Obstructive sleep apnea (OSA) is a common sleep‐related respiratory disorder. Around the world, more and more people are suffering from OSA. Because of the limitation of monitor equipment, many people with OSA remain undetected. Therefore, we propose a sleep‐monitoring model based on single‐channel electrocardiogram using a convolutional neural network (CNN), which can be used in portable OSA monitor devices. To learn different scale features, the first convolution layer comprises three types of filters. The long short‐term memory (LSTM) is used to learn the long‐term dependencies such as the OSA transition rules. The softmax function is connected to the final fully connected layer to obtain the final decision. To detect a complete OSA event, the raw ECG signals are segmented by a 10 s overlapping sliding window. The proposed model is trained with the segmented raw signals and is subsequently tested to evaluate its event detection performance. According to experiment analysis, the proposed model exhibits Cohen’s kappa coefficient of 0.92, a sensitivity of 96.1%, a specificity of 96.2%, and an accuracy of 96.1% with respect to the Apnea‐ECG dataset. The proposed model is significantly higher than the results from the baseline method. The results prove that our approach could be a useful tool for detecting OSA on the basis of a single‐lead ECG.

Diabetic retinopathy (DR) is an eye disease that damages the blood vessels of the eye. DR causes blurred vision or it may lead to blindness if it is not detected in early stages. DR has five stages, i.e., 0 normal, 1 mild, 2 moderate, 3 severe, and 4 PDR. Conventionally, many hand-on projects of computer vision have been applied to detect DR but cannot code the intricate underlying features. Therefore, they result in poor classification of DR stages, particularly for early stages. In this research, two deep CNN models were proposed with an ensemble technique to detect all the stages of DR by using balanced and imbalanced datasets. The models were trained with Kaggle dataset on a high-end Graphical Processing data. Balanced dataset was used to train both models, and we test these models with balanced and imbalanced datasets. The result shows that the proposed models detect all the stages of DR unlike the current methods and perform better compared to state-of-the-art methods on the same Kaggle dataset.

Atrial fibrillation is a common cardiac arrhythmia, and its incidence increases with age. Currently, numerous deep learning methods have been proposed for AF detection. However, these methods either have complex structures or poor robustness. Given the evidence from recent studies, it is not surprising to observe the limitations in the learning performance of these approaches. This can be attributed to their strictly homogenous conguration, which solely relies on the linear neuron model. The limitations mentioned above have been addressed by operational neural networks (ONNs). These networks employ a heterogeneous network configuration, incorporating neurons equipped with diverse nonlinear operators. Therefore, in this study, to enhance the detection performance while maintaining computational efficiency, a novel model named multi-scale Self-ONNs (MSSelf-ONNs) was proposed to identify AF. The proposed model possesses a significant advantage and superiority over conventional ONNs due to their self-organization capability. Unlike conventional ONNs, MSSelf -ONNs eliminate the need for prior operator search within the operator set library to find the optimal set of operators. This unique characteristic sets MSSelf -ONNs apart and enhances their overall performance. To validate and evaluate the system, we have implemented the experiments on the well-known MIT-BIH atrial fibrillation database. The proposed model yields total accuracies and kappa coefficients of 98% and 0.95, respectively. The experiment results demonstrate that the proposed model outperform the state-of-the-art deep CNN in terms of both performance and computational complexity.

In recent years, various models based on convolutional neural networks (CNN) have been proposed to solve the cardiac arrhythmia detection problem and achieved saturated accuracy. However, these models are often viewed as “blackbox” and lack of interpretability, which hinders the understanding of cardiologists, and ultimately hinders the clinical use of intelligent terminals. At the same time, most of these approaches are supervised learning and require label data. It is a time-consuming and expensive process to obtain label data. Furthermore, in human visual cortex, the importance of lateral connection is same as feed-forward connection. Until now, CNN based on lateral connection have not been studied thus far. Consequently, in this paper, we combines CNNs, lateral connection and autoencoder (AE) to propose the building blocks of lateral connection convolutional autoencoder neural networks (LCAN) for cardiac arrhythmia detection, which learn representations in an unsupervised manner. Concretely, the LCAN contains a convolution layer, a lateral connection layer, an AE layer, and a pooling layer. The LCAN detects salient wave features through the lateral connection layer. The AE layer and competitive learning is used to update the filters of the convolution network—an unsupervised process that ensures similar weight distribution for all adjacent filters in each convolution layer and realizes the neurons’ semantic arrangement in the LCAN. To evaluate the performances of the proposed model, we have implemented the experiments on the well-known MIT–BIH Arrhythmia Database. The proposed model yields total accuracies and kappa coefficients of 98% and 0.95, respectively. The experiment results show that the LCAN is not only effective, but also a useful tool for arrhythmia detection.

Due to the high cost and relatively low image quality of hyperspectral sensors, spectral super-resolution seeks to explore the mapping mechanisms between multispectral and hyperspectral images, with the goal of reconstructing high-quality hyperspectral images. In recent years, deep learning algorithms have achieved significant success in spectral super-resolution. However, most of these methods compute loss only at the final stage of the network, neglecting the intermediate generative processes, which leads to considerable spectral distortion in the reconstructed images. To address these issues, we propose a Progressive Conditional Diffusion Model (PCDM) for multi-stage spectral restoration. PCDM constructs a channel synthesis module that generates a ground truth set through band synthesis, and designs an image reconstruction module to ensure that the synthesized image in the next stage can effectively reconstruct the synthesized features from the previous stage. Multiple conditional diffusion models are then constructed based on the dataset. For each conditional diffusion model, the network parameters of the corresponding image reconstruction module are shared with the multispectral image for spectral up-sampling. The spectral-up-sampled multispectral features, combined with the output from the previous diffusion model, serve as a conditional matrix, which is input into the next diffusion model to obtain the final result. Experimental results on both synthetic and real datasets demonstrate that PCDM can effectively reconstruct hyperspectral images, showing robustness and outperforming state-of-the-art methods.