According to World Health Organization statistics, falls are the second leading cause of unintentional injury deaths worldwide. With older people being particularly vulnerable, detecting, and reporting falls have been the focus of numerous health technology studies. We screened 267 studies and selected 15 that detailed pervasive fall detection and alerting apps that used smartphone accelerometers. The fall datasets used for the analyses included between 4 and 38 participants and contained data from young and old subjects, with the recorded falls performed exclusively by young subjects. Threshold-based detection was implemented in six cases, while machine learning approaches were implemented in the other nine, including decision trees, k-nearest neighbors, boosting, and neural networks. All methods could ultimately achieve real-time detection, with reported sensitivities ranging from 60.4 to 99.3% and specificities from 74.6 to 100.0%. However, the studies had limitations in their experimental set-ups or considered a restricted scope of daily activities—not always representative of daily life—with which to define falls during the development of their algorithms. Finally, the studies omitted some aspects of data science methodology, such as proper test sets for results evaluation, putting into question whether reported results would correspond to real-world performance. The two primary outcomes of our review are: a ranking of selected articles based on bias risk and a set of 12 impactful and actionable recommendations for future work in fall detection.

The emerging field of digital phenotyping leverages the numerous sensors embedded in a smartphone to better understand its user's current psychological state and behavior, enabling improved health support systems for patients. As part of this work, a common task is to use the smartphone accelerometer to automatically recognize or classify the behavior of the user, known as human activity recognition (HAR). In this article, we present a deep learning method using the Resnet architecture to implement HAR using the popular UniMiB-SHAR public dataset, containing 11,771 measurement segments from 30 users ranging in age between 18 and 60 years. We present a unified deep learning approach based on a Resnet architecture that consistently exceeds the state-of-the-art accuracy and F1-score across all classification tasks and evaluation methods mentioned in the literature. The most notable increase we disclose regards the leave-one-subject-out evaluation, known as the most rigorous evaluation method, where we push the state-of-the-art accuracy from 78.24 to 80.09% and the F1-score from 78.40 to 79.36%. For such results, we resorted to deep learning techniques, such as hyper-parameter tuning, label smoothing, and dropout, which helped regularize the Resnet training and reduced overfitting. We discuss how our approach could easily be adapted to perform HAR in real-time and discuss future research directions.