Abstract:
The Internet of Things (IoT) has established itself as an indispensable part of current age of user centric connectivity. The domain of IoT covers a widespread spectrum of daily life applications such as smart healthcare facilities which have greatly benefitted with the evolutionary advancements in sensor technologies and IoT. This progress has led to innovative developments like Human Activity Recognition (HAR) systems, smart movement detectors, fitness tracking, ingestible sensors, personalized emergency response systems and Fall Detection Systems (FDS) etc. Fall detection is now a pertinent publicconcern because of high prevalence and detrimental impact of falls on the young and the elderly. A fall detection system gathers information from sensors to differentiate between falls and activities of daily life (ADLs). Hence, the integrity of collected data becomes imperative. A pressing challenge when dealing with wearable sensors to detect falls is that of unreliable data delivery and loss of information leading to missing values observed in data. This missingness can occur due to several reasons and has crucial effects on the performance of a fall detection system resulting in inaccurate, faulty outcomes. Dealing with such insufficient and incorrect sensor data becomes critical for patient health and safety. This research investigates the missing data problem in terms of IoT applications and in particular for sensor based fall detection. Moreover, the current imputation methods and proposed solutions are also analyzed. This analysis leads to the conclusion that current solutions for missing data problem in fall detection systems are very limited. The manuscript proposes a deep learning based fall detection solution to handle missing values and identify falls. This is achieved by a Recurrent Neural Network model, with underlying stacked bidirectional LSTM blocks, which treats fall detection as a sequence classification problem and exploits the patterns and intrinsic association between the variables in data.
