The deployment of machine learning in small, energy-constrained devices like wearable medical devices and IoT sensors is often limited by power consumption and security concerns. In-sensor computing, where sensing and processing occur within the same physical domain, offers a potential solution. This approach avoids the energy-intensive steps of signal transduction and digital processing typically found in conventional sensor systems. However, creating effective in-sensor computing devices has been hampered by the challenge of coupling computing functions with sensor functions efficiently. This research addresses this challenge by using a micro-electromechanical system (MEMS) accelerometer to perform both sensing and machine learning. The study demonstrates the feasibility and advantages of in-sensor computing by applying it to a complex real-world task: real-time identification of subtle human gait patterns. This task is chosen because it requires a fully integrated wearable device robust to variations in gait accelerations, walking speed, subject morphology, noise, and sensor imperfections, providing a rigorous test for the effectiveness of the in-sensor approach.

Previous research has explored the development of mechanical computing elements, such as memory cells and logic gates, built using contact switches or resonators. These components offer potential advantages in energy efficiency and speed over their electronic counterparts, particularly for applications with limited power budgets. The integration of these mechanical computing elements with sensors measuring various mechanical properties is a key area of interest, leading to the concept of in-sensor computing. Existing literature includes examples of resonating structures for speech recognition and metamaterial structures for shape identification, demonstrating the feasibility of in-materio computing. These devices perform complex data processing with minimal electronic components, offering a significant departure from traditional sensor systems. The central challenge, however, has been the development of an appropriate coupling between the computing and sensing functions within a single device.

The researchers developed a custom MEMS accelerometer that couples inertial movement of a suspended proof mass (for acceleration sensing) to the nonlinear oscillations of a doubly clamped beam (for computation). This leverages the concept of reservoir computing, using the nonlinear dynamics of the beam to perform data processing. A feedback mechanism enhances the dimensionality of acceleration data representation. The device was packaged with auxiliary electronics to form a miniaturized wearable system. The gait classification task involved four distinct gait patterns: normal (N), toe-out (TO), trunk-lean (TL), and toe-out and trunk-lean (TOTL). Ten healthy subjects participated, walking on a treadmill at various speeds while alternating between these patterns. Data from the MEMS device and a co-located reference accelerometer were acquired. For training, a four-fold cross-validation procedure with ridge regression was used to determine optimal weight vectors for each gait pattern. The trained weight vectors were then implemented in the MEMS device for real-time gait pattern classification. The performance of the MEMS-based in-sensor computing system was compared to conventional systems utilizing an echo-state network (ESN) and logistic regression (LR) algorithms applied to data from the reference accelerometer. Power consumption of the MEMS prototype and the conventional system were also compared. The power consumption of each subsystem within the MEMS prototype was analyzed, and the potential for power reduction through design optimization was explored.

The MEMS device successfully identified the four gait patterns in real-time with high accuracy. For the toe-out pattern, the true positive rate (TPR) exceeded 99%, and the false positive rate (FPR) was below 1%. For the trunk-lean pattern, which is a more challenging classification task, the TPR was approximately 90%, and the FPR was around 10%. The performance of the MEMS-based classifier was robust against variations between subjects and walking speeds. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve, a measure of classifier performance, was consistently high across different conditions. Direct comparison with the ESN and LR algorithms showed that the MEMS device performed similarly to or better than these conventional methods for the gait classification task. Furthermore, even with a relatively low-sensitivity sensor, the MEMS device effectively learned the sensor's limitations and achieved classification performance comparable to algorithms using data from a higher quality commercial accelerometer. Power consumption measurements revealed that the MEMS prototype, while currently comparable to the optimized ESN system (970 ± 10 mW vs. 280 ± 40 mW), can be significantly improved through design optimization (estimated to 94 mW, or even potentially below 12 mW with more complex modifications). The in-sensor system's data security feature was highlighted—only classification labels are transmitted, preserving user privacy by avoiding transmission of raw physiological data.

The successful application of in-sensor computing to the complex task of real-time gait pattern identification demonstrates the viability of this approach for wearable medical devices and edge computing applications. The results highlight the key advantages of the in-sensor approach: the ability to solve non-linear classification tasks effectively (linear classifiers failed this task), robustness to variations in data and non-ideal sensor behavior, small size, and low power consumption. These advantages stem from the co-integration of sensing and computing functions within a single device. The built-in data security provided by transmitting only classification labels is particularly significant for privacy-sensitive medical applications. This research underscores the potential of in-sensor computing to reduce data transmission rates, a critical factor for addressing data congestion and improving battery life in IoT applications. This work represents a substantial step forward in the development of energy-efficient, compact, and secure devices for diverse applications.

This research successfully demonstrated a MEMS-based in-sensor computing device for real-time human gait analysis. The device achieves high classification accuracy, comparable to or exceeding conventional methods, while offering significantly improved power efficiency and inherent data security. The findings highlight the potential of in-sensor computing for applications in wearable medicine and the Internet of Things. Future research directions could focus on further miniaturization, improved sensor sensitivity, and exploration of fully mechanical implementations of reservoir computing to achieve even lower power consumption.

The current MEMS accelerometer's sensitivity is not state-of-the-art. While the device successfully compensated for these limitations through learning, improvements to sensor performance could further enhance classification accuracy and reduce power consumption. The study focused on a specific population of healthy subjects, and further research is necessary to validate the device's performance across diverse populations with varying gait characteristics and conditions. The number of subjects involved in the study was relatively small, limiting the generalizability of findings. Finally, the reported power consumption is based on modeling and could vary slightly in practice.