A Novel Embedded Deep Learning Wearable Sensor for Fall Detection

The longer lifespans, the growing aging population, the prevalence of age-related illness, and the degeneration of the body's balance control system have led to extensive research to preserve independence and mobility in individuals. Accidental falls are a common occupational hazard and also a concerning problem due to age-related reduced stability and balance control, changes in reaction time, and postural disorders.

The advancement in body-wearable sensor technologies using electromechanical sensors to track, monitor, and detect body movements and activities, especially in older adults, could assist in case of necessity and prevent falls by prompt and timely fall detection.

The researchers developed an innovative deep-learning approach specifically designed for edge devices that detect falls. This device is based on a printed circuit board with the STM32U575xx microcontroller. It integrates an accelerometer, gyroscope, and pressure sensor to track movements and height variations.

The firmware, developed using STM32CubeIDE and a static C++ library, allows customization of parameters to enhance fall detection using deep learning models. The device underwent comprehensive testing on the SisFall and a self-collected dataset, which included a new fall, the syncope, and the device being kept in the right pocket, ensuring its robust performance.

All the signals were pre-processed to match the same sample frequency and reduce noise. The researchers also computed the Acceleration Vector Magnitude (AVM). The extracted features fed a Feed-Forward neural network selected due to its small memory footprint. The device continuously gathers data until the AVM signal exceeds 2.5g, triggering the deep learning algorithm. Data is collected for 5 seconds, with the fall event at its center, to improve model predictions.

The results showed a high accuracy of 98.05%, even without pressure data. However, false negatives were a concern, as some falls went unreported. But, adding pressure data increased accuracy and decreased false negatives. Combining the self-collected dataset with the SisFall one improved the results to an accuracy of 99.38%. The model adapted effectively to various datasets and had an inference time of 25ms.

Battery life tests revealed that decreasing the activation threshold reduced battery life, indicating that thresholds should be adjusted based on user activity level for best performance.

The results showed that the suggested strategy outperformed existing techniques like signal derivation, sacrifice memory efficiency, and computational simplicity in model accuracy and inference time. For example, the proposed system requires only 60 kb compared to 1.5 GB in convolutional neural networks (CNNs). It also has a 30ms inference time, which is far more efficient than CNN's 500ms or more.  Moreover, CNN's intrinsic drawback is its inability to adapt to different device orientations and sensor axes.

The work prioritizes minimum memory consumption, fast inference, and computational simplicity. However, the study faced a few limitations: the dataset was limited to younger participants, it lacked older individuals, and an adaptable automatic threshold was needed for different user groups, such as workers or older people.

The proposed fall detection model features dense layers, a minimal memory footprint, and low-power data collection, making it suitable for real-world applications and substantially impacting healthcare, individuals, and society.