• Wearable sensor
• Pervasive health-care
• Fall detection
• Barometer
• Wrist-worn sensor

Falls are a leading cause of injury and hospitalization among the elderly population. It has been estimated that about 30\% of people aged 65+ are subject to a fall each year. In light of the relevance of these issues in our ageing society, there has been great interest in finding automatic systems to promptly detect falls. Fall detection systems can help reduce adverse health outcomes, including both physical and psychological consequences.
Wearable devices embedding an accelerometer have been widely investigated as a means to automatically detect falls. However, usability and low-accuracy issues have hindered the adoption of such systems. In order to foster user-acceptance, the system should be comfortable to wear, unobtrusive and accurate. In this context, wrist-worn devices, such as smartwatch or smart bracelets, represent an attractive solution for a number of reasons. In particular, such devices can be worn during daily routines and provide continuous monitoring of falls. On the other hand, accurate recognition of falls exploiting a wrist-worn device is particularly challenging and has been scarcely investigated in the literature.
This work proposes an innovative fall detection method based on a wrist-worn device. For example, the device could be a smartwatch or smart bracelet, embedding accelerometer and barometer sensors. In the proposed method, barometric pressure is combined with acceleration information to achieve high sensitivity (rate of falls correctly detected) and high specificity (rate of potential false alarms that are correctly discarded). More precisely, relevant features are extracted from the barometer and accelerometer signal and then used as inputs to a machine learning classifier. The classifier is responsible for discriminating falls from normal activities.
To evaluate the method, we collected data from twelve volunteers. The volunteers executed some falls, and nine of them used the device during unsupervised daily activities. An off-line analysis shows that the proposed method is sound. The technique achieves promising results with a user-independent training approach. This actually means that the machine learning scheme can be trained with other users' data, which is an important factor as target users (older adults) cannot simulate falls to train the system. More precisely, the achieved average sensitivity is 92.35\% and the average specificity is 99.4\%. These results suggest that an unobtrusive fall detection system based on a single unit worn at the wrist is feasible. The use of the barometer in addition to the accelerometer has proved to be beneficial in order to achieve high accuracy.