Fei Dou
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Fei Dou has a total of 48 co-authors. Here are the top co-authors by number of shared publications:
- WenZhan Song
College of Engineering, University of Georgia, Athens, GA 30602, USA
6 shared publications - Xiang Zhang
Computer Science, University of North Carolina (UNC) at Charlotte, Charlotte, North Carolina, USA
5 shared publications - Jin Lu
University of Georgia
4 shared publications - Yida Zhang
University of Georgia, Athens, USA
4 shared publications - Yingjian Song
Electrical Computer Engineering, University of Georgia, Athens, Georgia, USA
4 shared publications - Bradley G. Phillips
The University of Georgia, USA, Athens, Georgia
3 shared publications
Latest Publications
- Open Access
- 26 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 111-112
Contact-Free Continuous ABP Waveform Reconstruction via Triaxial Bodyseismography
Continuous arterial blood pressure (ABP) waveform monitoring provides beat-to-beat hemodynamic information for clinical decision-making. However, invasive arterial catheterization limits unobtrusive long-term use. We develop a contact-free method to reconstruct continuous ABP waveforms from triaxial bodyseismography (BSG) signals, which capture bed vibrations induced by cardiac mechanical activity. The key challenge is recovering physiologically meaningful beat-level ABP morphology from indirect, noisy signals transmitted through the body-to-bed pathway. To address this challenge, we develop a dual-branch framework that fuses temporal BSG with wavelet-based time-frequency representations, combining local beat-level morphology with global spectral context. In leave-one-subject-out evaluation on 33 ICU patients, our method achieved a mean absolute error of 12.8 mmHg and a peak normalized cross-correlation of 0.896.
- 26 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 111-112
Open AccessContact-Free Continuous ABP Waveform Reconstruction via Triaxial Bodyseismography
Continuous arterial blood pressure (ABP) waveform monitoring provides beat-to-beat hemodynamic information for clinical decision-making. However, invasive arterial catheterization limits unobtrusive long-term use. We develop a contact-free method to reconstruct continuous ABP waveforms from triaxial bodyseismography (BSG) signals, which capture bed vibrations induced by cardiac mechanical activity. The key challenge is recovering physiologically meaningful beat-level ABP morphology from indirect, noisy signals transmitted through the body-to-bed pathway. To address this challenge, we develop a dual-branch framework that fuses temporal BSG with wavelet-based time-frequency representations, combining local beat-level morphology with global spectral context. In leave-one-subject-out evaluation on 33 ICU patients, our method achieved a mean absolute error of 12.8 mmHg and a peak normalized cross-correlation of 0.896.
- Open Access
- 30 September 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10 (3) , 1-47
LastAct: Trajectory-Guided Latest-Activity Localization for Real-Time Smart-Home Activity Recognition
Human Activity Recognition (HAR) from ambient sensors enables smart-home applications such as health monitoring and assisted living. In realistic deployments, however, sensor events arrive as a continuous stream and activity boundaries are unknown. Sliding-window inference therefore produces many windows that straddle transitions and contain mixed activities, creating boundary contamination that violates the pre-segmented instance assumption used by most benchmarks and models. Moreover, many pipelines under-use spatial context by treating sensor IDs as independent tokens. We present LastAct, a trajectory-centric framework for streaming smart-home HAR that targets the most recent activity under mixed windows while explicitly modeling spatial structure. LastAct projects sensor events onto the home floorplan to form a layout-aligned trajectory image sequence that preserves spatial continuity. A lightweight gate identifies contaminated windows, and a boundary localizer estimates the last transition to enable boundary-guided masking that emphasizes post-boundary evidence and suppresses stale context. For efficiency, we reuse a precomputed layout-aligned template cache to avoid repeated rendering. Empirically, across four public smart-home datasets under near-realistic mixed-activity protocols, LastAct achieves competitive or superior performance on pure windows and yields substantial Macro-F1 gains on cross/mixed windows, demonstrating improved robustness under near-realistic sliding-window regimes.
- 30 September 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10 (3) , 1-47
Open AccessLastAct: Trajectory-Guided Latest-Activity Localization for Real-Time Smart-Home Activity Recognition
Human Activity Recognition (HAR) from ambient sensors enables smart-home applications such as health monitoring and assisted living. In realistic deployments, however, sensor events arrive as a continuous stream and activity boundaries are unknown. Sliding-window inference therefore produces many windows that straddle transitions and contain mixed activities, creating boundary contamination that violates the pre-segmented instance assumption used by most benchmarks and models. Moreover, many pipelines under-use spatial context by treating sensor IDs as independent tokens. We present LastAct, a trajectory-centric framework for streaming smart-home HAR that targets the most recent activity under mixed windows while explicitly modeling spatial structure. LastAct projects sensor events onto the home floorplan to form a layout-aligned trajectory image sequence that preserves spatial continuity. A lightweight gate identifies contaminated windows, and a boundary localizer estimates the last transition to enable boundary-guided masking that emphasizes post-boundary evidence and suppresses stale context. For efficiency, we reuse a precomputed layout-aligned template cache to avoid repeated rendering. Empirically, across four public smart-home datasets under near-realistic mixed-activity protocols, LastAct achieves competitive or superior performance on pure windows and yields substantial Macro-F1 gains on cross/mixed windows, demonstrating improved robustness under near-realistic sliding-window regimes.
- Open Access
- 19 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 303-308
Learning a Simple Directional eHMI: A VR Study of AV-Pedestrian Interaction
External Human-Machine Interfaces (eHMIs) for automated vehicles must remain quickly interpretable if they are to scale to realistic traffic. This study examines a deliberately simple directional light-based windshield eHMI as a first step toward that goal. The eHMI uses a minimal light strip whose highlighted segment shifts based on the pedestrian’s relative position, aiming to convey yielding intent without text or complex visual elements. In a virtual-reality study with 30 participants, pedestrians encountered an automated vehicle with and without the eHMI; half received a brief onboarding video, while the others learnt through repeated exposure. Results suggest that the eHMI supported earlier crossing initiation and increased perceived safety and trust. Onboarding produced immediate benefits, whereas participants without onboarding showed gradual improvements over repeated encounters. These findings indicate that the signal is learnable in a controlled dyadic setting and motivate future VR studies of multi-pedestrian interaction, social influence, and traffic complexity.
- 19 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 303-308
Open AccessLearning a Simple Directional eHMI: A VR Study of AV-Pedestrian Interaction
External Human-Machine Interfaces (eHMIs) for automated vehicles must remain quickly interpretable if they are to scale to realistic traffic. This study examines a deliberately simple directional light-based windshield eHMI as a first step toward that goal. The eHMI uses a minimal light strip whose highlighted segment shifts based on the pedestrian’s relative position, aiming to convey yielding intent without text or complex visual elements. In a virtual-reality study with 30 participants, pedestrians encountered an automated vehicle with and without the eHMI; half received a brief onboarding video, while the others learnt through repeated exposure. Results suggest that the eHMI supported earlier crossing initiation and increased perceived safety and trust. Onboarding produced immediate benefits, whereas participants without onboarding showed gradual improvements over repeated encounters. These findings indicate that the signal is learnable in a controlled dyadic setting and motivate future VR studies of multi-pedestrian interaction, social influence, and traffic complexity.
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10 (2) , 1-31
Contactless Sleep Apnea Detection with Bodyseismography
Sleep apnea is a prevalent yet underdiagnosed sleep disorder. Existing monitoring approaches largely rely on wearable devices or smartphone-based systems, which can be uncomfortable, require active user engagement, or be sensitive to environmental conditions. In this paper, we present a feasibility study of a contactless, engagement-free sleep-apnea screening approach using an under-bed horizontal (head-to-foot) seismic sensor that captures micro vibrations from respiration, heartbeat, and movement. We show that the horizontal axis provides a clearer respiratory signature than the vertical axis, enabling respiration-focused signal analysis. From the seismic signal, we extract 15 features across three families—respiratory heartbeat, and movement—and formulate a minute-level three-class classification task: Normal, OSA+hypopnea , and CSA. Using a strict patient-independent 5-fold cross-validation protocol on 116 subjects, we achieve 85.5% balanced accuracy and 80.6% macro F1 with a random forest classifier; class-wise balanced accuracies are 96.3% (Normal), 75.0% (OSA+hypopnea), and 85.2% (CSA). Overall, this work demonstrates the feasibility of under-bed seismic sensing for window-level apnea-related state classification.
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10 (2) , 1-31
Open AccessContactless Sleep Apnea Detection with Bodyseismography
Sleep apnea is a prevalent yet underdiagnosed sleep disorder. Existing monitoring approaches largely rely on wearable devices or smartphone-based systems, which can be uncomfortable, require active user engagement, or be sensitive to environmental conditions. In this paper, we present a feasibility study of a contactless, engagement-free sleep-apnea screening approach using an under-bed horizontal (head-to-foot) seismic sensor that captures micro vibrations from respiration, heartbeat, and movement. We show that the horizontal axis provides a clearer respiratory signature than the vertical axis, enabling respiration-focused signal analysis. From the seismic signal, we extract 15 features across three families—respiratory heartbeat, and movement—and formulate a minute-level three-class classification task: Normal, OSA+hypopnea , and CSA. Using a strict patient-independent 5-fold cross-validation protocol on 116 subjects, we achieve 85.5% balanced accuracy and 80.6% macro F1 with a random forest classifier; class-wise balanced accuracies are 96.3% (Normal), 75.0% (OSA+hypopnea), and 85.2% (CSA). Overall, this work demonstrates the feasibility of under-bed seismic sensing for window-level apnea-related state classification.
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10 (2) , 1-44
Peak-Detector: Explainable Peak Detection via Instruction-Tuned Large Language Models in Physiological Signal
Accurate peak detection across diverse cardiac physiological signals, including the Electrocardiogram (ECG), Photoplethysmogram (PPG), Ballistocardiogram (BCG), and Bodyseismography (BSG), is fundamental for cardiovascular monitoring but is often hindered by artifacts and signal variability. Conventional algorithms are typically engineered with expert knowledge for a single signal modality, limiting their generalizability. Conversely, deep learning-based methods often lack interpretability, limiting transparency for expert verification and hindering expert-computer interaction. To address these limitations, we introduce Peak-Detector, a novel framework that leverages instruction-tuned Large Language Models (LLMs) for robust, cross-modal, and explainable peak detection. A core innovation of our framework is a “peak-representation” technique that transforms time-series data into a condensed format, preserving critical event information while significantly reducing signal length. This representation provides a crucial inductive bias, guiding the LLM to reason over physiologically meaningful events rather than raw, noisy data. The model is optimized through a two-stage process: supervised fine-tuning (SFT) followed by reinforcement learning (RL) with a multi-objective reward function. The model's self-explanation capabilities are cultivated by fine-tuning on a custom-built Peak-Explanation dataset. Across four modalities—ECG, PPG, BCG, and BSG—spanning seven datasets (six public benchmarks plus one real-world cohort), Peak-Detector demonstrates strong cross-modal performance, achieving best or tied-best detection under clinically relevant temporal tolerance. Beyond accuracy, the generated rationales surface failure modes and support verification and error analysis. Together, these results indicate a transparent and generalizable framework for trustworthy peak analysis and cardiovascular metric extraction.
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10 (2) , 1-44
Open AccessPeak-Detector: Explainable Peak Detection via Instruction-Tuned Large Language Models in Physiological Signal
Accurate peak detection across diverse cardiac physiological signals, including the Electrocardiogram (ECG), Photoplethysmogram (PPG), Ballistocardiogram (BCG), and Bodyseismography (BSG), is fundamental for cardiovascular monitoring but is often hindered by artifacts and signal variability. Conventional algorithms are typically engineered with expert knowledge for a single signal modality, limiting their generalizability. Conversely, deep learning-based methods often lack interpretability, limiting transparency for expert verification and hindering expert-computer interaction. To address these limitations, we introduce Peak-Detector, a novel framework that leverages instruction-tuned Large Language Models (LLMs) for robust, cross-modal, and explainable peak detection. A core innovation of our framework is a “peak-representation” technique that transforms time-series data into a condensed format, preserving critical event information while significantly reducing signal length. This representation provides a crucial inductive bias, guiding the LLM to reason over physiologically meaningful events rather than raw, noisy data. The model is optimized through a two-stage process: supervised fine-tuning (SFT) followed by reinforcement learning (RL) with a multi-objective reward function. The model's self-explanation capabilities are cultivated by fine-tuning on a custom-built Peak-Explanation dataset. Across four modalities—ECG, PPG, BCG, and BSG—spanning seven datasets (six public benchmarks plus one real-world cohort), Peak-Detector demonstrates strong cross-modal performance, achieving best or tied-best detection under clinically relevant temporal tolerance. Beyond accuracy, the generated rationales surface failure modes and support verification and error analysis. Together, these results indicate a transparent and generalizable framework for trustworthy peak analysis and cardiovascular metric extraction.
- Top Cited PapersOpen Access
- 12 May 2026
- Published by IOP Publishing in Progress in Biomedical Engineering
- Vol. 8 (2) , 022013
Cross-subject generalization for EEG decoding: a survey of deep learning methods
Deep learning for cross-subject electroencephalography (EEG) decoding is hindered by high inter-subject variability, which introduces a severe domain shift between training and unseen test subjects. This survey presents a comprehensive review of deep learning methodologies specifically engineered to address this cross-subject generalization challenge. To ground this analysis, we formalize the cross-subject setting as a multi-source domain problem and delineate the rigorous, subject-independent evaluation protocols required for valid assessment. Central to this survey is a systematic taxonomy of the current literature into discrete methodological families, including feature alignment, adversarial learning, feature disentanglement, and contrastive learning. We conclude by examining three critical elements for advancing robust, real-world decoding: the theoretical limitations of current methodologies, the structural value of subject identity, and the emergence of EEG foundation models.
- 12 May 2026
- Published by IOP Publishing in Progress in Biomedical Engineering
- Vol. 8 (2) , 022013
Top Cited PapersOpen AccessCross-subject generalization for EEG decoding: a survey of deep learning methods
Deep learning for cross-subject electroencephalography (EEG) decoding is hindered by high inter-subject variability, which introduces a severe domain shift between training and unseen test subjects. This survey presents a comprehensive review of deep learning methodologies specifically engineered to address this cross-subject generalization challenge. To ground this analysis, we formalize the cross-subject setting as a multi-source domain problem and delineate the rigorous, subject-independent evaluation protocols required for valid assessment. Central to this survey is a systematic taxonomy of the current literature into discrete methodological families, including feature alignment, adversarial learning, feature disentanglement, and contrastive learning. We conclude by examining three critical elements for advancing robust, real-world decoding: the theoretical limitations of current methodologies, the structural value of subject identity, and the emergence of EEG foundation models.
- Open Access
- 10 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 1029-1042
ADLGen: Synthesizing Symbolic, Event-Triggered Sensor Sequences for Smart-Home Human Activity Modeling
- 10 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 1029-1042
Open AccessADLGen: Synthesizing Symbolic, Event-Triggered Sensor Sequences for Smart-Home Human Activity Modeling
- Open Access
- 8 October 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Embedded Computing Systems
- Vol. 24 (6) , 1-32
Deep Q-Learning-Based Mobile Charger Path Planning in Wireless Powered Communication Networks
Wireless Powered Communication Network (WPCN) is a new paradigm to allow low-power wireless devices to exchange data packets and receive stable energy transfer from a power source and thus support autonomous and sustainable network operations without battery replacements. In recent years, we have witnessed the growing deployment of WPCNs in both industrial and consumer IoT systems to support time-triggered and event-triggered monitoring applications. In this article, we present a novel reinforcement learning (RL)-based on-demand path planning framework to plan the trajectory of a Mobile Charger (MC) and schedule the charging sequence of wireless devices to sustain the network operations. A modified Deep Q-learning approach is designed to charge the wireless devices by balancing between their residual energy level and the distance from the MC to the device. This approach minimizes the total distance that the MC travels while ensuring that individual residual energy of a given set of devices is above a designated threshold. Extensive experimental results from both the Gazebo-based high-fidelity simulation and Turtlebot-based physical testbed demonstrate that our approach outperforms the classic scheduling methods (e.g., Nearest Job Next and Earliest Deadline First), state-of-the-art scheduling methods(Extended Particle Swarm Optimization, Enhanced Teaching–Learning-Based Optimization Algorithm and Spatiotemporal Optimization for Charging Scheduling), learning-based methods (e.g., Proximal Policy Optimization and Advantage Actor-Critic) with similar sample sizes for training.
- 8 October 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Embedded Computing Systems
- Vol. 24 (6) , 1-32
Open AccessDeep Q-Learning-Based Mobile Charger Path Planning in Wireless Powered Communication Networks
Wireless Powered Communication Network (WPCN) is a new paradigm to allow low-power wireless devices to exchange data packets and receive stable energy transfer from a power source and thus support autonomous and sustainable network operations without battery replacements. In recent years, we have witnessed the growing deployment of WPCNs in both industrial and consumer IoT systems to support time-triggered and event-triggered monitoring applications. In this article, we present a novel reinforcement learning (RL)-based on-demand path planning framework to plan the trajectory of a Mobile Charger (MC) and schedule the charging sequence of wireless devices to sustain the network operations. A modified Deep Q-learning approach is designed to charge the wireless devices by balancing between their residual energy level and the distance from the MC to the device. This approach minimizes the total distance that the MC travels while ensuring that individual residual energy of a given set of devices is above a designated threshold. Extensive experimental results from both the Gazebo-based high-fidelity simulation and Turtlebot-based physical testbed demonstrate that our approach outperforms the classic scheduling methods (e.g., Nearest Job Next and Earliest Deadline First), state-of-the-art scheduling methods(Extended Particle Swarm Optimization, Enhanced Teaching–Learning-Based Optimization Algorithm and Spatiotemporal Optimization for Charging Scheduling), learning-based methods (e.g., Proximal Policy Optimization and Advantage Actor-Critic) with similar sample sizes for training.
- Open Access
- 7 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 195-198
Cross-Modal Translation and Alignment of Sensor Events for Layout-Aware Activity Modeling
- 7 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 195-198
Open AccessCross-Modal Translation and Alignment of Sensor Events for Layout-Aware Activity Modeling
- Open Access
- 9 June 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 9 (2) , 1-32
Multi-granularity Supervised Contrastive Learning with Online Adaptation for Contactless In-bed Posture Classification
In-bed postures offer valuable information about an individual's sleep quality and overall health conditions, particularly for patients with sleep apnea. However, current in-bed posture classification systems lack privacy-friendly and easy-to-install options. Furthermore, existing solutions do not consider variations between patients and are typically trained only once, neglecting the utilization of time consistency and unlabeled data from new patients. To address these limitations, this paper builds on a seismic sensor to introduce a novel sleep posture framework, which comprises two main components, namely, the Multi-Granularity Supervised Contrastive Learning (MGSCL) module and the ensemble Online Adaptation (oa) module. Unlike most existing contrastive learning frameworks that operate at the sample level, MGSCL leverages multi-granular information, operating not only at the sample level but also at the group level. The oa module enables the model to adapt to new patient data while ensuring time consistency in sleep posture predictions. Additionally, it quantifies model uncertainty to generate weighted predictions, further enhancing performance. Evaluated on a dataset of 100 patients collected at a clinical research center, MGSCLoa achieved an average accuracy of 91.67% and an average F1 score of 91.53% with only 40 seconds of labeled data per posture. In a Phase 2 evaluation with 11 participants over 13 nights in home settings, the framework reached an average accuracy of 85.37% and a weighted F1 score of 83.59% using just 3 minutes of labeled data per common posture for each participant. These results underscore the potential of seismic sensor-based in-bed posture classification for assessing sleep quality and related health conditions.
- 9 June 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 9 (2) , 1-32
Open AccessMulti-granularity Supervised Contrastive Learning with Online Adaptation for Contactless In-bed Posture Classification
In-bed postures offer valuable information about an individual's sleep quality and overall health conditions, particularly for patients with sleep apnea. However, current in-bed posture classification systems lack privacy-friendly and easy-to-install options. Furthermore, existing solutions do not consider variations between patients and are typically trained only once, neglecting the utilization of time consistency and unlabeled data from new patients. To address these limitations, this paper builds on a seismic sensor to introduce a novel sleep posture framework, which comprises two main components, namely, the Multi-Granularity Supervised Contrastive Learning (MGSCL) module and the ensemble Online Adaptation (oa) module. Unlike most existing contrastive learning frameworks that operate at the sample level, MGSCL leverages multi-granular information, operating not only at the sample level but also at the group level. The oa module enables the model to adapt to new patient data while ensuring time consistency in sleep posture predictions. Additionally, it quantifies model uncertainty to generate weighted predictions, further enhancing performance. Evaluated on a dataset of 100 patients collected at a clinical research center, MGSCLoa achieved an average accuracy of 91.67% and an average F1 score of 91.53% with only 40 seconds of labeled data per posture. In a Phase 2 evaluation with 11 participants over 13 nights in home settings, the framework reached an average accuracy of 85.37% and a weighted F1 score of 83.59% using just 3 minutes of labeled data per common posture for each participant. These results underscore the potential of seismic sensor-based in-bed posture classification for assessing sleep quality and related health conditions.