Sensor reliability remains a key bottleneck for realtime inference in earable and wearable systems, where resource constraints limit the use of computationally intensive faulthandling techniques. In this work, we present a lightweight and calibration efficient framework for sensor fault detection and reliability aware fusion in IMU-based earable systems. The proposed approach combines (i) a statistical detection module that characterizes sensor behavior using percentile-based descriptors derived from clean calibration data, and (ii) a modular ensemble mechanism that adaptively adjusts the contribution of each sensor stream based on detected reliability. Focusing on realtime fault mitigation through dynamic confidence modulation, our framework enables robust operation under degraded sensing conditions while preserving low computational overhead. We evaluate the system on three IMU-based datasets and demonstrate consistent fault detection performance across domains with minimal calibration requirements. In addition, we assess classification robustness under varying fault scenarios, including partial and full sensor degradations, showing that the proposed framework maintains stable performance compared to early fusion and single-modality baselines. Finally, we validate the deployment feasibility on microcontroller-class hardware, demonstrating that the full pipeline operates within strict memory and latency constraints typical of earable devices. These results highlight the potential of calibration-efficient, reliability-aware fusion strategies for enabling robust and scalable sensing in resource-constrained wearable systems.

This paper explores effective deep learning methods for sensor-based Human Activity Recognition (HAR), emphasizing their implementation in resource-limited wearable devices where Microcontroller Units (MCUs) impose restrictions on memory and processing capabilities. We aim to minimize both computational and memory requirements while ensuring high recognition accuracy. We offer a benchmark comparison of pruning and quantization optimization techniques versus lightweight models enhanced through attention mechanisms and knowledge distillation, from both recognition success and resource-efficiency angles. We evaluate two leading deep learning architectures, DeepConvLSTM and SqueezeNet, across four benchmark HAR datasets: Opportunity, Sensors, Wisdm, and Pamap2. For devices with limited memory capacity, we recommend using lightweight models that integrate attention mechanisms and knowledge distillation. We emphasize that quantization should be prioritized to enhance efficiency, with pruning acting as a secondary approach. Additionally, we provide practical guidelines for deploying optimized HAR models on resource-constrained wearable devices.

Color Vision Deficiency (CVD) affects nearly 8 d > 1.4 ), suggesting clinically meaningful gains for individuals with CVD. These findings highlight the potential of personalized AR interventions to improve visual accessibility and quality of life for people affected by CVD.

Silent communication is vital in high-noise environments such as construction sites, where conventional voice interfaces often fail. We present a silent speech recognition (SSR) system on the OpenEarable platform, combining ear canal pressure sensing with inertial measurement units (IMUs). We collected a dataset of 14 commands, including seven essential control words (e.g., Stop, Go) and their phonetically similar counterparts (e.g., Stop, Top), from 20 participants. The dataset is used for training various machine learning models. The results show that IMU-based sensing is the primary driver of robust, noise-resilient SSR, achieving 91.95% F1-score by capturing subtle jaw and head movements. Pressure-only sensing underperforms, while multimodal fusion with CNNs improves performance further. To enhance real-world deployability, we propose a binary-plus-multiclass pipeline that first filters intended commands from phonetically similar inputs. Our findings show that deeper CNNs and natural head gestures provide additional robustness. Deployment profiling confirms real-time inference (402–650 ms latency) on embedded hardware.

Wi-Fi-based human activity recognition is promising but often limited by costly retraining and poor scalability. This work introduces a consensus-based framework for distributed CSI sensing, enabling robust and scalable activity recognition with minimal training and communication overhead. Transmitter–receiver pairs are ranked based on their short training momentum and allocated by a central coordinator (e.g., a router) to monitor specific locations. Experiments across three locations and twelve participants show that a location-aware consensus approach matches optimally placed solutions (F1 =0.98) while improving robustness and temporal stability during dynamic activity flows.

Cycling, a sustainable active transportation, offers significant health, environmental, and societal benefits. Despite these advantages, cyclists are among the most vulnerable road users, with high fatality rates and many single-bicycle crashes (SBC) going unreported. Progress in cycling safety research is limited by the absence of comprehensive, real-world datasets capturing the interplay between cyclist physiology, bicycle motion, environmental perception, and contextual information. To address this, we present BicycleSense360, a modular, scalable, and real-time data acquisition system (DAS) for e-bicycles that unifies multi-modal sensing and is capable of supporting edge inference. The platform integrates synchronized data streams across five domains: (i) environmental perception; (ii) bicycle motion and motor telemetry; (iii) cyclist physiology; (iv) contextual information; and additionally includes a fifth domain v) bidirectional cyclist feedback, which supports real-time alerts and subjective experience logging. All sensors are orchestrated through a lightweight, socket-based publisher-subscriber architecture running on a Micro-controller unit. Unlike prior systems limited to isolated sensing or post-hoc analysis, BicycleSense360 ensures tight temporal alignment across data streams using a Last-In-First-Out (LIFO) truncation heuristic. We demonstrate synchronized multi-modal data collection under representative naturalistic cycling condition and visualize signal alignment across four domains.

Motion sensors integrated into wearable and mobile devices provide valuable information about the device users. Machine learning and, recently, deep learning techniques have been used to characterize sensor data. Mostly, a single task, such as recognition of activities, is targeted, and the data is processed centrally at a server or in a cloud environment. However, the same sensor data can be utilized for multiple tasks and distributed machine-learning techniques can be used without the requirement of the transmission of data to a centre. This paper explores Federated Transfer Learning in a Multi-Task manner for both sensor-based human activity recognition and device position identification tasks. The OpenHAR framework is used to train the models, which contains ten smaller datasets. The aim is to obtain model(s) applicable for both tasks in different datasets, which may include only some label types. Multiple experiments are carried in the Flower federated learning environment using the DeepConvLSTM architecture. Results are presented for federated and centralized versions under different parameters and restrictions. By utilizing transfer learning and training a task-specific and personalized federated model, we obtained a similar accuracy with training each client individually and higher accuracy than a fully centralized approach.

Visual distractions among cyclists significantly reduce their situational awareness, increasing the likelihood of accidents. This study introduces the use of an open-source OpenEarable device, equipped with onboard inertial measurement units (IMU), as an easy and non-invasive way to detect visual distractions by measuring head movements linked to such behaviors. Head movement data from 20 participants were collected during natural cycling situations using earable IMU sensors. Both classical machine learning and deep learning techniques are employed to analyze the data to detect visual distractions. Support Vector Machine (SVM) and Convolutional Neural Network (CNN) achieve weighted F1 scores of 0.85 and 0.87, and Cohen's Kappa scores of 0.74 and 0.59, respectively. These findings highlight the potential of earable devices in real-time distraction detection and establish a foundation for future wearable safety systems for cyclists.

The advancement and widespread adoption of computing technology has yielded services that could help mitigate the climate crisis. However, the retirement of obsolete equipment, the consumption of rare earth materials, and the escalating energy demands associated with massive data processing and cloud infrastructures have raised new environmental dilemmas. Existing design and development methodologies primarily focus on fulfilling functional requirements and improving performance. In this article, we argue that these methodologies must be augmented with sustainability considerations encompassing energy efficiency, material usage, longevity, and upgradability. Solutions at different layers of the system stack, from the physical to the application layer, must be integrated. Moreover, there should be a strong focus on the transparency of sustainability metrics across the whole computing continuum. Building on fruitful discussions at the International Lorentz Workshop on Future Computing for Digital Infrastructures, we advocate novel approaches in the design, development, and operation of the computing continuum.

Smart city development is a complex, transdisciplinary challenge that requires adaptive resource use and context-aware decision-making practices to enhance human functionality and capabilities while respecting societal and environmental rights, and ethics. There is an urgent need for action in cities, particularly to (i) enhance the health and wellbeing of urban residents while ensuring inclusivity in urban development (e.g., through the intelligent design of public spaces, mobility, and transportation) and (ii) improve resilience and sustainability (e.g., through better disaster management, planning of city logistics, and waste management). This paper aims to explore how neuroscientific and neurotechnological solutions can contribute to the development of smart cities, as experts in various fields underline that real-time sensing designs and control algorithms inspired by the brain could help build and plan urban systems that are healthy, safe, inclusive, and resilient. Motivated by the potential interplay between societal challenges and these emerging technologies, we provide an overview of state-of-the-art research through a bibliometric analysis of neurochallenges within the context of smart cities using terms and data extracted from the Scopus database between 2018 and 2022. The results indicate that smart city research remains fragmented and technology-driven, relying heavily on internet of things (IoT) and artificial intelligence (AI)-based technologies. Mostly, it also lacks careful integration and adoption tailored to societal goals and human-centric concerns. In this context, the article explores key research streams and discusses how to create new synergies and complementarities in the challenge-technology intersection. We conclude that realizing the vision of smart cities at the nexus of neuroscience, technology, urban space, and society requires more than just technological progress. Integrating the human dimension alongside various technological tools and systems is crucial. This necessitates better interdisciplinary collaboration and co-production of knowledge toward a hybrid intelligence, where synergies of education and research, technological innovation, and societal innovation are genuinely built. We hope the insights from this analysis will help orient neurotechnological interventions on urban living and ensure they are more responsive to societal and environmental challenges as well as to legal and ethical concerns.

Many real-world applications, such as smart homes, personal healthcare and fitness tracking, benefit from sensor-based human activity recognition (HAR), which identifies the patterns of human activities. Machine learning models are trained on the data collected from sensors, mostly the motion sensors, embedded in wearable devices. However, in this approach, a model cannot learn new tasks independently without total re-learning. The continual learning approach has emerged to tackle this problem. Various techniques have been proposed to enable continual learning, as it has been widely studied in computer vision. This paper suggests a framework for assessing how well different settings of a replay-based technique perform over a large HAR dataset under class incremental continual learning scenarios. Experimental results show that a larger sample size and random sampling method for replay data selection provide accuracy results which are close to the upper bound where all data is available at the start.

With their automatic feature extraction capabilities, deep learning models have become more widespread in sensor-based human activity recognition, particularly on larger datasets. However, their direct use on mobile and wearable devices is challenging due to the extensive resource requirements. Concurrently, attention-based models are emerging to improve recognition performance by dynamically emphasizing relevant parts of features and disregarding the irrelevant ones, particularly in the computer vision domain. This study introduces a novel application of attention mechanisms to smaller deep architectures, investigating whether smaller models can achieve comparable recognition performance to larger models in sensor-based human activity recognition systems while keeping resource usage at lower levels. For this purpose, we integrate the convolutional block attention module into a hybrid model, deep convolutional and long short-term memory network. Experiments are conducted using five public datasets in three model sizes: lightweight, moderate and original. The results show that applying attention to the lightweight model enables achieving similar recognition performances to the moderate-size model, and the lightweight model requires approximately 2–13 times fewer parameters and 3.5 times fewer flops. We also conduct experiments with sensor data at lower sampling rates and from fewer sensors attached to different body parts. The results show that attention improves recognition performance under lower sampling rates, as well as under higher sampling rates when model sizes are smaller, and mitigates the impact of missing data from one or more body parts, making the model more suitable for real-world sensor-based applications.

This paper investigates the potential of earables for real-time boxing gesture recognition. While prior research explores earables in sports, there is a gap in applying them to boxing, particularly for defensive manoeuvre recognition. We address this gap by exploring the capability of real-time Inertial Measurement Unit (IMU)-based boxing head gesture recognition using the open-source OpenEarable framework. We employ classical machine learning and dynamic time-warping (DTW) approaches. A dataset across left/right slips, rolls, and pullbacks is collected from a hobbyist boxer. Our results suggest that DTW combined with gesture templates derived from barycenter averaging achieves high gesture recognition accuracy. The implemented algorithm achieves a testing accuracy of 99% on the collected dataset. This performance is further validated in a real-world scenario, where the algorithm maintains an overall accuracy of 96%. Additionally, the system demonstrates robustness to variations in gesture execution speed and intensity.

While applying deep learning models has significantly enhanced the performance of sensor-based human activity recognition (HAR), their deployment on resource-limited mobile and wearable devices presents challenges. While reducing model size is an alternative, it often leads to performance degradation. This study introduces the integration of attention mechanisms and knowledge distillation techniques to enhance the recognition performance of lightweight sensor-based human activity recognition models. Employing three activity recognition datasets (Opportunity, Wisdm, and Sensors), initial experiments investigate the individual effects of response-based knowledge distillation and attention mechanisms on the performance of the lightweight model. Results demonstrate that attention surpasses distillation in recognition success. To further improve performance, we introduce two combined approaches: response and attention-based distillation and response-based distillation with attention directly applied to the lightweight model. Both approaches outperform attention alone, while the latter, by directly incorporating attention into the student without requiring an attention-based teacher, eliminates the need for a pre-trained attention-based teacher. Evaluating resource consumption, the performance-boosted lightweight model is compared to a moderate-size model, requiring approximately 3-4 times more flops and parameters. The improved lightweight model outperformed the moderate-size model across all datasets regarding recognition performance.

This study investigates the prediction of mental well-being factors—depression, stress, and anxiety—using the NetHealth dataset from college students. The research addresses four key questions, exploring the impact of digital biomarkers on these factors, their alignment with conventional psychology literature, the time-based performance of applied methods, and potential enhancements through multitask learning. The findings reveal modality rankings aligned with psychology literature, validated against paper-based studies. Improved predictions are noted with temporal considerations, and further enhanced by multitasking. Mental health multitask prediction results show aligned baseline and multitask performances, with notable enhancements using temporal aspects, particularly with the random forest (RF) classifier. Multitask learning improves outcomes for depression and stress but not anxiety using RF and XGBoost.

Earables, wearable devices worn around the ear, offer new possibilities for sports applications requiring precise head movement analysis, such as boxing. However, boxing-specific gesture recognition using IMU sensors integrated into earables remains underexplored. This work addresses this gap by investigating the potential of the open-source OpenEarable platform for real-time recognition of defensive boxing manoeuvres, including slipping, rolling and pulling back. We propose an extension to OpenEarable, integrating a Python server that leverages machine learning and dynamic time warping for gesture recognition. Furthermore, the web dashboard is enhanced to enable server communication and implement a gesture mirroring feature, providing real-time visual feedback. Real-time testing achieved a high accuracy of 96%, with feedback delivered within one second. All the system components are made available in a GitHub repository.

The traditional method to authenticate users on mobile devices or applications requires usernames and passwords. This approach authenticates the user only at the entry point of an application without providing continuous authentication during the whole session. This paper explores the use of behavioral biometrics, which involves tracking the unique movements of a user while interacting with a device, for continuous authentication on a mobile banking application. As a methodology, we use binary classification and explore the performance of deep learning algorithms. A dataset is collected from 45 participants using a mobile banking application in Turkey. We train four different types of deep architectures, including Multilayer Perceptron (MLP), LSTM, bi-directional LSTM, and convolutional LSTM. The dataset includes data from both touch screens and motion sensors. The results of the experiments reveal that MLP and the convolutional-LSTM algorithms achieve the best performance on raw data from both motion sensors and touch screens. Accuracy rates are over 99.85%, and FAR, FRR, and EER are below 0.5%.

Human activity recognition (HAR) enables the recognition of the activities of daily living using signals from motion sensors integrated into mobile and wearable devices. One of the challenges is the uniqueness of each individual with his/her different characteristics. A general model trained without user data may perform poorly on specific users. Another challenge is running deep learning (DL) models on mobile and wearable devices due to their limited resources. In this paper, to cope with these two challenges, we use transfer learning to build personalized models and model compression for running DL algorithms. We examine the impact of different DL architectures, the number of layers to be fine-tuned, the amount of user training data, and the transfer to new datasets on the performance of HAR. We compare the performance of the transferred models with general and user-specific models in terms of F1 score, training and inference time.

Human activity recognition (HAR) is a research domain that enables continuous monitoring of human behaviors for various purposes, from assisted living to surveillance in smart home environments. These applications generally work with a rich collection of sensor data generated using smartphones and other low-power wearable devices. The amount of collected data can quickly become immense, necessitating time and resource-consuming computations. Deep learning (DL) has recently become a promising trend in HAR. However, it is challenging to train and run DL algorithms on mobile devices due to their limited battery power, memory, and computation units. In this paper, we evaluate and compare the performance of four different deep architectures trained on three datasets from the HAR literature (WISDM, MobiAct, OpenHAR). We use the TensorFlow Lite platform with quantization techniques to convert the models into lighter versions for deployment on mobile devices. We compare the performance of the original models in terms of accuracy, size, and resource usage with their optimized versions. The experiments reveal that the model size and resource consumption can significantly be reduced when optimized with TensorFlow Lite without sacrificing the accuracy of the models.

The extent to which languages share properties reflecting the non-linguistic constraints of the speakers who speak them is key to the debate regarding the relationship between language and cognition. A critical case is spatial communication, where it has been argued that semantic universals should exist, if anywhere. Here, using an experimental paradigm able to separate variation within a language from variation between languages, we tested the use of spatial demonstratives—the most fundamental and frequent spatial terms across languages. In n = 874 speakers across 29 languages, we show that speakers of all tested languages use spatial demonstratives as a function of being able to reach or act on an object being referred to. In some languages, the position of the addressee is also relevant in selecting between demonstrative forms. Commonalities and differences across languages in spatial communication can be understood in terms of universal constraints on action shaping spatial language and cognition.