10 January 2019 in IEEE Transactions on Biomedical Circuits and Systems
Advancements in wireless sensor network technologies have enabled the proliferation of miniaturized body-worn sensors, capable of long-term pervasive biomedical signal monitoring. Remote cardiovascular monitoring has been one of the beneficiaries of this development, resulting in non-invasive, photoplethysmography (PPG) sensors being used in ambulatory settings. Wrist-worn PPG, although a popular alternative to electrocardiogram, suffers from motion artifacts inherent in daily life. Hence, in this paper, we present a novel deep learning framework (CorNET) to efficiently estimate heart rate (HR) information and perform biometric identification (BId) using only a wrist-worn, single-channel PPG signal collected in ambulant environment. We have formulated a completely personalized data-driven approach, using a four-layer deep neural network. Two convolution neural network layers are used in conjunction with two long short-term memory layers, followed by a dense output layer for modeling the temporal sequence inherent within the pulsatile signal representative of cardiac activity. The final dense layer is customized with respect to the application, functioning as: regression layer-having a single neuron to predict HR; classification layer-two neurons that identify a subject among a group. The proposed network was evaluated on the TROIKA dataset having 22 PPG records collected during various physical activities. We achieve a mean absolute error of 1.47 ± 3.37 beats per minute for HR estimation and an average accuracy of 96% for BId on 20 subjects. CorNET was further evaluated successfully in an ambulant use-case scenario with custom sensors for two subjects.
30 April 2020 in IEEE Sensors Journal
This paper presents a deep learning model ‘PP-Net’ which is the first of its kind, having the capability to estimate the physiological parameters: Diastolic blood pressure (DBP), Systolic blood pressure (SBP), and Heart rate (HR) simultaneously from the same network using a single channel PPG signal. The proposed model is designed by exploiting the deep learning framework of Long-term Recurrent Convolutional Network (LRCN), exhibiting inherent ability of feature extraction, thereby, eliminating the cost effective steps of feature selection and extraction, making less-complex for deployment on resource constrained platforms such as mobile platforms. The performance demonstration of the PP-Net is done on a larger and publically available MIMIC-II database. We achieved an average NMAE of 0.09 (DBP) and 0.04 (SBP) mmHg for BP, and 0.046 bpm for HR estimation on total population of 1557 critically ill subjects. The accurate estimation of HR and BP on a larger population compared to the existing methods, demonstrated the effectiveness of our proposed deep learning framework. The accurate evaluation on a huge population with CVD complications, validates the robustness of the proposed framework in pervasive healthcare monitoring especially cardiac and stroke rehabilitation monitoring.
18 February 2019 in IEEE Transactions on Biomedical Engineering
In this paper, we present a deep learning framework ‘Rehab-Net’ for effectively classifying three upper limb movements of the human arm, involving extension, flexion and rotation of the forearm which over the time could provide a measure of rehabilitation progress. The proposed framework, Rehab-Net is formulated with a personalized, light weight and low complex, customized CNN model, using 2-layers of Convolutional neural network (CNN), interleaved with pooling layers, followed by a fully-connected layer that classifies the three movements from tri-axial acceleration input data collected from the wrist. The proposed Rehab-net framework was validated on sensor data collected in two situations-a) seminaturalistic environment involving an archetypal activity of ‘making-tea’ with 4 stroke survivors and b) natural environment, where 10 stroke survivors were free to perform any desired arm movement for a duration of 120 minutes. We achieve an overall accuracy of 97.89% on semi-naturalistic data and 88.87% on naturalistic data which exceeded state-of-the-art learning algorithms namely, Linear Discriminant Analysis, Support Vector Machines, and k-means clustering with an average accuracy of 48.89%, 44.14% and 27.64%. Subsequently, a computational complexity analysis of the proposed model has been discussed with an eye towards hardware implementation. The clinical significance of this study is to accurately monitor the clinical progress of the rehabilitated subjects under the ambulatory settings.
1 March 2013 in IEEE Journal of Biomedical and Health Informatics
This paper introduces a low-complexity algorithm for the extraction of the fiducial points from the electrocardiogram (ECG). The application area we consider is that of remote cardiovascular monitoring, where continuous sensing and processing takes place in low-power, computationally constrained devices, thus the power consumption and complexity of the processing algorithms should remain at a minimum level. Under this context, we choose to employ the discrete wavelet transform (DWT) with the Haar function being the mother wavelet, as our principal analysis method. From the modulus-maxima analysis on the DWT coefficients, an approximation of the ECG fiducial points is extracted. These initial findings are complimented with a refinement stage, based on the time-domain morphological properties of the ECG, which alleviates the decreased temporal resolution of the DWT. The resulting algorithm is a hybrid scheme of time- and frequency-domain signal processing. Feature extraction results from 27 ECG signals from QTDB were tested against manual annotations and used to compare our approach against the state-of-the art ECG delineators. In addition, 450 signals from the 15-lead PTBDB are used to evaluate the obtained performance against the CSE tolerance limits. Our findings indicate that all but one CSE limits are satisfied. This level of performance combined with a complexity analysis, where the upper bound of the proposed algorithm, in terms of arithmetic operations, is calculated as 2.423 N +214 additions and 1.093 N +12 multiplications for N ≤ 861 or 2.553 N +102 additions and 1.093 N +10 multiplications for N > 861 ( N being the number of input samples), reveals that the proposed method achieves an ideal tradeoff between computational complexity and performance, a key requirement in remote cardiovascular disease monitoring systems.
13 February 2020 in IEEE Journal of Translational Engineering in Health and Medicine
The clinical assessment technology such as remote monitoring of rehabilitation progress for lower limb related ailments rely on the automatic evaluation of movement performed along with an estimation of joint angle information. In this paper, we introduce a transfer-learning based Long-term Recurrent Convolution Network (LRCN) named as ‘MyoNet’ for the classification of lower limb movements, along with the prediction of the corresponding knee joint angle. The model consists of three blocks- (i) feature extractor block, (ii) joint angle prediction block, and (iii) movement classification block. Initially, the model is end-to-end trained for knee joint angle prediction followed by transferring the knowledge of a trained model to the movement classification through transfer-learning approach making a memory and computationally efficient design. The proposed MyoNet was evaluated on publicly available University of California (UC) Irvine machine learning repository dataset of the lower limb for 11 healthy subjects and 11 subjects with knee pathology for three movements type-walking, standing with knee flexion movements and sitting with knee extension movements. The average mean absolute error (MAE) resulted in the prediction of joint angle for healthy subjects and subjects with knee pathology are 8.1 % and 9.2 % respectively. Subsequently, an average classification accuracy of 98.1 % and 92.4 % were achieved for healthy subjects and subjects with knee pathology, respectively. Interestingly, the significance of this study in itself is promising with substantial improvement in the performance compared to state-of-the-art methodologies. The clinical significance of such surface electromyography signals (sEMG) based movement recognition and prediction of corresponding joint angle system could be beneficial for remote monitoring of rehabilitation progress by the physiotherapist using wearables.
23 January 2018 in IEEE Transactions on Neural Systems and Rehabilitation Engineering
Surface electromyography (sEMG) data acquired during lower limb movements has the potential for investigating knee pathology. Nevertheless, a major challenge encountered with sEMG signals generated by lower limb movements is the intersubject variability, because the signals recorded from the leg or thigh muscles are contingent on the characteristics of a subject such as gait activity and muscle structure. In order to cope with this difficulty, we have designed a three-step classification scheme. First, the multichannel sEMG is decomposed into activities of the underlying sources by means of independent component analysis via entropy bound minimization. Next, a set of time-domain features, which would best discriminate various movements, are extracted from the source estimates. Finally, the feature selection is performed with the help of the Fisher score and a screen-plot-based statistical technique, prior to feeding the dimension-reduced features to the linear discriminant analysis. The investigation involves 11 healthy subjects and 11 individuals with knee pathology performing three different lower limb movements, namely, walking, sitting, and standing, which yielded an average classification accuracy of 96.1% and 86.2%, respectively. While the outcome of this study per se is very encouraging, with suitable improvement, the clinical application of such an sEMG-based pattern recognition system that distinguishes healthy and knee pathological subjects would be an attractive consequence.