Sleep apnea detection system and method

Abstract

A method (and system) for detecting sleep apnea includes determining a first characteristic, in particular a sleep stage, and detecting apnea events based on physiological sensor signals using a selection of a detection algorithm, the detection algorithm being dependent on the determined first characteristic. By taking the first characteristic (e.g., sleep stage) into account when performing an automated apnea detection process, the accuracy of apnea detection may be improved.

Description

Technical Field

[0001] The present invention relates to the detection of sleep apnea, wherein apnea is used as a general term to include obstructive sleep apnea, central sleep apnea and hypopnea episodes. Background Art

[0002] Sleep apnea is a general term used to designate two types of sleep breathing disorders, obstructive sleep apnea (OSA) and central sleep apnea (CSA). Although they are caused by different physiological phenomena, they are both associated with a reduction (hypopnea) or complete cessation (apnea) of airflow, leading to a decrease in blood oxygen saturation and ultimately an associated burst of cortical arousal and sympathetic activity, accompanied by an increase in heart rate and blood pressure.

[0003] Repetitive apnea and hypopnea events interrupt sleep continuity and reduce sleep time, which in part explains a defining symptom: excessive daytime sleepiness. Sleep apnea is associated with an increased risk of cardiac and cerebrovascular disease (e.g., hypertension, heart failure, arrhythmias, myocardial ischemia and infarction, pulmonary hypertension), kidney disease, metabolic abnormalities, insulin resistance and lipid abnormalities, stroke, dementia, and cognitive impairment in the elderly, as well as alterations in cerebral blood flow and cerebral autoregulation.

[0004] OSA is the most common type of sleep apnea and is caused by a complete or partial blockage of the upper airway. Normally, during sleep, the muscles of the tongue, mouth, and pharynx relax slightly, but not enough to block the airway. With OSA, the muscles relax too much. The tongue presses against the back of the upper airway, blocking airflow to the lungs.

[0005] During an OSA episode, the heart rate decreases and the blood oxygen saturation decreases. When the brain cannot get enough oxygen, this obstruction may cause arousals, waking the subject partially or completely from a sleep state due to resistance to airflow. Typically, the subject will gasp for air to restore airflow before returning to sleep. This is a cyclic pattern that typically repeats throughout the night (up to 100 times). OSA episodes typically last 20 to 40 seconds.

[0006] Although less common than OSA, CSA is still a related sleep breathing disorder that is more common in patients with other comorbidities, such as heart failure and neurological disease. CSA is characterized by cessation of airflow and respiratory effort (apnea), or decreased airflow and respiratory effort without clear evidence of partial obstruction of breathing (hypopnea).

[0007] To quantify the severity of sleep-disordered breathing, an index that includes both CSA and OSA apnea and hypopnea is used. The apnea-hypopnea index (AHI) reflects the average number of apneas and / or hypopneas per hour of observed sleep. In adults, an AHI ≤ 5 is considered normal. Mild apnea is characterized by an AHI between 5-15 events per hour, moderate apnea between 15-30 events per hour, and severe apnea is associated with AHI values ​​above 30 events per hour. People are often unaware of frequent awakenings during the night.

[0008] Regarding the prevalence of apnea, in adults between 30 and 60 years of age, approximately 9% of women and 24% of men have an AHI ≥ 5. Obesity is the strongest risk factor for apnea and is reflected by multiple parameters such as body mass index, neck circumference, and waist-to-hip ratio. Other risk factors include aging, gender, decreased pharyngeal muscle tone, tonsillar swelling, menopause, upper airway anatomy, smoking, alcohol consumption, ethnicity, and the presence of other cardiac or neurological diseases. There appears to be a direct relationship between the apnea epidemic and the obesity epidemic.

[0009] There are many signs associated with apnea, such as loud snoring, frequent awakenings from sleep to gasp for air to restore airflow, or a feeling of choking. Most people with apnea experience excessive daytime sleepiness or fatigue. Other complaints are related to insomnia and depression. The fragmented sleep caused by apnea may lead to decreased daytime cognitive performance and an increased risk of vehicle and workplace accidents.

[0010] Sleep apnea can be treated with continuous positive airway pressure (CPAP) applied through the nose with the help of a mask that the patient must wear at night. The traditional way to diagnose sleep apnea is to monitor the patient during a sleep study. Sleep studies are very expensive and require an overnight polysomnography (PSG) evaluation with the participant in a sleep lab. The polysomnogram will typically record multiple parameters, including:

[0011] Electroencephalogram (EEG) is used to monitor brain activity;

[0012] Electrooculogram (EOG) is used to monitor eye movements;

[0013] Electromyography (EMG) is used to monitor muscle tone;

[0014] An electrocardiogram (ECG) is used to monitor the electrical activity of the heart;

[0015] Respiratory inductance plethysmography or piezoelectric belts around the chest and abdomen are used to measure respiratory effort;

[0016] Nasal and oral thermistors or pressure sensors to measure airflow; and

[0017] Pulse oximetry is used to monitor changes in blood oxygen levels.

[0018] Underdiagnosis of sleep apnea is a major problem due to the limited availability of sleep laboratories and the high costs associated with sleep studies. It is reported that approximately 80%-85% of OSA patients are underdiagnosed.

[0019] Various algorithms for automated sleep apnea screening have been proposed. However, there are approximately 80 different sleep disorders. Therefore, detecting sleep apnea in patients with sleep disorders is considered challenging. There are known algorithms for detecting apnea and hypopnea episodes based on heart rate variability (HRV) features extracted from ECG signals. Apnea / hypopnea episodes have a significant impact on the instantaneous heart rate and hemodynamics. These episodes result in recurring heart rate patterns known as cyclical variations in heart rate (CVHR). CVHR peaks are due to a sudden increase in heart rate during the arousal phase that terminates an apnea or hypopnea event.

[0020] It is well known that undiagnosed sleep apnea is a significant risk factor for the development of cardiovascular diseases (such as hypertension, stroke and congestive heart failure), thinking disorders and diabetes. Therefore, it is desirable to perform early and simplified diagnosis of sleep apnea using a limited number of sensor sources. Summary of the invention

[0021] The invention is defined by the claims.

[0022] According to an example of one aspect of the present invention, a system for detecting sleep apnea is provided, comprising:

[0023] a physiological sensor for generating a sensor signal for detecting sleep apnea;

[0024] A processor adapted to:

[0025] determining, or receiving as input, an identification of, a first characteristic that is dependent on a subject's physiological condition during sleep, particularly a sleep stage; and

[0026] An apnea event is detected based on the sensor signal using a detection algorithm selected from a set of detection algorithms based on a first characteristic.

[0027] The system can provide automatic sleep apnea screening based on features extracted from physiological signals, such as features captured by a PPG sensor and / or an ECG sensor. The analysis of these features takes into account a first characteristic that depends on the physiological condition of the subject during sleep at that time. The extracted features include, for example, information related to heart rate variability (HRV).

[0028] In this way, an intermediate characteristic is obtained and used to select between a set of sleep apnea determination algorithms. Thus, for at least two non-awake sleep states, there are at least two sleep apnea determination algorithms. This provides a two-stage detection process, in which the subject's physiological condition is first classified, and the classification helps to more accurately determine apnea events.

[0029] The first characteristic, i.e. the sleep stage, may be input into the system, for example, based on manual inspection of the signal. However, the first characteristic detection is preferably also automatic. For example, the processor may be adapted to determine the sleep stage based on heart rate variability information derived from the sensor signal. The sleep stage is determined, for example, by a sleep stage classifier.

[0030] In recent years, automatic sleep stage detection has seen substantial development based on heart rate variability (HRV) features extracted from ECG signals and even PPG signals. The same physiological sensors that can automatically identify sleep stages can be used in the system of the present invention to enable detection of apnea events.

[0031] Compared to known methods for detecting apnea and hypopnea events based on HRV features extracted from physiological signals, the present invention enables a higher detection accuracy. In particular, information dependent on the subject's physiological condition during sleep, such as sleep stage specific information, is utilized in the process of apnea detection. In contrast, known solutions provide a generic classification scheme that is trained to find a mapping between apnea features extracted from physiological signals without taking into account sleep stage information or any other intermediate characteristics that may also be extracted from the physiological signals.

[0032] The present invention is based, for example, on training sleep stage specific classifiers based on features extracted from physiological signals in an offline training phase, which features maximize the separation of apnea classes within each sleep stage.

[0033] In order to detect an apnea event in an instance (ie, a time range) of a physiological signal, the sleep stage associated with the instance may first be estimated. Apnea detection is then based on applying an apnea classifier that is specific to the sleep stage.

[0034] The processor is, for example, suitable for:

[0035] Extracting heartbeat timing from sensor signals;

[0036] exporting a beat-to-beat interval time series; and

[0037] Heart rate variability information is extracted from the beat-to-beat interval time series.

[0038] It is known that heart rate variability information can be used as part of an automatic apnea detection algorithm, and heart rate variability can also be used for sleep stage detection. The present invention enables the same information source to be optimized for both types of detection to improve the accuracy of apnea detection.

[0039] The processor is, for example, suitable for:

[0040] Sampling the sensor signal into a sequence of time ranges;

[0041] extracting a first feature of the sensor signal from each time range; and

[0042] The sleep stage of each time range is determined according to the first feature.

[0043] Therefore, automatic sleep stage detection is performed based on sequential time windows.

[0044] As described above, the first feature used for sleep stage detection may include heart rate variability information.

[0045] The processor is, for example, also adapted to:

[0046] for each time range, selecting an apnea classification model for identifying sleep apnea from the sensor signal, the selection being dependent on a previously determined sleep stage;

[0047] extracting a second feature of the sensor signal from each time range; and

[0048] An apnea or non-apnea state is determined based on the second feature and based on the selected apnea classification model.

[0049] The apnea detection is performed using a model that depends on sleep stages.The second feature may also be extracted using a model that depends on previously identified sleep stages.

[0050] Thus, the second feature (for apnea detection) may be different from the first feature (for sleep stage detection) and furthermore, the second feature may be selected based on the previously identified sleep stage. Thus, the features most suitable for detecting apnea within a particular sleep stage are extracted and then input into a suitable apnea classification model.

[0051] As mentioned above, the second feature for apnea detection may also include heart rate variability information.

[0052] The present invention also provides a method for detecting sleep apnea, comprising:

[0053] generating a sensor signal for use in detecting sleep apnea;

[0054] determining, or receiving as input, an identification of, a first characteristic that is dependent on a physiological condition of the subject during sleep; and

[0055] An apnea event is detected based on the sensor signal by selecting a detection algorithm from a set of detection algorithms according to a first characteristic.

[0056] The first characteristic is, for example, a sleep stage, and the method may comprise determining the sleep stage from heart rate variability information derived from the sensor signal.

[0057] The method may include:

[0058] Extracting heartbeat timing from sensor signals;

[0059] exporting a beat-to-beat interval time series; and

[0060] Heart rate variability information is extracted from the beat-to-beat interval time series.

[0061] This heart rate variability can be used for sleep stage detection and apnea detection.

[0062] The sleep stage detection portion of the method may include:

[0063] Sampling the sensor signal into a sequence of time ranges;

[0064] extracting a first feature of the sensor signal from each time range; and

[0065] The sleep stage of each time range is determined according to the first feature.

[0066] The apnea detection portion of the method may include:

[0067] for each time range, selecting an apnea classification model for use in identifying sleep apnea from the sensor signal;

[0068] extracting a second feature of the sensor signal from each time range using a model that depends on the previously identified sleep stages; and

[0069] An apnea or non-apnea state is determined based on the second feature and based on the selected apnea classification model.

[0070] The present invention may be implemented at least partially in software.

[0071] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Examples of the present invention will now be described in detail with reference to the accompanying drawings, in which:

[0073] Figure 1 A typical hypnogram is shown indicating the various sleep stages and cycles in an adult sleep, wherein the stages are represented as a function of time;

[0074] Figure 2 A system according to an example of the present invention is shown;

[0075] Figure 3 shows how specific heart rate characteristics vary during different sleep stages and during apnea and non-apnea periods;

[0076] Figure 4 The system is shown as a wrist-worn sensor that is worn throughout the night;

[0077] Figure 5 Shown by Figure 2 The basic method of system execution;

[0078] Figure 6 Analytical results of the performance of a sleep stage-independent apnea detection system for sleep stage R are shown;

[0079] Figure 7 shows the results of the analysis of the performance of the apnea detection system according to the present invention for sleep stage R;

[0080] Figure 8 shows the analysis results of the performance of the sleep stage-independent apnea detection system for sleep stage N1;

[0081] Fig. 9 shows the analysis results of the performance of the apnea detection system according to the present invention for sleep stage N1;

[0082] Fig.10 Analytical results of the performance of the sleep stage-independent apnea detection system for sleep stage N2 are shown;

[0083] Fig.11 shows the analysis results of the performance of the apnea detection system according to the present invention for sleep stage N2;

[0084] Fig.12 shows the analysis results of the performance of the sleep stage-independent apnea detection system for sleep stage N3;

[0085] Fig.13 shows the analysis results of the performance of the apnea detection system according to the present invention for sleep stage N3;

[0086] Fig.14 shows analysis results of the performance of a sleep stage independent apnea detection system for sleep stage W; and

[0087] Fig.15 Results of an analysis of the performance of an apnea detection system according to the invention for sleep stage W are shown. DETAILED DESCRIPTION

[0088] The present invention will be described with reference to the accompanying drawings.

[0089] It should be understood that the detailed description and specific examples, although indicating exemplary embodiments of the device, system and method, are intended only for illustrative purposes and are not intended to limit the scope of the invention. These and other features, aspects and advantages of the device, system and method of the present invention will be better understood based on the following description, the appended claims and the accompanying drawings. It should be understood that the drawings are only schematic and are not drawn to scale. It should also be understood that throughout the drawings, the same reference numerals are used to represent the same or similar parts.

[0090] The present invention provides a method and system for detecting sleep apnea, wherein intermediate characteristics, in particular sleep stages, are determined, and apnea events are detected based on physiological sensor signals using a detection algorithm that depends on the determined sleep stages. By taking into account intermediate characteristics such as sleep stages when performing an automatic apnea detection process, the accuracy of apnea detection can be improved.

[0091] Figure 1 A typical hypnogram is shown indicating the various sleep stages and cycles in adult sleep, where the stages are represented as a function of time. This is reproduced from https: / / www.howsleepworks.com / types_cycles.html.

[0092] Throughout the night, sleep in a healthy adult progresses as a series of sleep cycles of non-REM and REM sleep. The average duration of a sleep cycle is about 90 minutes. Each sleep cycle follows multiple stages of non-REM sleep (divided into stages N1, N2, and N3). After a period of sleep stage N3, these stages may progress back to N1, or a period of waking up, or a brief period of REM sleep before returning to stages N1 to N3 in a new sleep cycle. Typically, the time spent in sleep stage N3 decreases as the night progresses, while the duration of REM sleep increases.

[0093] Figure 1 The sleep stages shown in the chart are defined by the American Academy of Sleep Medicine (AASM):

[0094] N1: Typically, the first stage of sleep usually lasts between 1 and 5 minutes. It involves the transition period from wakefulness to sleep. The brain waves emitted by the brain slow down.

[0095] N2: Characterized by a lower heart rate and lower core body temperature. Muscle activity decreases. This stage accounts for about 50% of sleep.

[0096] N3: Also called deep sleep or slow-wave sleep due to the typical low-frequency waves visible in EEG recordings. In this stage, brain waves are least comparable to the waking state. Muscle activity is further reduced. It is difficult to awaken a sleeper in a deep sleep state.

[0097] REM: Characterized by rapid eye movements. High-frequency waves occur. Although most muscles are paralyzed during this stage, brain waves are most similar to the waking state.

[0098] People with sleep apnea have reduced N3 and REM durations due to their sleep being fragmented by arousals caused by interrupted breathing. Typically, sleep stages oscillate between W (wakefulness) and N1.

[0099] As mentioned above, automatic sleep stage detection has witnessed substantial development in recent years based on heart rate variability (HRV) features extracted from ECG signals or even PPG signals. For example, see:

[0100] [1]Berry, Richard B. et al. "Rules for Scoring Respiratory Events in Sleep: Update of the 2007AASM Manual for the Scoring of Sleep and Associated Events: Deliberations of the Sleep Apnea Definitions Task Force of the American Academy of Sleep Medicine." Journal of Clinical Sleep Medicine: JCSM: Official Publication of the American Academy of Sleep Medicine 8.5(2012):597–619; and

[0101] [2] Fonseca P, Long

[0102] The present invention utilizes physiological sensors to provide automatic detection of apnea events. The detection algorithm takes into account sleep stage. Sleep stage can be manually input into the apnea detection algorithm, but in a preferred example, sleep stage detection is also automatic. In addition, the same sensor pattern as apnea detection can be used to automatically identify sleep stage.

[0103] Figure 2 A system according to an example of the present invention is shown.

[0104] The system includes a physiological sensor 10 that captures a sensor signal S from a human body. The signal is a signal that enables heart rate variability to be obtained. To this end, the sensor 10 may include a PPG sensor or an ECG sensor. The two sensing modalities may be used in combination to provide more robust heart rate monitoring information.

[0105] The sensor signals are provided to a processor 11 and the signal processing functions performed by the processor are described below.

[0106] The raw sensor signal S is processed by the sampling unit 12 according to the time length, and the sampling unit 12 samples the sensor signal according to the sequence of time ranges. The wth time range in the sensor signal S is represented by S W express.

[0107] The first feature extraction unit 14 is used for sleep stage detection. For a given time range w, the sensor signal S W Extract features F in the time range S W , thereby assigning a sleep stage to the time range w. This is the first feature set. The duration of the time range is in the order of minutes, for example 2 to 10 minutes, and the time ranges may overlap to define a moving time window.

[0108] Feature set F S W is input to the sleep stage detection unit 16. The feature extraction unit 14 may include additional pre-processing / post-processing steps such as data cleaning, feature normalization and transformation.

[0109] The sleep stage detection unit 16 selects from the set C SSelect the sleep stage, the set C S Contains labels for all possible sleep stages, for example defined as: C S ={W, N1, N2, N3, R}. W is the awake state, and the other sleep stages are outlined above. As explained further below, each state is associated with a different apnea model for detecting apnea events within each specific sleep stage.

[0110] Instead, a different set of sleep stages may be used, for example defining more or fewer non-REM stages.

[0111] For example, the set may be defined as Cs = {W, N1 and N2 combined, N3, REM}. In practice, it is difficult to separate N1 and N2 based on HRV characteristics alone, so in a system that implements a sleep stage classifier to determine which apnea model to use, this may be a combination of stages.

[0112] The further simplified set can be defined as S = {W, non-REM, REM}. In terms of HRV characteristics during sleep, the biggest differences will be in these terms.

[0113] Alternatively, the set C S Clusters of coherent HRV features can be included that define them without having to map to actual sleep stages. An apnea model can then be linked directly to each cluster and the actual sleep stage does not need to be identified. This approach also avoids the problem that the mapping between sleep stages and HRV is not completely clear due to differences between central nervous system activity (CNS, which regulates sleep) and autonomic nervous system activity (ANS, which is reflected by HRV). By trying to map from HRV features to sleep stages to a sleep apnea model, the best possible model may not be derived. By mapping directly from HRV to a sleep apnea model, the differences between the CNS and ANS can be avoided.

[0114] In the following description, it is assumed that sleep stage recognition exists.

[0115] The sleep stage classifier in the sleep stage detection unit 16 is based on the feature set F provided to the input S W Set C S Sleep stage category C S W is assigned to the time range w. The classifier for sleep stage detection is trained offline.

[0116] A set 18 of pre-trained classifiers for detecting sleep apnea is provided in memory. The set M of sleep stage specific apnea classifiers is given by M = {M W , MN1 , M N2 , M N3 , M R Each member of the set includes one or more sensor signal characteristics that can be used to identify apnea events within that particular sleep stage. Note that the classifier M W It may only be necessary to include an indication that apnea events are likely not to occur during wake periods, since these only occur during sleep.There are at least two other sleep stage specific apnea classifiers in the set M.

[0117] Based on the detected sleep stage C S W The apnea classifier model selection unit 20 selects the apnea classification model M in the set M. W .

[0118] The selected model is then used in the apnea detection unit 22 .

[0119] The apnea classification models in set M are trained offline.

[0120] With the first feature set F S W In parallel, for each time range w, the sensor signal S W Extract the second feature set F O W Specially for apnea detection. This is performed by the second feature extraction unit 24.

[0121] The types of features extracted within the time range for apnea detection may be fixed, but they may alternatively depend on the sleep stage class assigned to the time range w by the sleep stage detection unit 16. Figure 2 As shown in FIG. 1 , the assigned sleep stage class C S W is the input to the second feature extraction unit 24.

[0122] The features to be extracted for a specific sleep stage class are determined during the offline training phase.

[0123] In the apnea detection unit 22, an output is generated representing a binary detection or non-detection of apnea. Therefore, there is a set CO of only two class members, CO = {'apnea', 'non-apnea'}. For each time window, one of the set members is provided as output C O W The apnea detection unit 22 receives the feature F extracted for apnea detection. O W and sleep stage specific classification model MW as input.

[0124] As mentioned above, the extracted features are related to the heart rate variability information.

[0125] HRV characteristics can be divided into the following categories:

[0126] Linear characteristics in time domain;

[0127] Frequency band power characteristics;

[0128] Nonlinear characteristics describing the irregularity of the inter-beat interval (IBI) series;

[0129] Hilbert transform and discrete wavelet transform features.

[0130] Different examples of these features may provide the best correlation with sleep apnea events for different sleep stages. This correlation is determined during the general system training phase.

[0131] Therefore, each classification model M W， M N1 , M N2 , M N3 , M R is based on different sets of HRV features. Then, feature F O W These are the features to which a classification model can be applied.

[0132] The number of HRV features corresponding to a particular sleep stage that are ultimately used in each classification model can be pre-specified in an offline training phase. Moreover, the number of features (i.e., predictors) selected for classification can be automatically adjusted, for example, using an algorithm that determines which subset of available features yields the best classification performance.

[0133] For example, the training phase involves applying all available HRV features (e.g., there are 155 such features) and applying a feature selection process whose goal is to find a subset of features that maximizes the intervals between sleep apnea classes in a particular sleep stage. This means that the amount of features that will ultimately be used to detect sleep apnea events may vary in different sleep stages.

[0134] Depending on the type of classification model used to construct the sleep apnea classifier, a combination, weighted combination, or other function of HRV features can be made. Some classifiers can be interpreted as a function that extracts a certain set of features from the input, and the classification probability output by the function is based on a weighted combination of the selected features.

[0135] Furthermore, different examples of these features may provide better correlation with sleep stages than sleep apnea events. Therefore, different features are selected in the two feature extraction units. Again, this correlation is determined during the general system training phase.

[0136] For example, different HRV information will be best suited for reliable sleep stage detection (i.e., F S W ) and apnea detection (i.e. F O W ).

[0137] Different HRV information between different sleep stages is also related to apnea detection.

[0138] Figure 3 The distribution of specific HRV features (low frequency band power, y-axis) across different sleep stages (x-axis) for a dataset is shown, where all data samples are labeled as sleep apnea (left column of each pair) and non-sleep apnea (right column of each pair). It shows that the feature distribution is different for each sleep stage. This also shows that applying a sleep stage-specific sleep apnea classifier can achieve a larger classification interval. Since specific features contribute more or less to the prediction effect, the model built to maximize the separation of sleep apnea classes within each sleep stage (M W ...M R , etc.) are likely to be different (i.e., contain different features as a result of the resulting feature selection process, and possibly with different weights).

[0139] Note that Figure 2 The different processing sub-units are shown in order to explain the underlying processing. However, it is obvious that in practice all these units can be implemented by a single processor under the control of an appropriate program. It can be seen that the inputs to the processor are the sensor signals and information about the classifier M, which will be used to identify apnea at different sleep stages. However, the processing tasks may be distributed among multiple processors; some may be local to the sensor and others may be remote.

[0140] Sleep stages can be scored manually by a system operator based on visual analysis of PSG, EEG, EOG, and EMG signals, following established rules, such as following the manual of the American Academy of Sleep Medicine (AASM). In such an embodiment, the sleep stages are input to the processor, rather than the processor performing automatic sleep stage detection. In such an embodiment, the feature extraction unit 14 and the sleep stage detection processing block 16 for sleep stage detection are replaced by the sleep stages from this manual classification process input.

[0141] The system is preferably implemented as a wearable sensor, which can be worn during sleep. Figure 4 The system 30 is shown as a wrist-worn sensor (as seen in the top image) that is worn throughout the night (as shown in the bottom image).

[0142] The wearable device comprises a sensor 10 that captures a physiological signal 10 and a processing unit 11 or a portion of the processing unit 11. Some or all of the processing may be performed locally in the worn system, and some or all may be performed in a backend system that communicates with the worn system via a wireless communication channel.

[0143] In a preferred embodiment, a photoplethysmographic (PPG) sensor is used.

[0144] A PPG sensor is a pulse oximeter. While the purpose of this sensor is to obtain a measure of blood oxygen saturation, it can also detect changes in blood volume in the skin, thereby performing PPG sensing. By detecting changes in blood volume, a cyclic signal corresponding to the pulse is obtained. Therefore, PPG sensors (such as pulse oximeters) are often used to provide a measure of pulse rate.

[0145] The PPG sensor contains at least one LED and a light sensor. The LED and sensor are placed so that the LED directs light into the user's skin, which is reflected or transmitted and detected by the sensor. The amount of reflected / transmitted light is determined, among other things, by the perfusion of blood within the skin.

[0146] A PPG system includes, for example, a red LED, a near infrared LED, and a photodetector diode. The sensor is typically configured with an LED and a photodetector diode directly on the subject's skin.

[0147] The LED emits light of different wavelengths, which diffuses through the vascular bed of the skin and is received by a photodetector diode. The varying absorbance at each wavelength is measured, allowing the sensor to determine the absorbance due only to pulsating arterial blood, excluding, for example, venous blood, skin, bone, muscle, and fat. The resulting PPG signal can then be analyzed.

[0148] Other simpler versions of the system for obtaining PPG data may be used, including versions with a single light source of one or more wavelengths. The absorption or reflectance of the light is modulated by the pulsating arterial blood volume and detected using a photodetector.

[0149] In transmission pulse oximetry, the sensor device is placed on a thin part of the subject's body. Reflection pulse oximetry can be used as an alternative to transmission pulse oximetry. This method does not require a thin part of the person's body and is therefore well suited for more general applications, such as the wrist shown above.

[0150] The basic design of a PPG sensor is, for example, a contact sensor with a single wavelength light source, such as green light (550nm), for measuring the PPG signal. The light source is pulsed at a certain light output frequency, such as 128Hz. The sampling frequency of the optical sensor is higher, such as 256Hz, so that it measures during and between activations of the light source. This allows the system to distinguish between the light emitted by the LED and the ambient light, thereby filtering out the ambient light from the signal received during the light source pulses.

[0151] In other known proposals, PPG data can be obtained from camera images, wherein ambient light and / or other light sources are used to illuminate tissue (e.g. skin). Thus, PPG measurements can be performed even at a distance from the tissue, in which case the light source and / or detector are not in contact with the tissue, such as in camera-based measurements.

[0152] PPG data may be obtained at one or more wavelengths, such as any number of wavelengths typically between 1 and 10, although even more than 10 wavelengths may be used.

[0153] Equipment and techniques for obtaining PPG data are well known in the art, and indeed many different PPG sensors are commercially available. For example, they are used in devices that measure heart rate during exercise.

[0154] As mentioned above, the interesting features of the PPG (or other) signal are related to the heart rate variability (HRV). To this end, the processing unit locates the heartbeat onset of each cardiac cycle in the PPG signal. Then, a noise-robust estimation of the inter-beat interval (IBI) time series can be performed. HRV features are then derived from the extracted IBI time series for sleep stage detection and apnea detection.

[0155] As described above, multiple sensors can be used to enable a more robust determination of features of interest (e.g., HRV information). Additionally, other sensors such as accelerometers can be used, whose data can be used to identify and filter out information corrupted by motion artifacts, as well as to extract additional features that can be used to improve the accuracy of sleep stage classification. Other sensors can be used to extract respiration information. For example, pulse oximetry (SpO 2 ) derived features and can be obtained from a multi-wavelength PPG sensor. These features describe changes in oxygen saturation or desaturation in the blood, which also have a strong correlation with sleep apnea events.

[0156] Proxy measures of respiratory effort can also be automatically derived from the raw PPG signal. These measures can then be used to help differentiate between various types of apnea events, such as obstructive sleep apnea (where flow is interrupted but respiratory effort continues) versus central sleep apnea (where respiratory effort also ceases). This can be achieved without adding additional sensors, as it can be acquired entirely from the PPG sensing.

[0157] Furthermore, other features of the PPG signal may be interpreted, such as morphological characteristics of the PPG signal associated with sleep stage specific properties or apnea events.

[0158] Figure 5 shows the situation where the sleep stage is determined or input into the system by Figure 2 The method comprises:

[0159] In step 40, a sensor signal for detecting sleep apnea is generated;

[0160] In step 42, determining a sleep stage or receiving an identification of a sleep stage as input; and

[0161] In step 44 apnea events are detected based on the sensor signals using a detection algorithm that depends on the determined sleep stage.

[0162] This approach can be extended by considering the sleep stage as a more general first characteristic related to the subject's physiological parameters during sleep.

[0163] The system of the present invention has been tested by applying the processing to overnight polysomnographic recordings of subjects with sleep-disordered breathing and sleep movement disorders, including polysomnographic recordings.

[0164] As a source of physiological signals, PPG data were used, where the sensor was positioned on the subject's finger. Reference hypnograms from sleep epochs were obtained by expert review of sleep stages (automatically classified by a sleep scoring system). Reference apnea and other sleep disorder events were scored in the same procedure.

[0165] The dataset contains 55 sleep epochs with a total duration of 10,131 minutes, including 2,876 minutes of annotated apnea events.

[0166] The beat onset positions were derived from the finger PPG and converted into IBI time series. A total of 155 different HRV features were extracted from the IBI time series.

[0167] The offline feature selection and model training process provides a unique model for each sleep stage based on a subset of features that maximizes sleep apnea class separation for that sleep stage. These 155 features fall into the feature categories as described above.

[0168] Features are extracted from sliding windows (ie, time spans) with a length of 270 seconds and an overlap of 30 seconds.

[0169] Two experiments were conducted to verify the hypothesis that applying a sleep stage specific apnea classifier can achieve greater class separation compared to the use case of applying a non-sleep stage specific apnea detection model. To evaluate the accuracy of apnea detection in practice, a 10-fold cross validation was performed by splitting the data into training / testing parts so that the receiver operating characteristics (ROC) could be studied.

[0170] In the first configuration A, an apnea prediction model that does not take into account sleep stage specific information is cross-validated. In other words, in the k-th fold, a general classification model is trained that maps the input feature space of the training set to the corresponding reference apnea event labels. As a classifier, a logistic regression model is used. During the training phase, feature selection (recursive feature elimination) is applied to the k-th fold to enhance generalizability by reducing overfitting. This is part of the model building process.

[0171] From the highest ranked features, features that appear at least six times constitute the final feature set.

[0172] In the test phase, the test data in each stage is divided by the reference (ground truth) sleep stage assigned to the test epoch, and the apnea detection performance is evaluated for each sleep stage class. More specifically, a ROC curve is constructed for each k-th retrieval based on the set of apnea class probability estimates obtained by testing instances originating from a particular sleep stage class.

[0173] In the second configuration B, the above method was evaluated. The training data was classified by reference sleep stages. For each sleep stage category, a dedicated apnea event detection model was trained once in each k-th fold. During training, the same logistic regression classification scheme and feature selection strategy as in configuration A were applied.

[0174] During the testing phase, the focus was on investigating the improvement in apnea detection accuracy by applying sleep stage specific apnea detection models. To test whether an apnea event occurred during a given epoch, the reference stage associated with that epoch was used to select an appropriate sleep stage specific apnea detection model. In other words, a hypothetical perfect sleep stage classifier was simulated instead of a sleep stage classifier trained with actual feature data. Also in this case, for each k-th fold, a ROC curve was constructed from the obtained apnea class probability estimates. This performance evaluation scheme is equivalent to partitioning the test data by reference sleep stage and evaluating the apnea detection performance by each sleep stage class.

[0175] The results are shown as a curve of the ROC signal for the 10-fold, that is, the true positive rate versus the false positive rate. The graph also shows the line for the 10-fold, the unit gradient line (dashed line) and the average graph for all 10-fold (also dashed line).

[0176] Figure 6 Results are shown for configuration set A (apnea detection independent of sleep stage) for a test set containing test instances (epochs) originating from sleep stage R. Figure 7 The corresponding results for configuration B (apnea detection taking sleep stages into account), again for sleep stage R, are shown.

[0177] Figure 8 Results are shown for configuration set A (sleep stage independent apnea detection) for a test set containing test instances (epochs) originating from sleep stage N1. Fig. 9 The corresponding results for configuration B (apnea detection taking sleep stages into account) are shown, again for sleep stage N1.

[0178] Fig.10 Results are shown for configuration set A (apnea detection independent of sleep stage) for a test set containing test instances (epochs) originating from sleep stage N2. Fig.11 The corresponding results for configuration B (apnea detection taking sleep stages into account) are shown, again for sleep stage N2.

[0179] Fig.12 Results are shown for configuration set A (apnea detection independent of sleep stage) for a test set containing test instances (epochs) originating from sleep stage N3. Fig.13 The corresponding results for configuration B (apnea detection taking sleep stages into account) are shown, again for sleep stage N3.

[0180] Fig.14Results for configuration set A (apnea detection independent of sleep stage) are shown for a test set containing test instances (epochs) originating from sleep stage W. Fig.15 The corresponding results for configuration B (apnea detection taking sleep stages into account), again for sleep stage W, are shown.

[0181] Table 1 below summarizes the apnea prediction accuracy for each sleep stage category as measured by the mean area under the ROC curve (AUC), which was constructed from the 10-fold cross-validation responses obtained in configurations A and B.

[0182] Table 1

[0183]

[0184]

[0185] The results clearly show that applying a sleep stage specific apnea detection model outperforms apnea event detection accuracy compared to a system that applies a generic (non sleep stage specific) apnea detection system.

[0186] Among the tested sleep stages, the ROC curves obtained by validation configuration B generally have lower variance, and the curves in configuration A are generally closer to the diagonal line of the ROC space representing the random guess line. In addition, the AUC values ​​for each evaluated sleep stage category obtained in configuration B are generally higher compared to the results of configuration A.

[0187] In the example above, a distinction is made between apneas and non-apneas. Based on the different resulting HRV patterns, detection can be extended to differentiate between apneas and hypopneas, as well as OSA and CSA. For example, as mentioned above, since the presence or absence of respiratory effort is a key factor in distinguishing these two types of apneic events, other PPG features, particularly those related to alternative measures of respiratory effort, can be used to differentiate between OSA and CSA.

[0188] The detailed examples above relate to the identification of sleep stages. However, as mentioned above, the invention more generally relates to any intermediate characteristics which then enable selection between multiple more tailored apnea detection algorithms. As mentioned above, there may be clusters or groups of coherent HRV features which can then be mapped to separate algorithms for apnea detection. This would avoid the need for actual sleep stage classification. A two-stage apnea detection approach would still provide improved apnea detection accuracy.

[0189] As described above, the system utilizes a processor to perform data processing. The processor can be implemented in a variety of ways using software and / or hardware to perform the various functions required. The processor typically employs one or more microprocessors that can be programmed using software (e.g., microcode) to perform the required functions. The processor can be implemented as a combination of dedicated hardware for performing some functions and one or more programmed microprocessors and associated circuits for performing other functions.

[0190] Examples of circuits that may be used in various embodiments of the present disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field programmable gate arrays (FPGAs).

[0191] In various implementations, the processor may be associated with one or more storage media, such as volatile and non-volatile computer memory, such as RAM, PROM, EPROM, and EEPROM. The storage media may be encoded with one or more programs that perform the desired functions when run on one or more processors and / or controllers. The various storage media may be fixed within the processor or controller, or may be transportable so that one or more programs stored thereon can be loaded into the processor.

[0192] Those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed invention by studying the drawings, the disclosure and the claims. In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality. Although specific measures are recited in mutually different dependent claims, this does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be interpreted as limiting the scope.

Claims

1. A system for detecting sleep apnea, comprising: a physiological sensor (10) for generating a sensor signal for detecting sleep apnea; A processor (11) adapted to: determining a sleep stage or receiving as input an identification of a sleep stage; and Apnea events are detected based on the sensor signal using a second feature of the sensor signal and a sleep stage-specific classifier, wherein the sleep stage-specific classifier is selected from a set of sleep stage-specific classifiers depending on the sleep stage, the set of sleep stage-specific classifiers comprising at least two sleep stage-specific classifiers for different non-awake sleep states, wherein the second feature comprises heart rate variability information and each sleep stage-specific classifier is based on a different set of heart rate variability features.

2. The system according to claim 1, wherein: The processor is adapted to determine sleep stages based on the heart rate variability information derived from the sensor signal (S).

3. The system according to claim 1 or 2, wherein: The physiological sensor (10) includes a PPG sensor and / or an ECG sensor.

4. The system according to claim 1 or 2, wherein: The processor (11) is adapted to: extracting heartbeat timing from the sensor signal; exporting the beat-to-beat interval time series; and The heart rate variability information is extracted from the beat interval time series.

5. The system according to claim 1 or 2, wherein: The processor (11) is adapted to: sampling the sensor signal as a series of time ranges; extracting a first feature of the sensor signal from each time range; and The sleep stage for each time range is determined according to the first feature.

6. The system according to claim 5, wherein: The processor (11) is adapted to: for each time range, selecting the sleep stage specific classifier to be used for identifying sleep apnea from the sensor signal, the selection being dependent on a previously determined sleep stage; as well as The second feature of the sensor signal is extracted from each time range.

7. The system according to claim 6, wherein: The second feature is extracted using a model that depends on previously identified sleep stages.

8. The system according to claim 5, wherein: The first feature includes a heart rate variability feature.

9. A method for detecting sleep apnea, comprising: (40) generating a sensor signal for use in detecting sleep apnea; (42) determining a sleep stage of the subject during sleep or receiving as input an identification of the sleep stage; and (44) detecting apnea events based on the sensor signal using a second feature of the sensor signal and a sleep stage-specific classifier, wherein the sleep stage-specific classifier is selected from a set of sleep stage-specific classifiers depending on the sleep stage, the set of sleep stage-specific classifiers comprising at least two sleep stage-specific classifiers for different non-awake sleep states, wherein the second feature comprises heart rate variability information and each sleep stage-specific classifier is based on a different set of heart rate variability features.

10. The method according to claim 9, wherein: The method comprises determining the sleep stage based on the heart rate variability information derived from the sensor signal.

11. The method according to claim 9 or 10, comprising: extracting heartbeat timing from the sensor signal; Export beat-to-beat interval time series; and The heart rate variability information is extracted from the beat interval time series.

12. The method according to claim 9 or 10, wherein: Sampling the sensor signal into a sequence of time ranges; extracting a first feature of the sensor signal from each time range; and The sleep stage for each time range is determined according to the first feature.

13. The method according to claim 12, comprising: for each time range, selecting the sleep stage specific classifier to be used in identifying sleep apnea from the sensor signal; as well as The second feature of the sensor signal is extracted from each time range using a model that depends on previously identified sleep stages.

14. A computer program product storing a computer program comprising computer program code means adapted to implement the method according to any one of claims 9 to 13 when said program is run on said processor of the system according to any one of claims 1 to 8.

Images