Search

• The study compares the performance of 13 artificial intelligence-based fall detection algorithms when applied to accelerometer signals captured at the waist of the test subject with a wearable device. • Unlike most related studies, which just use a single data repository to test their proposals, the analysis employs up to 11 different public laboratory-based datasets, as well as two databases with signals measured during real falls, and four repositories with mobility samples gathered during long-term monitoring campaigns. • The study considers not only the typical intra-dataset evaluation procedure followed by the literature but also cross-dataset evaluation (i.e. using data from a different repository for testing than that considered for training). • The performance metrics show that the behavior of all algorithms significantly degrades under cross-dataset evaluation conditions, both when tested with samples from another dataset generated in the laboratory and with real fall or long-term monitoring samples. • The results show that the algorithms tend to overfit the patterns of emulated falls (obtained from laboratory based datasets) used for training. Thus they are not able to extrapolate their learning correctly to conventional movements or new fall patterns, which reflects in a high percentage of undetected falls and a high hourly rate of false alarms. • The obtained results clearly question the evaluation procedures of fall detectors used so far in the literature, highlighting the importance of testing this type of detectors in realistic scenarios with signals from real falls and actual daily life movements. The evaluation of fall detection systems based on wearables is controversial as most studies in the literature benchmark their proposals against falls that are simulated by experimental subjects under unrealistic laboratory conditions. In order to systematically investigate the suitability of this procedure, this paper evaluates a wide set of artificial intelligence algorithms used for fall detection, when trained with a large number of datasets containing acceleration samples captured during the emulation of falls and ordinary movements and then tested with the signals of both actual falls and long-term traces collected from the constant monitoring of users during their daily routines. The results, based on a large number of repositories, show a remarkable degradation in all performance metrics (sensitivity, specificity and false alarm hourly rate) with respect to the typical case in which the detectors are tested with the same types of laboratory movements for which they were trained. [ABSTRACT FROM AUTHOR]

• Most studies on wearable fall detection systems are evaluated with falls emulated in a laboratory. • Clinical trials show that wearable fall detectors underperform in actual application scenarios. • Two repositories containing real falls are compared with 18 well-known datasets of simulated falls. • Fall dynamics are characterized with 15 statistics, computed from the accelerometry components. • The comparison reveals the noteworthy discrepancies between real-world and mimicked falls. • Dynamics of real and fake falls may differ more than those of real falls and other daily activities. • These results call into question the evaluation procedures utilized by the literature to date. Automatic fall detection is one of the most promising applications of wearables in the field of mobile health. The characterization of the effectiveness of wearable fall detectors is hampered by the inherent difficulty of testing these devices with real-world falls. In fact, practically all the proposals in the literature assess the detection algorithms with 'scripted' falls that are simulated in a controlled laboratory environment by a group of volunteers (normally young and healthy participants). Aiming at appraising the adequacy of this method, this work systematically compares the statistical characteristics of the acceleration signals from two databases with real falls and those computed from the simulated falls provided by 18 well-known repositories commonly employed by the related works. The results show noteworthy differences between the dynamics of emulated and real-life falls, which undermines the testing procedures followed to date and forces to rethink the strategies for evaluating wearable fall detectors. [ABSTRACT FROM AUTHOR]

• Impact of sampling frequency on wearable fall detection system (based on accelerometer signals) is normally neglected. • Sampling rate is normally arbitrarily chosen although it can heavily impact on the battery lifetime of the wearable. • The paper assesses the influence of the sampling rate on the effectiveness of a deep learning classifier (a convolutional neural network) directly fed with an observation window of accelerometry signals. • The study reveals that a frequency of 15–20 Hz may be enough for a proper operation of a wearable (specificity/ sensitivity higher than 95%). • Results are confirmed when the power spectrum of the signals is investigated and when the series are low pass filtered. • The systematic study is based on the massive analysis of 15 public datasets (related literature typically utilizes just 1 or 2 datasets). Last decade has witnessed a major research interest on wearable fall detection systems. Sampling rate in these detectors strongly affects the power consumption and required complexity of the employed wearables. This study investigates the effect of the sampling frequency on the efficacy of the detection process. For this purpose, we train a convolutional neural network to directly discriminate falls from conventional activities based on the raw acceleration signals captured by a transportable sensor. Then, we analyze the changes in the performance of this classifier when the sampling rate is progressively reduced. In contrast with previous studies, the detector is tested against a wide set of public repositories of benchmarking traces. The quality metrics achieved for the different frequencies and the analysis of the spectrum of the signals reveal that a sampling rate of 20 Hz can be enough to maximize the effectiveness of a fall detector. [ABSTRACT FROM AUTHOR]