Search

Electrocardiogram (ECG) is one of the fundamental markers to detect different cardiovascular diseases (CVDs). Owing to the widespread availability of ECG sensors (single lead) as well as smartwatches with ECG recording capability, ECG classification using wearable devices to detect different CVDs has become a basic requirement for a smart healthcare ecosystem. In this paper, we propose a novel method of model compression with robust detection capability for CVDs from ECG signals such that the sophisticated and effective baseline deep neural network model can be optimized for the resource constrained micro-controller platform suitable for wearable devices while minimizing the performance loss. We employ knowledge distillation-based model compression approach where the baseline (teacher) deep neural network model is compressed to a TinyML (student) model using piecewise linear approximation. Our proposed ECG TinyML has achieved ~156x compression factor to suit to the requirement of 100KB memory availability for model deployment on wearable devices. The proposed model requires ~5782 times (estimated) less computational load than state-of-the-art residual neural network (ResNet) model with negligible performance loss (less than 1% loss in test accuracy, test sensitivity, test precision and test F1-score). We further feel that the small footprint model size of ECG TinyML (62.3 KB) can be suitably deployed in implantable devices including implantable loop recorder (ILR).

The primary risk factor of hypertension, is the lack of awareness caused by the unavailability of ubiquitous blood pressure (BP) measurement. In this study, we have investigated the BP estimation using the photoplethysmogram (PPG) signal and a suitable subject-specific mathematical model. The linear transfer function (LTF) technique was used to identify the subject-specific model. Firstly, we tried to identify the model considering arterial blood pressure (ABP) as input and PPG as output, and we achieved an average estimation accuracy (normalized root mean square, NRMSE) of 84.4%. Next, we fitted an inverse model, where ABP is the output, and PPG is the input, and we achieved an average estimation accuracy (NRMSE) of 84.7%. Finallly, We verified that the two identified models mentioned above are inverse of each other. In this study, we have used ABP and PPG signals of 10 (male = 7, female = 3) subjets from the MIMIC II database. The results are quite promising for the use of the PPG in the detection and diagnosis of cardiovascular diseases.

Ambulatory Photoplethysmogram (PPG) is a more user-friendly choice for continuous cardiac monitoring as compared to Electrocardiogram (ECG). However, wearable PPG is often prone to motion artefacts. In this paper, we propose a novel pipeline for motion-resistant beat-to-beat interval extraction from noisy PPG. Firstly, the effects of motion artefacts are minimized by using Adaptive Recursive-Least-Square (RLS) Filtering and Singular Spectrum Analysis (SSA). Next, the signal peaks are identified and their locations are corrected by weighted local interpolation. Finally, outlier peak-to-peak intervals are marked as incorrigible. Experimental validation on the training dataset of IEEE Signal Processing Cup 2015 reveals that the proposed method achieves 1.68% mean peak detection error rate and 11.32 milliseconds mean absolute error of detected beat-to-beat intervals. The metric values outperform those obtained by the state-of-the-art techniques by at least 12.58 and 5.74 times respectively.

State-of-the-art methods have reported various features for the non-invasive screening of Coronary Artery Disease (CAD). In this paper, we propose a novel approach to represent such features extracted from multiple physiological signals using hypergraph. Firstly, the biological and statistical interconnections among Photoplethysmogram (PPG) and Phonocardiogram (PCG) features are exploited by connecting them as hyperedges. Then, metadata features (age, weight and height) are connected using hyperedges with the rest of the features. Hypergraph based formalism provides greater flexibility in capturing the interrelationships among different features as compared to the graph counterpart. Finally, hypergraph laplacian as a derived feature is applied to classify CAD against non-CAD. The proposed method is validated on PPG and PCG data collected in a hospital setup. The results reveal 98% Sensitivity and 82% Specificity, leading to 92% classification accuracy.

Arterial blood pressure (ABP) is a vital hemodynamic signal to be monitored in general population- especially in critically ill patients. Pulse arrival time (PAT) has been widely used for the noninvasive measurement of systolic blood pressure (SBP) and diastolic blood pressure (DBP) and has shown excellent results for a healthy person as compared to the existing gold standard methods. In this study, we have used correlation analysis to study the feasibility of using PAT and heart rate (HR) for noninvasive BP (SBP and DBP) measurement for critically ill subjects. Physiological signals used for this study were collected from the MIMIC-II database of Physionet. Initially, we conducted a correlation analysis and then tried to estimate BP using linear regression. We found that the BP correlation with HR is more as compared with PAT. However, these correlations are not consistent. We have also found that PAT and HR can track the BP change as long as the correlation coefficient lies between ±0.7 and ±1. When the correlation falls below, this noninvasive BP estimation using PAT and HR needs to be revisited and improved for clinical applicability.

Sleep is a very important part of life. Lack of sleep or sleep disorder can cause a negative impact on day to day life and can have long term serious consequences. In this work, we propose an end-to-end trainable neural network for automated sleep arousal scoring. The network consists of two main parts. Firstly, a trend statistics network computes the moving average of the filtered signals at different scales. Secondly, we propose a channel invariant EEG network to detect the arousals in any Electroencephalography (EEG) channel. Finally, we combine the features from various channels through a convolution network and a bi-directional long short-term memory to predict the probability of arousal. Further, we propose an objective function that uses only respiratory effort related arousal (RERA) and non-arousal regions to optimize the network. We also propose a method to estimate the respiratory disturbance index (RDI) from the probability predicted by the network. Evaluation on Physionet Challenge 2018 database shows that the proposed method detects RERA with mean area under the precision-recall curve (AUPRC) of 0.50 in a 10-fold cross validation setup. The mean absolute error of RDI prediction is 6.11, while a two-class RDI severity prediction yields a specificity of 75% and a sensitivity of 83%.

Stress monitoring is important for mental wellbeing and early detection of related disorders. The current work is focused on stress detection from multiple non-invasive physiological signals like Electroencephalogram (EEG), Photoplethysmogram (PPG) and Galvanic Skin Response (GSR). We show that, compared to using only the well known EEG band powers in different frequencies for stress detection, an early fusion with GSR and PPG features shows a significant improvement. Maximum Relevance Minimum Redundancy (mRMR) based feature selection is used to identify the most suitable physiological features correlating with stress. A major contribution of this work lies in eliminating subject-specific bias to improve the classification accuracy. We use self-reported values of Valence, Arousal and Dominance to cluster subjects and build separate classification models specific to clusters. The proposed approach is validated on a publicly available dataset comprising 146 data instances from 10 subjects. The performances of Leave-One- Subject-Out cross validation (LOSOCV) in terms of mean Fscores are 0.61 using EEG features only, 0.64 using early fusion of EEG, GSR and PPG features and 0.69 by applying our clustering technique before fusion and classification.

Coronary Artery Disease (CAD) is an important problem in cardiac health and is a leading cause of human mortality. Prior arts have shown that features extracted from non-invasive Photoplethysmogram (PPG) signal are effective in classifying CAD. In this paper, we represent cardiac health as a graph (CHG) in order to exploit the dependencies of PPG features as well as the metadata features. We then compute spectral features from the eigenvalues of the graph Laplacian of CHG. Finally, k-means algorithm is employed for classifying the data into CAD and non-CAD. Unsupervised experiments on a cohort with 32 participants yields 88% accuracy and demonstrates advantage of the proposed formulation over a baseline and two state-of-the-art approaches.

This paper proposes a continuous and unsupervised approach of monitoring the arousal trend of an individual from his heart rate using Kalman Filter. The state-space model of the filter characterizes the baseline arousal condition. Deviations from this baseline model are used to recognize the arousal trend. A publicly available dataset, DECAF, comprising the physiological responses of 30 subjects while watching 36 movie clips inducing different emotions, is used to validate the proposed technique. For each clip, annotations of arousal given by experts per second are used to quantify the ground truth of arousal change. Experimental results suggest that the proposed algorithm achieves a median correlation of 0.53 between the computed and expected arousal levels which is significantly higher than that achievable by the state-of-the-art technique.

Phonocardiogram (PCG) or auscultation via a stethoscope forms the basis of preliminary medical screening. But PCG recorded in an uncontrolled environment is inherently noisy. In this paper we have derived novel features from the spectral domain and autocorrelation waveforms. These are used to identify the quality of a PCG recording and accepting only diagnosable quality recordings for further analysis. These features proved to be robust irrespective of variations in devices and in data collection protocols employed to ensure consistent data quality. A freely available, large, diverse, medical-grade PCG dataset was used for creating the training models. Results show that the proposed methodology yields an accuracy score of ~75% on our in-house PCG dataset, collected using a low-cost smartphone-based digital stethoscope.

Simple and non-invasive methods to estimate vital signs are very important for preventive healthcare. In this paper, we present a methodology to estimate Blood Pressure (BP) using Photoplethysmography (PPG). Instead of directly relating systolic and diastolic BP values with PPG features, our proposed methodology initially maps PPG features with some person specific intermediate latent parameters and later derives BP values from them. The 2-Element Windkessel model has been considered in the current context to estimate total peripheral resistance and arterial compliance of a person using PPG features, followed by linear regression for simulating arterial blood pressure. Experimental results, performed on a standard hospital dataset yielded absolute errors of 0.78±13.1 mmHg and 0.59 ± 10.23 mmHg for systolic and diastolic BP values respectively. Results also indicate that the methodology is more robust than the standard methodologies that directly estimate BP values from PPG signal.