IEEE Proof IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 38, NO. 6, NOVEMBER 2008 1 Correspondence Reducing the Uncertainty on Location Estimation of Mobile Users to Support Hospital Work Luis A. Castro and Jesus Favela Abstract—The nature of a context-aware application in hospital work demands a reliable and accurate location system. The activity for which this location information is needed determines to a great extent the relevancy of this contextual variable, since a minor error in delivering patient-based information can be critical. In this correspondence, we present an enhanced technique to infer the location of users in a hospital setting based on the strength of radio-frequency signals received by mobile devices that are used to train a neural network. The approach uses the neighbors surrounding the location to be estimated to track users continuously. This neighborhood eases the training and is used to simulate previous time instant guesses to reduce the location estimation error and alleviate the hopping trajectories of users. The results obtained by using this approach are in the order of 1.3 m for the average distance error during continuous motion. Index Terms—Computer applications, location estimation, neural net- work applications, pattern recognition. I. INTRODUCTION The use of electronic patient records and the adoption of handheld computers are two of the most important trends transforming the use of computers in hospitals. The deployment of wireless networks in hospitals is making it possible to integrate both technologies to pro- vide staff with access to Hospital Information Systems from anywhere within the hospital. Furthermore, this technological convergence opens the door for the deployment of location-aware applications that aid hospital workers to locate colleagues, artifacts, and adapt applications based on their current location [1] or activity [2]. For instance, depend- ing upon some contextual variables (e.g., time and people around), the personal digital assistant (PDA) of a physician approaching a patient’s bed might adapt itself to provide more immediate access to the patient’s medical record. The deployment of such a technology demands adequate location estimation techniques. In the last few years, the location estimation of mobile users has gained considerable attention, which has been favored, in some way, by the growth of research and development of location- based services. A great variety of location systems proposed so far are based on a specialized infrastructure or sensors that have to be deployed within the area of interest. One of the most well-known and widely spread commercial systems is the Global Positioning System (GPS). GPS uses triangulation to compute position from signals received by satellites with an approximate error of 10 m [3], which may be adequate for many open-air applications; however, this system does not work Manuscript received February 27, 2007; revised August 19, 2007. This work was supported in part by the Consejo Nacional de Ciencia y Tecnolog´ ıa (CONACYT) under Grant U-40799. The work of L. A. Castro was supported by a scholarship provided by CONACYT. This paper was recommended by Associate Editor P. Sanz. L. A. Castro is with Manchester Business School, University of Manchester, Manchester, M15 6PB, U.K. (e-mail: luis.castro@acm.org). J. Favela is with the Centro de Investigaci´ on Cient´ ıfica y de Educaci´ on Superior de Ensenada (CICESE), Ensenada 22860, Mexico (e-mail: favela@ cicese.mx). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TSMCC.2008.2001572 properly indoors, and it has some flaws in cities with tall buildings because the satellite signals can be obstructed. Several other approaches have also been proposed, mainly for in- door usage, based on technologies such as ultrasonic pulses [4], [5], in- frared [6], radio-frequency (RF) signals [7]–[12], high-precision video cameras [13], pressure sensors [14], RF identification (RFID) [15], and Global System for Mobile (GSM) networks [16], as well as a mixture of technologies through what is known as sensor fusion [17], and finally, ultra-wideband technology [18], which offers very high precision. However, these systems make use of additional specialized hardware, which usually means considerable installation and mainte- nance costs. Lastly, in an effort to spread location-based services and increase coverage, the use of GSM networks in a real-world setting has been explored with remarkable results [19]. In practice, the choice of any particular location system is determined by a tradeoff between the actual needs of the users and the cost. In a hospital setting, the personnel require information that is highly depen- dent on their location. They are constantly moving around to perform their daily work, which includes visiting patients, locating resources (e.g., medical records), or consulting with other specialists. As men- tioned previously, currently some very accurate estimation methods that require specialized infrastructure exist, but there are also tech- niques with limited accuracy using the WiFi network already installed in the hospital as well as the PDAs that increasingly assist hospital staff in their daily work. Nonetheless, these latter approaches are not accurate enough for hospital applications. Most of the location systems based on this approach provide an average distance error in the order of 2–4 m during continuous motion. Moreover, they often have lim- ited steadiness when approximating the location, i.e., the user seems to jump between two distant places in successive time instants. In this correspondence, we present an approach that enhances the accuracy and consistency of the estimations of the latter methods by using neural networks trained with information from the neighborhood surrounding the location to be estimated. This location system was tested and implemented in a mid-sized local hospital. Due to the public nature of the hospital, the bed concen- tration was considerable. However, even in the most populated rooms, a single bed is within an area of 3.6 m × 2 m; thus, a medical worker can be considered to be with a patient if he/she is within that boundary. One of our main interests was to support the daily activities of hospital staff as hospital workers apparently spend considerable time locating colleagues and patient-related documents [20]. Hence, for our system, we decided to set a target of a maximum distance error of 2 m to deliver location-aware information and 4 m for the purpose of locating people within the premises, since within this range, a person can be reached very easily, visually or by calling his or her name. II. INFERRING THE LOCATION OF USERS Of particular interest are those techniques that use existing wireless local area network (WLAN) infrastructure, since they are cost-effective. These methods use the RF signal strength to determine the location of a user; the signal strength is considered a function of the distance from the mobile device to multiple wireless access points. However, this rela- tionship becomes complex indoors, where the signal strength is affected significantly by walls, furniture, people, and other electronic devices in the environment. Usually, these methods include two phases: 1) offline phase where a radio map is created by sampling the signal 1094-6977/$25.00 © 2008 IEEE