We sought to use mobile phones and standard state-of-the-art machine learning to perform robust fall detection and classification. Instead of hand-picking the most relevant features, we used a large feature set, and had the relevant features selected by the algorithms. By applying classifiers such as SVM and SMLR, we showed that either popular technique performs well when large feature sets can be used, with near perfect accuracies in all cases. This was possible with the standard basic acceleration sensors found in almost all modern smartphones and could thus be implemented in phone apps.
Mobile phones are a convenient platform for recording movements, particularly measuring falls. They have built-in communication protocols that allow simple data logging to the device and wireless transmission. This permits real-time response or, in an experimental setting, compliance verification. Because mobile phones are widely adopted, compliance without verification is already high, as people are used to carrying them. Also, price is significantly reduced due to high production volume. Due to these advantages, mobile phones have the promise to provide a convenient, inexpensive, and objective means to track falls.
Previous work has shown how the accelerometers in mobile phones can be used to classify activities, including falls. Activities such as sitting, standing, walking, and running can be identified from mobile phone accelerations 
. Our use of SVM's is also supported by work showing activity recognition and fall detection that can be over 95% accurate in both cases 
; however, that work was done analyzing the movement of body position based on radio tags. Our experimental design is most similar to that of Lee and Carlisle 
(e.g. simulated falls, phones and dedicated accelerometers) although their accuracy was approximately 80%. This was primarily a result of using a simple threshold-based classification strategy relying primarily on the maximum and minimum accelerations to detect a fall. By comparison, we appear to be getting a lot of mileage out of using these machine learning algorithms. Tolkiehn and colleagues 
used a chest mounted accelerometer with a small feature set to obtain 84% detection accuracy. Uniquely, they also classified direction of fall with 94% accuracy. All of these previous fall detection and classification approaches either used smaller feature sets or simpler classification methods. We believe our work shows that the use of mobile phones for fall detection can be greatly improved when using modern machine learning strategies.
Although this study presents high accuracies, there are a number of issues that will still need to be addressed before such techniques could be used in a more applied setting. First, the mobile phones used here were placed in a standardized position. This allowed highly stereotypical measurements that aided accuracy ratings, but made the results less applicable to the way people carry their mobile phones every day – e.g. a smartphone in a pocket will certainly lead to lower accuracies due to the inconsistent ways it can be carried. Also, the fall-like events from the week-long recordings, though perhaps better than using certain simulated daily activities, may not have provided an adequate control data set. There is no guarantee that the fall-like events selected are a representative sample of potential misclassifications. This false-positive rate would have to be more directly assessed to consider viability for application. Although this result relies on standardized accelerometer positioning, and false positive rates for daily living were not fully explored, this does demonstrate a significant improvement in current fall detection and classification techniques.
Fall detection promises to be important in the context of healthcare. The impact on individuals has already been addressed in the introduction, but the injuries, psychological damage, and increased patterns of inactivity due to falls will also be an increasing burden to health care services 
. Over time this burden is expected to increase dramatically; currently there are approximately 40 million people over age 65, and this number is expected to reach 86.7 million in 2050 
. Most dramatically, the number of people aged 85 or older, the age most likely to suffer health consequences from a fall, is expected to triple by 2050. For this reason, there is an increasing incentive for the health care industry to pursue methods to minimize the number of falls, decrease the type of falls that are more likely to cause injury, and improve emergency response when falls do occur. We believe that any work that facilitates research addressing this issue can impact populations prone to falls, but also the health care infrastructure tasked to care for them.
This work is motivated by the possibility of using the fall detection algorithms in real-world scenarios where patients only have to carry phones with the downloaded app installed. This possibility depends on alleviating two critical, practical concerns – maximizing battery life and minimizing the number of false positives. On T-mobile G1's running android version 1.6 we were able to record directly to memory for approximately 10 hours without recharging depending on the quality of the battery and amount of movement. This is impractical for normal, daily use. Systems such as iFall 
and PerFallD 
applied techniques such as variable sampling rates, background services, simplified processing, and minimizing power-intensive features like screen use and access to storage. With these adjustments, and by using newer phones, the apps were able to run in the background over the course of a day. An additional practical concern is how the application responds to a potential fall during the day. For example, after a fall, a person may not be able to respond but may still require medical attention. The previously mentioned applications 
were able to detect when a fall occurs and respond appropriately, however with those algorithms the rate of false-positives may have been a significant issue. If a potential fall is detected, you may want to ask the user to respond; if they don't respond you may send a message to personal emergency contacts; and if no action is taken by anyone, emergency first responders could be contacted. The number of times each of these interventions occurs should be minimized; otherwise the application may not only be impractical from the user's perspective, but also too costly. By applying the techniques used here, along with improvements in battery management, we believe that fall detection comes one step closer to improving the immediate medical responsiveness after individuals with disabilities fall.
The methods we applied for fall detection and classification are important tools for improving patient outcomes from fall-related research. It is clear that mobile phones can be used for fall detection based on the mobile phone accelerometer readings that were used to acquire the data here. In one sense, the feature selection is actually simpler in this paper because it is automated – so a large feature set can be applied. This advances the application of modern machine learning classifiers such as support vector machines and regularized logistic regression. The rich features sets and state of the art machine learning classifiers are simple and straightforward to use in practice. Based on the high accuracies reported here for both fall detection and classification, we believe that these tools should be considered when addressing the ability to automatically detect falls. Such improvements have the possibility to impact not only fall-related research studies, but may also someday enable practical emergency responsiveness for fall related trauma.