With the rapid development of Internet of Things (IoT) technologies, traditional disease diagnoses carried out in medical institutions can now be performed remotely at home or even ambient environments, yielding the concept of the Internet of Health Things (IoHT). Among the diverse IoHT applications, inertial measurement unit (IMU)-based systems play a significant role in the detection of diseases in many fields, such as neurological, musculoskeletal, and mental. However, traditional numerical interpretation methods have proven to be challenging to provide satisfying detection accuracies owing to the low quality of raw data, especially under strong electromagnetic interference (EMI). To address this issue, in recent years, machine learning (ML)-based techniques have been proposed to smartly map IMU-captured data on disease detection and progress. After a decade of development, the combination of IMUs and ML algorithms for assistive disease diagnosis has become a hot topic, with an increasing number of studies reported yearly. A systematic search was conducted in four databases covering the aforementioned topic for articles published in the past six years. Eighty-one articles were included and discussed concerning two aspects: different ML techniques and application scenarios. This review yielded the conclusion that, with the help of ML technology, IMUs can serve as a crucial element in disease diagnosis, severity assessment, characteristic estimation, and monitoring during the rehabilitation process. Furthermore, it summarizes the state-of-the-art, analyzes challenges, and provides foreseeable future trends for developing IMU-ML systems for IoHT.

Robust and accurate human stride length estimation (SLE) using smartphone integrated inertial measurement units (IMU) is essential in pedestrian dead reckoning (PDR) and mobile health applications. However, the change of smartphone carrying mode (i.e. sensor location) in daily usage often leads to significant estimation errors. To address this problem, we propose a novel SLE framework called Mode-Independent Neural Network (MINN) using multi-source unsupervised domain adaptation (UDA) methods. First, we present a hierarchical neural network to extract spatio-temporal features based on multi-level ResNet and GRU. Then, we use adversarial training and a subclass classifier to build a UDA network that can extract mode-invariant features shared by the data from different modes. Finally, we integrate these architectures into an end-to-end learning framework. Through a systematic evaluation under the leave-one-out setting on two public SLE datasets, the MINN outperforms the state-of-the-art algorithms by achieving stride length error rates of 2.5% and 5.1% in supervised settings. We also evaluated the mode-independent adaptability of this model by performing single and multiple UDA tasks. The results demonstrate that the proposed MINN significantly improves the generalization of SLE model under new subjects or modes.

Robust and accurate attitude and heading estimation using Micro-Electromechanical System (MEMS) Inertial Measurement Units (IMU) is the most crucial technique that determines the accuracy of various downstream applications, especially pedestrian dead reckoning (PDR), human motion tracking, and Micro Aerial Vehicles (MAVs). However, the accuracy of the Attitude and Heading Reference System (AHRS) is often compromised by the noisy nature of low-cost MEMS-IMUs, dynamic motion-induced large external acceleration, and ubiquitous magnetic disturbance. To address these challenges, we propose a novel data-driven IMU calibration model that employs Temporal Convolutional Networks (TCNs) to model random errors and disturbance terms, providing denoised sensor data. For sensor fusion, we use an open-loop and decoupled version of the Extended Complementary Filter (ECF) to provide accurate and robust attitude estimation. Our proposed method is systematically evaluated using three public datasets, TUM VI, EuRoC MAV, and OxIOD, with different IMU devices, hardware platforms, motion modes, and environmental conditions; and it outperforms the advanced baseline data-driven methods and complementary filter on two metrics, namely absolute attitude error and absolute yaw error, by more than 23.4% and 23.9%. The generalization experiment results demonstrate the robustness of our model on different devices and using patterns.

The full-temperature bias stability (1σ) is a common indicator used to evaluate the performance of microelectromechanical systems (MEMS) gyroscopes over the entire temperature range. This paper proposes a full-temperature multi-parameter zero-rate output (ZRO) thermal drift compensation method based on a feedforward neural network (FFN). Notably, the proposed method eliminates the need for an external temperature sensor and is applicable to all MEMS gyroscopes. The model integrates four temperature-related parameters: resonant frequency, drive excitation voltage, quadrature suppression voltage, and reference source voltage. After hyperparameter optimization and training, the FFN model was validated on four gyroscopes with different performance levels to assess its generalization capability. Experimental results show that for Gyroscope 1, the full-temperature bias stability improved by 93.8% (reduced from 0.671°/s to 0.042°/s), and the full-temperature ZRO (3σ) suppression rate reached 97.9%. Notably, compared with the traditional multiple linear regression model, the proposed FFN demonstrates strong generalization, reducing the full-temperature bias stability (1σ) by 66%, 67%, 77%, and 44% across the four tested gyroscopes, respectively. Furthermore, implementation on an FPGA platform verifies the potential of the FFN model for real-time, in-situ temperature drift compensation.