Ventilatory Burden as a Measure of Obstructive Sleep Apnea Severity Is Predictive of Cardiovascular and All-Cause Mortality.

Rationale: The apnea-hypopnea index (AHI), used for the diagnosis of obstructive sleep apnea, captures only the frequency of respiratory events and has demonstrable limitations. Objectives: We propose a novel automated measure, termed "ventilatory burden" (VB), that represents the proportion of overnight breaths with less than 50% normalized amplitude, and we show its ability to overcome limitations of AHI. Methods: Data from two epidemiological cohorts (EPISONO [Sao Paolo Epidemiological Study] and SHHS [Sleep Heart Health Study]) and two retrospective clinical cohorts (DAYFUN; New York University Center for Brain Health) were used in this study to 1) derive the normative range of VB, 2) assess the relationship between degree of upper airway obstruction and VB, and 3) assess the relationship between VB and all-cause and cardiovascular disease (CVD) mortality with and without hypoxic burden that was derived using an in-house automated algorithm. Measurements and Main Results: The 95th percentiles of VB in asymptomatic healthy subjects across the EPISONO and the DAYFUN cohorts were 25.2% and 26.7%, respectively (median [interquartile range], VBEPISONO, 5.5 [3.5-9.7]%; VBDAYFUN, 9.8 [6.4-15.6]%). VB was associated with the degree of upper airway obstruction in a dose-response manner (VBuntreated, 31.6 [27.1]%; VBtreated, 7.2 [4.7]%; VBsuboptimally treated, 17.6 [18.7]%; VBoff-treatment, 41.6 [18.1]%) and exhibited low night-to-night variability (intraclass correlation coefficient [2,1], 0.89). VB was predictive of all-cause and CVD mortality in the SHHS cohort before and after adjusting for covariates including hypoxic burden. Although AHI was predictive of all-cause mortality, it was not associated with CVD mortality in the SHHS cohort. Conclusions: Automated VB can effectively assess obstructive sleep apnea severity, is predictive of all-cause and CVD mortality, and may be a viable alternative to the AHI.

A foundational transformer leveraging full night, multichannel sleep study data accurately classifies sleep stages.

Study Objectives: To investigate whether a foundational transformer model using 8-hour, multichannel data from polysomnograms can outperform existing artificial intelligence (AI) methods for sleep stage classification. Methods: We utilized the Sleep Heart Health Study (SHHS) visits 1 and 2 for training and validation and the Multi-Ethnic Study of Atherosclerosis (MESA) for testing of our model. We trained a self-supervised foundational transformer (called PFTSleep) that encodes 8-hour long sleep studies at 125 Hz with 7 signals including brain, movement, cardiac, oxygen, and respiratory channels. These encodings are used as input for training of an additional model to classify sleep stages, without adjusting the weights of the foundational transformer. We compared our results to existing AI methods that did not utilize 8-hour data or the full set of signals but did report evaluation metrics for the SHHS dataset. Results: We trained and validated a model with 8,444 sleep studies with 7 signals including brain, movement, cardiac, oxygen, and respiratory channels and tested on an additional 2,055 studies. In total, we trained and tested 587,944 hours of sleep study signal data. Area under the precision recall curve (AUPRC) scores were 0.82, 0.40, 0.53, 0.75, and 0.82 and area under the receiving operating characteristics curve (AUROC) scores were 0.99, 0.95, 0.96, 0.98, and 0.99 for wake, N1, N2, N3, and REM, respectively, on the SHHS validation set. For MESA, the AUPRC scores were 0.56, 0.16, 0.40, 0.45, and 0.65 and AUROC scores were 0.94, 0.77, 0.87, 0.91, and 0.96, respectively. Our model was compared to the longest context window state-of-the-art model and showed increases in macro evaluation scores, notably sensitivity (3.7% increase) and multi-class REM (3.39% increase) and wake (0.97% increase) F1 scores. Conclusions: Utilizing full night, multi-channel PSG data encodings derived from a foundational transformer improve sleep stage classification over existing methods.

SleepSIM: Conditional GAN-based non-REM sleep EEG Signal Generator.

Synthetic data generation has become increasingly popular with the increasing use of generative networks. Recently, Generative Adversarial Network (GAN) architectures have produced exceptional results in synthetic image generation. However, time series generation still needs to be studied. This paper proposes a Conditional GAN-based system to generate unique samples of non-REM sleep electroencephalographic (EEG) signals. The CGAN model had a 1-D Convolution Neural Network based architecture. The model was trained using real EEG from healthy controls. The trained model can generate an artificial 30-second epoch of non-REM sleep whose power spectrum is identical to that of a real sleep EEG.Clinical relevance- Sleep EEG simulation can be used to train and enhance the skillset of fellows and technicians in the sleep medicine field. Variations in EEG signals can be highly complex to model mathematically; however, here, we harness the power of deep learning, using generative models such as CGANs to train, model complex data distributions, and generate diverse and artificial but realistic EEG signals during non-REM sleep.

Detection and Characterization of Walking Bouts Using a Single Wrist-Worn Accelerometer in Free-living Conditions.

Detection and characterization of abnormalities of movement are important to develop a method for detecting early signs of Parkinson's disease (PD). Most of the current research in detection of characteristic reduction of movements due to PD, known as parkinsonism, requires using a set of invasive sensors in a clinical or controlled environment. Actigraphy has been widely used in medical research as a non-invasive data acquisition method in free-living conditions for long periods of time. The proposed algorithm uses triaxial accelerometer data obtained through actigraphy to detect walking bouts at least 10 seconds long and characterize them using cadence and arm swing. Accurate detection of walking periods is the first step toward the characterization of movement based on gait abnormalities. The algorithm was based on a Walking Score (WS) derived using the value of the auto-correlation function (ACF) for the Resultant acceleration vector. The algorithm achieved a precision of 0.90, recall of 0.77, and F1 score of 0.83 compared to the expert scoring for walking bout detection. We additionally described a method to measure arm swing amplitude.

LSTM based GAN Networks for Enhancing Ternary Task Classification Using fNIRS Data.

Brain activation patterns vary according to the tasks performed by the subject. Neuroimaging techniques can be used to map the functioning of the cortex to capture brain activation patterns. Functional near-infrared spectroscopy (fNIRS) is a neuroimaging technique increasingly used for task classification based on brain activation patterns. fNIRS can be widely used in population studies due to the technology's economic,non-invasive, and portable nature. The multidimensional and complex nature of fNIRS data makes it ideal for deep learning algorithms for classification. Most deep learning algorithms need a large amount of data to be appropriately trained. Generative networks can be used in such cases where a substantial amount of data is required. Still, the collection is complex due to various constraints. Conditional Generative Adversarial Networks (CGAN) can generate artificial samples of a specific category to improve the deep learning classifier's accuracy when the sample size is insufficient. The proposed system uses an LSTM based CGAN with an LSTM classifier to enhance the accuracy through data augmentation. The system can determine whether the subject's task is a Left Finger Tap, Right Finger Tap, or Foot Tap based on the fNIRS data patterns. The authors obtained a task classification accuracy of 90.2% for the LSTM based GAN combination.Clinical relevance- Acquiring medical data present practical difficulties due to time, money, labor, and economic cost. The deep learning-based model can better perform medical image classification than hand-crafted features when dealing with many data. GAN-based networks can be valuable in the medical field where collecting extensive data is not feasible. GAN-generated synthetic data can be used to improve the classification accuracy of classification systems.

Conditional-GAN Based Data Augmentation for Deep Learning Task Classifier Improvement Using fNIRS Data.

Functional near-infrared spectroscopy (fNIRS) is a neuroimaging technique used for mapping the functioning human cortex. fNIRS can be widely used in population studies due to the technology's economic, non-invasive, and portable nature. fNIRS can be used for task classification, a crucial part of functioning with Brain-Computer Interfaces (BCIs). fNIRS data are multidimensional and complex, making them ideal for deep learning algorithms for classification. Deep Learning classifiers typically need a large amount of data to be appropriately trained without over-fitting. Generative networks can be used in such cases where a substantial amount of data is required. Still, the collection is complex due to various constraints. Conditional Generative Adversarial Networks (CGAN) can generate artificial samples of a specific category to improve the accuracy of the deep learning classifier when the sample size is insufficient. The proposed system uses a CGAN with a CNN classifier to enhance the accuracy through data augmentation. The system can determine whether the subject's task is a Left Finger Tap, Right Finger Tap, or Foot Tap based on the fNIRS data patterns. The authors obtained a task classification accuracy of 96.67% for the CGAN-CNN combination.

Bi-Directional Gated Recurrent Unit Based Ensemble Model for the Early Detection of Sepsis.

Early prediction of sepsis is essential to give the patient timely treatment since each hour of delayed treatment has been associated with an increase in mortality. Current sepsis detection systems rely on empirical Clinical Decision Rules(CDR)s, which are based on vital signs that can be collected from the bedside. The main disadvantages of CDRs include questions of generalizability and performance variance when applied to the populations different from the groups used for derivation and often take years to develop and validate. This paper proposes a deep learning model using Bi-Directional Gated Recurrent Units(GRU), which uses a wide range of parameters that are associated with vitals, laboratory, and demographics of patients. The proposed model has an area under the receiver operating characteristic (AUROC) of 0.97, outperforming all the existing systems in the current literature. The model can handle the missing data, and irregular sampling intervals frequently present in medical records.Clinical relevance-The proposed model can be used to predict the onset of sepsis 6 hours ahead of time by the use of a machine learning algorithm. This proposed method outperforms the sepsis prediction machine learning models found in the current literature.

Automatic Detection of Respiratory Effort Related Arousals With Deep Neural Networks From Polysomnographic Recordings.

Sleep disorders have become more common due to the modern lifestyle and stress. The most severe case of sleep disorders called apnea is characterized by a complete breaking block, leading to awakening and subsequent sleep disturbances. The automatic detection of sleep arousals is still challenging. In this paper, a novel method is presented to detect non-apnea sources of arousals during sleep using Polysomnography(PSG) recordings. After the preprocessing, a sequence-to-sequence deep neural network (DNNs) consisting of a series of Bidirectional long short-term memory (Bi-LSTM) layer, and fully connected layers were trained to classify samples in the segments. Initially, three different models were prepared for different datasets. Finally, obtaining the classification result through an ensemble model consisting of the three trained models. The result shows that the area under the receiver precision-recall curve (AUPRC) is 0.59 for the test dataset exceeding the performance of the classifiers in the existing literature.Clinical relevance- Analyzing Polysomnographic recordings is a time consuming a critical process yet to identify sleep disorders. These recordings span several hours and contain different data streams that include EEG, EMG, etc. This paper proposes a system that can automatically detect respiratory effort-related arousals using a deep neural network from Polysomnographic Recordings. By automating this process with a machine learning-based solution that can eliminate the manual process while facilitating further improvements of the system with future data.