Objective: Smartwatches with photoplethysmographic (PPG) sensors are ideal for early atrial fibrillation (AF) detection through continuous monitoring. However, prior deep learning was limited either to controlled environments, to minimize motion artifacts, or to short duration data collection. Additionally, premature atrial/ventricular contractions (PAC/PVC), which often confound AF detection algorithms, remains understudied due to limited datasets. Current state-of-the-art methods achieve only 75% sensitivity for PAC/PVC class on minimally motion artifact corrupted PPG data, despite showing 97% AF detection accuracy. Methods: We addressed the above limitations using data from the recently completed NIH-funded Pulsewatch clinical trial which collected over two weeks of smartwatch PPG data from 106 subjects. Our computationally efficient 1D bi-directional Gated Recurrent Unit deep learning model incorporated multi-modal inputs (1D PPG, accelerometer, and heart rate data) to classify normal sinus rhythm, AF, and PAC/PVC. Results: Our model achieved an unprecedented 83% sensitivity for PAC/PVC detection while maintaining a high accuracy of 97.31% for AF detection, outperforming the best retrained state-of-the-art model by 20.81% and 2.55%, respectively. It was also 14 times more computationally efficient and 2.7 times faster. Testing on two external PPG datasets collected with a different smartwatch and a fingertip PPG sensor, our model demonstrated better generalizability with macro-averaged AUROC values of 96.22% and 94.17%, respectively. Conclusion: A light-weight multimodal input deep learning model can accurately distinguish PAC/PVC from AF, reducing false positive detection of AF. Significance: Accurate AF and PAC/PVC detection with minimal false positive detection can enhance clinical and public acceptance of smartwatch-based AF monitoring.

Most deep learning models of multiclass arrhythmia classification are tested on fingertip photoplethysmographic (PPG) data, which has higher signal-to-noise ratios compared to smartwatch-derived PPG, and the best reported sensitivity value for premature atrial/ventricular contraction (PAC/PVC) detection is only 75%. To improve upon PAC/PVC detection sensitivity while maintaining high AF detection, we use multi-modal data which incorporates 1D PPG, accelerometers, and heart rate data as the inputs to a computationally efficient 1D bi-directional Gated Recurrent Unit (1D-Bi-GRU) model to detect three arrhythmia classes. We used motion-artifact prone smartwatch PPG data from the NIH-funded Pulsewatch clinical trial. Our multimodal model tested on 72 subjects achieved an unprecedented 83% sensitivity for PAC/PVC detection while maintaining a high accuracy of 97.31% for AF detection. These results outperformed the best state-of-the-art model by 20.81% for PAC/PVC and 2.55% for AF detection even while our model was computationally more efficient (14 times lighter and 2.7 faster).

Prolonged exposure to cold air can impair reaction time and cognitive function, which can lead to serious consequences. One mitigation strategy is to develop models that can predict cognitive performance by tracking physiological metrics associated with cold stress. As females are evidenced to be more sensitive to cold exposure, this study investigated the relationship between physiological metrics and cognitive performance deterioration of female subjects under cold stress. Wearable electrodermal activity (EDA) and electrocardiogram (ECG) were collected from nineteen females who underwent five sessions of a cognitive task battery—assessing reaction time, memory, and attention—in a cold (10 °C) environment. Machine learning classifiers showed higher cognitive performance classification accuracies with heart rate variability (HRV) features than with EDA features. Particularly in detecting performance deterioration in a task associated with assessing short-term memory, our support vector machine classifier with HRV features showed an 82.4% accuracy, with a sensitivity of 84.2% and a specificity of 80.6%, whereas a 55.4% accuracy with a sensitivity of 44.7% and a specificity of 66.7% was obtained with EDA features. Our results demonstrate the feasibility of detecting performance deterioration from females who underwent cold exposure using wearable EDA and ECG, allowing for preventive measures to reduce risk in cold environments, especially for female military personnel.

Assessment of the sympathetic nervous system (SNS) is one of the major approaches for studying affective states. Skin nerve activity (SKNA) derived from high-frequency components of electrocardiogram (ECG) signals has been a promising surrogate for assessing the SNS. However, current SKNA analysis tools have shown high variability across study protocols and experiments. Hence, we propose a time-varying spectral approach based on SKNA to assess the SNS with higher sensitivity and reliability. We collected ECG signals at a sampling frequency of 10 KHz from sixteen subjects who underwent various SNS stimulations. Our spectral analysis revealed that frequency bands between 150 - 1,000 Hz showed significant increases in power during SNS stimulations. Using this information, we developed a time-varying index of sympathetic function measurement based on SKNA, termed, Time-Varying Skin Nerve Activity (TVSKNA). TVSKNA is calculated in three steps: time-frequency decomposition, reconstruction using selected frequency bands, and smoothing. TVSKNA indices exhibited generally higher Youden's J, balanced accuracy, and area under the receiver operating characteristic curve, indicating higher sensitivity. The coefficient of variance was lower with TVSKNA indices for most SNS tasks. TVSKNA can serve as a highly sensitive and reliable marker of quantitative assessment of sympathetic function, especially during emotion and stress.

This paper proposes an adjacent channel weight dependable recursive least square adaptive filter (ACWD-RLS) with variational mode decomposition (VMD) based artifact removal mechanism in separating the fetal ECG (FECG) components from the pregnant mother abdominal ECG (AECG). This approach requires the two abdominal, and a single thorax ECG to extract the FECG component present in the AECG. The algorithm uses three independent VMD decomposition algorithms in which one decomposes the thorax ECG while the other two decompose the abdominal ECGs of adjacent channels. The presence of baseline wander (BW) and powerline interference (PLI) is detected and eliminated from the modes obtained from each VMD algorithm. The work also proposes an ACWD-RLS filter that contains two sections of the RLS filter namely the main section and secondary section, where the weight update in the main section depends on the weight estimated in the secondary section. The performance of the VMD-based artifact removal algorithm in suppressing the BW and PLI artifacts was evaluated utilizing the MIT-BIH arrhythmia and MIT-BIH noise stress dataset, while the Synthetic dataset of Physionet and real-world Daisy dataset was utilized in the validation of the proposed ACWD-RLS approach in fetal ECG extraction. The proposed VMD-based BW and PLI artifact removal mechanism result in a correlation coefficient and output signal-to-noise ratio of 0.988 and 14.23 dB respectively with a signal to BW noise ratio of 5 dB. The evaluation results show that the algorithm yields a PDA of 95.54% and 97.96% in real-world Daisy and Synthetic datasets respectively.

Accurate estimation of sleepiness levels is crucial for managing sleep-related health risks and preventing cognitive degradation that can lead to accidents in the workplace. However, using machine learning (ML) to estimate these levels from speech remains challenging, with reported weak correlations (<0.40) with ground-truth sleepiness levels. Developing effective ML models requires high-quality sleepiness data from noticeably sleepy individuals, but collecting such data through prolonged sleep deprivation is both risky and costly. We propose that acted (feigned) sleepy speech can effectively represent realistic sleepiness and be used to train ML models for estimating sleepiness levels and detecting sleepiness-associated cognitive performance degradation. Our study demonstrates that: (1) human listeners perceive acted sleepy speech as sleepier than both non-sleepy and genuinely sleepy speech, and (2) ML models trained on acted speech can accurately estimate sleepiness levels in individuals who have been awake for 25 hours, achieving a correlation of 0.54 with ground-truth sleepiness levels while using fewer samples. Furthermore, the acted sleepy speech-based ML model detects cognitive performance degradation (F1 score = 0.80) in sleep-deprived individuals, outperforming models trained on real sleepy speech (F1 score = 0.32). Our approach provides efficient, effective, and scalable solutions to not only update benchmarks but also enhance the capabilities of real-world AI applications, from voice assistants to AI agents, ultimately supporting human health, workplace safety, and daily tasks.

This study addresses the challenges of arrhythmia detection, including atrial fibrillation (AF), using continuously collected smartwatch photoplethysmography (PPG) data. We propose a novel application of the densely connected convolutional neural network (DenseNet) for AF detection using smartwatch PPG data with effective differentiation from premature atrial contractions (PACs) and premature ventricular contractions (PVCs). While smartwatches offer a convenient platform for continuous monitoring, current research often relies on controlled settings using fingertip PPG and struggles to differentiate AF from other arrhythmias, particularly PAC/PVC $s$ . Unlike prior controlled-setting studies, our data collection includes 14 days of continuous smartwatch PPG data in real-world settings, which truly challenges the performance of arrhythmia classification algorithms. Notably, the proposed DenseNet model excelled at distinguishing AF (sensitivity: 0.950, specificity: 0.949), demonstrating promise for real-world continuous AF monitoring. The proposed DenseNet model also achieved high specificity (0.958) for the classification of PAC/PVCs despite the inherent difficulty in their identification from PPG data. These findings demonstrate the efficacy of DenseNet in accurately identifying AF and distinguishing it from other arrhythmias, highlighting its potential for widespread clinical application.

Continuous monitoring of non-invasive skin sympathetic nerve activity (SKNA) holds promise for understanding the sympathetic nervous system (SNS) dynamics in various physiological and pathological conditions. However, muscle noise artifacts present a challenge in accurate SKNA analysis, particularly in real-life scenarios. This study proposes a deep convolutional neural network (CNN) approach to detect and remove muscle noise from SKNA recordings obtained via ECG electrodes. Twelve healthy participants underwent controlled experimental protocols involving cognitive stress induction and voluntary muscle movements, while collecting SKNA data. Power spectral analysis revealed significant muscle noise interference within the SKNA frequency band (500-1000 Hz). A 2D CNN model was trained on the spectrograms of the data segments to classify them into baseline, stress-induced SKNA, and muscle noise-contaminated periods, achieving an average accuracy of 89.85% across all subjects. Our findings underscore the importance of addressing muscle noise for accurate SKNA monitoring, advancing towards wearable SKNA sensors for real-world applications.

Photoplethysmography (PPG) signals collected by wearables have been shown to be effective in accurate detection of atrial fibrillation (AF), provided that the data are devoid of motion and noise artifacts (MNA). Many studies have been previously conducted to detect AF arrhythmia using PPG data; however, the subjects were mostly in clinics or controlled settings with data collection lasting several minutes to at most several hours with minimal MNA. Our study, Pulsewatch, differs from previous AF studies in that PPG data from smartwatches prescribed to stroke survivors were continuously collected for two weeks in real-life conditions, which invariably included a significant amount of MNA. Our aim is to provide a framework for a novel use of a denoising autoencoder to reconstruct motion-artifact-removed PPG signals so that we can improve the AF detection per-formance and to increase the amount of analyzable data.We used more than 30,000 25-sec PPG segments from 129 subjects randomly selected from Pulsewatch and Stanford University's datasets. The training and testing datasets from these two databases came from smart-watches from different vendors with varying sampling frequencies and time duration of recordings in diverse and realistic settings. In this study, the highly corrupted PPG data were automatically detected and discarded, but those segments contaminated with low-to-moderate motion and noise artifacts (MNA) were subjected to a convolutional denoising autoencoder (CDA). To reconstruct the artifact-removed PPG segments, we proposed to employ two distinct CDA models for AF and non-AF data groups initially classified as AF or non-AF. Using the proposed approach, we significantly improved the performance of detecting occult AF. We achieved classifica-tion accuracy, sensitivity, and specificity of 91.02%, 91.54%, and 90.85%, respectively, for out-of-sample test data from both databases. By sanitizing data from low-to-moderate MNA, we were able to increase the usable data coverage by 21%.

Acute Stress Disorder (ASD) is a precursor disease to Post Traumatic Stress Disorder (PTSD), occurring within the first month after trauma and may serve as a vital timepoint for treating chronic PTSD. Current ASD diagnosis is resource intensive and may miss vulnerable patients in early prescreening steps, particularly female patients who are known to have higher rates of PTSD and a higher conversion rate from ASD to PTSD. We propose a new ASD diagnostic tool based on electrodermal activity (EDA) to measure changes in sympathetic nervous system (SNS) activity post-trauma. Using both traditional phasic decomposition and time-frequency decomposition methods, we are able to detect key differences in ASD and non-ASD groups within 72 hours of trauma. We compared several machine learning models to determine optimal diagnostic capability. Our logistic regression model was able detect ASD in female subjects with a balanced accuracy of 0.72 and an F1 score of 0.78. Our time-frequency decomposition techniques are highly sensitive to SNS changes post trauma. Future study on EDA dynamics in the early stages of trauma may help improve detection of ASD and increase PTSD treatment outcomes.

In recent years, skin sympathetic nerve activity (SKNA) extracted from electrocardiogram has gained attention as a novel noninvasive measure of the sympathetic nervous system (SNS), while electrodermal activity (EDA) has long served this purpose. SparsEDA is a sparse deconvolution technique originally developed for EDA to extract phasic drivers indicating the start of sympathetic burst responses. Our focus is on applying this method to preprocessed SKNA signals, justified by both SKNA and EDA signals' connection to sympathetic nerve activity and prior observed similarities. In a thermal-grill pain experiment, 16 subjects underwent six stimulations each to elicit SNS responses, with simultaneous recording of EDA and SKNA. We confirmed the method's accuracy in identifying stimuli initiation. Results were assessed for burst detection and accuracy of driver placement compared to annotated labels. The SKNA drivers achieved an RMSE of 0.42 from annotated stimulations, a 97% hit rate in detecting applied stimuli, and minimal false alarms (1.40 ± 1.76) during the 2-minute control period and interstimulus intervals.

Underwater recording of biosignals such as electrodermal activity has the potential to improve our current understanding of diving medicine. However, the deployment is currently limited by the required waterproofing of EDA electrodes. In this work, we showcase the development and verify our design of a new water-tolerant electrode for underwater EDA collection. In time, frequency, and time-frequency domains, our electrode design performs as well as or better than the standard Ag/AgCl electrodes, showing only a small baseline shift when submerged and produced signal differences no larger than those EDA collected at different sites on the same body. Further development is required to establish the longevity of our electrode and the usability in varied other real-world water environments.

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.

Pain is typically viewed as an internal perception that is difficult to objectively measure. Current methods attempt to quantify pain intensity using self-reported scores, but these are often subjective or difficult to administer. Three primary sensory nerve fibers carry pain and tactile information: A beta- and Ad-fibers carrying short touch and acute pain, respectively, and C-fibers carrying dull prolonged pain. Electrodermal activity (EDA), previously shown to be highly sensitive to pain and stress responses, may allow us to distinguish these different pain signals to more fully express measured pain. In this work, we use machine learning to classify EDA responses associated with these fibers during sine wave transcutaneous electrical nerve stimulation. Using several derived EDA features, we are able to develop models that obtain high accuracy (>75%) of classification across all three fiber types in this multiclass classification task. The ability to measure fiber-specific activation may help greatly improve pain detection and treatment research.

Accurate assessment of sleepiness is pivotal in managing the fatigue-associated risks stemming from sleep deprivation. Speech signals are easy to obtain, allowing detection of sleepiness anywhere. Previous machine learning (ML) studies using speech have not been successful in achieving reliable estimation of perceived sleepiness levels, which results in inaccurate sleepiness determination. In this paper, we propose that these challenges primarily stem from the inherent complexities of speech signals with inaccurate labels of sleepiness. Because the physical effects of sleepiness become pronounced after prolonged wakefulness, we conducted a 25-h sleep deprivation study. We collected electrodermal activity (EDA) and speech data from 30 subjects during speech production every 2 h over the 25-hour period, along with various sleepiness level labels—their cognitive impairment scores derived from the psychomotor vigilance test, their self-reported sleepiness scores, and the h awake scores. The data analysis compared EDA recorded during speech versus only the speech data and examined which approach provided better sleepiness level estimation and detection using ML. The ML result is that features derived from only EDA during speech production provided the most accurate sleepiness determination. Specifically, EDA ML models trained using the hours awake scores provided the best sleepiness level estimation, with 0.53 correlation, and better detection of sleepiness (which is related to cognitive performance deterioration), with 0.85 accuracy (0.80 sensitivity), when compared to ML features derived from speech, which obtained 0.40 correlation for sleepiness level estimation and 0.69 accuracy (0.59 sensitivity) for sleepiness detection. Moreover, the EDA data collected during speech production offered the best performance for sleepiness detection compared to EDA collected during other activities, such as visual vigilance (0.68 accuracy and 0.65 sensitivity). Given the potential of EDA data during speech production, this work demonstrates the promise of future wearable devices that could collect EDA data from speech activity, along with speech signals, for more advanced and accurate real-world sleepiness detection.

Early diagnosis of Acute Stress Disorder (ASD) is important, given its potential progression to post-traumatic system disorder (PTSD). The current diagnostic tool has some degree of subjectiveness in assessing emotional responses to trauma and the severity of stress reactions. To this end, we proposed a new method to detect ASD using machine learning with resting-state functional magnetic resonance imaging (rs-fMRI) data. We used 48 subjects of rs-fMRI data and PTSD Check List - Civilian Version (PCL-C) questionnaire from Advancing Understanding of RecOvery afteR traumA (AURORA) dataset. We extracted five frequency-domain features from each blood-oxygen-level dependent (BOLD) signal from 48 cortical and 21 subcortical regions. We also extracted four graph features from sparse inverse covariance matrices of the BOLD signals. Eighteen features appeared to be significantly different (p<0.05). Using these features, multi-layer perceptron showed accuracy 91.7%, sensitivity 96.8%, and specificity 82.4% using the leave-one-subject-out cross validation scheme. We found that the Right Accumbens and Lingual Gyrus has high effect size and substantial impact within the machine learning model.

We examined data from Naval Sea Systems Command grant project N0463A-12-C-001, “Hypercapnia: cognitive effects and monitoring”, with the objective of validating or repudiating heart rate variability (HRV) as a warning sign of cognitive impairment from diving gas narcosis or oxygen toxicity. We compared HRV feature scores to their temporally corresponding cognitive outcomes under normal and narcotizing conditions to identify specific HRV features associated with cognitive changes. N0463A-12-C-001 was conducted between 17 September 2013 and 29 January 2016 and employed NASA’s multi-attribute task battery (MATB-II) flight simulator to examine the independent effects of CO2, N2, and O2 partial pressure on diver performance at simulated depths up to 61 msw (200 fsw). We assessed the association of 23 distinct HRV features scores from 432 of the study’s analyzable exposure stages in relation to MATB-II’s four performance subclasses (motor, memory, attention, strategy) while controlling for exercise and CO2, N2, and O2 gas partial pressure. Performance decrements were associated with normalized high-frequency HRVfeatures (HFnu, p = 0.0016) and the number of pairs of successive R-R intervals that differed by more than 50 ms (NN50count1, p = 0.04). Secondary analysis with stratification restricted to non-exercise stages showed that several HRV parameters, including root mean square of the successive difference (RMSSD, p = 0.0015), width of Poincaré plot (p = 0.0017), NN50count1 (p = 0.0019), and standard deviation of normal-to-normal R peaks (p = 0.0082), were associated with performance impairment. The RMSSD association retained statistical significance after Bonferroni correction for multiple tests. HRV features collected from divers tested under narcotizing conditions of breathing gas partial pressure and exercise were associated with performance impairment.

BackgroundOxygen-rich breathing mixtures up to 100% are used in some underwater diving operations for several reasons. Breathing elevated oxygen partial pressures (PO2) increases the risk of developing central nervous system oxygen toxicity (CNS-OT) which could impair performance or result in a seizure and subsequent drowning. We aimed to study the dynamics of the electrodermal activity (EDA) and heart rate (HR) while breathing elevated PO2 in the hyperbaric environment (HBO2) as a possible means to predict impending CNS-OT.MethodsEDA is recorded during 50 subject exposures (26 subjects) to evaluate CNS-OT in immersed (head out of water) exercising divers in a hyperbaric chamber breathing 100% O2 at 35 feet of seawater (FSW), (PO2 = 2.06 ATA) for up to 120 min.Results32 subject exposures exhibit symptoms "definitely" or "probably" due to CNS-OT before the end of the exposure, whereas 18 do not. We obtain traditional and time-varying spectral indices (TVSymp) of EDA to determine its utility as predictive physio markers. Variations in EDA and heart rate (HR) for the last 5 min of the experiment are compared to baseline values prior to breathing O2. In the subset of experiments where "definite" CNS-OT symptoms developed, we find a significant elevation in the mean +/- standard deviation TVSymp value 57 +/- 79 s and median of 10 s, prior to symptoms.ConclusionsIn this retrospective analysis, TVSymp may have predictive value for CNS-OT with high sensitivity (1.0) but lower specificity (0.48). Additional work is being undertaken to improve the detection algorithm.

We propose a state-of-the-art deep learning approach for accurate electrocardiogram (ECG) signal analysis, addressing both waveform delineation and beat type classification tasks. For beat type classification, we integrated two novel schemes into the deep learning model, significantly enhancing its performance. The first scheme is an adaptive beat segmentation method that determines the optimal duration for each heartbeat based on RR-intervals, mitigating segmenting errors from conventional fixed-period segmentation. The second scheme incorporates relative heart rate information of the target beat compared to neighboring beats, improving the model’s ability to accurately detect premature atrial contractions (PACs) that are easily confused with normal beats due to similar morphology. Extensive evaluations on the PhysioNet QT Database, MIT-BIH Arrhythmia Database, and real-world wearable device data demonstrated the proposed approach’s superior capabilities over existing methods in both tasks. The proposed approach achieved sensitivities of 99.81% for normal beats, 99.08% for premature ventricular contractions, and 97.83% for PACs in beat type classification. For waveform delineation, we achieved F1-scores of 0.9842 for non-waveform, 0.9798 for P-waves, 0.9749 for QRS complexes, and 0.9848 for T-waves. It significantly outperforms existing methods in PAC detection while maintaining high performance across both tasks. The integration of aforementioned two schemes into the deep learning model improved the accuracy of normal sinus rhythms and arrhythmia detection.

Sleepiness results in an increased susceptibility to workplace accidents due to reduced cognitive functioning. Early detection of signs of sleepiness is crucial for mitigating work-related risks. Establishing a comprehensive database of sleepiness data is fundamental for developing effective sleepiness identification models. However, obtaining data on realistic sleepiness proves challenging in practice. This study explores a more efficient approach to sleepiness simulation to address this challenge. In this paper, we hypothesize that actively inducing feelings of sleepiness can precipitate sleepiness. In our experiment, ten participants simulated sleepiness while listening to a speech delivered by a tired, sleep-deprived individual. We measured self-reported levels of sleepiness and two widely-used physiological biomarkers-electrodermal activity (EDA) and electrocardiogram (ECG) signals-during this induced state. Our results shed light on the potential of this sleepiness simulation approach, unveiling physiological markers linked to induced sleepiness. By introducing this innovative sleepiness simulation method, our research advances building large sleepy datasets, which contribute to the proactive identification of sleepiness-related risks in the workplace.