As the healthcare industry continues to embrace digital transformation, the Internet of Medical Things (IoMT) emerges as a key technology. IoMT plays a critical role in revolutionizing healthcare delivery by enabling remote patient monitoring, personalized treatments, and efficient healthcare management. This survey offers a comprehensive overview of IoMT, elucidating its concepts and architectural framework. It explores its diverse applications and the challenges associated with its adoption. Additionally, it investigates key methods in lightweight, explainable artificial intelligence, discusses their applications in healthcare services, and outlines relevant evaluation metrics. Furthermore, it examines data security and privacy concerns in healthcare, presenting relevant methodologies. Finally, this review provides insights into the future of IoMT, considering existing challenges and opportunities for advancement.

With the rapid development of Artificial Intelligence of Things (AIoT) technology, its adoption in the field of smart healthcare is becoming increasingly pervasive. Leading cloud service providers like IBM Watson Health now offer neural network inference services tailored for smart healthcare applications - users simply need to send data to the server to get the diagnosis results. However, a growing concern arises regarding the potential compromise of user privacy. Currently, researchers propose the use of secure multi-party computation and homomorphic encryption techniques to address this issue. Nevertheless, further exploration and improvement are needed to mitigate the side effects, such as increased latency and challenges in meeting real-time monitoring requirements. In this paper, we propose a secure homomorphic encryption-based inference framework named SecureBadger for two typical medical inference scenarios: disease diagnosis based on image analysis and health monitoring with smart wearable devices. We design two inference modes—large-scale batch inference and small-scale low-latency inference. Additionally, different ciphertext packaging schemes are designed to enhance inference efficiency for different inference modes, different input data types and different network layers. Experimental evaluations are conducted on several datasets, and the results indicate that SecureBadger can significantly reduce the inference time overhead in both inference modes.

This study monitors neonatal sleep patterns using artificial intelligence through an innovative transfer learning approach integrating video and Electroencephalogram (EEG) data. Leveraging non-intrusive, camera-based technology, this method offers healthcare professionals an effective tool to evaluate and understand sleep quality in newborns. Such monitoring is essential for pediatricians, providing critical insights into the health and development of infants. Automated sleep/wake staging tools aid healthcare professionals in analyzing infant sleep patterns. Among these methods, camera-based approaches have gained prominence due to their non-intrusive and user-friendly characteristics, making them applicable for home monitoring. This study introduces a novel transfer learning technique for classifying neonatal sleep/wake stages. Our approach utilizes multiple color palettes, such as thermal, amber, grayscale, high contrast, hot metal, and red-blue, captured through a Fluke® (TiX 580) camera system. Continuous monitoring of neonatal sleep is essential for pediatricians to assess neonatal sleep quality comprehensively. Automated sleep/wake staging tools assist healthcare professionals in evaluating infant sleep patterns. Camera-based approaches have garnered significant attention among the various methods due to their non-intrusive and user-friendly nature, making them suitable for home use. In this paper, we propose a novel transfer learning approach for classifying neonatal sleep/wake staging using a combination of multiple color palettes, including thermal, amber, grayscale, high contrast, hot metal, and red-blue, recorded through a Fluke® (TiX 580) camera. The proposed method leverages the retraining of the last fully connected layer of well-established deep neural networks such as Visual Geometry Group 16 and 19, AlexNet, Inception-V3, ResNet-18, ResNet-50, and GoogLeNet to perform accurate sleep and wake stage classification. To enhance the precision of our approach, we integrate EEG data with video data obtained from neonatal subjects. The performance of transfer learning networks is validated using a leave-one-subject-out cross-validation strategy, ensuring robustness in classifying wake and sleep stages. In particular, the Inception-V3 model, when applied to the red-blue color palette video frames, demonstrates an impressive classification accuracy rate of 85.9%. Furthermore, we assessed the impact of including EEG data alongside video data, and even in this context, our approach maintains the same high accuracy of 85.9%. These findings underscore the robustness and effectiveness of our proposed method for neonatal sleep/wake staging classification, which is suitable for home-based monitoring and enhances its practicality and accessibility for pediatricians and caregivers.

Sharing Electronic Health Record (EHR) data is critical for improving medical decision-making and emergency care. However, EHRs contain sensitive patient information, making data breaches and misuse a significant concern during sharing. To tackle this challenge, we propose the Blockchain-Based Secure Access and Sharing (BCAS) scheme, a patient-centered, secure, and efficient solution for large-scale EHR access and sharing that intelligently combines blockchain with proxy re-encryption. Four smart contracts are designed for the BCAS to achieve effective user identity registration, precise virtual identity verification, strict access control, and secure data management. The adoption of hybrid on-chain and off-chain storage structures, coupled with the four smart contracts, effectively optimizes the efficiency of the blockchain in BCAS. Proxy re-encryption further enhances data confidentiality and streamlines EHR sharing among authorized users. The security analysis indicates that BCAS can resist DDoS, spoofing, and man-in-the-middle attacks, ensuring the privacy and security of patient EHR data. The performance evaluation demonstrates that the CPU usage of smart contracts does not exceed 15%. The on-chain and off-chain storage structures require less than 30 s for uploading and downloading an 800 MB file. Additionally, the proxy re-encryption scheme can process a 10 MB file in less than 43 ms. These results indicate that BCAS offers efficient performance without compromising data security and privacy. Compared with existing blockchain-based EHR sharing systems, BCAS offers superior performance, a more comprehensive solution, and increased resistance to attacks.

For more accessible and advanced health monitoring, the Body Area Network (BAN) design with semantic technologies offers efficient information sensing and communication in smart healthcare Artificial Intelligence of Things (AIoT). To address the critical challenges of effective communication and reduction of data transmission pressure in AIoT-BAN, a hybrid BAN system is proposed which enhances information processing and communication capabilities by leveraging semantic understanding and multimodal processing. It incorporates a semantic communication and sensing fusion framework, offloading based on the human Body Coupled Communication (BCC) channel, and multimodal semantic information integration to reduce data transmission pressure. The proposed method offers effective inclusive smart healthcare and daily health maintenance for the general public.

Currently, Cardiovascular Disease (CVD) remains a significant contributor to premature mortality and escalating health care expenses. Early and accurate detection is crucial for the successful treatment, intervention, and monitoring of heart health. Electrocardiograms (ECGs) are essential for diagnosing and monitoring cardiovascular diseases. However, the increasing demand for ECG signal detection, coupled with a shortage of specialized ECG doctors, has made automatic classification and diagnosis of ECG signals a prominent research area. Traditional ECG signal classification models often involve numerous parameters, rendering them unsuitable for resource-limited IoT devices in smart healthcare scenarios. In response, this paper proposes a novel lightweight ECG signal classification network based on the Transformer model, named DCETEN. Specifically, we introduce a lightweight Efficient Channel Attention (ECA) module, integrating it with Depthwise Separable Convolution (DSC) to design a One-dimensional Convolutional Neural Network (1D-CNN) that enhances feature extraction capabilities. Additionally, we fuse hand-crafted RR interval features and features learned by the Transformer to comprehensively capture the ECG signal characteristics. Finally, to make the proposed method suitable for resource-constrained IoT-based edge devices, we employ pruning techniques to reduce the number of model parameters. We validated the proposed model on the MIT-BIH Arrhythmia Database, achieving 99.84% accuracy and a 99.67% F1 score with low computational and memory requirements, making it suitable for deployment in smart healthcare settings with prevalent resource limitations.

To resolve the problems of unreliable transmission and low communication capacity caused by the time-frequency domain dual-selective fading of High Frequency (HF) channels, a constellation mapping fusion technology for HF Frequency Shift Chirp Modulation (FSCM) is proposed. On one hand, the FSCM technology is employed to compensate for the fast fading introduced by frequency domain expansion, and the spreading factor is used to adjust the transmission rate. On the other hand, based on the FSCM matching, by constructing a two-dimensional constellation mapping, the soft information and symbol weights of multipaths are extracted, thereby realizing single-channel RAKE fusion and multi-channel diversity fusion with symbol-by-symbol weighting. This effectively improves transmission reliability in HF dual-selective channels where the main and secondary paths switch. In typical HF scenarios defined by ITU-R F.1487, simulation results demonstrate that, compared with the traditional hard demodulation algorithms, the proposed soft information fusion algorithm can still ensure reliable transmission under low signal-to-noise ratio conditions, even with high multipath separation degrees. Moreover, through 24-h field tests, the practical applicability of the proposed algorithm is further verified.

With the rapid development of 5G-A technology, Wi-Fi-based indoor fingerprint localization has gained prominence for its high-precision indoor localization. However, conventional methods often suffer from noise interference and indistinct fingerprints in complex environments, reducing the accuracy of localization. To address these issues, the Feature Equalization Localization Method (FELoc) based on Channel State Information (CSI) is proposed in this paper. A unified preprocessing strategy named WaveICA is designed to enhance fingerprint quality by combining Hierarchical Thresholding Wavelet Denoising (HTWD) for noise suppression and Independent Component Analysis (ICA) for extracting independent features, which effectively improves the discriminability of CSI amplitude and phase data from multiple antennas. Additionally, Feature Equalization Localization Network (FELN) is designed for feature extraction and fusion, which employs dual-branch convolution with a Multiscale Channel-spatial Attention Module (MCAM) to enhance salient features and a novel Feature Equalization Module (FEM) to adaptively fuse phase and amplitude representations. Extensive experiments conducted in diverse indoor scenarios demonstrate that FELoc achieves superior localization performance, providing both higher accuracy and stronger robustness compared to other methods.

The integration of Digital Signal Processing (DSP) and Reinforcement Learning (RL) for optimal consensus control in Networked Multi-Agent Systems (NMASs) has garnered significant research attention. However, prior research encounters some limitations: 1) dependency on initial admissible control policies, and 2) systemic data redundancy arising from ineffective data governance in distributed architectures and slow convergence rates of conventional RL algorithms. To overcome these challenges, this paper proposes a Distributed Collaborative Iteration Adaptive Dynamic Programming (DCIADP) framework. The methodology reformulates the solution of Hamilton-Jacobi-Bellman (HJB) equations by integrating Value Iteration (VI) and Policy Iteration (PI) within a unified architecture, eliminating reliance on prior knowledge of system dynamics. Specifically, a dynamic factor is introduced to synergistically integrate the complementary strengths of VI and PI, achieving accelerated convergence while bypassing the initialization requirement for admissible policies. This innovation significantly mitigates computational overhead in distributed nodes during localized DSP operations. Furthermore, a self-tuning mechanism dynamically optimizes the factor, enhancing adaptability to heterogeneous network conditions. Through rigorous theoretical analysis, the proposed framework is proven to ensure asymptotic convergence and Lyapunov stability. Practical implementation is realized through actor-critic Neural Networks (NNs), incorporating an experience replay mechanism to exploit temporal correlation characteristics in networked data streams. This enables derivation of optimal control policies solely from transmitted network signals, independent of explicit system parameter knowledge. The framework thus establishes a resource-efficient adaptive control paradigm for bandwidth-constrained networked MASs. Finally, several numerical simulations validate the effectiveness and superiority of the proposed approach.

Blind encoder recognition methods have drawn much research interest in recent years, as they can play an important role in non-cooperative scenarios. This paper proposes an open-set blind recognition method for non-binary Low-Density Parity-Check (LDPC) codes with high performance. We derive the principle of obtaining a non-binary Parity-Check Matrix (PCM) from the received signal and further apply the Gauss-Jordan Column Elimination (GJCE) over GF( q). An iterative process involving GJCE, LDPC decoding, and PCM sparsification is applied in our proposed method. The proposed symbol-level blind recognition is shown to perform much better than the bit-level one. Experimental results demonstrate that our proposed method can completely reconstruct the PCMs at a bit error rate of 0.4% for the moderate-length non-binary LDPC codes.

Online hashing methods are receiving increasing attention in cross modal medical image retrieval research. However, existing online methods often lack the learning ability to maintain semantic correlation between new and existing data. This paper proposes an Online Semantic-similarity Cross Modal Hashing (OSCMH) learning framework that can incrementally learn compact binary hash codes of streaming data. Considering the dynamic nature of data streams in online hashing methods, the proposed learning framework is implemented under the Internet of Medical Things (IoMT) architecture, where IoMT enables real-time data transmission and processing through edge computing resources, effectively supporting the storage and analysis of medical image data. Then, a sparse representation of existing data based on online anchor datasets is designed to avoid semantic forgetting of the data and adaptively update hash codes, effectively maintaining semantic correlation between existing and arriving data, reducing information loss and improving training efficiency. Finally, an online discrete optimization method is proposed to solve the binary optimization problem by incrementally updating the hash function and optimizing the hash code on the stream data points. A large number of experiments on the benchmark datasets have shown that the algorithm proposed in this paper can effectively improve the retrieval efficiency in the field of medical images compared to existing online or offline hashing methods.

As commercial mega-constellations such as SpaceX’s Starlink and Amazon’s Kuiper system rapidly expand, traditional Tracking, Telemetry, and Command (TT&C) architectures encounter unprecedented scalability challenges. Thus, a novel scheme design and capacity analysis framework is proposed for networked TT&C system serving large-scale Low Earth Orbit (LEO) satellite constellations. The networked TT&C architecture is designed to leverage Medium Earth Orbit (MEO) satellites to monitor LEO constellations, providing TT&C service for both normal and anomalous LEO satellites through inter-satellite links. A capacity decomposition framework is introduced, dividing system capacity into access capacity and anomaly-handling capacity. Through stochastic geometry and queuing theory models, the maximum number of LEO satellites that can be supported under various operational constraints is quantified. It is demonstrated by simulation results that for Starlink Phase 1, if the second-generation O3b mPower system is utilized for TT&C, the system can support up to 59,497 LEO satellites under a 30-second threshold of polling waiting time, and up to 50,761 satellites under a 10-minute threshold of anomaly-handling time in the steady state. The analysis provides critical design guidelines for the deployment and management of large-scale LEO constellations, highlighting the trade-offs between system capacity and operational constraints such as polling waiting time and anomaly-handling waiting time.