Accelerometer-Based Fall Detection for Smartphones Bruno Aguiar * , Tiago Rocha * , Joana Silva * and Inˆ es Sousa * * Fraunhofer Portugal AICOS Rua Alfredo Allen, 455/461, 4200-135 Porto, Portugal Email: {bruno.aguiar, tiago.rocha, joana.silva, ines.sousa}@fraunhofer.pt Abstract—Falls are considered the main cause of fear and loss of independence among the elderly population and are also a major cause of morbidity, disability and health care utilization. In the majority of fall events external support is imperative in order to avoid major consequences. Therefore, the ability to automatically detect these fall events could help reducing the response time and significantly improve the prognosis of fall victims. This paper presents a unobtrusive smartphone based fall detection system that uses a combination of information derived from machine learning classification applied in a state machine algorithm. The data from the smartphone built-in accelerometer is continuously screened when the phone is in the user’s belt or pocket. Upon the detection of a fall event, the user location is tracked and SMS and email notifications are sent to a set of contacts. The accuracy of the fall detection algorithm here proposed is near 97.5% for both the pocket and belt usage. In conclusion, the proposed solution can reliably detect fall events without disturbing the users with excessive false alarms, presenting also the advantage of not changing the user’s routines, since no additional external sensors are required. Keywords—Fall Detection, Fall Classification, Ageing, Smart- phone, Inertial Sensors, Accelerometer, Classification Algorithms, ADL. I. I NTRODUCTION Falls are a major source of morbidity and mortality in older patients [1], representing 40% of all injury deaths [2]. More than one third of people aged over 65 years old falls each year [2]. Fall incidence is even more serious considering that worldwide the population is aging [1]. Falls are considered the main cause of fear and loss of independence among the elderly population [3]. The occurrence of moderate to severe injuries contributes not only to the decrease of mobility and balance but also to psychological damages. These factors increase the risk of falling resulting in fall recurrence, it has been discovered that previous fallers have a probability of two- thirds of falling again in the next year [3]. The personal, social and economical effects of fall injuries make this an important global health concern. Reliable fall detection and emergency assistance notification are essential to provide adequate care and to increase the quality of life, especially among the elderly. The delay between the fall and the intervention has been found to be related with morbidity and mortality rates [4]. Therefore, direct reporting to caregivers after fall detection can also improve medical care and reduce fall injuries [5]. Most of the commercial systems available in the market for fall detection are wrist bracelets or pendants that require the user to activate an alarm button in case of falling [6]. The system notifies a remote monitoring center that responds to the alarm. However, there are also some solutions for automatic fall detection [7], mainly based on wearable sensors [8], [9], [10], [11], [12]. Systems based on wearable devices use body-attached sensors such as accelerometers, gyroscopes or barometers to acquire kinetic data from human motion. Bourke et al. [8] correctly detected falls among activities of daily living (ADL) in a total of 480 movements, employing threshold-based fall-detection algorithms to data from a tri- axial accelerometer sensor [8] and a bi-axial gyroscope sensor [13] mounted on the trunk. In a real-time evaluation [14] of a system similar to that presented in [8], with users carrying the sensor continuously 8h per day during 5 days, the sensitivity dropped to 41% and the specificity was reduced to 86%. There is a lack of consensus concerning the most appropriate validation protocol and performance assessment of automatic fall detection systems [6]. Noury et al [4] addressed this issue by proposing a systematic validation protocol consisting of 10 different fall events and 10 different ADL events that should be repeated up to three times. More recently, some studies propose the combination of the accelerometer with other sensors to improve fall detection performance. Bianchi et al. [9] and Chen et al. [10] used the barometric pressure sensor to calculate the difference in altitude that occurs during the fall, to improve fall detection performance. Both studies obtained a sensitivity and specificity around 97% using the validation protocol proposed by Noury et al. [4]. In the last years, fall solutions based on the smartphone sensors [15], [16], [17] have been growing due to the de- velopment of inexpensive Micro Electro Mechanical System (MEMS) sensors and their inclusion in smartphones. Using smartphones avoids the need to acquire other wearable sensors and since users are perceiving them as personal, smartphones are less obtrusive items [7]. Despite the fact that elderlies today are not used to interact with smartphones, the future older users who have grown up with the technology will probably become an important market segment for this kind of applications. Although there are some smartphone applications for fall detection [18], most of them lack a representative dataset and consensual methodologies for the validation protocol [7]. iFall application [17] is an example of the use of the smartphone built-in accelerometer to detect fall events. The algorithm is based on acceleration magnitude thresholds, timeouts and long- lie detection. If a fall is detected the system sends a request for help to the caregivers. Abbate et al. [16] developed a waist mounted system implemented as a finite state machine followed by a classification engine using a neural network. The system achieved a sensitivity and specificity of 100%, however it was evaluated in a relatively reduced dataset with 44 fall events and 42 non-fall events. They also implemented the iFall algorithm and tested it on their dataset, obtaining a sensitivity of 82% a and specificity of 81%. Dai et al. proposed PerFallD [15], a pervasive fall detection system implemented 978-1-4799-2921-4/14/$31.00 ©2014 IEEE