Although various predictors and methods for BP estimation have been proposed, differences in study designs have led to difficulties in determining the optimal method. This study presents analyses of BP estimation methods using 2.4 million cardiac cycles of two commonly used non-invasive biosignals, electrocardiogram (ECG) and photoplethysmogram (PPG), from 1376 surgical patients. Feature selection methods were used to determine the best subset of predictors from a total of 42 including PAT, heart rate (HR), and various PPG morphology features, and BP estimation models constructed using linear regression (LR), random forest (RF), artificial neural network (ANN), and recurrent neural network (RNN) were evaluated. 28 features out of 42 were determined as suitable for BP estimation, in particular two PPG morphology features outperformed PAT, which has been conventionally seen as the best non-invasive indicator of BP. By modelling the low frequency component of BP using ANN and the high frequency component using RNN with the selected predictors, mean errors of 0.05 ± 6.92 mmHg for systolic BP, and -0.05 ± 3.99 mmHg for diastolic BP were achieved. External validation of the model using another biosignal database consisting of 334 intensive care unit patients led to similar results, satisfying three standards for accuracy of BP monitors. The results indicate that the proposed method can contribute to the realization of ubiquitous non-invasive continuous BP monitoring.
Reconstructing a standard 12-lead electrocardiogram (ECG) from signals received from electrodes packed into a patch-type device is a challenging task in the field of medical instrumentation. All attempts to obtain a clinically valid 12-lead ECG using a patch-type device were not satisfactory. In this study, we designed the hardware for a three-lead patch-type ECG device and employed a long short-term memory (LSTM) network that can overcome the limitations of the linear regression algorithm used for ECG reconstruction. The LSTM network can overcome the issue of reduced horizontal components of the vector in the electric signal obtained from the patch-type device attached to the anterior chest. The reconstructed 12-lead ECG that uses the LSTM network was tested against a standard 12-lead ECG in 30 healthy subjects and ECGs of 30 patients with pathologic findings. The average correlation coefficient of the LSTM network was found to be 0.95. The ability of the reconstructed ECG to detect pathologic abnormalities was identical to that of the standard ECG. In conclusion, the reconstruction of a standard 12-lead ECG using a three-lead patch-type device is feasible, and such an ECG is an equivalent alternative to a standard 12-lead ECG.
1 November 2019 in IEEE Journal of Biomedical and Health Informatics
Photoplethysmography (PPG) has become ubiquitous with the development of smart watches and the mobile healthcare market. However, PPG is vulnerable to various types of noises that are ever present in uncontrolled environments, and the key to obtaining meaningful signals depends on successful denoising of PPG. In this context, algorithms have been developed to denoise PPG, but many were validated in controlled settings or are reliant on multiple steps that must all work correctly. This paper proposes a novel PPG denoising algorithm based on bidirectional recurrent denoising auto-encoder (BRDAE) that requires minimal pre-processing steps and have the benefit of waveform feature accentuation beyond simple denoising. The BRDAE was trained and validated on a dataset with artificially augmented noise, and was tested on a large open database of PPG signals collected from patients enrolled in intensive care units as well as from PPG data collected intermittently during the daily routine of nine subjects over 24 h. Denoising with the trained BRDAE improved signal-to-noise ratio of the noise-augmented data by 7.9 dB during validation. In the test datasets, the denoised PPG showed statistically significant improvement in heart rate detection as compared with the original PPG in terms of correlation to reference and root-mean-squared error. These results indicate that the proposed method is an effective solution for denoising the PPG signal, and promises values beyond traditional denoising by providing PPG feature accentuation for pulse waveform analysis.
Non-invasive continuous blood pressure monitors are of great interest to the medical community due to their value in hypertension management. Recently, studies have shown the potential of pulse pressure as a therapeutic target for hypertension, but not enough attention has been given to non-invasive continuous monitoring of pulse pressure. Although accurate pulse pressure estimation can be of direct value to hypertension management and indirectly to the estimation of systolic blood pressure, as it is the sum of pulse pressure and diastolic blood pressure, only a few inadequate methods of pulse pressure estimation have been proposed. We present a novel, non-invasive blood pressure and pulse pressure estimation method based on pulse transit time and pre-ejection period. Pre-ejection period and pulse transit time were measured non-invasively using electrocardiogram, seismocardiogram, and photoplethysmogram measured from the torso. The proposed method used the 2-element Windkessel model to model pulse pressure with the ratio of stroke volume, approximated by pre-ejection period, and arterial compliance, estimated by pulse transit time. Diastolic blood pressure was estimated using pulse transit time, and systolic blood pressure was estimated as the sum of the two estimates. The estimation method was verified in 11 subjects in two separate conditions with induced cardiovascular response and the results were compared against a reference measurement and values obtained from a previously proposed method. The proposed method yielded high agreement with the reference (pulse pressure correlation with reference R ≥ 0.927, diastolic blood pressure correlation with reference R ≥ 0.854, systolic blood pressure correlation with reference R ≥ 0.914) and high estimation accuracy in pulse pressure (mean root-mean-squared error ≤ 3.46 mmHg) and blood pressure (mean root-mean-squared error ≤ 6.31 mmHg for diastolic blood pressure and ≤ 8.41 mmHg for systolic blood pressure) over a wide range of hemodynamic changes. The proposed pulse pressure estimation method provides accurate estimates in situations with and without significant changes in stroke volume. The proposed method improves upon the currently available systolic blood pressure estimation methods by providing accurate pulse pressure estimates.
To evaluate the feasibility of brainstem auditory evoked potential (BAEP) for rehabilitation prognosis prediction in patients with ischemic stroke, 181 patients were tested using the Korean version of the modified Barthel index (K-MBI) at admission (basal K-MBI) and discharge (follow-up K-MBI). The BAEP measurements were performed within two weeks of admission on average. The criterion between favorable and unfavorable outcomes was defined as a K-MBI score of 75 at discharge, which was the boundary between moderate and mild dependence in daily living activities. The changes in the K-MBI scores (discharge-admission) were analyzed by nonlinear regression models, including the artificial neural network (ANN) and support vector machine (SVM), with the basal K-MBI score, age, and interpeak latencies (IPLs) of the BAEP (waves I, I–III, and III–V). When including the BAEP features, the correlations of the ANN and SVM regression models increased to 0.70 and 0.64, respectively. In the outcome prediction, the ANN model with the basal K-MBI score, age, and BAEP IPLs exhibited a sensitivity of 92% and specificity of 90%. Our results suggest that the BAEP IPLs used with the basal K-MBI score and age can play an adjunctive role in the prediction of patient rehabilitation prognoses.
Silent speech interfaces (SSIs) decode articulatory activity without audible speech. Although accelerometers can capture facial and neck movements in compact wearable setups, accelerometer-specific sensor placement has not been systematically examined. This study evaluated all combinations of ten accelerometers across the lower face and neck in a 100 English word silent speech classification task with 20 participants. Subject-dependent classification was performed using a convolutional neural network followed by a bidirectional long short-term memory network. Accuracy improved as the number of sensors increased, reaching a plateau with five sensors. At this point, Sensors 1, 4, 5, 7, and 8 achieved the highest accuracy of 92.30 ± 4.64% and were selected as the efficient configuration. Regional coverage comparisons showed that configurations spanning broader functional articulatory regions outperformed those with more restricted coverage. Together, these results support regionally distributed placement as a design strategy for capturing complementary articulatory information in compact accelerometer-based SSI systems.