For the elderly, falls are among the most frequent causes of severe injuries or even death. Therefore it is highly desirable to develop methods for early recognition. Numerous indicators have been proposed and thoroughly validated, that allow medical sta to identify persons at a high risk of falling. However, these indicators su er from pragmatic drawbacks impeding their widespread application. To overcome this, we present the concept of a context-aware system, using low-cost accelerometers to collect motion data at peoples homes, from which a fall-risk prognosis is computed automatically. The underlying design principle can be generalized to other medical settings, opening up a promising eld for the application of context-based technologies.
Severe injuries from falls rank along with well-known diseases such as cancer
or hypertension as leading causes for the hospitalization of elderly people ([
The organization of the remaining paper is as follows: We brie y review existing work on fall-risk indicators and their drawbacks. Then we introduce our proposal, motivating our choice of sensor devices for the system and presenting the context-based key idea for the realization of the fall-risk prognosis module. Then we give some statistical background and conclude with an outline of future work. 2
Methodologically, the de facto standard in this area are prospective studies,
comprising the following stages: First, potential indicator variables are identi ed
heuristically (feature selection). After that, these variables are measured for
each member of an appropriate test population. This population is then tracked
during a prede ned follow-up period (one year by convention of the fall-risk
community, see [
Fall-risk indicators studied in this way can be divided into the following main categories:
Temporal parameters describing standardized tasks Gait parameters Cognitive Skill Tests Combinations of these factors
The rst category comprises a variety of parameters recorded during
standardized physical tasks; examples are muscle strength tests ([
As mentioned in the introduction, the overall vision is to provide a cheap system, that, by means of wearable sensor devices, permanently records data in persons everyday lives, without any further impairment. Then, by a recurrent o -line analysis of the gathered data (for instance once a month), a fall-risk prognosis is to be generated automatically, as if the person had attended a clinical fall-risk assessment.
Technically, the type of sensor devices and their positioning had to be determined, keeping in mind usability constraints. In the literature we found that 3D-accelerometers have successfully been used for a number of similar tasks (see [18],[19]). Furthermore, a sample rate of 40 Hz is generally considered to be sufcient to capture human motion ([20]). Regarding the integration respectively positioning of such a sensor, we chose wrist-worn clock-like devices. The major advantage of this is the high degree of familiarity persons have with that kind of device. This familiarity is important for at least the following to major reasons: First, user acceptance is increased and secondly, no intervention by medical sta is needed, because persons can handle the device on their own. However, as a fall-back solution, we keep the possibility to x the device to a belt, more closely to the center of mass of the human body.
So far, we have said nothing about the heart of the system, namely the fallrisk prognosis module intended to process the raw data. Subsequently, we will work out the challenges related to its realization and our proposed context-based solution. 3.1
One might be tempted to mechanically apply the approved method of
prospective study: Establishing statistical relationships between the data sampled by
the sensor device (sequences of three-dimensional acceleration vectors) and the
target variable indicating falls in the follow-up period, in order to implement
these ndings in the prognosis module. However, such an approach is likely to
fail, due to the absence of an obvious hypothesis about a potential interrelation.
To point out this crucial fact, we would like to oppose our situation to rather
clear settings like the proposal of [
As an answer to this challenge, we have searched for approved fall-risk indicators, which have already been prospectively validated are potentially reconstructible from the acceleration data With such indicators at hand, we argue that it is possible to establish fall prognoses in a "transitive" way:
Use recorded acceleration data to compute approved indicators
Use these indicators to establish fall prognosis As these parameters are measured under standardized conditions in appropriate laboratories, it is right here that context comes into play. The above sketched procedure can only be applied to data sampled during intervals where the laboratory conditions were approximately met. This consideration leads to the following three stage process, being the conceptual backbone of our system:
In an obvious way, one gains a general scheme by abstracting from the concrete setting of fall-risk prognosis. This scheme can be customized towards di erent medical scenarios, involving costly laboratory-based methods. Exemplarily, we cite early recognition of movement disorders in patients a ected by multiple sclerosis ([21]) and motion assessment of Parkinson patients ([22]). In other words, our proposal can be seen as a rst blueprint for a promising class of applications for context-based technologies.
However, we focus back on fall-risk assessment and proceed by brie y introducing the two approved fall-risk indicators we claim to be reconstructible from the acceleration data. 3.2
In [
M GC :=
PnN=11(tl(n + 1)
tl(n)) N
1 GV := s
PnN=11(M GC
The con rmed hypothesis is again that a higher value of this quantity indicates
uncertainty in walking and thus a higher risk of falling.
The second indicator is the so called Postural Sway (see [
This section contains a brief sketch of the statistical modeling underlying the system, the studies that must be performed in order to both validate these models and evaluate the overall system. 4.1
For the rst stage, given a nite sequences of vectors recorded during a given time interval, we must segment this interval into a nite number of subintervals labeled with elements of a nite set of activities (concretely: "walking", "standing","something else").
For the reconstruction of time stamps of steps, we must further segment the intervals labeled with "walk" into subintervals labeled with "Swing Phase Left" and "Swing Phase Right"; the desired time stamps are determined by the boundaries of the subintervals.
Both cases lead to the same mathematical model: Vector sequences are to be mapped to activity traces, e.g. functions from the time interval into the set of activities; subintervals are delimited by the discontinuity points of these functions. As there is no intuitive way of nding a deterministic functional relationship, the problem must be tackled statistically, for example by estimating a regression functional. Similar problems arise also in many other settings and have received some attention in the literature (under the name of "activity classi cation/recognition"). For a systematic review, we refer to [23]. On the one hand, we simply adopt the state-of-the art being described in [23], which makes use of standard machine learning theories, implemented in the well-known Weka tool kit for instance. On the other hand, we have for the rst time provided a thorough theoretical justi cation of this approach, in terms of su cient conditions under which the stochastic guarantees (yielded by standard theories) to nd a classi er near to the supremal accuracy carry over to the non-standard case we are faced with. This work will be published in the near future and is out of the scope of this contribution.
Finally, reconstruction of Postural Sway can be done by twofold numerical integration over the intervals labeled with "standing". 4.2
The suitability of the theory sketched in the previous section must be validated by means of patient studies, as having an estimator for nding functionals near to the optimal accuracy with high probability does not mean that this optimum is "good" enough for the application at hand. To this aim, our project partner Sophia AG Bamberg, who, among other services, provides technical security equipment for elderly people, recruits elderly probands among his clients. Excluded are persons su ering from special diseases like Parkinson. This is not a real restriction, because it is a priori known that these people are at a high risk of falling For the validation of the rst stage, we are currently letting these probands, wearing our accelerometer, execute everyday tasks, observed by a supervisor who logs the actual activity traces. These samples are aggregated and "fed into" the learning procedures of Weka. According to our analysis conducted so far, we can state that it is possible to reach 90-95 per cent (cross-fold evaluation) accuracy after an individual training phase of 15 minutes for each person. Opposed to the great bene ts of the system, this e ort seems well justi ed. For the validation of the step time reconstruction, persons will be asked to walk, while being equipped with both the accelerometer and insoles like the ones used by Hausdor . This again yields samples for a supervised training and cross-fold evaluation.
For validation of Postural Sway reconstruction, we will ask probands (again equipped with the accelerometer) to stand on a special platform designed to measure trunk oscillations. Afterwards, we can numerically process (twofold integration) the sequence of acceleration samples and appropriately compare the reconstructed oscillation to the actual one measured directly. 4.3
After the (hopefully) successful validation and ne-tuning of the single processing stages, an evaluation of the composite system's fall-risk prognosis must be carried out in the eld. To this aim, a group of probands will be asked to wear the sensor devices at home, making available large amounts of data that will be used to compute fall-risk predictions. These will be compared to the prognoses of an approved fall-risk assessment tool, serving as a gold standard. 5
We have proposed the concept of a context-based system for fall-risk assessment of elderly persons at home, along with a detailed outline of our development methodology. Beyond the perspective of making fall-risk assessment a ordable for a much larger part of the elderly population than it has been achieved up to now, our approach can be tailored towards other medical settings. In the future, we plan to explore parametric statistic models for the key problem of activity classi cation. Furthermore, in case of positive results of the planned evaluation, it would be desirable to proceed to a prospective evaluation. For this stage, we (resp. our project partners) are planning to come up with a fully functional prototype, including wireless data transmission to a base station connected to some suitable WAN. This prototype is supposed to serve as a basis for a commercial product.
We would like to thank our project partner Sophia AG for their funding and the Medical Valley EMN e.V. for creating a stimulating context for research. 12. Verghese, J., Holtzer, R., Lipton, R.B., Wang, C.: Quantitative gait markers and incident fall risk in older adults. Journal of Gerontology 64(8) (Aug 2009) 896{901 13. Delbaere, K., Close, J.C.T., Heim, J., Sachdev, P.S., Brodaty, H., Slavin, M.J., Kochan, N.A., Lord, S.R.: A multifactorial approach to understanding fall risk in older people. Journal of the American Geriatrics Society 58(9) (Sep 2010) 1679{ 1685 14. Shumway-Cook, A., Baldwin, M., Polissar, N.L., Gruber, W.: Predicting the probability for falls in community-dwelling older adults. Physical Therapy 77(8) (Aug 1997) 812{819 15. Tinetti, M.E., Williams, T.F., Mayewski, R.: Fall risk index for elderly patients based on number of chronic disabilities. The American Journal of Medicine 80(3) (Mar 1986) 429{434 16. Koepke, S., Meyer, G.: The tinetti test: Babylon in geriatric assessment. Zeitschrift fuer Gerontologie und Geriatrie 39(4) (Aug 2006) 288{291 17. Lajoie, Y., Girard, A., Guay, M.: Comparison of the reaction time, the berg scale and the abc in non-fallers and fallers. Archives of Gerontology and Geriatrics 35(3) (2002) 215{225 18. Mathie, M.J., Coster, A.C.F., Lovell, N.H., Celler, B.G.: Accelerometry: providing an integrated, practical method for long-term, ambulatory monitoring of human movement. Physiological Measurement 25(2) (2004) R1 19. Aminian, K., Robert, P., Buchser, E., Rutschmann, B., Hayoz, D., Depairon, M.: Physical activity monitoring based on accelerometry: validation and comparison with video observation. Medical and Biological Engineering and Computing 37 (1999) 304{308 10.1007/BF02513304. 20. Sun, M., Hill, J.: A method for measuring mechanical work and work e ciency during human activities. Journal of Biomechanics 26 (1993) 229 { 241 21. Corradini, M.L., Fioretti, S., Leo, T., Piperno, R.: Early recognition of postural disorders in multiple sclerosis through movement analysis: a modeling study. IEEE Transactions on Biomedical Engineering 44(11) (1997) 1029{1038 22. Hass, C.J., Waddell, D.E., Fleming, R.P., Juncos, J.L., Gregor, R.J.: Gait initiation and dynamic balance control in parkinson's disease. Archives of Physical Medicine and Rehabilitation 86 (2005) 2172 { 2176 23. Preece, S., Goulermas, J., Kenney, L., Howard, D., Meijer, K., Crompton, R.: Activity identi cation using body-mounted sensors-a review of classi cation techniques. Physiological Measurement 30 (2009)