Rapid eye movement sleep disorder detection to facilitate selective screening of Parkinson's disease

Abstract

Detection of fast eye movement (REM) sleep and associated abnormalities is important for identifying various nervous system diseases. The present disclosure relates to the detection of time intervals during which a subject is in REM sleep by utilizing electroencephalogram (EEG) data from selective spectral frequencies. The techniques as disclosed herein may use one or more EEG electrodes, and more particularly, single channel EEG signals provide a simple and cost-effective solution. The disclosed techniques may pre-process the EEG data, extract features and / or derived features for one or more time intervals from selective spectral bands (e.g., delta and gamma). These extracted features may be normalized and clustered to further determine REM, non-REM, and waking time intervals. By using one or more non-EEG sensors to detect the presence of muscular tension of the identified REM sleep intervals, and by performing sleep pattern analysis, potential REM intervals can be verified to detect REM behavioral disorders.

Description

Cross-reference to related applications

[0001] This application claims priority and benefit to U.S. Provisional Application No. 63 / 505,339, filed May 31, 2023, the entire contents of which are incorporated herein by reference for all purposes. Background Technology

[0002] Rapid eye movement (REM) sleep behavior disorder (RBD) is a sleep disorder characterized by a lack of muscle paralysis during REM sleep, leading to the reenactment of dream content through vocalization or body movement. RBD can be identified as a potential prodromal symptom of Parkinson's disease (PD) and other neurodegenerative diseases, such as multiple system atrophy (MSA) and Lewy body dementia (LBD). Parkinson's disease is a progressive neurological disorder primarily associated with motor symptoms such as tremor, rigidity, and bradykinesia. However, non-motor symptoms, including sleep disturbances, often precede the onset of motor symptoms by a considerable period (e.g., spanning several years). New observations suggest a potential link between RBD and the development of Parkinson's disease. This association underscores the importance of early detection and intervention for RBD as a proactive strategy for identifying high-risk individuals for neurodegenerative diseases such as Parkinson's disease, thereby facilitating timely clinical management.

[0003] Routine methods for diagnosing RBD can rely on clinical assessment, questionnaires, and polysomnography (PSG), which involves monitoring physiological signals, including electroencephalography (EEG). For example, EEG electrodes can capture EEG signals, including neural signals, which can then be processed to detect different types of brain waves that indicate different non-REM sleep stages. Routine methods involving only EEG may be insufficient for detecting RBD because brain activity patterns during REM sleep can be similar to those observed when awake, making it difficult to distinguish between the two states. Furthermore, RBD episodes can occur during REM sleep but may not always present with obvious EEG abnormalities, further complicating detection through EEG analysis alone. Incorporating the reasons and actions of the subject's speech (e.g., capturing visual cues and behavior through video recordings, audio, or speech analysis) can provide additional context and insights, particularly for sleep disorders such as RBD.

[0004] However, incorporating video and / or audio data into sleep monitoring can be costly, time-consuming, and inconvenient, posing a barrier to accessibility for participants. The acquisition and maintenance of specialized equipment, as well as technical expertise in data acquisition and analysis, can increase the overall cost and complexity of implementing a comprehensive sleep monitoring solution. Furthermore, the accessibility of sleep monitoring facilities equipped with these advanced technologies may be limited in certain geographic areas or healthcare settings, leading to disparities in healthcare services and diagnostic capabilities. Addressing these challenges may require the development of more affordable and widely used sleep monitoring technologies that can provide reliable insights into sleep disorders while minimizing the burden on healthcare institutions. Summary of the Invention

[0005] Certain aspects and features of this disclosure relate to the identification of various sleep stages, including rapid eye movement (REM) sleep, via one or more electroencephalogram (EEG) signals. This disclosure may also identify REM behavioral disorders (RBDs) by utilizing one or more non-EEG sensors to detect muscle tone for the identified REM sleep and by performing sleep analysis. One or more first signals from one or more EEG electrodes (e.g., at least one active electrode, a reference electrode, and a potential ground electrode) may be received over a period of time. During this period, at least one of these EEG electrodes may be placed on the subject. In a preferred embodiment, the one or more first signals comprise single-channel EEG data. One or more second signals from non-EEG electrodes (e.g., electromyography (EMG) sensors or accelerometers) may also be received, wherein at least one of these non-EEG electrodes may be located on, near, or inside the subject during the period of time. This disclosure may also preprocess these one or more first signals from the EEG data (e.g., manually or automatically, or in combination thereof, removing artifacts, normalizing one or more time intervals for power differences across time to determine appropriate frequency bands suitable for feature extraction). Features and / or derived features for one or more time intervals can be extracted from selective spectral bands (e.g., delta and gamma). These extracted features can be further normalized (e.g., z-scores or quantile transformations to bring different features or derived features to the same scale) and clustered to determine time intervals corresponding to REM, wakefulness, and non-REM (e.g., any of sleep stages 1 to 4). REM sleep can be characterized by vivid dreams and high brain activity similar to wakefulness. Therefore, by utilizing one or more non-EEG sensors to detect muscle tone—that is, detecting whether the subject's muscles during the REM time window are in a relaxed state depicting normal REM sleep with muscle dyspnea or have muscle tone (showing muscle tension indicating preparation for movement or posture)—predictive subsets of multiple time intervals corresponding to potential REM time intervals can be further validated to detect REM behavioral disorders (RBD).

[0006] Additionally, sleep analysis can be performed via one or more modeling techniques to validate subsets of multiple time intervals corresponding to REM stages of sleep. One or more modeling techniques can be trained to learn the temporal structure of different sleep stages within a given time period. In some aspects, one or more modeling techniques can be configured to identify smooth transitions between different sleep stages by analyzing one or more adjacent time intervals within each time interval of a subset of multiple time intervals corresponding to REM stages of sleep. For example, if a specific time interval from a subset of multiple time intervals is surrounded by a "stage 2" window, one or more modeling techniques can be configured to predict that the specific window (or time interval) corresponds to REM sleep. Furthermore, one or more modeling techniques can also be configured to identify a gradual increase in consecutive REM stages of sleep from subsets of multiple time intervals corresponding to REM stages of sleep over a period of time.

[0007] Based on predictions from non-EEG sensors, the presence of muscle tone is depicted during predicted subsets of multiple time intervals corresponding to REM states, and sleep analysis is performed by one or more modeling techniques, which can predict whether a subject will be diagnosed with RBD. In response to this prediction, an output indicating the predicted outcome of RBD can be displayed. For example, if a subject is diagnosed with RBD, one or more actions can be triggered. These actions may include, but are not limited to, issuing alarms, notifying relevant authorized parties (e.g., healthcare professionals, relatives, and / or the subject) to conduct a comprehensive medical evaluation to confirm a diagnosis of neurodegenerative diseases such as Parkinson's disease, and arranging a consultation with a sleep specialist for a thorough neurological examination.

[0008] EEG signals can be characterized by different frequency bands, each associated with a specific cognitive and physiological state. For example, stage 3 of non-REM sleep can be associated with the delta band, characterized by slow waves (or frequencies) with high amplitude. Similarly, gamma can be characterized by high frequencies in EEG signals with low amplitude, associated with higher information processing and perception, such as REM sleep. The extracted feature set and / or derived features associated with one or more of these frequency bands can include, for example, delta power, gamma power, standard deviation, maximum amplitude, gamma power / delta power, time derivative of delta, and time derivative of gamma power / delta power.

[0009] One or more modeling techniques can model continuous data and learn the probabilities of transitions between sleep stages. In some cases, one or more modeling techniques include Hidden Markov Models (HMMs) and / or Recurrent Neural Networks (RNNs).

[0010] In some cases, during or before preprocessing, one or more first signals received from EEG electrodes and one or more second signals received from non-EEG sensors over a period of time can be divided into multiple time intervals.

[0011] In some embodiments, a system is provided that includes one or more data processors and a nontransitory computer-readable storage medium containing instructions that, when executed on the one or more data processors, cause the one or more data processors to perform part or all of the methods disclosed herein.

[0012] In some embodiments, a computer program product tangibly embodied in a non-transitory machine-readable storage medium includes instructions configured to cause one or more data processors to perform some or all of the methods disclosed herein.

[0013] In some embodiments, a system is provided that includes one or more means for performing part or all of the one or more methods or processes disclosed herein.

[0014] The terms and expressions used are descriptive and not limiting. Their use is not intended to exclude any equivalents of the shown and described features or portions thereof, but it should be recognized that various modifications can be made within the scope of the claimed invention. Therefore, it should be understood that although the claimed invention has been specifically disclosed through embodiments and optional features, modifications and variations of the concepts disclosed herein can be made by those skilled in the art, and such modifications and variations are considered to be within the scope of the invention as defined by the appended claims. Attached Figure Description

[0015] Various embodiments are described below with reference to the accompanying drawings. It should be noted that the drawings are not drawn to scale, and elements with similar structures or functions are represented by similar reference numerals throughout the drawings. It should also be noted that the drawings are intended only to facilitate the description of embodiments. They are not intended to be an exhaustive description of this disclosure, nor should they be construed as limiting the scope of this disclosure.

[0016] Figure 1 An exemplary block diagram is shown illustrating the performance of feature extraction for determining one or more sleep stages of a subject from one or more electroencephalogram (EEG) channels.

[0017] Figure 2 An exemplary overview is shown for determining one or more sleep stages of a subject from one or more EEG channels.

[0018] Figure 3Example EEG waveforms of subjects recorded during rapid eye movement (REM) and awake states are shown for different time windows.

[0019] Figure 4 An example overview of how REM behavioral disorders (RBD) are determined by utilizing one or more non-EEG sensors is shown.

[0020] Figure 5 An example description of a subject's normal sleep cycle is shown.

[0021] Figure 6 An example architecture for predicting RBD is shown, based on some aspects of this disclosure.

[0022] Figure 7 An example process flow is shown for determining RBD by utilizing one or more EEG channels, additional non-EEG sensors, and sleep analysis.

[0023] Figure 8 These are example illustrations of computer systems in which various embodiments of the present disclosure may be implemented. Detailed Implementation

[0024] Detecting rapid eye movement (REM) sleep and associated abnormalities is crucial for identifying various neurological and psychiatric disorders. Examples include narcolepsy characterized by abnormal REM patterns, REM behavioral disorder (RBD) characterized by loss of muscle tone during REM sleep, Parkinson's disease characterized by RBD, Lewy body dementia (LBD) associated with RBD, and Alzheimer's disease, depression, and mood disorders associated with non-REM sleep disorders. This invention relates to detecting specific time intervals of REM sleep in subjects using electroencephalogram (EEG) data from selective spectral frequencies. The disclosed technique allows for the preprocessing of this EEG data to extract features and / or derived features from one or more time intervals from selective frequency bands (e.g., delta and gamma). These extracted features can be further normalized to map and cluster on the same scale to further identify various sleep stages, including REM, non-REM, and wakefulness time intervals. By using one or more non-EEG sensors to detect the presence of muscle tone in subjects within identified REM sleep intervals, and by performing sleep analysis, it is possible to further verify whether potential REM intervals can be used to detect REM behavioral disorders (RBD).

[0025] The techniques disclosed herein, using one or more EEG electrodes, and more specifically, single-channel EEG signals (i.e., at least one active electrode, a reference electrode, and a potential ground electrode), provide a simple and cost-effective solution. Detecting various stages of sleep (including REM) using single-channel EEG signals can be challenging compared to conventional multi-channel EEG setups. This is because single-channel EEG signals may have limited spatial resolution, limited coverage of the EEG signal, and sleep stages can be characterized by complex and dispersed patterns of neural activity across different brain regions, lacking generalization ability—that is, the variability in sleep patterns (and the resulting EEG signals) between individuals.

[0026] In some embodiments, this technique can capture brain activity by strategically placing a pair of EEG electrodes on the scalp to identify various time intervals indicative of REM sleep. When a subject (e.g., an animal or human) falls asleep, sleep transitions into an intermediate sleep stage and can enter a sleep state called REM, characterized by high-frequency, low-power EEG activity. EEG signals can follow a 1 / f distribution, meaning that high-frequency signals in the EEG tend to have smaller amplitudes and therefore lower spectral power, and vice versa. The EEG data can be preprocessed (or pre-processed) to remove artifacts, which may be caused by, for example, muscle activity, motion, or high-frequency noise. These artifacts can be removed manually or automatically using various techniques, including filtering, independent component analysis (ICA), principal component analysis (PCA), or template matching. EEG signals are typically examined in time intervals of similar or different lengths; that is, when the EEG signal is used to analyze sleep, it can be segmented (before or after artifact removal) into multiple time intervals (or windows), where each time interval represents a portion of the data in the EEG sequence. For further analysis, scanning windows and sliding windows can be used to separate EEG data into time series increments.

[0027] For each time window, the differential EEG signal can be normalized for one or more time windows to account for power differences across time, in order to determine appropriate frequency band feature extraction. This normalization can reveal low-power, statistically significant power variations at different frequency bands (e.g., delta and gamma). It should be understood that any frequency band (or range) can be utilized. In some cases, delta and gamma bands are used to extract features such as delta power, gamma power, standard deviation, maximum amplitude, and / or derived features such as delta power / gamma power, time derivative of delta, time derivative of gamma power / delta power, etc., as robust features for extracting REM intervals. These derived features can be clustered based on common spectral features to group similar sleep time intervals. To perform clustering efficiently, features (or derived features) including, for example, two or more of these features can be normalized again to a comparable scale. After clustering, each cluster can represent a similar set of spectral features, and sleep stages can be assigned to multiple time intervals based on (predefined or dynamic) thresholds. By analyzing the distribution of normalized derived features within each cluster, thresholds can be defined to separate different clusters. For example, if the power exceeds a certain threshold, the window can be assigned as a "potential REM" window. Other time windows can be assigned to non-REM states, i.e., "Phase 1", "Phase 2", "Phase 3", or "Phase 4" or awake states.

[0028] More information on the processing of EEG data to facilitate REM detection can be found in U.S. Application No. 13 / 129,185, filed November 16, 2009, the entire contents of which are incorporated herein by reference for all purposes.

[0029] In one aspect of this disclosure, one or more non-EEG sensors can be used to further validate potential REM sleep zones, thereby avoiding false positives. Incorporating such multimodal data—that is, non-EEG sensors (e.g., electromyography (EMG) and / or accelerometers), alone or in combination—can utilize different physiological characteristics associated with that sleep stage to provide an improved assessment of REM sleep detection. During normal REM sleep, a natural decrease in muscle tone occurs, a state known as hypotonia, in which muscles relax and show minimal electrical activity when detected by sensors (e.g., EMG sensors). This muscle tone can represent a baseline level of muscle tension or stiffness at rest, contributing to maintaining posture, joint stability, and motor readiness, thus making it easier for limbs or the body to move quickly. In RBD, hypotonia is partially or completely absent, resulting in increased muscle tone, or even manifesting as body movement rather than a relaxed muscle state. One or more non-EEG sensors can be optionally or additionally attached to a wearable device (e.g., a hat or headband) or integrated within the wearable device. In some cases, a single device (e.g., an adhesive patch, cap, headband, etc.) can be configured to include or receive (e.g., mechanically connected) one or more non-EEG sensors. In other cases, one or more EMG sensors are used to detect above-normal muscle tone, which can characterize the continuous and passive contractions of muscles during REM sleep. Measuring muscle tone during REM sleep using one or more EMG sensors can serve as an indicator of synucleinopathy or its early stages.

[0030] In some other cases, one or more non-EEG sensors, including accelerometers, can be used to capture body movements during REM sleep. Data from these one or more non-EEG sensors can be extracted for the same time interval as the potential REM time interval. This extracted (or segmented) data can be preprocessed, i.e., normalized (for consistency and noise reduction), and feature extraction, such as calculating power, RMS value, or entropy, can be performed. Data from one or more non-EEG sensors (e.g., EMG sensors and / or accelerometers) (preprocessed, partially preprocessed, or raw) can be combined with the potential REM time interval. Data from these sensors can be fed into one or more predictive models, such as machine learning models, configured to detect aberrant body movements during REM sleep (i.e., kicking, punching, fidgeting, talking or shouting, violent twisting, jumping out of bed, etc.) and / or aberrant muscle tone (i.e., hypertonia, the opposite of normal REM sleep, where muscles relax due to atony), thereby reliably detecting RBD. For example, sensor data from one or more accelerometers and one or more EMG sensors can be aggregated within each mode and fed into separate predictive models trained to detect the corresponding motion.

[0031] During REM sleep, the use of non-EEG sensors (e.g., accelerometers) can trigger false alarms due to irrelevant movements, such as those of a bed partner or muscle twitches caused by external factors. EMG sensors are typically mounted on muscles to capture muscle tone; therefore, movements from a partner are unlikely to generate significant tension in the muscles. Furthermore, sensitivity thresholds can be adjusted to filter out subtle muscle tension or movements in order to distinguish between real signals and noise. Additionally, sensor fusion, i.e., combining EMG and accelerometer (or electrocardiogram (ECG)) data, can provide cross-validation, reducing false alarms by cross-referencing data from different sensors.

[0032] Performing sleep cycle analysis can involve identifying abnormalities during REM sleep and specific patterns indicating disease, and may also help detect RBD. For example, given long sleep durations, such as approximately 8 hours of nighttime sleep, REM windows tend to be prolonged and occur more frequently in the latter half of total sleep. In the first half of sleep, more time can be spent in stage 3 or stage 4 (deep sleep), and less time is spent in REM sleep. Additionally, given that humans and animals typically transition smoothly between different sleep stages, it is possible to predict whether, for each “potential REM” window, the window corresponds to the subject being awake or in REM sleep. For example, if a “potential REM” window (or a consecutive set of “potential REM” windows) is surrounded by a “stage 2” window, one or more modeling techniques can be configured to predict that the window (or consecutive set of windows) corresponds to REM sleep. Similarly, if a “potential REM” window (or a consecutive set of “potential REM” windows) is surrounded by a “stage 1” window, one or more modeling techniques can be configured to predict that the window corresponds to wakefulness.

[0033] To perform sleep analysis, one or more modeling techniques can be trained to learn the sequence structure encompassing various sleep stages throughout the entire sleep process. Modeling techniques such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), or hidden Markov models (HMMs) can be used for sleep pattern analysis and temporal analysis. These techniques can model sequential data to capture temporal context by automatically estimating relationships between consecutive time intervals. This modeling technique can learn temporal structure and account for the probability of transitions between sleep stages. By training one or more models on labeled sleep data segmented into aggregated time intervals or overlapping windows of time intervals (which may or may not include EEG data and / or data from one or more non-EEG sensors), statistical relationships and probabilities of transitions between different sleep states based on observed sequences can be learned, and this information can be used to correct inconsistencies. It should be understood that a unified model or different models can be trained to perform various aspects of sleep analysis. Once a model is trained, it can be applied to new input segmented sleep data to validate the sequence of sleep stages. The trained model can look for anomalous patterns and transitions between sleep stages. In RBD, transitions into or out of REM may occur more frequently or in unusual ways, such as frequent interruptions of REM by wakefulness or other phases. Based on observed patterns, such networks can be trained to validate each interval labeled as a potential REM, wakefulness, or non-REM state.

[0034] Predictions from one or more modeling techniques capable of performing sleep analysis and predictive models capable of detecting muscle tone during the REM window can be combined to obtain a final prediction. This combination can be performed using various techniques, including simple averaging, weighted averaging, majority voting, or other similar techniques. After identifying the RBD based on the prediction, one or more actions can be performed. Such predictions can be integrated into a system that performs one or more actions related to both clinical interventions and safety measures to improve the subject's health and reduce the risk of injury. These actions may include, but are not limited to, issuing alerts, notifying relevant authorized parties (such as healthcare professionals, relatives, and / or the subject) to conduct a full medical evaluation to confirm a diagnosis for a neurodegenerative disease such as Parkinson's disease, and arranging a consultation with a sleep specialist for a thorough neurological examination.

[0035] Figure 1 An exemplary block diagram is shown illustrating the performance of feature extraction for determining a subject's sleep stages from one or more electroencephalogram (EEG) channels. For example, EEG data 102 may be received from a single channel or multiple channels. This EEG data 102 may optionally be processed to remove artifacts, where artifacts refer to any portion of the EEG data that distorts the data to be received (e.g., motion data in the EEG signal). These artifacts may occur, for example, due to high-frequency noise from muscle activity (such as clenching teeth or head movements), periodic interference from the electrical activity of the heart, or other environmental artifacts (such as electromagnetic interference), thus affecting the accuracy of sleep stage observation. These artifacts may be removed from the EEG data 102 at 104, for example, by manual removal, i.e., by visually examining the EEG signal (and / or observing the subject in parallel) and rejecting segments with large fluctuations or sudden changes that may be artifacts, or by automatically filtering out the EEG data 102 during, before, or after segmentation 106 using filtering (e.g., DC filtering) or data smoothing techniques.

[0036] The EEG data (or signal) 102 can also be preprocessed using component analysis, i.e., by decomposing the EEG signal into independent components, identifying and removing artifacts based on spatial and temporal features (i.e., at 104). EEG artifacts can also be removed by estimating the artifact shadow space using methods such as PCA and projecting the EEG data 102 onto an orthogonal subspace for artifact removal 104. In other cases, template matching can be performed, which identifies and removes known artifact patterns by comparing the EEG data 102 with a predefined template. Additionally, wavelet transform can be applied, decomposing the EEG data 102 into different frequency components and removing artifacts in specific frequency bands.

[0037] After (or before) artifact processing, EEG data 102 (or EEG signal) segmentation 106 can be performed, which divides the EEG time series data 102 into multiple time intervals (similar or different lengths) via various separation techniques, where each time interval is a part of the EEG sequence data. A scan window 106a can be used to segment the time intervals into different sub-parts. For example, a one-hour time frame of the received EEG signal can be scanned in one-minute increments (i.e., a one-minute scan window), resulting in 60 discrete time intervals. The scan window 106a can use a sliding window 106b, where portions of the sliding window 106b can have overlapping time series. For example, a one-hour time frame of the received EEG signal can be scanned with a one-minute scan window starting every 30 seconds (i.e., a 30-second sliding window), resulting in a 30-second overlapping one-minute scan window. The scan window 106a and the sliding window 106b can be used to separate the EEG data 102 into time series increments. In one aspect, the EEG signal can be adjusted by normalization 110 to account for power differences. For this purpose, the power spectrum 108 can be calculated, for example, by calculating the power spectral density of each interval of the EEG data 102. The power can be calculated using various techniques, such as multi-taper transform, Fourier transform, or wavelet transform, and then any form of normalization 110 can be performed by weighting the spectral power across one or more time intervals. The normalized power of each time interval at one or more frequencies across time can help determine the appropriate frequency window for extracting information. This normalization 110 can reveal low power and statistically significant power shifts at one or more frequencies (e.g., delta bands, gamma bands, etc.).

[0038] EEG signals can be characterized by different frequency bands, each associated with specific cognitive and physiological states. For example, the delta band, typically ranging from [0.5–4] Hz, includes slow-wave or high-amplitude frequencies. Deep sleep that supports recovery processes, such as stage 3 of non-REM sleep, can be associated with this band. Similarly, the theta band, ranging from approximately [4–8] Hz, includes intermediate frequencies and amplitudes. Light sleep, such as stages 1 and 2 of non-REM sleep, drowsiness, meditation, or similar states, may be associated with this band. The alpha band, ranging from approximately [8–12] Hz, can characterize intermediate frequencies with amplitudes lower than the delta and theta bands. Various states, such as relaxation, wakefulness, or closed eyes, can be associated with the alpha band. Additionally, the alpha band can facilitate the transition between wakefulness and sleep. Following alpha, the beta band, approximately in the [12-30] Hz range, can be characterized by higher frequencies and lower amplitudes, which can be associated with active thinking, focus, wakefulness, or similar activities. The gamma band, with relatively higher frequencies, is approximately in the [30-100] Hz range, characterized by very high frequencies and low amplitudes in EEG signals. The gamma band can be associated with higher information processing and perception; for example, REM sleep can be characterized by vivid dreams and high levels of wakefulness-like brain activity. By processing these spectral characteristics of the spectral bands, various sleep stages can be assigned to segmented time intervals.

[0039] Within these frequencies, any frequency band can be revealed and used for analysis. After establishing a suitable frequency window, features can be calculated for each time interval. These features may include low-frequency power (e.g., delta power and theta), high-frequency power (e.g., gamma power), standard deviation, maximum amplitude (e.g., maximum value of peak absolute value), etc. The calculated features for each time interval can be further calculated to create derived features, such as gamma power / delta power, time derivative of delta, time derivative of gamma power / delta power, etc. The time derivative can be calculated over previous and consecutive time intervals. These derived features can be clustered based on these spectral features or normalized features to group time intervals of similar sleep. To perform clustering 202 efficiently, normalized features including, for example, two or more of these parameters can preferably be on comparable scales. Therefore, normalization 110 can be performed again on the calculated information across time intervals to bring the different derived features to the same scale. Various data normalization techniques can be performed, including z-scores, min-max scaling, quantile transformations, logarithmic transformations, and other similar techniques. In some cases, normalization is performed via z-scores, a statistical technique used to standardize the range of an independent variable (or feature). It may involve transitional features to make the mean of the feature zero and the standard deviation one. By applying z-scores, different derived features of spectral power data (such as delta power and gamma power / delta power) can be scaled to a common range, thereby eliminating bias.

[0040] Figure 2An exemplary overview of determining a subject's sleep stages from one or more EEG channels (or a single-channel EEG signal) 102 is shown. In some cases, normalized features (such as delta power, gamma power, standard deviation, maximum amplitude of gamma or delta) can be input into classification techniques (e.g., supervised machine learning techniques including decision trees, support vector machines (SVM), random forests, or custom neural networks trained on labeled sleep data) or clustering techniques 202, such as k-means clustering, hierarchical clustering, or Gaussian mixture models (GMM). Furthermore, component analysis such as PCA or ICA can also be used to determine the parameter space (e.g., the type of information used) in the clusters. After clustering 202, each cluster can represent a similar group of spectral features and an optional threshold 204, and sleep stages can be assigned to multiple time intervals. By analyzing the distribution of normalized derived features within each cluster, thresholds can be defined to separate different clusters. For example, if, for a given cluster, the alpha power is below a certain threshold and / or the beta power exceeds a certain threshold (predefined or dynamic, e.g., by calculating the mean and standard deviation), that cluster can be assigned as “awake.” The labeled time intervals within each cluster can then represent the subject’s sleep state during those time intervals.

[0041] In some aspects of this disclosure, exemplary system 100 can analyze (single-channel) EEG signal 102 within multiple time windows to extract features that can be further processed (by exemplary system 200) to identify potential REM time intervals 208. Features can be extracted by applying a differential transform to the EEG signal (e.g., a first or second time derivative to highlight changes in the EEG signal), and then detecting power in the gamma band. If the power exceeds a specified threshold, the window can be classified (as shown at 206) as a “REM” window. Other time windows can be considered “awake” or non-REM windows, with non-REM windows assigned to various sleep stages, such as “stage 1,” “stage 2,” “stage 3,” and “stage 4.”

[0042] Additionally, artifact information (e.g., motion data, poor signal data, etc.) can also be used in classification or clustering 202. For the detected REM sleep intervals, artifacts (e.g., motion data, poor signal data, etc.) can also be used for sleep state classification. For example, artifacts can be used to analyze whether the time interval or window initially assigned to the sleep state should be reassigned to a new sleep state due to adjacent artifact data. For example, a time window or wakefulness window assigned to a REM state with previous motion artifacts can be reassigned to a wakefulness state. Furthermore, for example, an artifact window with a subsequent stage 3 (SWS) window can be reassigned to a stage 3 state because that time window is more likely to represent a large stage 3 sleep window rather than a large motion artifact, which is more common during wakefulness. In this way, artifact data can be utilized, for example, in data smoothing techniques.

[0043] Assigning sleep stages from stage 1 to stage 4 can involve analyzing EEG signals recorded during sleep and identifying specific patterns associated with each sleep stage. Automated algorithms, such as the disclosed techniques, can perform EEG waveform analysis based on waveform characteristics (such as amplitude, frequency, and duration of sleep stages) throughout the recorded sleep. For a sleep cycle, "stage 1" sleep is the lightest non-REM sleep stage and can occur during the transition from wakefulness to deep sleep. During stage 1 sleep, EEG signals may exhibit theta waves, which are slower in frequency compared to alpha waves observed during wakefulness. This stage can be characterized by drowsiness, muscle relaxation, and occasional muscle twitches. Following stage 1, "stage 2" sleep is a deeper stage of non-REM sleep and can constitute the majority of the sleep cycle in healthy adults. It can be characterized by the presence of sleep spindle waves (i.e., short bursts of high-frequency brain activity) and K-complex waves (i.e., large bursts of slow waves) in the EEG signal. Stage 2 can be associated with further relaxation of muscle tone and decreased heart rate and body temperature. In some examples, stages 1 and 2 are associated with intermediate sleep (IS) because these stages represent lighter stages of non-REM sleep. The IS state tends to act as a transitional state between REM and SWS. Stage 3 sleep, also known as slow-wave sleep (SWS) or deep sleep, can be characterized by high-amplitude and low-frequency EEG signals (i.e., the presence of slow delta waves in the EEG signal). It can be considered the deepest and most restorative stage of sleep, during which the body undergoes physiological repair and recovery. Therefore, stage 3 sleep can significantly affect physical health, immune function, and cognitive function. Stage 4 sleep can be characterized by the maximum proportion of delta waves in the EEG signal (e.g., over 50%), and it is another term for the deepest stage of SWS. These two stages (i.e., stages 3 and 4) are often collectively classified as SWS based on similar characteristics. During stage 4 sleep, the body may experience minimal physiological activity, and the wakefulness threshold may be highest.

[0044] Figure 3 Examples of EEG waveforms from subjects recorded during REM and wakefulness at different time windows are shown. REM sleep can be characterized by a very low-amplitude, “wake-like” EEG signal 302 (e.g., approximately ±30 µV) with higher gamma power than non-REM sleep. The EEG waveform 302 during REM sleep can be very similar to the wakeful EEG waveform 304 seen during wakefulness, hence the term “wake-like.” Figure 3As shown. During REM sleep, the eyes begin to move rapidly behind closed eyelids, and brain activity may increase, similar to patterns observed by subjects when awake. Additionally, because REM sleep can be characterized, for example, by vivid, complex, and often narrative dreams, the high levels of cognitive activity can become similar to the active thought processes and consciousness experienced in the waking state. In some cases, this similarity can make it challenging to distinguish REM sleep from wakefulness using EEG alone. Both REM sleep and wakefulness can exhibit similar characteristics, such as low-amplitude and high-frequency EEG waves. Therefore, additional non-EEG sensors, such as accelerometers or electromyography (EMG) sensors, can be used to further improve the accuracy of detecting REM sleep intervals by avoiding false positives.

[0045] Figure 4 Example block diagram 400 is shown for detecting potential REM behavioral disorders (RBDs) from a subset of potential REM time intervals 208. REM sleep abnormalities can be important indicators of neurodegenerative diseases such as Parkinson's disease and synucleinopathies. A key indicator can be RBD, which typically occurs during REM sleep as muscle weakness (i.e., loss of muscle tone), leading to dream enactment behaviors such as talking, shouting, impulsive, or violent actions. These conditions can often precede the motor symptoms of Parkinson's disease by many years, thus aiding in early diagnosis. Other conditions can include altered REM sleep patterns, such as a shortened percentage of REM sleep, a shortened REM latency (i.e., the duration of the altered entry into the first REM sleep period), and autonomic dysfunction, including irregularities or abnormalities in heart rate and breathing patterns, which can also be considered noteworthy indicators. To monitor these indications, one or more non-EEG sensors 402 can be used to accurately assess REM sleep intervals and avoid false positives. Such non-EEG sensors may include, but are not limited to, electromyography (EMG) sensors, electrooculography (EOG) sensors, accelerometers, heart rate monitors such as electrocardiograms (ECGs) that can detect increased heart rate variability during REM sleep compared to non-REM sleep, respiratory monitors for measuring irregular breathing patterns, and thermocouples for measuring changes in skin temperature.

[0046] In one scenario, one or more EMG sensors can be used to detect muscle tone during REM sleep when muscle tone is expected to be minimal or absent (due to lack of tone), an indicator of synucleinopathy or its early stages. Lack of muscle tone typically occurs during REM sleep, resulting in low EMG signal activity (as opposed to signals indicating brief twitches). In wakefulness, muscle activity is relatively high and more continuous. Therefore, non-EEG sensors 402 placed, for example, on the chin, face, eyes, or neck, such as EMG electrodes for capturing muscle tone, muscle movement (or neural signals sent to the muscles), can help confirm the lack of muscle tone, a characteristic of REM sleep, and distinguish it from muscle movement attributed to wakefulness. Small muscle movements (e.g., on the face) can be detected even if the subject moves another part of their body. One or more non-EEG sensors 402 can be optionally or additionally attached to a wearable device (e.g., a hat or headband) or integrated within a wearable device. In some cases, a single device (e.g., an adhesive patch, hat, headband, etc.) can be configured to include or receive (e.g., mechanically connected) one or more non-EEG sensors. Alternatively or additionally, one or more EOG sensors can be used to determine eye movements during REM sleep, as rapid and irregular eye movements are likely strongly attributable to REM sleep. Wakefulness, compared to REM sleep, may involve more controlled and purposeful eye movements. To detect REM, EOG sensors can be placed near the eyes to further confirm the accuracy of the detected REM sleep intervals.

[0047] In some cases, the non-EEG sensor 402 may include one or more accelerometers, which may be placed and / or configured to detect body movement. The non-EEG sensor (such as an accelerometer) may be a non-invasive device and receives measurements along multiple axes (e.g., x, y, and z) to measure the acceleration of body movement along these axes. When used during REM sleep monitoring, the accelerometer can provide data on movement frequency, intensity, duration, and other similar properties that can help identify aberrant motor activity during REM sleep. For example, an increase in the frequency of high-intensity movements during REM sleep intervals can be flagged for the diagnosis of potential RBD in the subject. Since REM sleep is attributed to minimal body movement (excluding mild twitching) caused by muscle weakness, accelerometer data can be used to distinguish between relative stillness attributed to normal REM sleep and body movement typically attributed to wakefulness (in healthy patients). One or more accelerometers may be placed or attached to various parts of the body, such as the wrist, ankle, and torso, to capture body movements during REM sleep that indicate RBD. Data collected by the accelerometer can be segmented and analyzed according to the same time intervals predicted for REM sleep to analyze movement patterns. This combination of multimodal data (including signals from EEG sensors and one or more non-EEG sensors 402, individually or in combination) can leverage different physiological characteristics associated with that sleep stage to provide an improved assessment of REM sleep monitoring.

[0048] Using one or more non-EEG sensors 402 may make it difficult to predict whether a person is exhibiting movements inconsistent with REM sleep. For example, one type of non-EEG sensor 402 may be configured to detect motion signals that may reflect movements that are not actually related to the monitored subject. For example, movements of a bed partner or spontaneous muscle twitches or spasms that may occur due to external causes. These movements may be captured by these non-EEG sensors and may be misinterpreted as movements from the monitored subject, leading to inaccurate or erroneous results.

[0049] In some cases, EMG sensors can be placed directly on the muscles of the monitored subject, making it unlikely that movements from a bed partner will produce noticeable muscle activity in the subject. Furthermore, various signal processing or filtering techniques can be used to identify the true signal from this noisy signal. By adjusting the sensitivity threshold, these sensors can be tuned to ignore minute or irrelevant movements. Sensor fusion can also be used to reduce false alarms; that is, cross-validation can be performed by combining data from different modalities (or sensors). For example, using an EMG and an accelerometer, if both sensors detect movement simultaneously, it may indicate a true signal.

[0050] To validate a subset of the potential REM time interval 208 for further determination of the RBD, one or more prediction models 406 can be used, which acquire data from one or more non-EEG sensors 402. EMG sensors can be configured and / or positioned to capture muscle activity, which can provide informative signals to distinguish between REM and non-REM sleep stages (or wakefulness). Other non-EEG sensors, such as accelerometers for capturing body motion, EOG sensors for eye movements, and / or heart rate monitors, such as ECG, can also provide additional useful data for the prediction task. Validating REM and identifying the RBD using EMG sensor data and / or other non-EEG sensors during the potential REM window (or time interval) 208 can employ a structured approach involving one or more techniques. For example, data acquired from non-EEG sensors 402 can be extracted within the same time interval as the potential REM time interval. This extracted (or segmented) sensor data can be normalized for consistency and noise reduction, which may include artifact removal and baseline drift reduction. From these one or more non-EEG sensor data 402, features can be extracted, such as amplitude and power (e.g., calculating root mean square (RMS) values), activity index (i.e., the number of bursts of activity above a certain threshold), tetany (i.e., sustained low-level muscle activity), and phasing (i.e., short bursts or pulse responses) measurements of muscle activity. Additionally, statistical metrics such as entropy, mean, standard deviation, and variability can also provide valuable insights into how the sensor data changes over time.

[0051] These features from one or more non-EEG sensors 402 can be combined with a latent REM time interval 208 and fed into one or more predictive models trained on labeled data to detect corresponding motion. Segmented data from each sensor corresponding to the REM interval can be aggregated to form a single observation as input to the predictive model. Alternatively, segmented data from each of the one or more sensors can be individually fed into one or more predictive models for detecting muscle tone, aberrant motor activity, and / or motion detection during REM sleep. Such a predictive model 406 may include a machine learning model configured to capture muscle weakness (i.e., lack of muscle tone, which can indicate muscle relaxation) and / or body movement during REM sleep, thereby reliably detecting RBD. In some cases, preprocessed data (e.g., normalized and / or denoised) from one or more non-EEG sensors 402 can be directly combined with the latent REM time interval 208 and fed into one or more predictive models 406. These predictive models 406 may include, but are not limited to, traditional models such as decision trees, random forests, support vector machines (SVMs), or deep learning models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or combinations thereof, such as ensemble methods (i.e., combining predictions from multiple models that take input from one or more sensors to improve overall performance, for example, combining random forests, SVMs, or deep learning to capture aspects of data from one or more sensors).

[0052] To further verify the accuracy of the detected REM sleep stages, an analysis of sleep structure compared to normal or expected sleep can be performed. Figure 5An example description of a subject's normal sleep cycle 500 is shown. Observing the individual stages of sleep cycle 500 and the transitions between sleep stages within total sleep also helps in the prediction of RBD. Sleep typically progresses through sleep cycles in finite durations (e.g., approximately 8 hours of nighttime sleep) against the main sleep time (e.g., approximately 90–120 minutes) and repeats throughout the entire sleep cycle. Each cycle typically includes a transition from light sleep (i.e., stage 1 502, which can be characterized by theta waves) to wakefulness. This stage typically lasts for a small portion of the sleep cycle (e.g., a few minutes) and progresses to stage 2 504, which is also attributed to lighter sleep but is relatively stable than stage 1 502 and can last beyond stage 1 502 (e.g., approximately 10–25 minutes, showing sleep spindles and K-complexes). Following stage 2 504, sleep can transition to stages 3 and 4 506 (i.e., deep sleep characterized by high-amplitude, low-frequency delta waves). This stage can last for a considerable period of the initial sleep cycle (e.g., 20–40 minutes), and its duration decreases as sleep progresses. After deep sleep (i.e., stages 3 and 4, 504), subjects typically transition back to stage 2, 504 before entering REM sleep, 508, due to, for example, low-amplitude, mixed-frequency EEG patterns resembling wakefulness, REM, muscle weakness, and vivid dreams. The REM window 508 can be relatively short in the initial sleep cycle (e.g., approximately 10–20 minutes) and can lengthen in subsequent cycles. Understanding these transitions in sleep stages can help validate sleep stages, particularly REM sleep used to detect neurodegenerative diseases.

[0053] Additionally, sleep pattern analysis can be performed, which involves observing adjacent windows of predicted REM intervals, such as... Figure 5 The sleep sequence 510 is shown. The REM window typically follows stage 2 504, and occasionally a short cycle of stage 1 502, especially if the subject wakes up during a brief phase. Alternatively, the subject can return to stage 2 504, maintaining the continuity of the sleep cycle 500 and preparing for the next deep sleep stage (i.e., stages 3 and 4) 506. To verify REM intervals, examining these preceding and following stages of the REM window 508 may be helpful.

[0054] Figure 6An example architecture for predicting REM behavioral disorders (RBD) from supplementary data is shown. Supplementary data may include sleep pattern analysis 602 and / or sleep duration analysis 604. In RBD, the absence of normal muscle tone may cause the subject to physically enact dreams. Predicting RBD from observed sleep cycles 500 can involve identifying abnormalities during REM sleep and specific patterns indicative of the disorder. This observation can provide a framework for understanding anticipated patterns and transitions within the sleep architecture. For example, it can be combined with sleep pattern analysis 602 as described above. Sleep pattern analysis 602 can involve observing adjacent windows for potential REM time intervals 208. To perform sleep pattern analysis 602 and consider typical smooth transitions between different sleep stages for subjects (i.e., humans and animals), smoothing algorithms, modeling techniques, or statistical analyses can be used to predict whether the window corresponds to the subject being awake or in REM sleep for each potential REM time interval 208. For example, if a potential REM window 208 (or a consecutive set of potential REM windows) is surrounded (or followed by) a "Stage 2" window, a smoothing algorithm, modeling technique, or statistical analysis can be configured to predict that the window (or a series of consecutive windows) corresponds to REM sleep. Similarly, if a potential REM window 208 (or a consecutive set of potential REM windows) is surrounded by a "Stage 1" window, a smoothing algorithm, modeling technique, or statistical analysis can be configured to predict that the window corresponds to wakefulness. Furthermore, if a particular window shows an abrupt transition from one stage to another, such as directly from deep sleep to REM, or vice versa, it may indicate incorrectly labeled REM sleep.

[0055] For sleep pattern analysis 602, smoothing techniques, such as Gaussian filters, can mitigate abrupt changes by assigning weights to each adjacent window based on its distance from a specific observation window (center point). Kalman filters can use weighted averages to predict or forecast the next sleep state based on the current sleep state and observed sleep state data to minimize uncertainty. Alternatively, moving average filters can calculate the average label of a series of adjacent windows to smooth transitions between sleep stages and reduce noise. For example, if a particular window is labeled as stage 1 and surrounded by consecutive windows of stage 3, it might represent a brief period of wakefulness or a misclassified interval. Smoothing techniques can merge stage 1 into the surrounding windows, i.e., stage 3.

[0056] Alternatively, or in combination with, overall sleep duration analysis 604 can also be observed, which may involve a predictable extension of the REM window for each consecutive sleep cycle. For example, given a long sleep duration, such as approximately 8 hours of nighttime sleep, the REM window 508 tends to be extended and occurs more frequently in the latter half of the total sleep. In the first half of sleep, more time may be spent on stage 3 or stage 4 (deep sleep) 506, and less time on REM sleep 508. After the first half, in the second half, the duration of the REM window 508 may be extended, deep sleep 506 may be reduced, and more time may be spent on stage 2 504 and REM state 508.

[0057] Modeling techniques such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), or hidden Markov models (HMMs) can be used for sleep pattern analysis 602 and time analysis 604. These techniques can model sequential data to capture temporal context by automatically estimating relationships between consecutive time intervals. This modeling technique can learn temporal structure and take into account the probability of transitions between sleep stages. By aggregating time intervals or overlapping windows of time intervals (which may or may not include EEG data 102 and / or data from one or more non-EEG sensors 402), the model is trained on labeled sleep data and can learn statistical relationships and probabilities of transitions between different sleep states based on the observed sequence, and use this information to correct inconsistencies. It should be understood that a unified model or different models can be trained to perform sleep pattern analysis 602 and sleep time analysis 604. Once the model is trained, it can be applied to newly input segmented sleep data to validate the sequence of sleep stages. The trained model can look for anomalous patterns in sleep intervals. In RBD, transitions into or out of REM may occur more frequently or in unusual ways, such as frequent interruptions of REM in awake or other phases. Such a network can be trained to classify each interval as REM, awake, or non-REM based on extracted features.

[0058] In one aspect, modeling techniques can be equipped with, for example, nighttime time segmentation, which may involve dividing total sleep into two parts, namely the first half and the second half, and appending this information to the segmented time intervals (e.g., binary variables (or bits) can indicate whether each time window belongs to the first or second half). This appending allows more attention to be focused on analyzing the distribution of REM sleep throughout the entire sleep cycle, for performing sleep duration analysis 604. Predictions from modeling techniques that can perform sleep pattern analysis 602 and sleep duration analysis 604, along with a predictive model 406 that can detect muscle tone during REM window 508, can be combined to obtain a final prediction 606. This combination can be performed using various techniques, including simple averaging, weighted averaging, majority voting, or other similar techniques.

[0059] For example, if a subject has been identified with RBD based on predictive model 406, one or more actions can be performed. Such predictions can be integrated into a system that can perform one or more actions associated with both clinical interventions and safety measures to improve the subject's health and reduce the risk of injury. These actions may include, but are not limited to, issuing alerts, notifying relevant authorized parties (e.g., healthcare professionals, relatives, and / or the subject) to conduct a full medical evaluation to confirm a diagnosis for a neurodegenerative disease such as Parkinson's, and arranging a consultation with a sleep specialist for a thorough neurological examination.

[0060] Figure 7An example process flow 700 is shown for determining REM behavioral disorder (RBD) by utilizing one or more channels of an EEG sensor, and for further determining REM behavioral disorder (RBD) by utilizing one or more non-EEG sensors to detect muscle tone and perform sleep analysis for identified REM sleep. The boxes in process flow 700 are shown in a specific order, but this order can be modified; for example, some boxes may be performed before others, while some boxes may be performed simultaneously. The boxes can be performed by hardware or software, or a combination thereof. Process 700 may include, at box 702, receiving one or more first signals from EEG electrodes placed on the subject over a period of time. EEG data may be received via a single channel (e.g., at least one active electrode, a reference electrode, and a potential ground electrode) or multiple channels. At box 704, the received data may be preprocessed or further adjusted, for example, for artifact removal, normalization, or frequency weighting of selective bands such as gamma bands to extract REM sleep-related information. Features and / or derived features for one or more time intervals can be extracted from these selective spectral bands. At box 706, these extracted features can be further normalized (e.g., z-score) and clustered to determine time intervals corresponding to REM, wakefulness, and non-REM (e.g., any one of sleep stages 1 to 4). In addition to one or more EEG signals, one or more second signals from non-EEG electrodes or sensors (e.g., electromyography (EMG) or accelerometers) can be received. These second signals can be combined to examine the presence of muscle tone and / or subject movement during a predicted subset of (potential) REM sleep.

[0061] At box 708, predictions regarding whether a subject has loss of muscle tone and / or whether a subject exhibits any movement during the predicted subset of time intervals corresponding to multiple time intervals of the REM phase can be performed. For this validation, one or more predictive models can be used, acquiring data from one or more non-EEG sensors, such as EMG sensors (and / or accelerometers for anomalous motor activity), which capture muscle tone or muscle tension, indicating muscle readiness to perform action, which is important for distinguishing between REM and non-REM sleep phases (or wakefulness). A second signal acquired from the non-EEG sensors can be extracted for the same time interval as the potential REM time interval. The data extracted from one or more non-EEG sensors can be normalized and / or filtered to remove noise and can be combined with the potential REM time interval. One or more predictive models, such as machine learning models, can acquire this input and capture muscle tone (or muscle activity) or body movement during REM sleep, resulting in reliable detection of RBD. Alternatively, features can be derived from one or more non-EEG sensor data and then combined with the potential REM time interval data and fed into the predictive model.

[0062] Additionally, at box 710, sleep analysis can be performed via one or more modeling techniques to validate subsets of multiple time intervals corresponding to REM stages of sleep. The modeling techniques can be trained to learn the temporal structure of different sleep stages within a given time period. In some aspects, one or more modeling techniques can be configured to identify smooth transitions between different sleep stages by analyzing adjacent time intervals of a subset of multiple time intervals corresponding to REM sleep stages. For example, if a specific time interval from a subset of multiple time intervals is surrounded by a "stage 2" window, the modeling technique can be configured to predict that the specific window (or time interval) corresponds to REM sleep. Furthermore, one or more modeling techniques can also be configured to identify a gradual increase in consecutive REM stages of sleep from a subset of multiple time intervals corresponding to REM stages of sleep over a period of time.

[0063] At box 712, a diagnosis of RBD can be predicted based on predictions from non-EEG sensors regarding whether the subject moves during predicted subsets of multiple time intervals corresponding to REM states, and sleep analysis performed by one or more modeling techniques. In response to this prediction, at box 714, an output indicating the prediction result for RBD can be displayed. For example, if the subject is diagnosed with RBD, one or more actions that may involve preventative measures and medical interventions can be triggered. These actions may include, but are not limited to, issuing an alarm, notifying relevant authorized parties (e.g., healthcare professionals, relatives, and / or the subject) to conduct a full medical evaluation to confirm the diagnosis for a neurodegenerative disease such as Parkinson's disease, and arranging a consultation with a sleep specialist for a thorough neurological examination.

[0064] Figure 8 This is an example illustration of a computer system 800 in which various embodiments of the present disclosure may be implemented. For example, the techniques described above for determining rapid eye movement (REM) behavior disorder (RBD) from one or more EEG signals recorded from a subject's sleep can be implemented in computer-executable instructions (e.g., organized in program module 804). Program module 804 may include routines, programs, objects, components, and data structures that perform tasks and implement data types for implementing the techniques described above. The functions described herein can be performed at least in part by one or more hardware logic components.

[0065] In order to provide additional context for its various aspects, Figure 8 The following description is intended to provide a brief, general description of a suitable computer system 800 in which various aspects can be implemented. While the above description is in the general context of computer-executable instructions that can run on one or more computers, those skilled in the art will recognize that novel implementations may also be implemented in combination with other program modules and / or as a combination of hardware and software. The computer system 800 for implementing the various aspects includes a processing unit 808 having one or more processors (also referred to as microprocessors), a computer-readable storage medium (wherein the medium is any physical device or material capable of storing and retrieving data electronically and / or optically), such as a data storage unit 810 (computer-readable storage media / mediums also include disks, optical disks, solid-state drives, external storage systems, and flash drives), and a system bus 812. The system bus 812 may provide interfaces to system components, including but not limited to system memory 814, to the processing unit 808. Such a system bus 812 may be any of several types of bus architectures, which may use any of a variety of commercially available bus architectures, further interconnected to a memory bus (with or without a controller) and a peripheral bus (e.g., PCI, PCIe, AGP, LPC, etc.).

[0066] Figure 8 An example configuration of a typical computer is shown, which can be other commercially available microprocessors, such as single-processor, multi-processor, single-core, and multi-core units for processing and / or storage circuitry. Furthermore, those skilled in the art will understand that this novel system and method can be implemented using other computer system configurations, including minicomputers, mainframes, and personal computers (e.g., desktops, laptops, tablets, etc.), handheld computing devices, microprocessor-based or programmable consumer electronics, etc., each configuration potentially co-coupled to one or more associated devices.

[0067] In some aspects, computer system 800 may be one of several computers employed in a data center and / or computing resources (hardware and / or software) to support cloud computing services for portable and / or mobile computing systems, such as wireless communication devices, cellular phones, and other mobile-capable devices. For example, cloud computing services include, but are not limited to, Infrastructure as a Service (IaaS), Platform as a Service (PaaS), Software as a Service (SaaS), Storage as a Service (SaS), Desktop as a Service (DAS), Data as a Service (DAS), Security as a Service (SAS), and API (Application Programming Interface) as a Service (API) as a Service (AS). In some cases, system memory 814 may include computer-readable storage (physical storage) media, such as volatile memory (e.g., random access memory (RAM) 816) and non-volatile memory (e.g., ROM 818). The Basic Input / Output System (BIOS) may be stored in non-volatile memory and includes basic routines that facilitate data and signal communication between components within computer system 800, such as during startup. Volatile memory may also include high-speed RAM, such as static RAM for caching data.

[0068] By way of example and not limitation, system memory 814 may also include program module 804 (which may include client applications, web browsers, middleware applications, relational database management systems (RDBMS), etc.), program data 806, and operating system 802. For example, operating system 802 may include various versions of Microsoft Windows®, Apple Macintosh®, and / or Linux operating systems, various commercially available UNIX® or UNIX-like operating systems (including but not limited to various GNU / Linux operating systems, Google Chrome OS, etc.), and / or mobile operating systems such as iOS, Windows® Phone, Android OS, BlackBerry® OS, and Palm® OS. All or part of operating system 802, program module 804, and / or program data 806 may also be cached in memory, such as volatile memory and / or non-volatile memory (RAM 816 or ROM 818). It should be understood that the disclosed structure can be implemented using various commercially available operating systems or combinations of operating systems (e.g., virtual machines).

[0069] In some other examples, computer system 800 may have additional features or functions. For example, computer system 800 may also include additional data storage devices (removable and / or non-removable), such as, for example, disks, optical discs, or magnetic tapes. Computer-readable media may include at least two types of computer-readable media: computer storage media and communication media. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data.

[0070] System memory 814 and data storage 810, including removable and non-removable storage, are examples of computer storage media. In addition to RAM 816 and ROM 818, computer storage media include, but are not limited to, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store target information and can be accessed by computer system 800. Furthermore, computer-readable media may include computer-executable instructions that, when executed by processing unit 808, perform the various functions and / or operations described herein. In contrast, communication media may embody computer-readable instructions, data structures, program modules, or other data in modulated data signals (such as carrier waves) or other transmission mechanisms.

[0071] The computer system 800 may also include one or more input devices 820, such as a keyboard, mouse, pen, voice input device, touch input device, etc. It may also include one or more output devices 822, such as a display, speaker, printer, etc. These devices are well known in the art and will not be discussed in detail here. The computing device 800 may also include one or more network interfaces 824 to establish communication that allows the computer system 800 to communicate with other systems or devices, such as via a network. These networks may include wired networks and wireless networks. Here, the computer system 800 is an example of a suitable device or system and is not intended to impose any limitation on the scope or functionality of the various embodiments described.

[0072] Other known computer systems, environments, and / or configurations that may be suitable for use with the embodiments include, but are not limited to, personal computers, server computers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set-top boxes, game consoles, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the systems or devices described above. For example, some or all components of computer system 800 may be implemented in a cloud computing environment to allow resources and / or services to be selectively available to user devices via a computer network.

[0073] Although specific aspects have been described, various modifications, alterations, alternative constructions, and equivalents are possible. The embodiments are not limited to operation within certain specific data processing environments but can operate freely in multiple data processing environments. Additionally, although certain aspects have been described using a specific set of transactions and steps, it should be apparent to those skilled in the art that this is not intended to be limiting. While some flowcharts describe operations as sequential processes, many operations can be executed in parallel or concurrently. Furthermore, the order of operations can be rearranged. A process may have additional steps not included in the diagram. The various features and aspects of the foregoing can be used individually or in combination.

[0074] Furthermore, while certain aspects have been described using specific combinations of hardware and software, it should be recognized that other combinations of hardware and software are also possible. Some aspects can be implemented solely in hardware, solely in software, or using a combination thereof. The various processes described herein can be implemented on the same or different processors in any combination.

[0075] When a device, system, component, or module is described as being configured to perform certain operations or functions, this configuration can be achieved, for example, by designing electronic circuitry to perform the operation, by programming programmable electronic circuitry (such as a microprocessor) to perform the operation, such as by executing computer instructions or code, or by a processor or core programmed to execute code or instructions stored on a non-transitory storage medium, or any combination thereof. Processes can communicate using various technologies, including but not limited to conventional technologies for inter-process communication; different pairs of processes can use different technologies, or the same pair of processes can use different methods at different times.

[0076] Specific details are set forth in this disclosure to provide a thorough understanding of these aspects. However, these aspects can be practiced without these specific details. For example, well-known circuits, processes, algorithms, structures, and techniques have been shown without unnecessary detail to avoid obscuring these aspects. This description provides only exemplary aspects and is not intended to limit the scope, applicability, or configuration of other aspects. Rather, the foregoing description of these aspects can provide those skilled in the art with enabling descriptions for implementing the various aspects. Various changes can be made to the function and arrangement of the elements.

[0077] Therefore, the specification and drawings are to be regarded as illustrative, not restrictive. However, it will be apparent that additions, deletions, omissions, and other modifications and alterations may be made therein without departing from the broader spirit and scope set forth in the claims. Thus, although specific aspects have been described, these aspects are not intended to be restrictive. Various modifications and equivalents are within the scope of the appended claims.

Claims

1. A computer-implemented method, comprising: Received over a period of time: One or more first signals from one or more EEG electrodes, wherein at least one of the one or more EEG electrodes is placed on the subject during the time period; One or more second signals from one or more non-EEG electrodes, wherein at least one of the one or more non-EEG electrodes is placed on, near or inside the subject during the time period; The one or more first signals are preprocessed to extract a feature set associated with one or more frequency bands of the one or more first signals; For each of the multiple time intervals within the aforementioned time period, a state corresponding to the time interval is predicted based on the feature set, wherein the state corresponds to any one of one or more sleep stages or wakefulness states, and wherein for a subset of the multiple time intervals, the state is a rapid eye movement (REM) stage of sleep. For the subset of multiple time intervals corresponding to the predicted REM stage of the sleep, based on the one or more second signals within the time intervals, predict whether the subject has muscle tone during the subset of the multiple time intervals; Sleep analysis is performed via one or more modeling techniques to verify subsets of the plurality of time intervals corresponding to REM stages of the sleep, wherein the one or more modeling techniques are trained to learn the temporal structure of the one or more sleep stages within the time interval. Based on predictions from one or more second signals during subsets of the plurality of time intervals and the sleep analysis, the subject is predicted to have REM behavioral disorder. Output a prediction about whether the subject has REM behavioral disorder; as well as An action is triggered when the subject is predicted to have REM behavioral disorder.

2. The computer-implemented method of claim 1, wherein the one or more modeling techniques are configured to identify a smooth transition for the subset of the plurality of time intervals corresponding to the REM phase of the sleep by analyzing one or more adjacent time intervals of each time interval in the subset of the plurality of time intervals.

3. The computer-implemented method of claim 1, wherein the one or more modeling techniques are configured to identify a gradual increase in consecutive REM stages of sleep from a subset of the plurality of time intervals corresponding to REM stages of sleep within the time period.

4. The computer-implemented method according to claim 1, wherein the one or more modeling techniques include Hidden Markov Models (HMMs) or Recurrent Neural Networks (RNNs).

5. The computer-implemented method of claim 1, wherein the one or more first signals comprise single-channel EEG data.

6. The computer-implemented method of claim 1, wherein the feature set associated with the one or more frequency bands includes one or more of delta power, gamma power, standard deviation, maximum amplitude, gamma power / delta power, time derivative of delta, and time derivative of gamma power / delta power.

7. The computer-implemented method according to claim 1, further comprising: The one or more first signals and the one or more second signals are divided into the plurality of time intervals.

8. A system comprising: One or more data processors; as well as A non-transitory computer-readable storage medium containing instructions that, when executed on the one or more data processors, cause the one or more data processors to perform operations, including: Received over a period of time: One or more first signals from one or more EEG electrodes, wherein at least one of the one or more EEG electrodes is placed on the subject during the time period; One or more second signals from one or more non-EEG electrodes, wherein at least one of the one or more non-EEG electrodes is placed on, near or inside the subject during the time period; The one or more first signals are preprocessed to extract a feature set associated with one or more frequency bands of the one or more first signals; For each of the multiple time intervals within the aforementioned time period, a state corresponding to the time interval is predicted based on the feature set, wherein the state corresponds to any one of one or more sleep stages or wakefulness states, and wherein for a subset of the multiple time intervals, the state is a rapid eye movement (REM) stage of sleep. For the subset of multiple time intervals corresponding to the predicted REM stage of the sleep, based on the one or more second signals within the time intervals, predict whether the subject has muscle tone during the subset of the multiple time intervals; Sleep analysis is performed via one or more modeling techniques to verify subsets of the plurality of time intervals corresponding to REM stages of the sleep, wherein the one or more modeling techniques are trained to learn the temporal structure of the one or more sleep stages within the time interval. Based on predictions from one or more second signals during subsets of the plurality of time intervals and the sleep analysis, the subject is predicted to have REM behavioral disorder. Output a prediction about whether the subject has REM behavioral disorder; and An action is triggered when the subject is predicted to have REM behavioral disorder.

9. The system of claim 8, wherein the one or more modeling techniques are configured to identify a smooth transition for the subset of the plurality of time intervals corresponding to the REM stage of the sleep by analyzing one or more adjacent time intervals of each time interval in the subset of the plurality of time intervals.

10. The system of claim 8, wherein the one or more modeling techniques are configured to identify a gradual increase in consecutive REM stages of sleep from a subset of the plurality of time intervals corresponding to REM stages of sleep within the time period.

11. The system of claim 8, wherein one or more modeling techniques include Hidden Markov Models (HMMs) or Recurrent Neural Networks (RNNs).

12. The system of claim 8, wherein the one or more first signals comprise single-channel EEG data.

13. The system of claim 8, wherein the feature set associated with the one or more frequency bands includes one or more of delta power, gamma power, standard deviation, maximum amplitude, gamma power / delta power, time derivative of delta, and time derivative of gamma power / delta power.

14. The system according to claim 8, further comprising: The one or more first signals and the one or more second signals are divided into the plurality of time intervals.

15. A computer program product tangibly embodied in a non-transitory machine-readable storage medium, the computer program product comprising instructions configured to cause one or more data processors to perform operations including: Received over a period of time: One or more first signals from one or more EEG electrodes, wherein at least one of the one or more EEG electrodes is placed on the subject during the time period; One or more second signals from one or more non-EEG electrodes, wherein at least one of the one or more non-EEG electrodes is placed on, near or inside the subject during the time period; The one or more first signals are preprocessed to extract a feature set associated with one or more frequency bands of the one or more first signals; For each of the multiple time intervals within the aforementioned time period, a state corresponding to the time interval is predicted based on the feature set, wherein the state corresponds to any one of one or more sleep stages or wakefulness states, and wherein for a subset of the multiple time intervals, the state is a rapid eye movement (REM) stage of sleep. For the subset of multiple time intervals corresponding to the predicted REM stage of the sleep, based on the one or more second signals within the time intervals, predict whether the subject has muscle tone during the subset of the multiple time intervals; Sleep analysis is performed via one or more modeling techniques to verify subsets of the plurality of time intervals corresponding to REM stages of the sleep, wherein the one or more modeling techniques are trained to learn the temporal structure of the one or more sleep stages within the time interval. Based on predictions from one or more second signals during subsets of the plurality of time intervals and the sleep analysis, the subject is predicted to have REM behavioral disorder. Output a prediction about whether the subject has REM behavioral disorder; as well as An action is triggered when the subject is predicted to have REM behavioral disorder.

16. The computer program product of claim 15, wherein the one or more modeling techniques are configured to: By analyzing one or more adjacent time intervals within each of the subsets of the plurality of time intervals, a smooth transition for the subsets of the plurality of time intervals corresponding to the REM stage of the sleep is identified; and Within the time period, a gradual increase in consecutive REM stages of sleep is identified from the subset of the plurality of time intervals corresponding to the REM stages of the sleep.

17. The computer program product of claim 15, wherein one or more modeling techniques include Hidden Markov Models (HMMs) or Recurrent Neural Networks (RNNs).

18. The computer program product of claim 15, wherein the one or more first signals comprise single-channel EEG data.

19. The computer program product of claim 15, wherein the feature set associated with the one or more frequency bands includes one or more of delta power, gamma power, standard deviation, maximum amplitude, gamma power / delta power, time derivative of delta, and time derivative of gamma power / delta power.

20. The computer program product according to claim 15, further comprising: The one or more first signals and the one or more second signals are divided into the plurality of time intervals.

Images