Abstract— This paper presents a medical remote monitoring application which aims at detecting falls. The detection system is based on three modalities: a wearable sensor, infrared sensors and a sound analysis module. The sound analysis is presented briefly. The multimodal fusion is made using the Dempster Schaffer theory through Evidential Network. A first evaluation of the use of data mining techniques in order to extract blindly data representatives is proposed. These representatives are used to continuously increase the system performances. The system is evaluated on a local recorded data base. I. INTRODUCTION Life expectancy is currently increasing in the majority of countries. In 2007, life expectancy was about 79 years (women) and 71 (men) in the world and about 84 years (women) and 77 (men) in France (source: INSEE 2007). Figure 1 - Life expectancy in the world (ONU, 2007) In France, 4.5% of men and 8.9% of women aged more than 65 years has had an accident. Among the elderly, 61 % of the accidents occur at home, and 54% take place in the house. In France also 2 million elderly people fall every year. The consequences of these falls are: 10000 deaths, 30-55 % contusions, 3-13 % fractures, dislocations of an articulation, and shock. The total cost is estimated at 1,034 billion € which represent 10 % of the total health costs in France. P. A. Cavalcante, J. Boudy and B; Dorizzi are with TSP, 9 Rue Charles Fourier, 91000 evry (phone: +331 60 764663; e-mail: paulo.cavalcante@it- sudparis.eu). M. Herbin and F. Blanchard are with CReSTIC, Université de Reims Champagne Ardenne, Chaussée du port - BP541 51012 Châlons-en- Champagne (e-mail: michel.herbin@univ-reims.fr). D. Istrate, M.A. Sehili are with ESIGETEL, 1 Rue du Port de Valvins, 77210 Avon (e-mail: dan.istrate@esigetel.fr). The most well known medical remote monitoring systems developed and evaluated up to now [1,2] include distress situation identification and especially the fall detection feature. Different sensors are used: wearable sensors (accelerometers, magnetic sensors, etc.), video analysis, infrared sensors and sound analysis. Generally a multi-sensor fusion is proposed in order to provide more accurate and reliable information. The potential possible benefits of multi- sensor fusion are the redundancy and complementarity of information. The fusion of redundant information can reduce the overall uncertainty. Moreover, the data from multiple heterogeneous sensors of the medical remote systems present uncertainty and lack of confidence [1, 2]. Among multi-sensor fusion techniques, we have Fuzzy logic, Bayesian methods [3] and the Theory of Evidences based on the Dempster-Shafer theory [4], which are commonly used to process and estimate degrees of uncertainty in the fusion process. These theories are based on graphical representations: Bayesian Networks and Evidential Networks (EN) [5]. This paper proposes and evaluates an evidential network to detect fall situations through a heterogeneous multi- sensors fusion. The original aspect of this work consists of the coupling with a data mining algorithm which enables the continuous adaptation of the EN configuration to the real data. This research is being conducted under the European Project CompanionAble 1 an internationally active group dedicated to carrying out leading-edge research in computer vision and signal processing for human-machine communication, including patient home-care, gesture-based interaction, biometry, video surveillance. The resulting evidential network will be further applied in the French National Project Sweet-Home which aims at supplying a user-friendly (sound/speech and touch) interface to domotics. II. PROPOSED MEDICAL REMOTE MONITORING PLATFORM A medical remote monitoring platform was developed at Telecom SudParis elaborated with the close collaboration of ESIGETEL and U558-INSERM. This distress situation detection platform is composed of three modalities: infrared sensors (GARDIEN [7]), a wearable sensor (RFPAT [8]) and sound analysis (ANASON [6]). The platform comprises two simulated rooms architecture filled with the modalities sensors (more details in [9]). RFPAT system was designed for remote monitoring of vital and actimetric signals recorded on the person. This system is composed of a wearable terminal carried by the patient that can automatically identify distress situations 1 www.companionable.net First Steps in Adaptation of an Evidential Network for Data Fusion in the Framework of Medical Remote Monitoring P.A. Cavalcante, M.A. Sehili, M.Herbin, D. Istrate, F. Blanchard, J. Boudy and B. Dorizzi