario and

whether performance of the overall system is sufﬁcient for daily use.

These studies will also serve as basis for evaluation of the SQL-based, object-

relational integration approach, to clarify whether the Local-as-View concept is viable

or has to be extended with regard to the intended coordination and communication plat-

form. An evolution strategy for the integrated system will be developed in parallel, to

ensure its usability and maintainability in the future.

Acknowledgements

MARIKA [3] is funded by the state of Mecklenburg-Vorpommern, Germany, within the

scope of the “Landesforschungsschwerpunkt Mobile Assistenzsysteme” main research

initiative.

References

1. Roush, R.E.: Smart home technology for aging in place longer and better. Technical report,

Roy M. and Phyllis Gough Hufﬁngton Center on Aging (2004)

2. Adamer, K., Bannach, D., Klug, T., Lukowicz, P., Sbodio, M., Tresman, M., Zinnen, A.,

Ziegert, T. In: Developing a Wearable Assistant for Hospital Ward Rounds: An Experience

Report. (2008) 289–307

3. Landesforschungsschwerpunkt: http://marika.lfs-ma.de/ (2008)

4. Bricon-Souf, N., Anceaux, F., Bennani, N., Dufresne, E., Watbled, L.: A distributed coordi-

nation platform for home care: analysis, framework and prototype. International Journal of

Medical Informatics 74 (2005) 809–25 PMID: 16024285.

5. Karopka, T., Bruder, I., van den Berg, N., Hoffmann, W.: Ict architecture for a community

medicine nurse project. In: Proceedings of the First International Conference on Health

Informatics. Volume 2. (2008) 26–30

6. H

agglund, M., Scandurra, I., Mostr

om, D., Koch, S.: Integration architecture of a mobile

virtual health record for shared home care. Studies in Health Technology and Informatics

116 (2005) 340–345 PMID: 16160282.

7. Intel: Products and solutions for health and healthcare, mobile clinical assistant (mca).

http://www.intel.com/healthcare/ps/mca/ (2008)

8. Rivera-illingworth, F., Callaghan, V., Hagras, H.: Detection Of Normal and Novel Be-

haviours In Ubiquitous Domestic Environments. The Computer Journal (2007) bxm078

9. Hightower, J., LaMarca, A., Smith, I.E.: Practical lessons from place lab. IEEE Pervasive

Computing 5 (2006) 32–39

10. Bao, L., Intille, S.S.: Activity recognition from user-annotated acceleration data. In Fer-

scha, A., Mattern, F., eds.: Pervasive. Volume 3001 of Lecture Notes in Computer Science.,

Springer (2004) 1–17

11. Lester, J., Choudhury, T., Kern, N., Borriello, G., Hannaford, B.: A hybrid discrimina-

tive/generative approach for modeling human activities. In Kaelbling, L.P., Safﬁotti, A.,

eds.: IJCAI, Professional Book Center (2005) 766–772

12. Patterson, D.J., Fox, D., Kautz, H., Philipose, M.: Fine-grained activity recognition by ag-

gregating abstract object usage. iswc 0 (2005) 44–51

13. Hein, A., Kirste, T.: Towards recognizing abstract activities: An unsupervised approach.

In: Proceedings of the 2nd Workshop on Behaviour Monitoring and Interpretation, BMI’08,

Kaiserlautern, Germany, September 23, 2008, Kaiserlautern (2008)

14. Sheth, A.P.: Changing focus on interoperability in information systems: From system, syn-

tax, structure to semantics. In: Interoperating Geographic Information Systems, Springer US

(1999) 5–29

15. Halevy, A.Y.: Answering queries using views: A survey. The VLDB Journal 10 (2001)

270–294

16. Lenzerini, M.: Data integration: a theoretical perspective. In: Proceedings of the twenty-ﬁrst

ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, Madison,

Wisconsin, ACM (2002) 233–246

17. Busse, S.: Modellkorrespondenzen f

ur die kontinuierliche Entwicklung mediatorbasierter

Informationssysteme. Logos-Verl., Berlin (2002)

18. Xu, L., Embley, D.W.: Combining the best of global-as-view and local-as-view for data

integration. In: ISTA. (2004) 123—136

19. Boyd, M., Kittivoravitkul, S., Lazanitis, C., McBrien, P., Rizopoulos, N.: Automed: A bav

data integration system for heterogeneous data sources. In: Proceedings of the 16th Inter-

national Conference on Advanced Information Systems Engineering, CAiSE 2004. (2004)

82–97

20. ISO/IEC: Iso/iec 9075-9:2008 - information technology – database languages – sql – part 9:

Management of external data (sql/med). http://www.iso.org/ (2001)

21. T

urker, C.: SQL 1999 und SQL 2003: Objektrelationales SQL, SQLJ und SQL/XML. 1 edn.

Dpunkt Verlag (2003)

22. Lindner, W.: Konzepte f

ur ein offenes, objektrelationales Multimedia-Datenbanksystem. 1.

auﬂ. edn. Diplomica, Hamburg (2002)