This paper introduces a novel wearable solution for continuous respiratory monitoring through electrocardiogramderived respiration (EDR) using custom-designed, dual-sided grid-patterned inkjet-printed (IJP) flexible dry electrodes and real-time smartphone-based analysis. The proposed electrode design reduces silver ink usage while maintaining signal quality and wearer comfort. We first compared ECG signal quality across gel, one-sided, and two-sided grid-patterned electrodes. Our mobile application, CardioHelp, processes ECG signals in real time to extract respiratory waveforms and continuously updates the respiration rate. EDR performance was validated against a commercial respiration belt across four activity conditions in five healthy adults. Bland-Altman and statistical analyses revealed minimal bias (mean difference$<0.5$bpm), MAE$\leq 0.36$bpm, and RMSE≤0.38 bpm. These results confirm robust and reliable performance. This integrated solution provides an affordable and practical approach to continuous cardiorespiratory monitoring in everyday life.

This paper introduces a novel edge-computing wearable device for real-time beat-by-beat electrocardiogram (ECG) classification. Early detection of heart disease can prevent and improve the patient's health and minimize the load on healthcare professionals. The proposed wearable integrates the pre-trained artificial intelligence (AI) model with the device firmware to detect abnormal ECG beats. The wearable contains a custom printed circuit board (PCB) connected to a commercial Bluetooth system on chip (SoC) to process ECG signals. The AI-integrated firmware is programmed into 32 bit ARM® Cortex™ based SoC. Data pre-processing, feature extraction, and inferencing are done on the SoC. The smart wearable is tested by passing the MIT-BIH test dataset through the wearable system as sample real-time data. The accuracy of the proposed device is assessed by testing normal and 4 abnormal ECG beats. The ANN model, after testing, provides an accuracy of 90.6 % with a precision of 95.6% and a recall of 95.1 %. The system consumes low power, a maximum of 128 m W, and offers a low latency of 6 ms for inferencing. The wearable is also tested on 5 subjects. The proposed wearable is smart, low-powered, and suitable for real-time regular cardiac monitoring. The performance indicates the device can effectively detect abnormal heart rhythms.

Detecting respiratory disorders can be improved with real-time diagnostic solutions, where monitoring respiration rates and patterns can act as early indicators for various cardiorespiratory diseases. We present a cost-effective edge-computing method utilizing wearable technology for respiratory disorder detection. Our system employs a wearable device with an IMU sensor for real-time signal transmission to a custom smart-phone app which enables thorough visualization of respiratory signals, continuous monitoring of respiration rates, and rapid alarms for breathing anomalies, alongside ECG functionalities. A novel approach is introduced for respiratory pattern detection that includes a pre-trained AI model for apnea detection and classification of normal, bradypnea, and tachypnea patterns from non-apnea signals. During model building using the Apnea-ECG dataset, the proposed hyper-feature algorithm for apnea detection demonstrates excellent performance, achieving accuracies of 92.33%, 94.89%, and 97.66% for Chest, Abdominal, and Nasal respiration signals, respectively. With an average inference time of 5 ms for respiratory event detection, the classifier achieves outstanding accuracy rates of 91.24%, 92.42%, and 93.26% for these signals on the validation dataset after being implemented and tested at the edge device. Real-time data analysis from 10 subjects further underscores the system's potential for continuous respiratory and cardiac monitoring in real-world scenarios.

Successful integration of deep neural networks (DNNs) or deep learning (DL) has resulted in breakthroughs in many areas. However, deploying these highly accurate models for data-driven, learned, automatic, and practical machine learning (ML) solutions to end-user applications remains challenging. DL algorithms are often computationally expensive, power-hungry, and require large memory to process complex and iterative operations of millions of parameters. Hence, training and inference of DL models are typically performed on high-performance computing (HPC) clusters in the cloud. Data transmission to the cloud results in high latency, round-trip delay, security and privacy concerns, and the inability of real-time decisions. Thus, processing on edge devices can significantly reduce cloud transmission cost. Edge devices are end devices closest to the user, such as mobile phones, cyber–physical systems (CPSs), wearables, the Internet of Things (IoT), embedded and autonomous systems, and intelligent sensors. These devices have limited memory, computing resources, and power-handling capability. Therefore, optimization techniques at both the hardware and software levels have been developed to handle the DL deployment efficiently on the edge. Understanding the existing research, challenges, and opportunities is fundamental to leveraging the next generation of edge devices with artificial intelligence (AI) capability. Mainly, four research directions have been pursued for efficient DL inference on edge devices: 1) novel DL architecture and algorithm design; 2) optimization of existing DL methods; 3) development of algorithm–hardware codesign; and 4) efficient accelerator design for DL deployment. This article focuses on surveying each of the four research directions, providing a comprehensive review of the state-of-the-art tools and techniques for efficient edge inference.

A premature ventricular contraction (PVC) disrupt the normal heart rhythm and indicate underlying cardiac disease. We aim to detect these PVC beats from electrocardiogram (ECG/EKG) data by automatically classifying these ECG beats with high accuracy in real-time. In this study, we used MIT BIH Long-Term Electrocardiogram Database (ltdb) dataset from the PhysioNet database. We extract signal-specific features and signal-independent features and combine them for feature ranking. We use principal component analysis (PCA), elastic net regularization (ENR), univariate filter of constant, quasi constant and duplicate feature removal (CQCDFR) and analysis of variance test (ANOVA) for feature selection. We take the top 10 features for four methods and classify them separately. The machine learning model explored is the random forest classifier. In our analysis, elastic net regularization performed best in terms of accuracy in cardiac patients. We further use the feature with the best accuracy in four algorithms to test sensitivity, specificity, accuracy, precision, f1-score to evaluate statistics. The overall accuracy of elastic net regularization for classifying the highest first 8 feature data is 97.8%. The sensitivity was 94.7% and the specificity was 99.6%. The accuracy rate is 99.6%, and the F1 score is 97.1%. The method can accurately detect ECG beats and analyze categories for real-time cardiac monitoring for feedback to the use patient. Efficient feature selection minimizes the number of features used and reduces the power consumption of the monitoring device.

Artificial Intelligence-enabled applications on edge devices have the potential to revolutionize disease detection and monitoring in future smart health (sHealth) systems. In this study, we investigated a minimalist approach for the severity classification, severity estimation, and progression monitoring of obstructive sleep apnea (OSA) in a home environment using wearables. We used the recursive feature elimination technique to select the best feature set of 70 features from a total of 200 features extracted from polysomnogram. We used a multi-layer perceptron model to investigate the performance of OSA severity classification with all the ranked features to a subset of features available from either Electroencephalography or Heart Rate Variability (HRV) and time duration of SpO2 level. The results indicate that using only computationally inexpensive features from HRV and SpO2, an area under the curve of 0.91 and an accuracy of 83.97% can be achieved for the severity classification of OSA. For estimation of the apnea-hypopnea index, the accuracy of RMSE = 4.6 and R-squared value = 0.71 have been achieved in the test set using only ranked HRV and SpO2 features. The Wilcoxon-signed-rank test indicates a significant change (p < 0.05) in the selected feature values for a progression in the disease over 2.5 years. The method has the potential for integration with edge computing for deployment on everyday wearables. This may facilitate the preliminary severity estimation, monitoring, and management of OSA patients and reduce associated healthcare costs as well as the prevalence of untreated OSA.

Patch sensors are slowly becoming ubiquitous in the market for wearable devices. They can collect and send data unobtrusively, which is beneficial to several applications in healthcare, sports, and fitness industries. Patch sensor by inkjet printing (IJP) with functional materials on flexible substrates allows design flexibility and portability at a very low cost. In this paper, preliminary results on characterization of IJP resistors printed with a variety of inks on substrates such as paper and polyimide have been reported, followed by a discussion on printed breath-rate and ECG sensors. The results promise a plethora of possibilities in printed circuits.

Brain activities commonly recorded using the electroencephalogram (EEG) are contaminated with ocular artifacts. These activities can be suppressed using a robust independent component analysis (ICA) tool, but its efficiency relies on manual intervention to accurately identify the independent artifactual components. In this paper, we present a new unsupervised, robust, and computationally fast statistical algorithm that uses modified multiscale sample entropy (mMSE) and Kurtosis to automatically identify the independent eye blink artifactual components, and subsequently denoise these components using biorthogonal wavelet decomposition. A 95% two-sided confidence interval of the mean is used to determine the threshold for Kurtosis and mMSE to identify the blink related components in the ICA decomposed data. The algorithm preserves the persistent neural activity in the independent components and removes only the artifactual activity. Results have shown improved performance in the reconstructed EEG signals using the proposed unsupervised algorithm in terms of mutual information, correlation coefficient, and spectral coherence in comparison with conventional zeroing-ICA and wavelet enhanced ICA artifact removal techniques. The algorithm achieves an average sensitivity of 90% and an average specificity of 98%, with average execution time for the datasets ( N = 7) of 0.06 s ( SD = 0.021) compared to the conventional wICA requiring 0.1078 s ( SD = 0.004). The proposed algorithm neither requires manual identification for artifactual components nor additional electrooculographic channel. The algorithm was tested for 12 channels, but might be useful for dense EEG systems.

In this paper, a new metric for the decision rule of the space-time block (STB) codes having imperfect estimate of the channel fading parameters is derived. Estimation errors of these fading parameters have high impact on the performance of the STB code as the decision rule is dependent on these fading parameters. By including the variance of the channel estimation error in the decoder metric derivation of the STB codes, we derive an exact probability distribution function (pdf) of the received signals conditioned on the estimated channel parameters and transmitted symbol sequences. A modified decision rule is easily found from this and compared with the state-of-the-art scheme proposed by Tarokh. Performance comparison shows that significant improvement is achieved in terms of error rate and system complexity, especially for quadrature amplitude modulation (QAM). The derived decision rule converges to the ideal case for no estimation error and to Tarokh's decision rule for high signal-to-noise ratio (SNR).