Enhancing the accuracy of shock advisory algorithms in automated external defibrillators during ongoing cardiopulmonary resuscitation using a cascade of CNNEDs

Delivery of continuous cardiopulmonary resuscitation (CPR) plays an important role in the out-of-hospital cardiac arrest (OHCA) survival rate. However, to prevent CPR artifacts being superimposed on ECG morphology data, currently available automated external defibrillators (AEDs) require pauses in CPR for accurate analysis heart rhythms. In this study, we propose a novel Convolutional Neural Network-based Encoder-Decoder (CNNED) structure with a shock advisory algorithm to improve the accuracy and reliability of shock versus non-shock decision-making without CPR pause in OHCA scenarios. Our approach employs a cascade of CNNEDs in conjunction with an AED shock advisory algorithm to process the ECG data for shock decisions. Initially, a CNNED trained on an equal number of shockable and non-shockable rhythms is used to filter the CPR-contaminated data. The resulting filtered signal is then fed into a second CNNED, which is trained on imbalanced data more tilted toward the specific rhythm being analyzed. A reliable shock versus non-shock decision is made when both classifiers from the cascade structure agree, while segments with conflicting classifications are labeled as indeterminate, indicating the need for additional segments to analyze. To evaluate our approach, we generated CPR-contaminated ECG data by combining clean ECG data with 52 CPR samples. We used clean ECG data from the CUDB, AFDB, SDDB, and VFDB databases, to which 52 CPR artifact cases were added, while a separate test set provided by the AED manufacturer Defibtech LLC was used for performance evaluation. The test set comprised 20,384 non-shockable CPR-contaminated segments from 392 subjects, as well as 3744 shockable CPR-contaminated samples from 41 subjects with coarse ventricular fibrillation (VF) and 31 subjects with rapid ventricular tachycardia (rapid VT). We observed improvements in rhythm analysis using our proposed cascading CNNED structure when compared to using a single CNNED structure. Specifically, the specificity of the proposed cascade of CNNED structure increased from 99.14% to 99.35% for normal sinus rhythm and from 96.45% to 97.22% for other non-shockable rhythms. Moreover, the sensitivity for shockable rhythm detection increased from 90.90% to 95.41% for ventricular fibrillation and from 82.26% to 87.66% for rapid ventricular tachycardia. These results meet the performance thresholds set by the American Heart Association and demonstrate the reliable and accurate analysis of heart rhythms during CPR using only ECG data without the need for CPR interruptions or a reference signal.

Tracking Tidal Volume from Holter and Wearable Armband Electrocardiogram Monitoring

A novel method for tracking the tidal volume (TV) from electrocardiogram (ECG) is presented. The method is based on the amplitude of ECG-derived respiration (EDR) signals. Three different morphology-based EDR signals and three different amplitude estimation methods have been studied, leading to a total of 9 amplitude-EDR (AEDR) signals per ECG channel. The potential of these AEDR signals to track the changes in TV was analyzed. These methods do not need a calibration process. In addition, a personalized-calibration approach for TV estimation is proposed, based on a linear model that uses all AEDR signals from a device. All methods have been validated with two different ECG devices: a commercial Holter monitor, and a custom-made wearable armband. The lowest errors for the personalized-calibration methods, compared to a reference TV, were -3.48% [-17.41% / 12.93%] (median [first quartile / third quartile]) for the Holter monitor, and 0.28% [-10.90% / 17.15%] for the armband. On the other hand, medians of correlations to the reference TV were higher than 0.8 for uncalibrated methods, while they were higher than 0.9 for personal-calibrated methods. These results suggest that TV changes can be tracked from ECG using either a conventional (Holter) setup, or our custom-made wearable armband. These results also suggest that the methods are not as reliable in applications that induce small changes in TV, but they can be potentially useful for detecting large changes in TV, such as sleep apnea/hypopnea and/or exacerbations of a chronic respiratory disease.

Analysis of sympathetic responses to cognitive stress and pain through skin sympathetic nerve activity and electrodermal activity

We explored the non-invasive evaluation of the sympathetic nervous system (SNS) by employing two distinct physiological signals: skin sympathetic nerve activity (SKNA), extracted from electrocardiogram (ECG) signals, and electrodermal activity (EDA), a well-studied marker in the context of the SNS assessment. Our investigation focused on cognitive stress and pain; two conditions closely associated with the SNS. We sought to determine if the information and dynamics of EDA could be derived from the novel SKNA signal. To this end, ECG and EDA signals were recorded simultaneously during three experiments aimed at sympathetic stimulation, Valsalva maneuver (VM), Stroop test, and thermal-grill pain test. We calculated the integral area under the rectified SKNA signal (iSKNA) and decomposed the EDA signal to its phasic component (EDAphasic). An average delay of more than 4.6 s was observed in the onset of EDAphasic bursts compared to their corresponding iSKNA bursts. After shifting the EDAphasic segments by the extent of this delay and smoothing the corresponding iSKNA bursts, our results revealed a strong average correlation coefficient of 0.85±0.14 between the iSKNA and EDAphasic bursts, indicating a noteworthy similarity between the two signals. We also reconstructed the EDA signals with time-varying sympathetic (TVSymp) and modified TVSymp (MTVSymp) methods. Then we extracted the following features from iSKNA, EDAphasic, TVSymp, and MTVSymp signals: peak amplitude, average amplitude (aSKNA), standard deviation (vSKNA), and the cumulative duration during which the signals had higher amplitudes than a specified threshold (HaSKNA). A strong average correlation of 0.89±0.18 was found between vSKNA and subjects' self-rated pain levels during the pain test. Our statistical analysis also included applying Linear Mixed-Effects Models to check if there were significant differences in features across baseline and different levels of SNS stimulation. We then assessed the discriminating power of the features using Area Under the Receiver Operating Characteristic Curve (AUROC) and Fisher's Ratio. Finally, using all the four EDA features, a multi-layer perceptron (MLP) classifier reached the classification accuracies 95.56%, 89.29%, and 67.88% for the VM, Stroop, and thermal-grill pain control and stimulation classes. On the other hand, the highest classification accuracies based on SKNA features were achieved using K-nearest neighbors (KNN) (98.89%), KNN (89.29%), and MLP (95.11%) classifiers for the same experiments. Our comparative analysis showed the feasibility of SKNA as a novel tool for assessing the SNS with accurate classification capability, with a faster onset of amplitude increase in response to SNS activity, compared to EDA.

Noise Reduction in Photoplethysmography Signals Using a Convolutional Denoising Autoencoder With Unconventional Training Scheme

OBJECTIVE

We propose an efficient approach based on a convolutional denoising autoencoder (CDA) network to reduce motion and noise artifacts (MNA) from corrupted atrial fibrillation (AF) and non-AF photoplethysmography (PPG) data segments so that an accurate PPG-signal-derived heart rate can be obtained. Our method's main innovation is the optimization of the CDA performance for both rhythms using more AF than non-AF data for training the AF-specific CDA model and vice versa for the non-AF CDA network.

METHODS

To evaluate this unconventional training scheme, our proposed network was trained and tested on 25-sec PPG data segments from 48 subjects from two different databases-the Pulsewatch dataset and Stanford University's publicly available PPG dataset. In total, our dataset contains 10,773 data segments: 7,001 segments for training and 3,772 independent segments from out-of-sample subjects for testing.

RESULTS

Using real-life corrupted PPG segments, our approach significantly reduced the average heart rate root mean square error (RMSE) of the reconstructed PPG segments by 45.74% and 23% compared to the corrupted non-AF and AF data, respectively. Further, our approach exhibited lower RMSE, and higher sensitivity and PPV for detected peaks compared to the reconstructed data produced by the alternative methods.

CONCLUSION

These results show the promise of our approach as a reliable denoising method, which should be used prior to AF detection algorithms for an accurate cardiac health monitoring involving wearable devices.

SIGNIFICANCE

PPG signals collected from wearables are vulnerable to MNA, which limits their use as a reliable measurement, particularly in uncontrolled real-life environments.

Technology engagement is associated with higher perceived physical well-being in stroke patients prescribed smartwatches for atrial fibrillation detection

Background

Increasing ownership of smartphones among Americans provides an opportunity to use these technologies to manage medical conditions. We examine the influence of baseline smartwatch ownership on changes in self-reported anxiety, patient engagement, and health-related quality of life when prescribed smartwatch for AF detection.

Method

We performed a post-hoc secondary analysis of the Pulsewatch study (NCT03761394), a clinical trial in which 120 participants were randomized to receive a smartwatch-smartphone app dyad and ECG patch monitor compared to an ECG patch monitor alone to establish the accuracy of the smartwatch-smartphone app dyad for detection of AF. At baseline, 14 days, and 44 days, participants completed the Generalized Anxiety Disorder-7 survey, the Health Survey SF-12, and the Consumer Health Activation Index. Mixed-effects linear regression models using repeated measures with anxiety, patient activation, physical and mental health status as outcomes were used to examine their association with smartwatch ownership at baseline.

Results

Ninety-six participants, primarily White with high income and tertiary education, were randomized to receive a study smartwatch-smartphone dyad. Twenty-four (25%) participants previously owned a smartwatch. Compared to those who did not previously own a smartwatch, smartwatch owners reported significant greater increase in their self-reported physical health (β = 5.07, P < 0.05), no differences in anxiety (β = 0.92, P = 0.33), mental health (β = -2.42, P = 0.16), or patient activation (β = 1.86, P = 0.54).

Conclusions

Participants who own a smartwatch at baseline reported a greater positive change in self-reported physical health, but not in anxiety, patient activation, or self-reported mental health over the study period.

Design and Evaluation of Deep Learning Models for Continuous Acute Pain Detection Based on Phasic Electrodermal Activity

The current method for assessing pain in clinical practice is subjective and relies on self-reported scales. An objective and accurate method of pain assessment is needed for physicians to prescribe the proper medication dosage, which could reduce addiction to opioids. Hence, many works have used electrodermal activity (EDA) as a suitable signal for detecting pain. Previous studies have used machine learning and deep learning to detect pain responses, but none have used a sequence-to-sequence deep learning approach to continuously detect acute pain from EDA signals, as well as accurate detection of pain onset. In this study, we evaluated deep learning models including 1-dimensional convolutional neural networks (1D-CNN), long short-term memory networks (LSTM), and three hybrid CNN-LSTM architectures for continuous pain detection using phasic EDA features. We used a database consisting of 36 healthy volunteers who underwent pain stimuli induced by a thermal grill. We extracted the phasic component, phasic drivers, and time-frequency spectrum of the phasic EDA (TFS-phEDA), which was found to be the most discerning physiomarker. The best model was a parallel hybrid architecture of a temporal convolutional neural network and a stacked bi-directional and uni-directional LSTM, which obtained a F1-score of 77.8% and was able to correctly detect pain in 15-second signals. The model was evaluated using 37 independent subjects from the BioVid Heat Pain Database and outperformed other approaches in recognizing higher pain levels compared to baseline with an accuracy of 91.5%. The results show the feasibility of continuous pain detection using deep learning and EDA.

Anxiety, patient activation, and quality of life among stroke survivors prescribed smartwatches for atrial fibrillation monitoring

Background

The detection of atrial fibrillation (AF) in stroke survivors is critical to decreasing the risk of recurrent stroke. Smartwatches have emerged as a convenient and accurate means of AF diagnosis; however, the impact on critical patient-reported outcomes, including anxiety, engagement, and quality of life, remains ill defined.

Objectives

To examine the association between smartwatch prescription for AF detection and the patient-reported outcomes of anxiety, patient activation, and self-reported health.

Methods

We used data from the Pulsewatch trial, a 2-phase randomized controlled trial that included participants aged 50 years or older with a history of ischemic stroke. Participants were randomized to use either a proprietary smartphone-smartwatch app for 30 days of AF monitoring or no cardiac rhythm monitoring. Validated surveys were deployed before and after the 30-day study period to assess anxiety, patient activation, and self-rated physical and mental health. Logistic regression and generalized estimation equations were used to examine the association between smartwatch prescription for AF monitoring and changes in the patient-reported outcomes.

Results

A total of 110 participants (mean age 64 years, 41% female, 91% non-Hispanic White) were studied. Seventy percent of intervention participants were novice smartwatch users, as opposed to 84% of controls, and there was no significant difference in baseline rates of anxiety, activation, or self-rated health between the 2 groups. The incidence of new AF among smartwatch users was 6%. Participants who were prescribed smartwatches did not have a statistically significant change in anxiety, activation, or self-reported health as compared to those who were not prescribed smartwatches. The results held even after removing participants who received an AF alert on the watch.

Conclusion

The prescription of smartwatches to stroke survivors for AF monitoring does not adversely affect key patient-reported outcomes. Further research is needed to better inform the successful deployment of smartwatches in clinical practice.

Genetic data visualization using literature text-based neural networks: Examples associated with myocardial infarction

Data visualization is critical to unraveling hidden information from complex and high-dimensional data. Interpretable visualization methods are critical, especially in the biology and medical fields, however, there are limited effective visualization methods for large genetic data. Current visualization methods are limited to lower-dimensional data and their performance suffers if there is missing data. In this study, we propose a literature-based visualization method to reduce high-dimensional data without compromising the dynamics of the single nucleotide polymorphisms (SNP) and textual interpretability. Our method is innovative because it is shown to (1) preserves both global and local structures of SNP while reducing the dimension of the data using literature text representations, and (2) enables interpretable visualizations using textual information. For performance evaluations, we examined the proposed approach to classify various classification categories including race, myocardial infarction event age groups, and sex using several machine learning models on the literature-derived SNP data. We used visualization approaches to examine clustering of data as well as quantitative performance metrics for the classification of the risk factors examined above. Our method outperformed all popular dimensionality reduction and visualization methods for both classification and visualization, and it is robust against missing and higher-dimensional data. Moreover, we found it feasible to incorporate both genetic and other risk information obtained from literature with our method.

Deep cross-modal feature learning applied to predict acutely decompensated heart failure using in-home collected electrocardiography and transthoracic bioimpedance

BACKGROUND

Deep learning has been successfully applied to ECG data to aid in the accurate and more rapid diagnosis of acutely decompensated heart failure (ADHF). Previous applications focused primarily on classifying known ECG patterns in well-controlled clinical settings. However, this approach does not fully capitalize on the potential of deep learning, which directly learns important features without relying on a priori knowledge. In addition, deep learning applications to ECG data obtained from wearable devices have not been well studied, especially in the field of ADHF prediction.

METHODS

We used ECG and transthoracic bioimpedance data from the SENTINEL-HF study, which enrolled patients (≥21 years) who were hospitalized with a primary diagnosis of heart failure or with ADHF symptoms. To build an ECG-based prediction model of ADHF, we developed a deep cross-modal feature learning pipeline, termed ECGX-Net, that utilizes raw ECG time series and transthoracic bioimpedance data from wearable devices. To extract rich features from ECG time series data, we first adopted a transfer learning approach in which ECG time series were transformed into 2D images, followed by feature extraction using ImageNet-pretrained DenseNet121/VGG19 models. After data filtering, we applied cross-modal feature learning in which a regressor was trained with ECG and transthoracic bioimpedance. Then, we concatenated the DenseNet121/VGG19 features with the regression features and used them to train a support vector machine (SVM) without bioimpedance information.

RESULTS

The high-precision classifier using ECGX-Net predicted ADHF with a precision of 94 %, a recall of 79 %, and an F1-score of 0.85. The high-recall classifier with only DenseNet121 had a precision of 80 %, a recall of 98 %, and an F1-score of 0.88. We found that ECGX-Net was effective for high-precision classification, while DenseNet121 was effective for high-recall classification.

CONCLUSION

We show the potential for predicting ADHF from single-channel ECG recordings obtained from outpatients, enabling timely warning signs of heart failure. Our cross-modal feature learning pipeline is expected to improve ECG-based heart failure prediction by handling the unique requirements of medical scenarios and resource limitations.

False Atrial Fibrillation Alerts from Smartwatches are Associated with Decreased Perceived Physical Well-being and Confidence in Chronic Symptoms Management

Wrist-based wearables have been FDA approved for AF detection. However, the health behavior impact of false AF alerts from wearables on older patients at high risk for AF are not known. In this work, we analyzed data from the Pulsewatch (NCT03761394) study, which randomized patients (≥50 years) with history of stroke or transient ischemic attack to wear a patch monitor and a smartwatch linked to a smartphone running the Pulsewatch application vs to only the cardiac patch monitor over 14 days. At baseline and 14 days, participants completed validated instruments to assess for anxiety, patient activation, perceived mental and physical health, chronic symptom management self-efficacy, and medicine adherence. We employed linear regression to examine associations between false AF alerts with change in patient-reported outcomes. Receipt of false AF alerts was related to a dose-dependent decline in self-perceived physical health and levels of disease self-management. We developed a novel convolutional denoising autoencoder (CDA) to remove motion and noise artifacts in photoplethysmography (PPG) segments to optimize AF detection, which substantially reduced the number of false alerts. A promising approach to avoid negative impact of false alerts is to employ artificial intelligence driven algorithms to improve accuracy.

Differentiating between stress- and EPT-induced electrodermal activity during dental examination

Dental pain invokes the sympathetic nervous system, which can be measured by electrodermal activity (EDA). In the dental clinic, accurate quantification of pain is needed because it could enable optimized drug-dose treatments, thereby potentially reducing drug addiction. However, a confounding factor is that during pain there is also lingering residual stress, hence, both contribute to the EDA response. Therefore, we investigated whether EDA can differentiate stress from pain during dental examination. The use of electrical pulp test (EPT) is an ideal approach to tease out the dynamics of stress and mimic pain with lingering residual stress. Once the electrical sensation is felt and reaches a critical current threshold, the subject removes the probe from their tooth, hence, this stage of data represents largely EPT stimulus and the residual stress-induced EDA response is smaller. EPT was performed on necrotic and vital teeth in fifty-one subjects. We defined four different data groups of reactions based on each individual's EPT intensity level expectation based on the visual analog scale (VAS) of their baseline trial, as follows: mild stress, mild stress + EPT, strong stress, and strong stress + EPT. EDA-derived features exhibited significant difference between residual lingering stress + EPT groups and stress groups. We obtained 84.6% accuracy with 76.2% sensitivity and 86.8% specificity with multilayer perceptron in differentiating between pure-stress groups vs. stress + EPT groups. Moreover, EPT induced much greater EDA amplitude and faster response than stress. Our finding suggests that our machine learning approach can discriminate between stress and EPT stimulation in EDA signals.

The use of electrodermal activity in pulpal diagnosis and dental pain assessment

AIMS

To explore whether electrodermal activity (EDA) can serve as a complementary tool for pulpal diagnosis (Aim 1) and an objective metric to assess dental pain before and after local anaesthesia (Aim 2).

METHODOLOGY

A total of 53 subjects (189 teeth) and 14 subjects (14 teeth) were recruited for Aim 1 and Aim 2, respectively. We recorded EDA using commercially available devices, PowerLab and Galvanic Skin Response (GSR) Amplifier, in conjunction with cold and electric pulp testing (EPT). Participants rated their level of sensation on a 0-10 visual analogue scale (VAS) after each test. We recorded EPT-stimulated EDA activity before and after the administration of local anaesthesia for participants who required root canal treatment (RCT) due to painful pulpitis. The raw data were converted to the time-varying index of sympathetic activity (TVSymp), a sensitive and specific parameter of EDA. Statistical analysis was performed using Python 3.6 and its Scikit-post hoc library.

RESULTS

Electrodermal activity was upregulated by the stimuli of cold and EPT testing in the normal pulp. TVSymp signals were significantly increased in vital pulp compared to necrotic pulp by both cold test and EPT. Teeth that exhibited intensive sensitivity to cold with or without lingering pain had increased peak numbers of TVSymp than teeth with mild sensation to cold. Pre- and post-anaesthesia EDA activity and VAS scores were recorded in patients with painful pulpitis. Post-anaesthesia EDA signals were significantly lower compared to pre-anaesthesia levels. Approximately 71% of patients (10 of 14 patients) experienced no pain during treatment and reported VAS score of 0 or 1. The majority of patients (10 of 14) showed a reduction of TVSymp after the administration of anaesthesia. Two of three patients who experienced increased pain during RCT (post-treatment VAS > pre-treatment VAS) exhibited increased post-anaesthesia TVSymp.

CONCLUSIONS

Our data show promising results for using EDA in pulpal diagnosis and for assessing dental pain. Whilst our testing was limited to subjects who had adequate communication skills, our future goal is to be able to use this technology to aid in the endodontic diagnosis of patients who have limited communication ability.

A Smartwatch System for Continuous Monitoring of Atrial Fibrillation in Older Adults After Stroke or Transient Ischemic Attack: Application Design Study

BACKGROUND

The prevalence of atrial fibrillation (AF) increases with age and can lead to stroke. Therefore, older adults may benefit the most from AF screening. However, older adult populations tend to lag more than younger groups in the adoption of, and comfort with, the use of mobile health (mHealth) apps. Furthermore, although mobile apps that can detect AF are available to the public, most are designed for intermittent AF detection and for younger users. No app designed for long-term AF monitoring has released detailed system design specifications that can handle large data collections, especially in this age group.

OBJECTIVE

This study aimed to design an innovative smartwatch-based AF monitoring mHealth solution in collaboration with older adult participants and clinicians.

METHODS

The Pulsewatch system is designed to link smartwatches and smartphone apps, a website for data verification, and user data organization on a cloud server. The smartwatch in the Pulsewatch system is designed to continuously monitor the pulse rate with embedded AF detection algorithms, and the smartphone in the Pulsewatch system is designed to serve as the data-transferring hub to the cloud storage server.

RESULTS

We implemented the Pulsewatch system based on the functionality that patients and caregivers recommended. The user interfaces of the smartwatch and smartphone apps were specifically designed for older adults at risk for AF. We improved our Pulsewatch system based on feedback from focus groups consisting of patients with stroke and clinicians. The Pulsewatch system was used by the intervention group for up to 6 weeks in the 2 phases of our randomized clinical trial. At the conclusion of phase 1, 90 trial participants who had used the Pulsewatch app and smartwatch for 14 days completed a System Usability Scale to assess the usability of the Pulsewatch system; of 88 participants, 56 (64%) endorsed that the smartwatch app is "easy to use." For phases 1 and 2 of the study, we collected 9224.4 hours of smartwatch recordings from the participants. The longest recording streak in phase 2 was 21 days of consecutive recordings out of the 30 days of data collection.

CONCLUSIONS

This is one of the first studies to provide a detailed design for a smartphone-smartwatch dyad for ambulatory AF monitoring. In this paper, we report on the system's usability and opportunities to increase the acceptability of mHealth solutions among older patients with cognitive impairment.

TRIAL REGISTRATION

ClinicalTrials.gov NCT03761394; https://www.clinicaltrials.gov/ct2/show/NCT03761394.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

RR2-10.1016/j.cvdhj.2021.07.002.

Design and Validation of a Multimodal Wearable Device for Simultaneous Collection of Electrocardiogram, Electromyogram, and Electrodermal Activity

Bio-signals are being increasingly used for the assessment of pathophysiological conditions including pain, stress, fatigue, and anxiety. For some approaches, a single signal is not sufficient to provide a comprehensive diagnosis; however, there is a growing consensus that multimodal approaches allow higher sensitivity and specificity. For instance, in visceral pain subjects, the autonomic activation can be inferred using electrodermal activity (EDA) and heart rate variability derived from the electrocardiogram (ECG), but including the muscle activation detected from the surface electromyogram (sEMG) can better differentiate the disease that causes the pain. There is no wearable device commercially capable of collecting these three signals simultaneously. This paper presents the validation of a novel multimodal low profile wearable data acquisition device for the simultaneous collection of EDA, ECG, and sEMG signals. The device was validated by comparing its performance to laboratory-scale reference devices. N = 20 healthy subjects were recruited to participate in a four-stage study that exposed them to an array of cognitive, orthostatic, and muscular stimuli, ensuring the device is sensitive to a range of stressors. Time and frequency domain analyses for all three signals showed significant similarities between our device and the reference devices. Correlation of sEMG metrics ranged from 0.81 to 0.95 and EDA/ECG metrics showed few instances of significant difference in trends between our device and the references. With only minor observed differences, we demonstrated the ability of our device to collect EDA, sEMG, and ECG signals. This device will enable future practical and impactful advances in the field of chronic pain and stress measurement and can confidently be implemented in related studies.

Optimized Signal Quality Assessment for Photoplethysmogram Signals Using Feature Selection

OBJECTIVE

With the increasing use of wearable healthcare devices for remote patient monitoring, reliable signal quality assessment (SQA) is required to ensure the high accuracy of interpretation and diagnosis on the recorded data from patients. Photoplethysmographic (PPG) signals non-invasively measured by wearable devices are extensively used to provide information about the cardiovascular system and its associated diseases. In this study, we propose an approach to optimize the quality assessment of the PPG signals.

METHODS

We used an ensemble-based feature selection scheme to enhance the prediction performance of the classification model to assess the quality of the PPG signals. Our approach for feature and subset size selection yielded the best-suited feature subset, which was optimized to differentiate between the clean and artifact corrupted PPG segments.

CONCLUSION

A high discriminatory power was achieved between two classes on the test data by the proposed feature selection approach, which led to strong performance on all dependent and independent test datasets. We achieved accuracy, sensitivity, and specificity rates of higher than 0.93, 0.89, and 0.97, respectively, for dependent test datasets, independent of heartbeat type, i.e., atrial fibrillation (AF) or non-AF data including normal sinus rhythm (NSR), premature atrial contraction (PAC), and premature ventricular contraction (PVC). For independent test datasets, accuracy, sensitivity, and specificity rates were greater than 0.93, 0.89, and 0.97, respectively, on PPG data recorded from AF and non-AF subjects. These results were found to be more accurate than those of all of the contemporary methods cited in this work.

SIGNIFICANCE

As the results illustrate, the advantage of our proposed scheme is its robustness against dynamic variations in the PPG signal during long-term 14-day recordings accompanied with different types of physical activities and a diverse range of fluctuations and waveforms caused by different individual hemodynamic characteristics, and various types of recording devices. This robustness instills confidence in the application of the algorithm to various kinds of wearable devices as a reliable PPG signal quality assessment approach.

A Deep Convolutional Autoencoder for Automatic Motion Artifact Removal in Electrodermal Activity Signals: A Preliminary Study

Automatic motion artifact (MA) removal in electrodermal activity (EDA) signals is a major challenge because of the aperiodic and irregular characteristics of EDA. Given the lack of a suitable MA removal algorithm, a substantial amount of EDA data is typically discarded, especially during ambulatory monitoring. Current methods for MA removal in EDA are feasible when data are corrupted with low magnitude artifacts. In this study, we propose a more data-driven deep convolutional autoencoder (DCAE) for automated motion artifact removal in EDA signals. The DCAE was trained using several publicly available datasets. We used both Gaussian white noise (GWN) and real-life induced MA data records collected in a laboratory setting to corrupt the clean EDA signals. We compared the performance of our DCAE network with three state-of-the-art methods using the performance metrics the signal-to-noise ratio (SNR) improvement (SNR_imp), and the mean squared error (MSE). The proposed DCAE provided significantly higher SNR_impand lower MSE compared to three other methods for both synthetically and real-life induced MA. While the work is preliminary, this work illustrates a promising approach which can potentially be used to remove many different types of MA.

Language-Independent Sleepy Speech Detection

Prolonged sleepiness can lead to impairment of cognitive and physical performance and may cause unfortunate accidents. Speech signals are easily accessible using a simple microphone or other means, hence, automated approaches for accurate sleepiness detection from speech signals are desired to prevent degradation in human performance and accidental injury. Sleepiness is known to affect acoustic patterns of speech so that they are different from those of normal speech, and this change is also independent of the language being spoken. To date, there have been no studies examining linguistic-independent sleepy speech detection. We used two different languages, English and German, to detect sleepy speech, where the former was used to train/validate and the latter to test the effectiveness of machine and deep learning models. Specifically, we trained ResNet50, a deep learning model, and five machine learning models with relevant vocal features. Speech data segments from three English-speaking subjects were used for training the model and segments from an English-speaking subject were used for validation. We then tested ResNet50 and the five different machine-learning models using speech data segments from one German-speaking subject. Deep learning far outperformed all of the machine learning approaches. The accuracy, sensitivity, specificity, and geometric mean values were found to be 0.96, 0.92, 0.99, and 0.95, respectively, using ResNet50 on the test data. Our preliminary results suggest that sleepiness can be accurately detected independently from linguistic speech. Clinical Relevance-It is not known if sleepiness can be detected regardless of the language spoken. Our results show the feasibility of accurate sleepiness detection using deep learning even when tested with a different language than trained on.

Multi-level Pain Quantification using a Smartphone and Electrodermal Activity

Appropriate prescription of pain medication is challenging because pain is difficult to quantify due to the subjectiveness of pain assessment. Currently, clinicians must entirely rely on pain scales based on patients' assessments. This has been alleged to be one of the causes of drug overdose and addiction, and a contributor to the opioid crisis. Therefore, there is an urgent unmet need for objective pain assessment. Furthermore, as pain can occur anytime and anywhere, ambulatory pain monitoring would be welcomed in practice. In our previous study, we developed electrodermal activity (EDA)-derived indices and implemented them in a smartphone application that can communicate via Bluetooth to an EDA wearable device. While we previously showed high accuracy for high-level pain detection, multi-level pain detection has not been demonstrated. In this paper, we tested our smartphone application with a multi-level pain-induced dataset. The dataset was collected from fifteen subjects who underwent four levels of pain-inducing electrical pulse (EP) stimuli. We then performed statistical analyses and machine-learning techniques to classify multiple pain levels. Significant differences were observed in our EDA-derived indices among no-pain, low-pain, and high-pain segments. A random forest classifier showed 62.6% for the balanced accuracy, and a random forest regressor exhibited 0.441 for the coefficient of determination. Clinical Relevance - This is one of the first studies to present a smartphone application for detecting multiple levels of pain in real time using an EDA wearable device. This work shows the feasibility of ambulatory pain monitoring which can potentially be useful for chronic pain management.

Improving the Accuracy of R-Peak Detection in a Wearable Armband Device for Daily Life Electrocardiogram Monitoring Using a Deep Convolutional Denoising Encoder-Decoder Network

Continuous long-term heart rate (HR) monitoring using wearable devices is desirable to aid in the diagnosis of many health-related conditions. Recently, we have developed an armband device that does not use obstructive leads, has dry electrodes which are convenient for long-term electrocardiogram (ECG) recording, and has been shown to be an effective alternate approach for continuous ECG monitoring. However, motion artifacts (MA) due to electromyogram (EMG) contractions are acknowledged as the major challenge of an armband. In this study, we used a deep convolutional neural network denoising encoder-decoder (CNNDED) to enhance the accuracy of R-peak detection in MA-corrupted ECG recordings obtained by an armband device. We collected simultaneous 24-hour ECG recordings using both the armband device and a Holter monitor on 10 subjects. Each 10-sec ECG segment was converted to a time-frequency representation and subsequently used as the input to CNNDED. During the training process, the model learned to accentuate the location of R peaks by amplifying their values in each ECG beat and suppressing the remaining waveforms. For the training output, the model used the R-peak location information from the simultaneously collected Holter ECG data, which were considered as the reference. The performance of CNNDED was evaluated on an independent test data set using the standard performance metrics. The mean relative error of the estimated HR with respect to the Holter data was 17.5 and 7.3 beats/min, pre- and post-CNNDED, respectively. The mean relative difference of the root mean square of successive difference values were 0.23 and 0.06 before and after applying CNNDED, respectively. Although further study is needed, the current preliminary results suggest that CNNDED can improve detection of R peaks even when they are completely buried in the presence of EMG artifacts.

A Deep Convolutional Autoencoder for Automatic Motion Artifact Removal in Electrodermal Activity

OBJECTIVE

This study aimed to develop a robust and data driven automatic motion artifacts (MA) removal technique from electrodermal activity (EDA) signal.

METHODS

we proposed a deep convolutional autoencoder (DCAE) approach for automatic MA removal in EDA signals. Our model was trained using several publicly available datasets that were collected using a wide variety of stimuli to cause EDA reactions; the sample size was large (N=385 subjects). We trained and validated our DCAE network using both Gaussian white noise (GWN) and realistic MA data records collected using a novel circuitry in our lab. We further evaluated and compared the performance of our DCAE model with the existing methods on two independent and unseen datasets called Chon lab motion artifact dataset II (CMAD II) and central nervous system oxygen toxicity dataset (CNS-OT).

RESULTS

Our DCAE model showed significantly higher signal-to-noise-power-ratio improvement (SNR_imp) and lower mean squared error (MSE) when compared with that of the three previous methods (averaged SNR_imp=35.25 dB, and MSE=0.028 on the MA-corrupted data). Moreover, the reconstructed EDAs from the CMAD II dataset had a mean correlation value of 0.78 (statistically significantly higher when compared with other methods) with the reference clean data from the motionless hand, whereas the raw MA-corrupted data had a correlation value of only 0.68.

CONCLUSION

The results presented in the paper indicates that our DCAE can remove MAs with higher intensity where the existing methods fails.

SIGNIFICANCE

Proposed DCAE model can be used to recover a significant amount of otherwise discarded EDA data.

Comparison of Electrodermal Activity from Multiple Body Locations Based on Standard EDA Indices' Quality and Robustness against Motion Artifact

The most traditional sites for electrodermal activity (EDA) data collection, palmar locations such as fingers or palms, are not usually recommended for ambulatory monitoring given that subjects have to use their hands regularly during their daily activities, and therefore, alternative sites are often sought for EDA data collection. In this study, we collected EDA signals (n = 23 subjects, 19 male) from four measurement sites (forehead, back of neck, finger, and inner edge of foot) during cognitive stress and induction of mild motion artifacts by walking and one-handed weightlifting. Furthermore, we computed several EDA indices from the EDA signals obtained from different sites and evaluated their efficiency to classify cognitive stress from the baseline state. We found a high within-subject correlation between the EDA signals obtained from the finger and the feet. Consistently high correlation was also found between the finger and the foot EDA in both the phasic and tonic components. Statistically significant differences were obtained between the baseline and cognitive stress stage only for the EDA indices computed from the finger and the foot EDA. Moreover, the receiver operating characteristic curve for cognitive stress detection showed a higher area-under-the-curve for the EDA indices computed from the finger and foot EDA. We also evaluated the robustness of the different body sites against motion artifacts and found that the foot EDA location was the best alternative to other sites.

Diabetes Distal Peripheral Neuropathy: Subtypes and Diagnostic and Screening Technologies

Diabetes distal symmetrical peripheral neuropathy (DSPN) is the most prevalent form of neuropathy in industrialized countries, substantially increasing risk for morbidity and pre-mature mortality. DSPN may manifest with small-fiber disease, large-fiber disease, or a combination of both. This review summarizes: (1) DSPN subtypes (small- and large-fiber disease) with attention to clinical signs and patient symptoms; and (2) technological diagnosis and screening for large- and small-fiber disease with inclusion of a comprehensive literature review of published studies from 2015-present (N = 66). Review findings, informed by the most up-to-date research, advance critical understanding of DSPN large- and small-fiber screening technologies, including those designed for point-of-care use in primary care and endocrinology practices.

Autonomic Nervous System characterization in hyperbaric environments considering respiratory component and non-linear analysis of Heart Rate Variability

OBJECTIVES

an evaluation of Principal Dynamic Mode (PDM) and Orthogonal Subspace Projection (OSP) methods to characterize the Autonomic Nervous System (ANS) response in three different hyperbaric environments was performed.

METHODS

ECG signals were recorded in two different stages (baseline and immersion) in three different hyperbaric environments: (a) inside a hyperbaric chamber, (b) in a controlled sea immersion, (c) in a real reservoir immersion. Time-domain parameters were extracted from the RR series of the ECG. From the Heart Rate Variability signal (HRV), classic Power Spectral Density (PSD), PDM (a non-linear analysis of HRV which is able to separate sympathetic and parasympathetic activities) and OSP (an analysis of HRV which is able to extract the respiratory component) methods were used to assess the ANS response.

RESULTS

PDM and OSP parameters follows the same trend when compared to the PSD ones for the hyperbaric chamber dataset. Comparing the three hyperbaric scenarios, significant differences were found: i) heart rate decreased and RMSSD increased in the hyperbaric chamber and the controlled dive, but they had the opposite behavior during the uncontrolled dive; ii) power in the OSP respiratory component was lower than power in the OSP residual component in cases a and c; iii) PDM and OSP methods showed a significant increase in sympathetic activity during both dives, but parasympathetic activity increased only during the uncontrolled dive.

CONCLUSIONS

PDM and OSP methods could be used as an alternative measurement of ANS response instead of the PSD method. OSP results indicate that most of the variation in the heart rate variability cannot be described by changes in the respiration, so changes in ANS response can be assigned to other factors. Time-domain parameters reflect vagal activation in the hyperbaric chamber and in the controlled dive because of the effect of pressure. In the uncontrolled dive, sympathetic activity seems to be dominant, due to the effects of other factors such as physical activity, the challenging environment, and the influence of breathing through the scuba mask during immersion. In sum, a careful description of the changes in all the possible factors that could affect the ANS response between baseline and immersion stages in hyperbaric environments is needed for better interpretation of the results.

Seizures Caused by Exposure to Hyperbaric Oxygen in Rats Can Be Predicted by Early Changes in Electrodermal Activity

Hyperbaric oxygen (HBO₂) is breathed during undersea operations and in hyperbaric medicine. However, breathing HBO₂ by divers and patients increases the risk of central nervous system oxygen toxicity (CNS-OT), which ultimately manifests as sympathetic stimulation producing tachycardia and hypertension, hyperventilation, and ultimately generalized seizures and cardiogenic pulmonary edema. In this study, we have tested the hypothesis that changes in electrodermal activity (EDA), a measure of sympathetic nervous system activation, precedes seizures in rats breathing 5 atmospheres absolute (ATA) HBO₂. Radio telemetry and a rodent tether apparatus were adapted for use inside a sealed hyperbaric chamber. The tethered rat was free to move inside a ventilated animal chamber that was flushed with air or 100% O₂. The animal chamber and hyperbaric chamber (air) were pressurized in parallel at ~1 atmosphere/min. EDA activity was recorded simultaneously with cortical electroencephalogram (EEG) activity, core body temperature, and ambient pressure. We have captured the dynamics of EDA using time-varying spectral analysis of raw EDA (TVSymp), previously developed as a tool for sympathetic tone assessment in humans, adjusted to detect the dynamic changes of EDA in rats that occur prior to onset of CNS-OT seizures. The results show that a significant increase in the amplitude of TVSymp values derived from EDA recordings occurs on average (±SD) 1.9 ± 1.6 min before HBO₂-induced seizures. These results, if corroborated in humans, support the use of changes in TVSymp activity as an early "physio-marker" of impending and potentially fatal seizures in divers and patients.

Automated Condition-Based Suppression of the CPR Artifact in ECG Data to Make a Reliable Shock Decision for AEDs during CPR

Cardiopulmonary resuscitation (CPR) corrupts the morphology of the electrocardiogram (ECG) signal, resulting in an inaccurate automated external defibrillator (AED) rhythm analysis. Consequently, most current AEDs prohibit CPR during the rhythm analysis period, thereby decreasing the survival rate. To overcome this limitation, we designed a condition-based filtering algorithm that consists of three stop-band filters which are turned either 'on' or 'off' depending on the ECG's spectral characteristics. Typically, removing the artifact's higher frequency peaks in addition to the highest frequency peak eliminates most of the ECG's morphological disturbance on the non-shockable rhythms. However, the shockable rhythms usually have dynamics in the frequency range of (3-6) Hz, which in certain cases coincide with CPR compression's harmonic frequencies, hence, removing them may lead to destruction of the shockable signal's dynamics. The proposed algorithm achieves CPR artifact removal without compromising the integrity of the shockable rhythm by considering three different spectral factors. The dataset from the PhysioNet archive was used to develop this condition-based approach. To quantify the performance of the approach on a separate dataset, three performance metrics were computed: the correlation coefficient, signal-to-noise ratio (SNR), and accuracy of Defibtech's shock decision algorithm. This dataset, containing 14 s ECG segments of different types of rhythms from 458 subjects, belongs to Defibtech commercial AED's validation set. The CPR artifact data from 52 different resuscitators were added to artifact-free ECG data to create 23,816 CPR-contaminated data segments. From this, 82% of the filtered shockable and 70% of the filtered non-shockable ECG data were highly correlated (>0.7) with the artifact-free ECG; this value was only 13 and 12% for CPR-contaminated shockable and non-shockable, respectively, without our filtering approach. The SNR improvement was 4.5 ± 2.5 dB, averaging over the entire dataset. Defibtech's rhythm analysis algorithm was applied to the filtered data. We found a sensitivity improvement from 67.7 to 91.3% and 62.7 to 78% for VF and rapid VT, respectively, and specificity improved from 96.2 to 96.5% and 91.5 to 92.7% for normal sinus rhythm (NSR) and other non-shockables, respectively.

Time-varying Spectral Index of Electrodermal Activity to Predict Central Nervous System Oxygen Toxicity Symptoms in Divers: Preliminary results

The most effective method to mitigate decompression sickness in divers is hyperbaric oxygen (HBO₂) pre-breathing. However, divers breathing HBO₂ are at risk for developing central nervous system oxygen toxicity (CNS-OT), which can manifest as symptoms that might impair a diver's performance, or cause more serious symptoms like seizures. In this study, we have collected electrodermal activity (EDA) signals in fifteen subjects at elevated oxygen partial pressures (2.06 ATA, 35 FSW) in the "foxtrot" chamber pool at the Duke University Hyperbaric Center, while performing a cognitive stress test for up to 120 minutes. Specifically, we have computed the time-varying spectral analysis of EDA (TVSymp) as a tool for sympathetic tone assessment and evaluated its feasibility for the prediction of symptoms of CNS-OT in divers. The preliminary results show large increase in the amplitude TVSymp values derived from EDA recordings ~2 minutes prior to expert human adjudication of symptoms related to oxygen toxicity. An early detection based on TVSymp might allow the diver to take countermeasures against the dire consequences of CNS-OT which can lead to drowning.Clinical Relevance-This study provides a sensitive analysis method which indicates a significant increase in the electrodermal activity prior to human expert adjudication of symptoms related to CNS-OT.

A Preliminary Study on Automatic Motion Artifact Detection in Electrodermal Activity Data Using Machine Learning

The electrodermal activity (EDA) signal is a sensitive and non-invasive surrogate measure of sympathetic function. Use of EDA has increased in popularity in recent years for such applications as emotion and stress recognition; assessment of pain, fatigue, and sleepiness; diagnosis of depression and epilepsy; and other uses. Recently, there have been several studies using ambulatory EDA recordings, which are often quite useful for analysis of many physiological conditions. Because ambulatory monitoring uses wearable devices, EDA signals are often affected by noise and motion artifacts. An automated noise and motion artifact detection algorithm is therefore of utmost importance for accurate analysis and evaluation of EDA signals. In this paper, we present machine learning-based algorithms for motion artifact detection in EDA signals. With ten subjects, we collected two simultaneous EDA signals from the right and left hands, while instructing the subjects to move only the right hand. Using these data, we proposed a cross-correlation-based approach for non-biased labeling of EDA data segments. A set of statistical, spectral and model-based features were calculated which were then subjected to a feature selection algorithm. Finally, we trained and validated several machine learning methods using a leave-one-subject-out approach. The classification accuracy of the developed model was 83.85% with a standard deviation of 4.91%, which was better than a recent standard method that we considered for comparison to our algorithm.

Preliminary Analysis of the Risk Factor Identification Embedding Model for Cardiovascular Disease

Cardiovascular Disease (CVD) is responsible for a large part of healthcare costs every year, but susceptibility to it is affected by complex biological and physiological variables including patients' genetics and lifestyles. There has not been much work to develop a framework that incorporates these important and clinically relevant risk factors into a comprehensive model for CVD research. Moreover, the data labeling required to do so, such as annotating gene functions, is an extremely challenging, tedious, and time-consuming process. In this work, our goal was to develop and validate a risk factor embedding model, which incorporates genotype, phenotype without pre-labeled information to identify various risk factors of CVD. We hypothesize that (1) the knowledge background that does not require data labeling could be gathered from published abstract data, (2) the phenotype, genotype risk factors could be represented in an embedding vector space. We collected 1,363,682 published abstracts from PubMed using the keyword "heart" and 19,264 human gene names, then trained our model using the collected abstracts. We evaluated our CVD risk factor identification model using both intrinsic and extrinsic evaluations: for the intrinsic evaluation, we examined whether or not the captured top-10 words and genes have references related to the input query "myocardial infarction", as one of CVDs, and our model correctly identified them. For the extrinsic evaluation, we used our model to the dimensionality reduction task for classifications, and our method outperformed other popular methods. These results show the feasibility of our approach for disease-associated risk factors of CVD which incorporates genotype, phenotype.Clinical Relevance-Our model provides a comprehensive tool to incorporate various risk factors without any a priori data labeling knowledge for CVD. Our approach shows a potential to provide discovered knowledge that contributes to better understanding and treatment of CVD.

Female-male Differences Should be Considered in Physical Pain Quantification based on Electrodermal Activity: Preliminary Study

Objective pain quantification is an important but difficult goal. Electrodermal activity (EDA) has been widely explored for this purpose, given its reported sensitivity to pain. However, cognitive stress can hinder successful estimation of physical pain when using EDA signals. We collected EDA signals from ten subjects (5 male and 5 female) undergoing pain stimulation, and calculated phasic, tonic, and frequency-domain features. Each subject experienced pain with and without stress. Three low and three high pain sessions were induced using two thermal grills (low-level for visual analog scale [VAS] 4 or 5 and high-level for VAS 7 or more). The Stroop test was performed for inducing cognitive stress. Significant differences between EDA features of painless and pain segments were observed. Significant differences between no pain and stress were also observed. Furthermore, we compared differences in EDA features between females and males under pain and cognitive stress. Frequency-domain EDA features of pain increased with stress for both females and males. Frequency-domain features derived from females also showed higher standard deviation than did those derived from males. We performed machine learning analysis and evaluated the models using leave-one-subject-out cross-validation. We obtained balanced accuracies of 63.5%, 72.4%, and 53.2% (combined, male, and female) when using training data of the same sex and 47.6%, 57.4%, and 42.7% (combined, male, and female) when using different sex for training.Clinical Relevance-Our preliminary results suggest that sex of patients should be considered to increase the accuracy of pain quantification based on EDA in the presence of cognitive stress.

Validation of Spectral Indices of Electrodermal Activity with a Wearable Device

Electrodermal activity (EDA) has been found to be a highly sensitive, accurate and non-invasive measure of the sympathetic nervous system's activity and has been used to extract biomarkers of various pathophysiological conditions including stress, fatigue, epilepsy, and chronic pain. Recently, various robust signal processing techniques have been developed to obtain more reliable and accurate indices that capture the meaningful characteristics of the EDA using data collected from laboratory-scale devices. However, EDA also has the potential to monitor such physiological conditions in active ambulatory settings, for which the developed tools must be deployed in wearable devices. In this paper, we studied the feasibility of obtaining the highly-sensitive spectral indices of EDA using a wearable device. EDA signals were collected from left hand fingers using a wearable device and a laboratory-scale reference device, while N=18 subjects underwent the Head up Tilt test and the Stroop test to stimulate orthostatic and cognitive stress, respectively. We computed two time-domain indices, the skin conductance level (SCL) and nonspecific skin conductance responses (NS.SCRs), and two spectral indices, the normalized sympathetic components of the EDA (EDASympn), and the time-varying EDA index of sympathetic control (TVSymp). The results showed similar performances for EDASympn and TVSymp indices across both devices. While spectral indices obtained from both devices performed similarly in response to orthostatic and cognitive stress, time-domain exhibited large variation when obtained by the wearable device. Further research is required to develop and refine such devices, as well as the indices used to analyze EDA results.Clinical Relevance- This study proves the feasibility of obtaining spectral indices of EDA using a wearable device, which can be used to develop wearable tools to detect pain, stress, fatigue, between others.

Sensitive Physiological Indices of Pain Based on Differential Characteristics of Electrodermal Activity

OBJECTIVE

Electrodermal activity (EDA) has been widely used to assess human response to stressful stimuli, including pain. Recently, spectral analysis of EDA has been found to be more sensitive and reproducible for assessment of sympathetic arousal than traditional indices (e.g., tonic and phasic components). However, none of the aforementioned analyses incorporate the differential characteristics of EDA, which could be more sensitive to capturing fast-changing dynamics associated with pain responses.

METHODS

We have tested the feasibility of using the derivative of phasic EDA and the modified time-varying spectral analysis of EDA. Sixteen subjects underwent four levels of pain stimulation using electric stimulation. Five-second segments of EDA were used for each level of stimulation, and pre-stimulation segments were considered stimulation level 0. We used support vector machines with the radial basis function kernel and multi-layer perceptron for three different scenarios of stimulation-level classification tasks: five stimulation levels (four levels of stimulation plus no stimulation); low, medium, and high pain stimulation (stimulation levels 0-1, 2, and 3-4, respectively); and high stimulation levels (stimulation levels 3-4) vs. no stimulation.

RESULTS

The maximum balanced accuracies were 44% (five stimulation levels), 63% (for low, medium, and high pain stimulation), and 87% (sensitivity 83% and specificity 89%, for high stimulation vs. no stimulation).

CONCLUSION

The differential characteristics of EDA contributed highly to the accuracy of pain stimulation level detection of the classifiers. The external validity dataset was not considered in the study.

SIGNIFICANCE

Our approach has the potential for accurate pain quantification using EDA.

Atrial Fibrillation Prediction from Critically Ill Sepsis Patients

Sepsis is defined by life-threatening organ dysfunction during infection and is the leading cause of death in hospitals. During sepsis, there is a high risk that new onset of atrial fibrillation (AF) can occur, which is associated with significant morbidity and mortality. Consequently, early prediction of AF during sepsis would allow testing of interventions in the intensive care unit (ICU) to prevent AF and its severe complications. In this paper, we present a novel automated AF prediction algorithm for critically ill sepsis patients using electrocardiogram (ECG) signals. From the heart rate signal collected from 5-min ECG, feature extraction is performed using the traditional time, frequency, and nonlinear domain methods. Moreover, variable frequency complex demodulation and tunable Q-factor wavelet-transform-based time-frequency methods are applied to extract novel features from the heart rate signal. Using a selected feature subset, several machine learning classifiers, including support vector machine (SVM) and random forest (RF), were trained using only the 2001 Computers in Cardiology data set. For testing the proposed method, 50 critically ill ICU subjects from the Medical Information Mart for Intensive Care (MIMIC) III database were used in this study. Using distinct and independent testing data from MIMIC III, the SVM achieved 80% sensitivity, 100% specificity, 90% accuracy, 100% positive predictive value, and 83.33% negative predictive value for predicting AF immediately prior to the onset of AF, while the RF achieved 88% AF prediction accuracy. When we analyzed how much in advance we can predict AF events in critically ill sepsis patients, the algorithm achieved 80% accuracy for predicting AF events 10 min early. Our algorithm outperformed a state-of-the-art method for predicting AF in ICU patients, further demonstrating the efficacy of our proposed method. The annotations of patients' AF transition information will be made publicly available for other investigators. Our algorithm to predict AF onset is applicable for any ECG modality including patch electrodes and wearables, including Holter, loop recorder, and implantable devices.

Objective pain stimulation intensity and pain sensation assessment using machine learning classification and regression based on electrodermal activity

An objective measure of pain remains an unmet need of people with chronic pain, estimated to be 1/3 of the adult population in the United States. The current gold standard to quantify pain is highly subjective, based upon self-reporting with numerical or visual analog scale (VAS). This subjectivity complicates pain management and exacerbates the epidemic of opioid abuse. We have tested classification and regression machine learning models to objectively estimate pain sensation in healthy subjects using electrodermal activity (EDA). Twenty-three volunteers underwent pain stimulation using thermal grills. Three different "pain stimulation intensities" were induced for each subject, who reported the "pain sensation" right after each stimulus using a VAS (0-10). EDA data were collected throughout the experiment. For machine learning, we computed validated features of EDA based on time-domain decomposition, spectral analysis, and differential features. Models for estimation of pain stimulation intensity and pain sensation achieved maximum macroaveraged geometric mean scores of 69.7% and 69.2%, respectively, when three classes were considered ("No," "Low," and "High"). Regression of levels of stimulation intensity and pain sensation achieved R² values of 0.357 and 0.47, respectively. Overall, the high variance and inconsistency of VAS scores led to lower performance of pain sensation classification, but regression was better for pain sensation than stimulation intensity. Our results provide that three levels of pain can be quantified with good accuracy and physiological evidence that sympathetic responses recorded by EDA are more correlated to the applied stimuli's intensity than to the pain sensation reported by the subject.

Real-Time High-Level Acute Pain Detection Using a Smartphone and a Wrist-Worn Electrodermal Activity Sensor

The subjectiveness of pain can lead to inaccurate prescribing of pain medication, which can exacerbate drug addiction and overdose. Given that pain is often experienced in patients' homes, there is an urgent need for ambulatory devices that can quantify pain in real-time. We implemented three time- and frequency-domain electrodermal activity (EDA) indices in our smartphone application that collects EDA signals using a wrist-worn device. We then evaluated our computational algorithms using thermal grill data from ten subjects. The thermal grill delivered a level of pain that was calibrated for each subject to be 8 out of 10 on a visual analog scale (VAS). Furthermore, we simulated the real-time processing of the smartphone application using a dataset pre-collected from another group of fifteen subjects who underwent pain stimulation using electrical pulses, which elicited a VAS pain score level 7 out of 10. All EDA features showed significant difference between painless and pain segments, termed for the 5-s segments before and after each pain stimulus. Random forest showed the highest accuracy in detecting pain, 81.5%, with 78.9% sensitivity and 84.2% specificity with leave-one-subject-out cross-validation approach. Our results show the potential of a smartphone application to provide near real-time objective pain detection.

Deep Neural Network Approach for Continuous ECG-Based Automated External Defibrillator Shock Advisory System During Cardiopulmonary Resuscitation

Background Because chest compressions induce artifacts in the ECG, current automated external defibrillators instruct the user to stop cardiopulmonary resuscitation (CPR) while an automated rhythm analysis is performed. It has been shown that minimizing interruptions in CPR increases the chance of survival. Methods and Results The objective of this study was to apply a deep-learning algorithm using convolutional layers, residual networks, and bidirectional long short-term memory method to classify shockable versus nonshockable rhythms in the presence and absence of CPR artifact. Forty subjects' data from Physionet with 1131 shockable and 2741 nonshockable samples contaminated with 43 different CPR artifacts that were acquired from a commercial automated external defibrillator during asystole were used. We had separate data as train and test sets. Using our deep neural network model, the sensitivity and specificity of the shock versus no-shock decision for the entire data set over the 4-fold cross-validation sets were 95.21% and 86.03%, respectively. This result was based on the training and testing of the model using ECG data in both the presence and the absence of CPR artifact. For ECG without CPR artifact, the sensitivity was 99.04% and the specificity was 95.2%. A sensitivity of 94.21% and a specificity of 86.14% were obtained for ECG with CPR artifact. In addition to 4-fold cross-validation sets, we also examined leave-one-subject-out validation. The sensitivity and specificity for the case of leave-one-subject-out validation were 92.71% and 97.6%, respectively. Conclusions The proposed trained model can make shock versus nonshock decision in automated external defibrillators, regardless of CPR status. The results meet the American Heart Association's sensitivity requirement (>90%).

A robust ECG denoising technique using variable frequency complex demodulation

BACKGROUND AND OBJECTIVE

Electrocardiogram (ECG) is widely used for the detection and diagnosis of cardiac arrhythmias such as atrial fibrillation. Most of the computer-based automatic cardiac abnormality detection algorithms require accurate identification of ECG components such as QRS complexes in order to provide a reliable result. However, ECGs are often contaminated by noise and artifacts, especially if they are obtained using wearable sensors, therefore, identification of accurate QRS complexes often becomes challenging. Most of the existing denoising methods were validated using simulated noise added to a clean ECG signal and they did not consider authentically noisy ECG signals. Moreover, many of them are model-dependent and sampling-frequency dependent and require a large amount of computational time.

METHODS

This paper presents a novel ECG denoising technique using the variable frequency complex demodulation (VFCDM) algorithm, which considers noises from a variety of sources. We used the sub-band decomposition of the noise-contaminated ECG signals using VFCDM to remove the noise components so that better-quality ECGs could be reconstructed. An adaptive automated masking is proposed in order to preserve the QRS complexes while removing the unnecessary noise components. Finally, the ECG was reconstructed using a dynamic reconstruction rule based on automatic identification of the severity of the noise contamination. The ECG signal quality was further improved by removing baseline drift and smoothing via adaptive mean filtering.

RESULTS

Evaluation results on the standard MIT-BIH Arrhythmia database suggest that the proposed denoising technique provides superior denoising performance compared to studies in the literature. Moreover, the proposed method was validated using real-life noise sources collected from the noise stress test database (NSTDB) and data from an armband ECG device which contains significant muscle artifacts. Results from both the wearable armband ECG data and NSTDB data suggest that the proposed denoising method provides significantly better performance in terms of accurate QRS complex detection and signal to noise ratio (SNR) improvement when compared to some of the recent existing denoising algorithms.

CONCLUSIONS

The detailed qualitative and quantitative analysis demonstrated that the proposed denoising method has been robust in filtering varieties of noises present in the ECG. The QRS detection performance of the denoised armband ECG signals indicates that the proposed denoising method has the potential to increase the amount of usable armband ECG data, thus, the armband device with the proposed denoising method could be used for long term monitoring of atrial fibrillation.

Development and Validation of an Automated Algorithm to Detect Atrial Fibrillation Within Stored Intensive Care Unit Continuous Electrocardiographic Data: Observational Study

Background

Atrial fibrillation (AF) is the most common arrhythmia during critical illness, representing a sepsis-defining cardiac dysfunction associated with adverse outcomes. Large burdens of premature beats and noisy signal during sepsis may pose unique challenges to automated AF detection.

Objective

The objective of this study is to develop and validate an automated algorithm to accurately identify AF within electronic health care data among critically ill patients with sepsis.

Methods

This is a retrospective cohort study of patients hospitalized with sepsis identified from Medical Information Mart for Intensive Care (MIMIC III) electronic health data with linked electrocardiographic (ECG) telemetry waveforms. Within 3 separate cohorts of 50 patients, we iteratively developed and validated an automated algorithm that identifies ECG signals, removes noise, and identifies irregular rhythm and premature beats in order to identify AF. We compared the automated algorithm to current methods of AF identification in large databases, including ICD-9 (International Classification of Diseases, 9th edition) codes and hourly nurse annotation of heart rhythm. Methods of AF identification were tested against gold-standard manual ECG review.

Results

AF detection algorithms that did not differentiate AF from premature atrial and ventricular beats performed modestly, with 76% (95% CI 61%-87%) accuracy. Performance improved (P=.02) with the addition of premature beat detection (validation set accuracy: 94% [95% CI 83%-99%]). Median time between automated and manual detection of AF onset was 30 minutes (25th-75th percentile 0-208 minutes). The accuracy of ICD-9 codes (68%; P=.002 vs automated algorithm) and nurse charting (80%; P=.02 vs algorithm) was lower than that of the automated algorithm.

Conclusions

An automated algorithm using telemetry ECG data can feasibly and accurately detect AF among critically ill patients with sepsis, and represents an improvement in AF detection within large databases.

Novel Density Poincaré Plot Based Machine Learning Method to Detect Atrial Fibrillation From Premature Atrial/Ventricular Contractions

OBJECTIVE

Detection of Atrial fibrillation (AF) from premature atrial contraction (PAC) and premature ventricular contraction (PVC) is difficult as frequent occurrences of these ectopic beats can mimic the typical irregular patterns of AF. In this paper, we present a novel density Poincaré plot-based machine learning method to detect AF from PAC/PVCs using electrocardiogram (ECG) recordings.

METHODS

First, we propose the generation of this new density Poincaré plot which is derived from the difference of the heart rate (DHR) and provides the overlapping phase-space trajectory information of the DHR. Next, from this density Poincaré plot, several image processing domain-based approaches including statistical central moments, template correlation, Zernike moment, discrete wavelet transform and Hough transform features are used to extract suitable features. Subsequently, the infinite latent feature selection algorithm is implemented to rank the features. Finally, classification of AF vs. PAC/PVC is performed using K-Nearest Neighbor, Support vector machine (SVM) and Random Forest (RF) classifiers. Our method is developed and validated using a subset of Medical Information Mart for Intensive Care (MIMIC) III database containing 10 AF and 10 PAC/PVC subjects. Results- During the segment-wise 10-fold cross-validation, SVM achieved the best performance with 98.99% sensitivity, 95.18% specificity and 97.45% accuracy with the extracted features. In subject-wise scenario, RF achieved the highest accuracy of 91.93%. Moreover, we further validated the proposed method using two other databases: wearable armband ECG data and the Physionet AFPDB. 100% PAC detection accuracy was obtained for both databases without any further training.

CONCLUSION

Our proposed density Poincaré plot-based method showed superior performance when compared with four existing algorithms; thus showing the efficacy of the extracted image domain-based features.

SIGNIFICANCE

From intensive care unit's ECG to wearable armband ECGs, the proposed method is shown to discriminate PAC/PVCs from AF with high accuracy.

Atrial Fibrillation Detection During Sepsis: Study on MIMIC III ICU Data

Sepsis is defined by life-threatening organ dysfunction during infection and is one of the leading causes of critical illness. During sepsis, there is high risk that new-onset of atrial fibrillation (AF) can occur, which is associated with significant morbidity and mortality. As a result, computer aided automated and reliable detection of new-onset AF during sepsis is crucial, especially for the critically ill patients in the intensive care unit (ICU). In this paper, a novel automated and robust two-step algorithm to detect AF from ICU patients using electrocardiogram (ECG) signals is presented. First, several statistical parameters including root mean square of successive differences, Shannon entropy, and sample entropy were calculated from the heart rate for the screening of possible AF segments. Next, Poincaré plot-based features along with P-wave characteristics were used to reduce false positive detection of AF, caused by the premature atrial and ventricular beats. A subset of the Medical Information Mart for Intensive Care (MIMIC) III database containing 198 subjects was used in this study. During the training and validation phases, both the simple thresholding as well as machine learning classifiers achieved very high segment-wise AF classification performance. Finally, we tested the performance of our proposed algorithm using two independent test data sets and compared the performance with two state-of-the-art methods. The algorithm achieved an overall 100% sensitivity, 98% specificity, 98.99% accuracy, 98% positive predictive value, and 100% negative predictive value on the subject-wise AF detection, thus showing the efficacy of our proposed algorithm in critically ill sepsis patients. The annotations of the data have been made publicly available for other investigators.

Smartphone Based Human Activity Recognition with Feature Selection and Dense Neural Network

For the past few years, smartphone based human activity recognition (HAR) has gained much popularity due to its embedded sensors which have found various applications in healthcare, surveillance, human-device interaction, pattern recognition etc. In this paper, we propose a neural network model to classify human activities, which uses activity-driven hand-crafted features. First, the neighborhood component analysis derived feature selection is used to choose a subset of important features from the available time and frequency domain parameters. Next, a dense neural network consisting of four hidden layers is modeled to classify the input features into different categories. The model is evaluated on publicly available UCI HAR data set consisting of six daily activities; our approach achieved 95.79% classification accuracy. When compared with existing state-of-the-art methods, our proposed model outperformed most other methods while using fewer features, thus showing the importance of proper feature selection.

Denoising Wearable Armband ECG Data Using the Variable Frequency Complex Demodulation Technique

We propose a novel electrocardiogram (ECG) denoising technique using the variable frequency complex demodulation (VFCDM) algorithm. We used VFCDM to perform the sub-band decomposition of the noise-contaminated ECG to remove the noise components so that accurate QRS complexes could be identified. The ECG quality was further improved by removing baseline drift and smoothing via adaptive mean filtering. The proposed method was validated on the MIT-BIH arrhythmia database (MITDB) and wearable armband ECG data. For the former, we added Gaussian white noise to the ECG signals at different signal-to-noise ratios and the denoising performance of the proposed method was compared with other denoising techniques. The proposed approach showed superior denoising performance compared to the other methods. We compared the QRS complex detection performance of the noisy to the denoised armband ECG. The performance of the proposed denoising method on the armband ECG resulted in comparable QRS complex detection as that obtained when using Holter monitor ECG signals. This demonstrates that the proposed algorithm can significantly increase the amount of usable armband ECG data, which would otherwise have been discarded due to electromyogram contamination especially during arm movements. Hence, the proposed algorithm has the potential to enable long-term monitoring of atrial fibrillation using the armband without the discomfort of skin irritation often experienced with Holter monitors.Clinical Relevance- The proposed ECG denoising method can significantly increase the ECG quality of wearable ECG devices, which are more susceptible to muscle and motion artifacts.

Digital Image Processing Features of Smartwatch Photoplethysmography for Cardiac Arrhythmia Detection

The aim of our work is to design an algorithm to detect premature atrial contraction (PAC), premature ventricular contraction (PVC), and atrial fibrillation (AF) among normal sinus rhythm (NSR) using smartwatch photoplethysmographic (PPG) data. Novel image processing features and two machine learning methods are used to enhance the PAC/PVC detection results of the Poincaré plot method. Compared with support vector machine (SVM) methods, the Random Forests (RF) method performs better. It yields a 10-fold cross validation (CV) averaged sensitivity, specificity, positive predicted value (PPV), negative predicted value (NPV), and accuracy for PAC/PVC labels of 63%, 98%, 83%, 94%, and 93%, respectively, and a 10-fold CV averaged sensitivity, specificity, PPV, NPV, and accuracy for AF subjects of 92%, 96%, 85%, 98%, and 95%, respectively. This is one of the first studies to derive image processing features from Poincaré plots to further enhance the accuracy of PAC/PVC detection using PPG recordings from a smartwatch.

Preliminary Results on Density Poincare Plot Based Atrial Fibrillation Detection from Premature Atrial/Ventricular Contractions

Detection of Atrial fibrillation (AF) from premature atrial contraction (PAC) and premature ventricular contraction (PVC) is challenging as frequent occurrences of these ectopic beats can mimic the typical irregular patterns of AF. In this paper, we present a preliminary study of using density Poincare plot based machine learning method to detect AF from PAC/PVCs using electrocardiogram (ECG) recordings. First, we propose creation of this new density Poincare plot which is derived from the difference of the heart rate. Next, from this density Poincare plot, template correlation and discrete wavelet transform are used to extract suitable image-based features, which is followed by infinite latent feature selection algorithm to rank the features. Finally, classification of AF vs PAC/PVC is performed using K-Nearest Neighbor, discriminant analysis and support vector machine (SVM) classifiers. Our method is developed and validated using a subset of Medical Information Mart for Intensive Care (MIMIC) III database containing 8 AF and 8 PAC/PVC subjects. Both 10-fold and leave-one-subject-out cross validations are performed to show the robustness of our proposed method. During the 10-fold cross-validation, SVM achieved the best performance with 99.49% sensitivity, 94.51% specificity and 97.29% accuracy with the extracted features while for the leave-one-subject-out, the highest overall accuracy is 90.91%. Moreover, when compared with two state-of-the-art methods, the proposed algorithm achieves superior AF vs. PAC/PVC discrimination performance.Clinical Relevance-This preliminary study shows that with the help of density Poincare plot, AF can be separated from PAC/PVC with better accuracy.

Pain Detection using a Smartphone in Real Time

We developed an objective real-time pain detection method using a smartphone and a wrist-worn wearable device to collect electrodermal activity (EDA) signals. Recently, various researchers have developed pain management applications. However, they rely on subjective self-reported pain scores or the video camera of a smartphone to detect pain, but the latter method's accuracy needs further improvement. In our work, we use a wrist-worn EDA device which transmits data via Bluetooth to a smartphone. A smartphone application was developed to analyze the EDA data so that near real-time processed pain detection information can be displayed. The analysis of EDA is based on estimating time-varying spectral power in the frequency range (0.08-0.24 Hz) associated with the sympathetic nervous system. This time-varying characterization of EDA is termed TVSymp. In this work, we also examined whether removing baseline EDA fluctuations from TVSymp would provide more accurate results. This was carried out by taking the moving average of the EDA response prior to stimulus and subtracting that value from the EDA response post stimulus. This approach is termed modified TVSymp (MTVSymp). Pain stimuli were induced in ten subjects using a thermal grill, which gives intense pain perception without damaging skin tissues. We compared both TVSymp and MTVSymp in detecting pain induced by the thermal grill using machine learning approaches. We found the accuracy of pain detection of TVSymp and MTVSymp to be 80% and 90%, respectively.

Premature Atrial and Ventricular Contraction Detection using Photoplethysmographic Data from a Smartwatch

We developed an algorithm to detect premature atrial contraction (PAC) and premature ventricular contraction (PVC) using photoplethysmographic (PPG) data acquired from a smartwatch. Our PAC/PVC detection algorithm is composed of a sequence of algorithms that are combined to discriminate various arrhythmias. A novel vector resemblance method is used to enhance the PAC/PVC detection results of the Poincaré plot method. The new PAC/PVC detection algorithm with our automated motion and noise artifact detection approach yielded a sensitivity of 86% for atrial fibrillation (AF) subjects while the overall sensitivity was 67% when normal sinus rhythm (NSR) subjects were also included. The specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy values for the combined data consisting of both NSR and AF subjects were 97%, 81%, 94% and 92%, respectively, for PAC/PVC detection combined with our automated motion and noise artifact detection approach. Moreover, when AF detection was compared with and without PAC/PVC, the sensitivity and specificity increased from 94.55% to 98.18% and from 95.75% to 97.90%, respectively. For additional independent testing data, we used two datasets: a smartwatch PPG dataset that was collected in our ongoing clinical study, and a pulse oximetry PPG dataset from the Medical Information Mart for Intensive Care III database. The PAC/PVC classification results of the independent testing on these two other datasets are all above 92% for sensitivity, specificity, PPV, NPV, and accuracy. The proposed combined approach to detect PAC and PVC can ultimately lead to better accuracy in AF detection. This is one of the first studies involving detection of PAC and PVC using PPG recordings from a smartwatch. The proposed method can potentially be of clinical importance as this enhanced capability can lead to fewer false positive detections of AF, especially for those NSR subjects with frequent episodes of PAC/PVC.