Why we made it: People with dementia are stated to be 4-5 times as likely to fall down and injure themselves than people without dementia. Without help, they would be in serious trouble. Assistance from clinical caregivers can only come when they are effectively and immediately notified of the patient's need for help. However, current methods of fall detection often involve unreliable, unscalable, and highly intrusive methods such as ceiling mounted camera monitoring. We aim to bring effective monitoring to anywhere, indoors or outdoors.
What we made: The solution involves a lightweight Android application that runs on a smartphone meant to be carried by a patient. It monitors the patient's status in real time, and alerts a caregiver web server and the caregiver's mobile phone in the event the patient falls and needs help. In a request for help, important information such as whether or not the patient is in a dark place (correlated with higher fall likelihood) and the patient's location is sent to the web server also. Moreover, the application monitors and alerts the caregiver should the patient remove the device from their body.
How we made it: The input data for detecting a fall comes primarily from real time sampling of the phone's built in accelerometer. Labeled time series data is used to train a temporal convolution neural network on a computer. The parameters of network is attached to the Android application and run in real time, alerting the web server and the caregiver's phone in the event of a fall.
What was difficult: Obtaining sufficient and statistically unbiased training data for the network proved to be very difficult. The key issue is similar to the classical Bayesian inference problem where a diagnosis method with high specificity and selectivity nonetheless raises many false positives when the disease being tested for is very rare in nature. This is similar to our issue where during any given analysis timeframe (eg 1-2 seconds), the probability of a fall occurring is minuscule. Accounting for this and tuning the network to minimize false positives and false negatives proved to be very challenging.
What we are proud of: Ultimately, we obtained a network that runs well in real time. When a person walks regularly or up and down stairs, or stands up and sits down, there are very few false positives. However, when a person falls, it appears to successfully raise the alarm very reliably.
What we learned: Need data! To train any neural net well, we need a tremendous amount of training data that is statistically unbiased.
What's next? The very expandable and scalable nature of the neural network allows us to, with additional labeled training data, add more discriminative capacity to tell when a person is falling versus other activities like running that might put them in danger, and alert the caregivers accordingly.
Log in or sign up for Devpost to join the conversation.