Early detection of obstructive sleep apnea (OSA) is extremely necessary to control its’ rising prevalence worldwide. Conventional diagnostic method like polysomnography (PSG) is uncomfortable, intrusive, and costly, thus, limiting its’ easy accessibility among people. To address this, we propose a novel deep learning method for detecting OSA using photoplethysmography (PPG) signal, a non-invasive method that is commonly available in wearable devices. We introduce a novel methodology using Multivariate Long Short-Term Memory-Fully Convolutional Network (MLSTM-FCN) model that effectively captures both temporal dependencies and local features in PPG signals for OSA detection. A new windowing technique was introduced to ensure apneic events are centered within each window to enhance the model’s ability to detect delayed physiological responses to apnea. The model was trained and evaluated on the Multi-Ethnic Study of Atherosclerosis (MESA) dataset, achieving an improvement of 11.3% in accuracy over the state-of-the-art method. The method obtained an accuracy of 93.44%, precision of 0.94, recall of 0.91, and an F1-score of 0.93. These results demonstrate the potential of our method in accurately identifying OSA events. This method offers a unobtrusive, comfortable, and cost-effective alternative to traditional diagnostic tools, making it suitable for long-term, home-based monitoring.

Obstructive sleep apnea (OSA) has become a serious health concern with increasing morbidity worldwide. Even though polysomnography is widely used by physicians for diagnosing OSA, the process is costly, time-consuming, and uncomfortable for patients. This increases the demand for developing unobtrusive, cost-effective, and reliable solutions for detecting OSA and reducing patient discomfort. Several machine learning-based and some deep learning based methods using extracted ECG features for OSA detection from ECG are found to be less reliable due to the manual feature extraction process, very few studies(included in the comparison table of below section) have used only deep learning methods for OSA detection from ECG signals. In this study, we proposed a novel deep learning method that leverages convolutional neural networks (CNN) and long short-term memory (LSTM) networks to learn spatial and temporal features for detecting OSA from ECG data. Our model was trained and evaluated on the publicly available MESA and Apnea-ECG datasets to assess its robustness. In a prudent data windowing step, we center the apnea data within each data window, enabling the model to better learn apnea patterns and resulting in achieving an accuracy of 95.4%, 92.2%, Specificity of 95.1%, 92.8%, Sensitivity of 94.8%, 91.6% and F1 score of 94.1%, 92.1% respectively on MESA and Apnea-ECG dataset. Our model yields an increase of 0.57% and 9.31% in specificity, 1.87% and 50.47% in sensitivity, 4.89% in F1-score, and 3.47% and 17.77% in accuracy against the best result we found using the Apnea-ECG and MESA dataset respectively. The results show our proposed model outperforms state-of-the-art methods on the MESA dataset and achieves equally good results on the Apnea-ECG dataset. Implementation of our model on Jetson orin AGX board gains comparable results with the above-stated accuracy showing the possible practical use of our methodology. These findings highlight the model’s stability, robustness, and high accuracy in detecting OSA.

The interpretability of machine learning model is crucial in healthcare as it fosters reliability and supports clinicians in making informed decisions based on predictions. However, recent approaches for vital parameter estimation often act as black boxes, lacking clarity in their predictive reasoning. Therefore, accurate estimation of vital parameters is crucial not only for timely diagnosis and effective patient monitoring but also for ensuring that clinicians can trust and understand the model’s predictions. In this study, we introduce XVPE-Net, an explainable framework aimed at improving the state-of-the-art VPE-Net model for estimating vital parameters: heart rate (HR), respiratory rate (RR), systolic blood pressure (SBP), and diastolic blood pressure (DBP). Our model enhances interpretability by visually highlighting key parts of the PPG signal that significantly influence VPE-Net’s predictions of vital parameters. We test our model using 1,000 PPG segments from 38 subjects in the MIMIC-III dataset, which is available in the Physionet repository. The results show strong predictive capabilities, along with better transparency and reliability. This study helps us understand how the model makes decisions and supports the use of explainable AI in health monitoring, contributing to the development of more reliable models for medical applications.

Cardiovascular diseases (CVDs) are the major cause of global morbidity and mortality. CVDs can be preliminarily diagnosed by analyzing a patient’s electrocardiogram (ECG), which requires long-term continuous ECG monitoring to suitably detect the onset of the disease. However, ECG data acquisition involves multiple lead attachments and requires regular intervention by an expert, thereby making the process cumbersome and inappropriate for continuous health monitoring due to limited portability and discomfort caused to the patients. Nowadays, automated ECG measurement techniques are gaining popularity in wearable health monitoring applications to seamlessly identify cardiac abnormalities even in a home environment. On the other hand, photoplethysmography (PPG) signals can be acquired from the wrist or fingertip of a patient by using a lead-less patch-less set-up that can be easily integrated with smart wearable devices. Therefore, to address the aforementioned demerits associated with ECG devices, a few researchers have fostered the idea of reconstructing ECG from photoplethysmogram (PPG) signals to generate simple yet effective CVD monitoring methodologies. Hence, in this paper, we propose P2E-LGAN, a hybrid generative adversarial network (GAN) based framework for generating ECG from PPG. The proposed network is evaluated on a benchmark database combined with ECG and PPG data. The inclusion of LSTM in the GAN network reduces the root mean square (RMSE), mean absolute error of heart rate (MAE(HR)) and percentage root mean square difference (PRD) by 35.7%, 37.2% and 9.8% respectively. Individual graphical analysis and performance evaluation of different metrics with state-of-the-art methods demonstrate the effectiveness of the proposed framework for reconstructing ECG from PPG.