Method, computing device, and wearable device for sleep stage detection

Abstract

To provide a method for detecting sleep stages including rapid eye movement (REM) and non-REM (NREM) based on machine learning.SOLUTION: A method for sleep stage detection includes the steps of: receiving a first vital sign feature including a plurality of first feature values corresponding to each of the plurality of periods; and executing a first logistic regression operation, for each intermediate period of the plurality of periods, to calculate a first indicated value representative of a sleep stage of the intermediate period based on a corresponding first feature value and first feature values of a preceding period and a subsequent period of the plurality of periods.SELECTED DRAWING: Figure 7

Description

[Technical field]

[0001] The present disclosure relates generally to methods, computing devices, and wearable devices for sleep stage detection, and more particularly to methods, computing devices, and wearable devices for rapid eye movement (REM) sleep detection based on machine learning. Methods for detecting sleep stages including rapid eye movement (REM) and non-rapid eye movement (NREM) sleep stages, Chairs and wearable devices. [Background technology]

[0002] It is known that a person's physical and mental state is affected by the quality of their sleep. Quality sleep is essential for maintaining health, happiness, and good mood. During sleep, humans typically alternate between rapid eye movement (REM) and non-REM (NREM) stages. Several studies have shown that the REM stage plays an important role in regulating mood and memory consolidation. Furthermore, it has been shown that depressive disorders are closely related to REM sleep dysregulation, such as long REM sleep duration and short REM latency. ,Monitoring sleep stages can provide important information for the ,analysis of sleep behavior and stress management, as well as the ,treatment of depressive disorders, elderly care and performance ,analysis.,The quality of the analysis and therefore the quality of the ,treatment depends heavily on the collected data, both in terms of ,quality and quantity., ,To meet these requirements, the monitoring devices used are ,typically hardware intensive and complex.

[0003] The current gold standard for sleep research is polysomnography (PSG). This requires a number of sensors, including electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), electrocardiography (ECG), and respiration sensors. PSG provides more information about sleep studies. However, they are highly immobile and expensive, making them impractical for large-scale observations.

[0004] U.S. Patent No. 9,655,559 describes an automated sleep stage measurement system using wearable sensors. The disclosed arrangement suffers from several drawbacks, including poor accuracy and susceptibility to poor sensor readings.

[0005] A method, computing device, and wearable device for sleep stage detection that addresses at least one of the shortcomings of the prior art and / or is useful to the public. It is desirable to provide methods, computing devices, and wearable devices that provide new and / or alternative options. Summary of the Invention

[0006] According to one aspect, a method of sleep stage detection includes receiving a first vital sign feature including a plurality of first feature values ​​corresponding to respective epochs of a plurality of epochs; and receiving, for each intermediate epoch of the plurality of epochs, a first The indicator value is calculated based on the corresponding first characteristic value and the preceding and succeeding epochs of the plurality of epochs. and performing a first logistic regression operation to calculate a first feature value for the period.

[0007] The above-described embodiment is highly advantageous. For example, for each intermediate period of the plurality of periods, a first indication value representative of a sleep stage of each intermediate period of the plurality of periods can be calculated by performing a first logistic regression operation based on first feature values ​​of the preceding, intermediate and subsequent periods of the plurality of periods. This is advantageous in that not only are feature values ​​from preceding and intermediate epochs used, but also feature values ​​from subsequent epochs are used to achieve improved accuracy. In one embodiment, the method does not require a lot of computing resources to be implemented, either in hardware form, software form, or a combination of both, since the amount of data and processing required is minimal. Thus, the method is particularly suited to implementation using wearable devices, but is not limited thereto.

[0008] The first vital sign feature comprises a convolution of heart rate, pulse shape variability, and heart rate variability. These vital sign features may be related to one of the high frequency powers detected during sleep. It has been shown to be related to the sleep stage and is therefore useful for sleep stage detection.

[0009] Preferably, the first logistic regression operation includes weighting and a sigmoid operation. The logistic regression operation is less complex and more interpretable than a non-linear model. In the field of machine learning, it is known that complex models have a higher risk of overfitting compared to simple models. Therefore, by using the logistic regression operation, the sleep stage detection according to an embodiment is suitable for execution, for example, on a wearable device. The relatively complex algorithm used in conjunction with the convolutional neural network for extracting the HRV PSD including three dimensions may be executed, for example, on a server of a cloud service. Extraction of the HRV PSD.

[0010] The method further comprises: receiving a second vital sign characteristic of the subject during each of the plurality of epochs; a second indication value representative of a sleep stage of the intermediate period based on the corresponding second characteristic value and second characteristic values ​​of the preceding and succeeding periods of the plurality of periods; performing a second logistic regression operation to calculate the performing a further logistic regression calculation to determine a sleep stage for each intermediate period among the first and second indication values ​​based on the corresponding first and second indication values.

[0011] In the above embodiment, for example, a plurality of vital sign features are received, a plurality of logistic regression calculations are performed based on the respective vital sign features to calculate respective indication values, and further logistic regression calculations are performed based on these indication values ​​to determine the sleep stage of each intermediate period. By taking into account the information of the plurality of vital sign features, an improved result of sleep stage detection can be achieved. By taking into account the plurality of vital sign features and performing further logistic regression calculations, the sleep stage can be detected with a better accuracy.

[0012] One of the logistic regression operations may have a weighting value that is different from a weighting value of another of the logistic regression operations.

[0013] Preferably, the method further comprises: receiving a third vital sign feature including a feature value of the plurality of time periods; for each intermediate time period among the plurality of time periods, a third indication value is calculated based on the corresponding third characteristic value and the third characteristic values ​​of the preceding and succeeding time periods among the plurality of time periods. and performing a third logistic regression operation, and a logistic regression calculation including: a sleep stage of each intermediate period among the plurality of periods is calculated by: The determining step is further based on a third indication of the degree of freedom of movement of the subject.

[0014] This configuration may be used, for example, where the corresponding feature value of the third vital sign feature is not compensated; This is advantageous because the third reading may not represent a sleep stage. Such a third indication, when used in the further logistic regression calculation, may provide a sleep stage indication. Therefore, by using the third indication only when it represents a sleep stage, accurate sleep stage detection results may be achieved. Cut.

[0015] According to another aspect, a method of creating a model for logistic regression includes receiving vital sign features associated with baseline sleep stage information; generating a cross-validation set from the vital sign features and the baseline sleep stage information, the cross-validation set including a plurality of sub-samples; computing each of the sub-samples with respect to other of the sub-samples using a model with each of a plurality of machine learning parameter sets; and associating one of the parameter sets with the model based on a comparison of the baseline sleep stage information to a result of the computation. The law is provided.

[0016] In one embodiment, machine learning techniques can be used to efficiently identify correlations between vital sign features and baseline sleep stage information to determine a parameter set for use in the model. For example, the parameter set may be derived for a first sleep period. a first set of parameters retrieved during a second sleep period; In another example, a first set of parameters may be retrieved for a first sleep period, a second set of parameters may be retrieved for a second sleep period, and both the first and second sleep periods may be retrieved and / or stored and programmed into a trend model to identify trends in a user's sleep patterns. A third set of parameters may be retrieved for a third sleep period, and the third sleep period may be retrieved and / or stored and programmed into a trend model. That is, the parameter set for one sleep period may differ from the parameter set for another sleep period. In another example, a first set of parameters may be retrieved for a first sleep period, a second set of parameters may be retrieved for a second sleep period, and both the first and second sleep periods may be retrieved and / or stored and programmed into a trend model to identify trends in a user's sleep patterns. A third set of parameters may be retrieved for a third sleep period, and the third sleep period may be retrieved and / or stored and programmed into a trend model. The third sleep period is separate in space from the first and second sleep periods. and / or can be stored and programmed into a trend model.

[0017] The vital sign features may relate to one of heart rate, pulse shape variability, and convolved high frequency power of heart rate variability. These vital sign features may relate to sleep stages. It has been shown to be relevant and thus useful for sleep stage detection. In another embodiment, the vital sign features may also relate to the sum of the high frequency power of heart rate variability, the sum of the low frequency power of heart rate variability, the LF / HF ratio of heart rate variability, and the standard deviation of the heart rate. These vital sign features may be derived from current and / or historical data from another device and thus useful for sleep stage detection.

[0018] Preferably, the method further comprises receiving another vital sign feature associated with said reference sleep stage information, and said cross-validation set is further generated from said another vital sign feature. By taking into account the another vital sign feature, the cross-validation set contains more information that can be used to accurately detect sleep stages.

[0019] Slicing the vital sign features by at least three consecutive epochs. The at least three consecutive periods may include a middle period among the consecutive periods. The at least three consecutive epochs may include a preceding epoch, the intermediate epoch, and a succeeding epoch of the consecutive epochs. By considering the succeeding epochs, the sleep stage can be detected more accurately. In the described embodiment, the at least three consecutive epochs include eleven epochs. Additionally, the at least three consecutive periods are within a total window period of at least 30 minutes. The sleep data may be longer than one hour in duration (e.g. For example, 7 to 8 hours. In one embodiment, the sleep data is divided into hourly periods or segments. "Slicing" the data allows the data to resemble sleep cycles and therefore allows efficient calculation of sleep stages.

[0020] The vital sign features may be normalized. Additionally, the vital sign features may be compensated for missing parts. For example, missing parts may be compensated for using max pooling. Such techniques are useful for reducing the adverse effect of such missing parts on the detection results that may lead to false positive or false negative detections. Alternatively, such missing parts may be replaced with random values, average values, nearest values, linear interpolation, etc., depending on the implementation.

[0021] According to another aspect, there is provided a method for extracting heart rate variability features, comprising: convolving a high frequency portion of a heart rate variability power spectral density with a convolution filter associated with a convolutional neural network model to generate a plurality of convolved values ​​representing respective patterns of heart rate variability; and selecting one of the convolved values ​​based on a result of an activation function operation performed on the convolved values.

[0022] In one disclosed embodiment, the method further comprises detecting low power in the high frequency portion of the REM The high frequency portion of the power spectral density of heart rate variability and rapid eye movement (REM) ) is taken into account.

[0023] The selected convolution value preferably corresponds to the maximum value resulting from the executed activation function, the maximum value corresponding to REM and / or NREM sleep stages advantageously providing a coordinated approach to sleep stage derivation.

[0024] Moreover, the high frequency portion is preferably in the range of 0.15 Hz to 0.4 Hz. Dynamic power spectral density has been shown to be highly relevant to sleep stages in this frequency range, and by considering the spectral information in this frequency range, the resulting heart rate variability features can better indicate sleep stages.

[0025] The power spectral density can be normalized across its high and low frequency parts. Normalization throws away the scale of the data, but in return provides a consistent power spectral density for adaptation.

[0026] Preferably, the activation function operation relates to one of a rectified linear function, a soft real, and a sigmoid function.

[0027] According to another aspect of the present disclosure, a method of creating a model for extracting heart rate variability features is disclosed, comprising receiving vital sign features representative of heart rate variability power spectral density in association with baseline sleep stage information, and creating a model including a convolution filter using machine learning based on the vital sign features with reference to the baseline sleep stage information. By using such a model to derive a physiological signal, high frequency components of heart rate variability serve as a relatively consistent basis for sleep stage detection.

[0028] The vital sign features are preferably derived from physiological signals for 3 minute intervals within each 5 minute interval, which can reduce power consumption without significantly affecting the accuracy of sleep stage detection. Alternatively, the vital sign features may be derived from physiological signals for an epoch that is shorter than the epoch.

[0029] Preferably, said vital signs features are normalized.

[0030] The machine learning preferably includes a convolutional neural network. A convolutional neural network (CNN) algorithm for training a model can correlate information in both HRV PSD and sleep stages. The HRV PSD can be compared to the circadian rhythm over a period of time. This can be seen as an image that may contain patterns of HRV changes in the wavenumber domain. Alternating low and high power peaks in the HRV PSD that correspond to REM and NREM sleep stage patterns. A turn can be observed in the experimental data. Using the CNN algorithm, It is possible to extract information from images like this.

[0031] In one disclosed embodiment, one advantage of using a CNN algorithm is that the CNN algorithm The goal is to process the HRV PSD into a calculated value related to the REM / NREM sleep stages by optimizing convolution filters, activation functions, data pooling methods, and neural network weights that can minimize or reduce the prediction error. In other words, the CNN algorithm The HRV PSD is summarized by training a model to calculate values ​​representing sleep stages or by other methods. The present disclosure in one exemplary configuration advantageously processes data in at least three dimensions, including frequency, time domain, and power density of HRV.

[0032] The collected or acquired information can be used to eventually apply other information acquired from another device, along with information collected from the wearable device. Conversely, information such as HR, HRV, and pulse shape variability can be acquired from any type of sensor and used to measure the mechanical Learning algorithms can refine the information further.

[0033] According to another aspect, a method of deriving medical dosing includes: A method is disclosed that includes receiving a sleep stage signal between the first and second time periods, receiving and comparing sleep stage information between the first and second time periods, and providing baseline sleep stage information from which normative medical dosing occurs.

[0034] According to another aspect, a method of assessing responsiveness to a medical medication comprises: receiving sleep stage information for a first time period; and comparing said sleep stage information for said first time period with said sleep stage information for said second time period. receiving sleep stage information for a second time period for detecting a sleep state of the first and second time periods; and evaluating responsiveness to a medical medication administered during said period based on results of said comparison.

[0035] Exemplary said medical medication may relate to at least one of anticonvulsants, benzodiazepines, and emazepam.

[0036] According to another aspect of the present disclosure, a computer-readable medium is disclosed that includes instructions for causing a processor to perform any of the methods described above.

[0037] Preferably, the instructions are configured to be implemented in firmware. Due to their simplicity, at least some of the methods described above are particularly well suited for implementation in the form of instructions in firmware, although this is not limited thereto.

[0038] According to another aspect of the present disclosure, a computing device is disclosed that includes a processor and a storage device that includes instructions for causing the processor to perform any of the methods described above.

[0039] The storage device is preferably a firmware chip. As mentioned above, at least some of the methods described above can be implemented using firmware due to their simplicity and low hardware requirements.

[0040] According to another aspect of the present disclosure, a wearable device is disclosed that includes a storage device containing instructions for causing a processor to perform any of the methods described above.

[0041] The storage device is preferably a firmware chip. As mentioned above, at least some of the methods described above may be implemented using firmware.

[0042] The wearable device is preferably in the form of a wristwatch, which can facilitate data collection over long periods of time.

[0043] It is contemplated that features relating to one embodiment may be applicable to other embodiments. [Brief description of the drawings]

[0044] Exemplary embodiments will now be described with reference to the accompanying drawings. [Figure 1] 1 illustrates an exemplary system block diagram of a computing device according to one embodiment of the present disclosure. [Diagram 2] 2 shows a flowchart of a method of sleep stage detection performed by the computing device of FIG. 1. [Figure 3A] 2 shows a flowchart of a method for creating a logistic regression model executed by the computing device of FIG. 1 . [Figure 3B]1 shows an example of generating a training dataset. [Figure 4] 2 shows a flowchart of a method for extracting heart rate variability features performed by the computing device of FIG. 1; [Diagram 5] 2 shows a flowchart of a method for creating a model for extracting heart rate variability features executed by the computing device of FIG. 1; [Figure 6] 1. (a)-(c) show various scenarios of data error compensation performed by the computing device of FIG. [Figure 7] 2 illustrates a typical scenario of multiple logistic regression calculations performed by the computing device of FIG. 1 followed by a further logistic regression calculation and a sleep stage calculation including three vital sign features. [Figure 8] 2 illustrates an exemplary scenario of the computing device of FIG. 1 performing a logistic regression operation in a sliding window manner. [Figure 9] 2 shows line graphs of various signals at various stages of the calculations performed by the computing device of FIG. 1; [Figure 10] Figure 2 shows a Bland-Altman plot of the percentage error of sleep stage detection results obtained using the method. [Figure 11] 1 shows histograms of REM percentage and frequency for patients with major depressive disorder. [Figure 12A] 2 illustrates an embodiment of the computing device of FIG. 1 in the form of a watch or wearable device. [Figure 12B] FIG. 12B shows a schematic diagram of a head assembly of the wearable device of FIG. 12A. [Figure 12C] 12B shows a schematic cross-sectional view of the head assembly of FIG. 12A. [Figure 13] 12B illustrates an example of wireless transmission of calculated sleep stage information from the wearable device of FIG. 12A to another computing device in a typical situation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0045] FIG. 1 illustrates an exemplary computing device suitable for implementing some method embodiments of the present disclosure. FIG. 1 shows a system block diagram of a computing device 100. The computing device 100 includes a processor 101, a firmware chip 102, a storage device 103, and a physiological sensor 104. (e.g., an electrocardiogram sensor, a photoplethysmogram sensor, or a combination thereof) and a communication interface 105 (wired or wireless) operatively associated with the processor 101. It should be noted that any type of sensor capable of physiological measurements may be used. In one embodiment, the computing device 100 is a small device ( In another embodiment, the computing device 100 is in the form of a computer (e.g., a server). In one embodiment, the computing device 100 is in the form of a mobile device. Instructions that can be executed by processor 101 can be stored in firmware chip 102, storage device 103, or a combination thereof.

[0046] FIG. 2 illustrates a flow chart of a method 200 of sleep stage detection according to one embodiment of the present disclosure. Instructions embodying the method 200 may be stored and executed as described above. In an embodiment, the computing device 100 is a wearable sleep stage tracking device. In step 201, the processor 101 selects any suitable body part ( Physiological signals are collected from sensors 104 attached to the wrist, upper arm, index finger, neck, or head. Receive the number.

[0047] In step 202, the processor 101 derives from the physiological signal a vital sign representing the heart rate (HR), the convolved high frequency power of the heart rate variability (Conv_HF), and the pulse shape variability (PSV). In another embodiment, the vital sign feature may represent additional physiological features. The vital sign feature may embody a typical form of a series of feature values ​​each corresponding to a sleep stage in a corresponding epoch. In other words, each vital sign feature includes a plurality of feature values ​​each corresponding to a plurality of epochs. For example, taking the convoluted high frequency power of heart rate variability as an example, each feature value of the vital sign feature may represent a unique pattern of variability in each epoch. However, in another embodiment, the processor 101 may extract any number (e.g., one, two, or four) of vital sign features from the physiological signal.

[0048] Pulse shape variation (PSV) can be referred to as the standard deviation of a vital sign feature derived from the pulse shape of an input physiological signal (e.g., a PPG signal) and can be calculated, for example, based on pulse time interval information, skewness, magnitude, integral and differential information, frequency content, or derivatives thereof.

[0049] In step 203, the processor 101 compensates for missing data portions in the vital sign features, if such data portions exist. Missing data portions may be caused, for example, by poor sensor signal quality. Such compensation can reduce the impact of the missing data portions on the detection results. Compensation for missing data portions in the vital sign features is described in detail below in relation to FIG. 6. Compensation for missing data portions can include extraction of at least two series of inputs. Alternatively, compensation for missing data portions can include extraction of a set of other series of inputs. Another alternative is to extract at least two series of inputs. and a set of alternative combinations of inputs. In step 204, the processor 101 normalizes the vital sign features, whether compensated or not.

[0050] In step 205, the processor 101 selects at least one model (e.g., a trained The processor 101 then uses the model of the sleep stage of each epoch to calculate (e.g., determine, detect, or classify) the sleep stage of each epoch from the processed vital sign features. A machine learning classifier function is implemented to calculate the sleep stage for each epoch from the vital sign features. Based on the results of the calculation, the sleep stage is determined to be one of rapid eye movement (REM) and non-REM (NREM). This calculation step is described in conjunction with Figures 7, 8, and 9. The at least one model may be stored in the firmware chip 102 or in the storage device 103. However, the at least one model may be created, developed, or trained using machine learning on any computing platform, which will be described in more detail below. If the models are stored independently on the cloud server, the retrieval of the at least one model by the processor 101 is based on the processor 101 communicating, in one configuration, via a communication interface 105 that connects or links to the cloud server. The vital sign features may, in an alternative embodiment, be provided by the processor 101 to the cloud server that executes the at least one model for the calculation of the sleep stages. The processor 101 of the computing device 100 may then request (e.g., retrieve) the sleep stage calculation results from the cloud server. In one embodiment, the sleep stage calculation results may be transferred to another device, which may be programmed to combine or extract the sleep stage calculation results with information in the other device for calculation. In other words, some or all of the burden of the sleep stage calculation may be shifted from the processor 101 to an associated processing platform, such as a cloud server. Each The calculated or detected sleep stages are converted into a sleep stage hypnosis by the processor 101 or the like. can be processed into grams.

[0051] For purposes of step 205, for machine learning classification (e.g., logistic regression) A model is used. The creation (e.g., training) of the model is described below in relation to FIG. 3A. Once a new model is trained (i.e., a parameter set is associated with the model), the associated parameter set, which may include a set of constants, can be stored in the firmware chip 102 of the computing device 100. The model training / relearning or adaptive modeling can be performed in real-time, and if the parameter set is not stored in the firmware chip, it may need to be retrieved externally, for example from an associated server device that stores the parameter set. The parameter set can be transferred to a server hosting a cloud service to enable future long-term recording, trend analysis, and tracking of the user's sleep patterns. In contrast, existing PSG technology, due to its non-portable nature, does not allow such an efficient way of doing things. However, it cannot be used to check the user's sleep patterns.

[0052] It can therefore be seen that machine learning algorithms are used for sleep stage detection based on vital sign features. Machine learning algorithms can be implemented using available machine learning tools. Logistic regression can be implemented, for example, using the MATLAB® programming language. Machine learning algorithms (e.g., logistic regression) identify input features that can best separate the input features into two classes. A linear boundary is searched for that is a linear combination of the features. This can be done by reducing the error of the system in classifying the training data. The resulting output is a set of linear combinations of the features that are bypassed as represented by said one of the parameter sets. This is taken as the model. This model can be applied to a new set of inputs to detect the classification of the subject, i.e., REM or NREM sleep stage in the context of Figure 2. Machine learning can also be used to derive vital sign features, as described below.

[0053] FIG. 3A illustrates a method 300 for creating a logistic regression model according to one embodiment of the present disclosure. 1 shows a flowchart of a computing device 100 according to an embodiment of the present invention. , takes the form of a server device 100 configured to perform machine learning operations. The resulting model can be used for the purposes of step 205.

[0054] In step 301, the processor 101 receives a physiological signal via the communication interface 105 in association with the baseline sleep stage information. Alternatively, the physiological signal may be associated with the baseline sleep stage information. The baseline sleep stage information and physiological The biological signals may be received simultaneously or otherwise. The baseline sleep stage information is provided by a reference device, which may be a device that has been tested and approved by an agency.

[0055] The baseline sleep stage information may be derived, for example, from a medical device approved by the FDA to provide approximately 89% agreement to the gold standard of sleep stage detection.

[0056] In step 302, the processor 101 in this embodiment extracts vital sign features from the physiological signal, which represent heart rate (HR), convolved high frequency power of heart rate variability (Conv_HF), and pulse shape variability (PSV), respectively. The lucifer characteristics may represent additional physiological characteristics of other kinds.

[0057] In step 303, the processor 101 processes the vital sign features by slicing at least three consecutive epochs, in any suitable order, having a total span of a window period of at least one hour in this embodiment, compensating for any missing data portions, and normalizing the slices by the standard score of the slice. Alternatively, the total span of the window period may be at least 30 minutes. In another embodiment, multiple total spans of different window periods may be mixed. For example, a window period of one hour may be at least 30 minutes. There may be at least four window periods of 1 hour and at least six window periods of 30 minutes, resulting in a total span of at least 7 hours. For a total span of at least 8 hours, the blending may include, for example, at least five window periods of 1 hour and at least six window periods of 30 minutes. It may be.

[0058] In step 304, the processor 101 combines the processed vital sign features with the baseline sleep stage information to generate a dataset of multiple subsamples. A cross-validation set for training the model and a blind test set for evaluating the model can be derived from the dataset. The cross-validation set is a set of (k) equally sized subsamples. The input data set contains a subsample, or is divided into (k) equally sized subsamples.

[0059] In step 305, the processor 101 calculates (e.g., determines, predicts, or estimates) each of the subsamples in the cross-validation set with reference to the other subsamples (k-1) in the cross-validation set using a model with each of the multiple machine learning parameter sets. In other words, each subsample is calculated by the processor 101 using the other subsamples (k-1). This computation process is such that each subsample of the cross-validation set is a computation target for computation by processor 101 using other subsamples of the cross-validation set. Each machine learning parameter set is repeated (k) times to serve as a set of Therefore, the one subsample is considered to be an unknown sample for the model. The first subsample may be referred to as a target subsample of knowledge, while the remaining subsamples may be referred to as training subsamples that serve as a basis for computing the target subsamples. With this configuration, each subsample may be used once as a target subsample without "polluting" the training subsamples. Therefore, the computed sleep stage information represents generalized performance rather than overfitted performance.

[0060] Following the calculations, the processor 101 generates a calculated target size for each parameter set. The sample samples are combined to form calculated sleep stage information for comparison with the baseline sleep stage information.

[0061] In step 306, the processor 101 compares the calculated sleep stage information with the reference sleep stage information. Based on the comparison with the model, the model is associated with one of the parameter sets. The parameter set that results in calculated sleep stage information that most closely resembles the quasi-sleep stage information is associated with the model to achieve the highest calculation accuracy (or lowest calculation error).

[0062] FIG. 3B shows an example of the generation of a training data set, showing two tables, denoted as T1 and T2, respectively. In this example, referring to table T1, The training data set for training the logistic regression model of the vital sign feature was the 11 feature values ​​(X t-5 ~X t+5 ) are expressed together as time t-5 to t-1 (the leading period), t In another example, a training dataset for training a logistic regression model of vital sign features is created by extracting feature values ​​at t+1 to t+5 (subsequent periods) and t+2 to t+3 (intermediate periods). t-2 ~X t+2 ) together The times t-2 to t-1 (the leading period), t (the intermediate period), and t+1 to t+2 (the trailing period) are expressed together. The feature values ​​at time t are extracted from the corresponding The semi-sleep stage is the cell (S t ) The algorithm then calculates the 11 feature values ​​(X t-5 ~X t+5 ) and the corresponding reference sleep stages (St), and then the algorithm treats the rearranged feature values ​​and the corresponding reference sleep stages as a training data set. The training data set is shown in Table T2. For each of the 11 feature values ​​(X t-5 ~X t+5 ) is added to the feature value (X t-5 ~X t+5 ) during REM or NREM Standard sleep stages (S t A linear relationship can be calculated that represents the REM and NREM states at a particular time by defining weights and calculating the corresponding REM and NREM states at a particular time. In another embodiment, the time taken to extract the feature values ​​for each vital sign feature can be extended to be longer.

[0063] FIG. 4 may be performed independently or as part of step 202 of method 200. 4 shows a flowchart of a method 400 for extracting heart rate variability features that can be performed. In one embodiment, the computing device 100 is a wearable sleep stage tracking device. It takes the form of chair 100.

[0064] In step 401, the processor 101 receives A heart rate variability power spectral density is extracted from the physiological signal based on pulse interval detection.

[0065] In step 402, the processor 101 normalizes the power spectral density by dividing the power at each frequency by the total power across the low and high frequency portions of the spectrum, where the low frequency portion ranges from 0.04 Hz to 0.15 Hz and the high frequency portion ranges from 0.15 Hz to 0.4 Hz.

[0066] In step 403, the processor 101 convolves the high frequency portion of the power spectral density with a convolution filter to generate a plurality of convolution values ​​representing respective HRV features. Each convolution value represents a pattern of high and low power in the high frequency portion of the HRV power spectral density. The convolution filter is related to a convolution neural network model described below.

[0067] In step 404, the processor 101 performs an activation function on the convolution values ​​and selects one of the convolution values ​​based on the result of the activation function. The results show the similarity relationship between the pattern of each convolution value and the actual sleep stages. The selected convolution value serves as a vital sign feature representing the convolved high frequency power of HRV (Conv_HF).

[0068] FIG. 5 illustrates a method for creating a model for extracting heart rate variability features according to one embodiment of the present disclosure. 5 shows a flowchart of a method 500 for computing devices according to this embodiment. The 100 is a desktop personal computer with ample processing resources for machine learning operations. The system takes the form of a computer 100.

[0069] In step 501, the processor 101 receives a physiological signal via the communication interface 105 in association with the baseline sleep stage information. Alternatively, the physiological signal may be associated with the baseline sleep stage information. The baseline sleep stage information and physiological The biological signals may be received simultaneously or otherwise. The baseline sleep stage information is provided by a reference device, which may be a device that has been tested and approved by an authority. Alternatively, the baseline sleep stage information may be provided by a wearable device that has been tested and approved by an authority.

[0070] In step 502, the processor 101 extracts vital sign features representing heart rate variability power spectral density from each interval of the physiological signal of at least 3 minutes (e.g., 5 minutes). The reason for dividing the signal into intervals of at least 3 minutes is that this step divides the signal into several periods. In other words, the extraction of vital sign features from a physiological signal is performed for each of a plurality of intervals by dividing the physiological signal into a plurality of intervals, each of which has a duration of at least 3 minutes. If no steps are performed, the extracted HRV features will be based on a specific time period or duration in the session. is considered to represent the signal for an entire sleep session, rather than a period. In this embodiment, each interval is 5 minutes, and in each interval, only 3 minutes of the corresponding portion of the physiological signal are used to derive or extract the vital sign features. In other words, the vital sign features are derived from the physiological signal for 3 minutes of each 5 minute interval. This is the case for 5 minutes. During each interval, the components related to feature extraction only work for 3 minutes, and therefore consume In other embodiments, each interval may be other than 5 minutes. Frequently, the corresponding portion of the physiological signal used for feature extraction is longer than 3 minutes. It may be so.

[0071] In step 503, the processor 101 normalizes the power spectral density by dividing the power at each frequency by the total power across the low and high frequency portions of the spectrum, where the low frequency portion ranges from 0.04 Hz to 0.15 Hz and the high frequency portion ranges from 0.15 Hz to 0.4 Hz.

[0072] In step 504, the normalized HRV PSD and the reference sleep stages of the training dataset are input to a machine learning algorithm, in particular a convolutional neural network. A model with convolutional filters, activation functions, and data pooling with acceptable or best accuracy is selected and used to extract the Conv_HF feature, for example, in method 400. Used for.

[0073] In particular, we develop models for convolutional neural networks, CNNs, and programming. The model can be trained using algorithms using existing toolboxes such as "mxnet" based on the language "R". First, a training set containing the HRV PSD (features) and associated REM / NREM sleep stages (labels) is input to the CNN algorithm. The CNN algorithm performs training by calculating the best convolutional filters and neural network weights that minimize the classification or detection error. The training process starts with a random initialization of the model filters and weights. Then, a backpropagation process is performed to calculate the gradient towards a smaller error and adjust the weights at each iteration. To obtain a model that reduces or minimizes the error without overfitting, the model is trained using appropriate model parameters (e.g. size of convolutional filters, type of activation function, data pool, etc.) that minimize the error on the validation set. The model is fine-tuned by determining the convolutional filters, activation functions, and data pooling method (convolutional filter, activation function, and data pooling method) and used to derive the HRV PSD into Conv_HF features.

[0074] As discussed above in relation to steps 203 and 303, data errors (e.g., missing data portions) may occur. If a significant error is detected, the processor 101 in this embodiment compensates for the data error. If the data errors are corrected in this way, the compensated vital sign features are used in the subsequent steps of sleep stage detection. However, if the data errors remain, the epoch is treated as "unknown" to maintain the reliability of the sleep stage detection. Figures 6(a)-6(c) show the compensation by processor 101 for the vital sign features. 1 illustrates various aspects of data error compensation in accordance with the present invention. A vital sign feature is illustrated in the form of a series of feature values, each corresponding to a respective time period.

[0075] In FIG. 6(a), processor 101 detects no missing feature values ​​in vital signs feature 601. No compensation is performed and the feature values ​​of the vital signs feature are used in a subsequent step (e.g., algorithmic operation) 602. In FIG. 6(b), processor 101 detects three missing feature values ​​of zero in vital signs feature 603, two of which are consecutive. Processor 101 performs a feature compensation by selecting, for each intermediate feature value having a preceding feature value and a succeeding feature value, the maximum feature value of these three feature values. , compensate for the missing feature values ​​and generate a new set of compensated vital sign features 605. Such a technique is sometimes called “max pooling” 604. The compensated vital sign features 605 are then used in a subsequent step 606 (e.g., algorithmic computation). In feature 607, we found four missing feature values ​​of zero, three of which were consecutive. The processor 101 compensates for missing feature values ​​in a similar manner using max pooling 608, resulting in a compensated vital signs feature 609. In this situation, three consecutive missing feature values ​​will result in the compensated vital signs feature 609 having an additional zero after compensation. Since the data error still exists, the processor 101 In order to maintain the reliability of the stage detection, the errors still present in step 610 ( The system responds by saying that the sleep stage is unknown for this epoch (in this case represented by a feature value of "0") due to the presence of a feature value of "0" (shown as a shaded block in the figure). Correcting other forms of data errors It should be noted that compensation may also be available.

[0076] Step 205 of the method 200 described above involves calculating the sleep stages for each epoch using the model. The processor 101, in step 205, calculates the sleep stages for each intermediate epoch as a set of respective feature values ​​v1-v2. 11 Three vital sign features 701a-701c (e.g., heart rate, pulse shape variability, and convoluted high frequency of heart rate variability, respectively) may be represented by the vital sign features 701a-701c (only eleven are shown for each vital sign feature 701a-701c in FIG. 7). A typical situation in which an indication value 702a-702c is calculated from the power of the corresponding feature value v6 and the adjacent feature values ​​v1-v5, v7-v8, v9-v100, v11-v120, v13-v140, v15-v160, v17-v180, v19-v200, v21-v220, v22-v300, v23-v400, v24-v250, v31-v320, v32-v400, v33-v500 11 Based on The calculation is performed for each of the intermediate periods. The adjacent periods are those periods that are temporally adjacent to each of the intermediate periods. In this case, the adjacent periods are the five preceding periods (corresponding to the respective feature values ​​v1 to v5) that are the five periods immediately preceding the intermediate period (corresponding to the feature value v6) and the five periods immediately following the intermediate period. The five subsequent periods (each of which has a feature value v7 to v 11 However, in other embodiments, the adjacent epochs include any number (e.g., one, two, or three) of preceding epochs. and subsequent periods. Each indicator 702a-702c describes or indicates a sleep stage for a respective intermediate period.

[0077] In this embodiment, each logistic regression calculation 701a'-701c' is performed by the processor 101 using a respective model representing the weights and the sigmoid function (or That is, for each of the vital sign features 701a-701c, the processor 101 executes a respective logistic regression operation 701a'-701c' to obtain a time The indicator values ​​702a to 702c for each intermediate period are represented by the corresponding feature value v6 and the feature values ​​v1 to v5, v7 to v8 for the adjacent periods. 11 The calculation is based on the above. The indications 702a-702c describe or represent the sleep stage of the corresponding intermediate epoch. More specifically, for each model, the model includes a number of weights and a sigmoid function. The number of weights corresponds to the total number of preceding, intermediate, and following epochs, which is 11 in this embodiment. In this embodiment, the models use the same sigmoid function for each set of weights. The set of weights for one model may be different from the set of weights for another model. As used herein, the term "weight" means "weight value." As described above, the weights for each vital sign feature represent a relationship (linear combination) of feature values ​​of the vital sign features that best represents or describes, or ideally represents or describes, the sleep stage of a particular epoch. Thus, each such linear combination of one vital sign feature is equivalent to such linear combination of another vital sign feature (HR, Conv_HF, PSV, etc.). It may be the same or it may be different.

[0078] In the same example of FIG. 7, after calculating the indicator values ​​702a-702c for the intermediate periods, the process The sensor 101 determines the intermediate sleep stage based on the corresponding indication values ​​702a to 702c. Then, the processor 101 performs a further logistic regression calculation 702'. The further logistic regression calculation 702' uses a model that is independent of or separate from the models of the logistic regression calculations 701a' to 701c'. Specifically, the processor 101 performs a further logistic regression calculation 702' for each of the intermediate time indication values ​​702a to 702c. A tick regression operation is performed to calculate a REM value 703 for each intermediate epoch. For each intermediate epoch, the processor 101 determines that if the REM value exceeds a predetermined threshold (e.g., 0.5), the epoch corresponds to REM, otherwise it corresponds to NREM. If data is missing as described in relation to Figure 6, the period is determined as unknown sleep. Each of the logistic regression calculations 701a' to 701c' (see FIG. 7) is determined to correspond to the The first stage of the logistic operation is shown as the regression. While this can be done, the further logistic operation 702' (indicated in FIG. 7 by the label "further regression") can be considered a second stage logistic operation. can be done.

[0079] Each of the logistic regression operations 701a' to 701c' calculates a time-dependent logarithm of a physiological signal (e.g., a PPG signal) depending on the feature value of each vital sign feature extracted from the input physiological signal. It can be seen that the logistic regression calculation 702' calculates the respective indicator values ​​702a-702c for the periods. Then, a further logistic regression calculation 702' calculates the respective indicator values ​​702a-702c for the periods. , and calculates the REM value 703 for that epoch. In this embodiment, a further logistic regression operation 702' uses the same sigmoid operation as operations 701a'-701c'. However, the weighting is different.

[0080] It should be noted that a further logistic regression calculation 702' can be performed with at least two of the indicators 702a-702c, i.e., if one of the indicators 702a-702c at a particular time is not correct, e.g., because compensation was not successful or was not performed, the explanation may be omitted. If the first indicator value is not significant or is inconclusive (e.g., the value is zero), then a further logistic regression calculation 702' is performed based on the remaining two of the indicator values ​​702a-702c for that particular time period. If the one of the indicators 702a-702c describes a sleep stage at a particular time, In other words, the further logistic regression calculation 702' requires at least two indicators that are explanatory of the corresponding time, and further non-explanatory Such indicators can be omitted from further logistic regression calculations 702'. This prevents or reduces the accuracy of sleep stage detection from being adversely affected by indications that do not account for the sleep stage, which may be caused by unsuccessful sensor detection. To facilitate this, the logistic regression operations 701a'-701c' may use models that are independent of each other or distinct from each other.

[0081] The process of FIG. 7 uses a sliding window of logistic regression in this embodiment. The process is carried out in a window (11 periods) fashion, with each intermediate period being carried out in sequence. See Figure 8. In response to the above, the processor 101 performs a logistic calculation for each of the vital sign features 701a to 701c. The block regression calculations 701a' to 701c' are performed to calculate indication values ​​702a to 702c for the intermediate periods of the vital sign features 701a to 701c. The normalized feature values ​​are shown in line graph 801. Of all the epochs represented by the series of blocks 802, intermediate epochs for the calculation of the sleep stages are represented by the unshaded blocks 802b and the diagonally shaded blocks 802c, respectively. The normalized feature values ​​of line graph 801 range from -1 to 2 in this embodiment, but may have other ranges in other embodiments. Each of the first five periods (considered to be non-intermediate) represented by block 802a in , are predetermined to have a feature value of zero and are identified as NREM. This corresponds to the fact that a sleep session begins with an NREM period, which typically ranges from 30 to 70 minutes. In the sliding window approach, the processor 101 performs the above-described process. , and calculates indicator values ​​702a-702c for each successive intermediate period toward the end of the sleep session (represented by the last of the diagonally shaded blocks 802c). The calculated indicator values ​​702a-702c are illustrated by a line graph 803.

[0082] To perform calculations for the last five epochs with fewer than five subsequent epochs (indicated by diagonally shaded block 802c), the vital signs features 701a-701c are extended by five extended epochs (not shown) having feature values ​​that mirror the feature values ​​of the last five epochs. The dashed portion at the end of the line graph 801 (indicated by "AA" in FIG. 8) corresponds to the extended epoch and corresponds to the last five epochs preceding the extended epoch. 8. The solid line portion shown in FIG. 8 is a mirror image of the solid line portion shown in FIG. 8 (indicated by "BB" in FIG. 8). The dashed line portion ("AA") is the reverse of the solid line portion ("BB") of the last five periods that led to it. Thus, the last five epochs at the end of the sleep session (i.e., the diagonally shaded blocks 802c) can be considered as inverted or mirrored extensions. , which also serve as intermediate epochs for purposes of calculation. In the manner described above, readings 702a-702c are calculated for each intermediate epoch of vital sign features 701a-701c. In other embodiments, the calculation of readings 702a-702c for each intermediate epoch can take into account any equal or unequal number of preceding and succeeding epochs. For example, the number of preceding and succeeding epochs may be two, in which case five epochs are used in calculating the readings for each intermediate epoch.

[0083] FIG. 9 shows a line graph 901 of three vital sign features (HR, Conv_HF, and PSV), a line graph 902 of readings calculated from the vital sign features using a logistic regression operation, a line graph 903 of REM values ​​calculated from the readings using a further logistic regression operation, a line graph 904 of sleep stage information derived from line graph 903 based on a threshold, and a line graph 905 of baseline sleep information for comparison with line graph 904. For each epoch, if the corresponding REM value exceeds a threshold (e.g., 0.5), the epoch is determined to correspond to REM, otherwise it is determined to correspond to NREM. Depending on the implementation, the threshold may be other than 0.5.

[0084] FIG. 10 is a Bland-Altman plot showing the percentage error of REM values ​​calculated according to the embodiment of FIG. 1 against baseline sleep stage information over 50 nights. The x-axis shows the mean error rate. The horizontal line indicates the mean difference of 0.09%, while the dashed lines above and below the horizontal line indicate the confidence interval of the error of -12.10 and 12.28, respectively. The closer the horizontal line is to zero and the closer the dots are to the horizontal line, the more accurate the calculated REM value is.

[0085] FIG. 11 shows first and second rows 1110, 1120 each including three histograms of REM percentage determined in the manner described above for each subject. The subjects in this example are sufferers of depressive disorder, and in other applications may be healthy individuals wishing to adjust their sleep patterns. Each row 1110, 1120 includes a first histogram 1111, 1121 (labeled "Period 1") corresponding to a first period of time, a second histogram 1112, 1122 (labeled "Period 2") corresponding to a second period of time after the first period, and a third histogram 1113, 1123 (labeled "Period 3") corresponding to a third period of time after the second period. The subjects in the first row 1110 had three histograms of REM percentage determined in the manner described above for the first period of time. Subjects in the second row 1120 receive an initial dose of antispasmodic agent before the first period, between the first and second periods, and between the second and third periods. Subjects in the second row 1120 receive an initial dose of antispasmodic agent before the first period, an increased dose of antispasmodic agent between the first and second periods, and a further increased dose of antispasmodic agent between the second and third periods. Findings that the patient continues to experience frequent occurrences of sleep in the high REM fraction range Each of the first, second, and third periods represents a complete subsequence of sleep in this situation. Each period represents one sleep cycle (i.e., an entire sleep session). In other applications, each period may represent more sleep cycles. Different subjects may have complete cycles of sleep of different lengths.

[0086] The information represented by histograms 1111-1113 and 1121-1123 is the REM state experienced by the subject. The results may serve as a basis for determining a normative medical medication or dose to be administered to the subject to adjust the range and frequency of sleep disturbance. The normative medical medication or dose may relate to anticonvulsants, benzodiazepines, emazepam, or any receptor agonist that can relieve insomnia.

[0087] Specifically, a method for assessing responsiveness to a medical medication according to one embodiment includes: receiving sleep stage information for a period of time and comparing the sleep stage information for a first time period to the sleep stage information for a second time period; receiving sleep stage information for a second time period; and determining whether sleep stage information is performed during the first and second time periods. and evaluating the responsiveness of the subject to the medical medication based on the results of the comparison. In the example subject of the first row 1110, during the second and third time periods, during the first and second time periods, An assessment of the subject's responsiveness to unadjusted medical medication administered in the first and second periods can be performed by receiving and comparing sleep stage information for the first and second periods. In one embodiment, during the second and third time periods, an assessment of the subject's responsiveness to the increased doses of the medical medication administered during the first and second time periods may be performed by receiving and comparing sleep stage information from the first and second time periods. and as a basis for maintaining or adjusting any further medical medication administered during the third period. It may be useful to assess the subject's responsiveness to further medical treatment after the third period. Thus, a method of deriving a medical dose according to one embodiment includes: receiving sleep stage information for a period of time and comparing the sleep stage information for a first time period to the sleep stage information for a second time period; Receiving sleep stage information for a second time period and providing medical medication based on the comparison. It will be understood that this includes

[0088] FIG. 12A illustrates an embodiment of the computing device 100 of FIG. 1 in the form of a watch or wearable device 1200. With further reference to FIGS. 12B and 12C, The device 1200 includes, among other components, two physiological sensors 1211 ( 1, a physiological sensor 104 of FIG. 1), a lighting element 1212, and a wireless communication module. The head assembly 1210 includes a physiological sensor 1211 and a lighting element 1212. The physiological sensor 1211 and the lighting element 1212 cooperate to form a physiological sensor 104 operatively associated with the processor 101. The wireless communication module 1213 functions to provide wireless functionality for the communication interface 105 associated with the processor 101. The firmware chip 102 and / or storage The device 103 may store instructions for causing the processor 101 to execute any of the steps of the methods 200-500. For example, for purposes of step 201, the physiological sensor 1211 and the lighting element 1212 may measure physiological signals of a user wearing the wearable device 1200. In addition, when the wearable device 1200 executes step 205, the calculated sleep stage for each epoch can be provided wirelessly by the processor 101 via the wireless communication module 1213. The sleep stages calculated and transmitted in this manner may constitute sleep stage information. With further reference to FIG. 13, another computing device 1300 (e.g., the one shown in FIG. 1) may be used to A computing device (such as a personal computer or server device having components similar to those of the wearable device 1200) can wirelessly receive such sleep stage information for the first and second time periods from the wearable device 1200 and compare the received sleep stage information. The computing device 1300 can then provide medical medication based on the results of the comparison, or evaluate responsiveness to medical medication administered during the first and second time periods based on the results of the comparison.

[0089] An alternative embodiment including a single vital sign feature (e.g., a first vital sign feature 701a) In an embodiment, the intermediate logistic regression resulting from the corresponding first logistic regression calculation 701a′ Because the first indication of time 702a represents a mid-time sleep stage, further logistical steps are The regression calculation 702' may be omitted and the first indication value 702a may serve as the REM value 703 for determining the sleep stage.

Claims

1. 1. A method of sleep stage detection, comprising: receiving a first vital sign feature including a plurality of first feature values ​​corresponding to respective epochs of the plurality of epochs; For each intermediate period among the plurality of periods, a first The indicator value is calculated based on the corresponding first characteristic value and the preceding and succeeding epochs of the plurality of epochs. performing a first logistic regression operation to calculate a logistic regression function based on the first feature value at the time; A method comprising:

2. The first vital sign feature comprises a convolution of heart rate, pulse shape variability, and heart rate variability. The method of claim 1 , wherein the at least one high frequency power is one of the at least one high frequency powers.

3. The method of claim 1 or 2, wherein the first logistic regression operation includes a weighting and a sigmoid operation.

4. receiving a second vital sign feature including a plurality of second feature values ​​corresponding to respective epochs among the plurality of epochs; For each intermediate period among the plurality of periods, a second sleep stage representative of the intermediate period is obtained. The indicator value is compared with a corresponding second characteristic value and the preceding and succeeding epochs of the plurality of epochs. performing a second logistic regression operation to calculate based on the second feature values ​​for the epochs; performing a further logistic regression operation to determine a sleep stage for each intermediate epoch of the plurality of epochs based on the corresponding first and second indicators; The method of any one of claims 1 to 3, further comprising:

5. The method of claim 4 , wherein one of the logistic regression operations has a weight value that is different from a weight value of another of the logistic regression operations.

6. receiving a third vital sign feature including a plurality of third feature values ​​corresponding to respective epochs among the plurality of epochs; For each intermediate time period of the plurality of time periods, a third indication value is determined based on the corresponding third characteristic value and the third characteristic values ​​of the preceding and succeeding time periods of the plurality of time periods. performing a third logistic regression operation to calculate Further comprising: The further logistic regression calculation is, when the corresponding third indication value represents a sleep stage of each intermediate period of the plurality of periods, The method of claim 4 or 5 further comprises determining based on the corresponding third indication value. The method described above.

7. 1. A method of creating a model for logistic regression, comprising the steps of: receiving vital sign features associated with baseline sleep stage information; generating a cross-validation set from the vital sign features and the reference sleep stage information, the cross-validation set including a plurality of sub-samples; calculating each of the subsamples with reference to other of the subsamples using a model with each of a plurality of machine learning parameter sets; Based on a comparison of the baseline sleep stage information and the results of the calculation, associating one of them with said model; A method comprising:

8. The method of claim 7 , wherein the vital sign feature relates to one of heart rate, pulse shape variability, and convolved high frequency power of heart rate variability.

9. Receiving another vital sign feature associated with the baseline sleep stage information. Further comprising: The cross-validation set is further generated from the additional vital sign features. Or the method according to item 8.

10. The vital sign features are sliced ​​by at least three consecutive epochs. The method according to any one of claims 7 to 9.

11. The at least three consecutive periods include, with respect to an intermediate period among the consecutive periods, 11. The method of claim 10, including a preceding time period, the intermediate time period, and a succeeding time period in a series.

12. 12. The method of claim 10 or 11, wherein the at least three consecutive periods comprises eleven periods. method.

13. The at least three consecutive epochs correspond to a total window period of at least 30 minutes. The method according to any one of claims 10 to 12, wherein the method is configured to:

14. The method according to any one of claims 1 to 13, wherein the vital signs features are normalized. 。

15. 15. The method according to claim 1, wherein the vital signs feature is compensated for missing parts. The method according to claim 5.

16. The method of claim 15, wherein the missing portions are compensated for using max pooling.

17. 1. A method for extracting heart rate variability features, comprising: generating a plurality of convolved values ​​representing respective patterns of heart rate variability by convolving a high frequency portion of the heart rate variability power spectral density with a convolution filter related to a convolutional neural network model; selecting one of the convolved values ​​based on a result of an activation function operation performed on the convolved values; A method comprising:

18. The method of claim 17 , wherein the selected convolution value corresponds to a maximum value resulting from the executed activation function.

19. 19. The method according to claim 17 or 18, wherein the high frequency part is in the range of 0.15 Hz to 0.4 Hz. method.

20. The method according to any one of claims 17 to 19, wherein the power spectral density is normalised across the high and low frequency parts of the power spectral density.

21. The activation function calculation is selected from the group consisting of a normalized linear function, a soft real function, and a sigmoid function. The method according to any one of claims 17 to 20, wherein

22. 1. A method of creating a model for extracting heart rate variability features, comprising: receiving vital sign features representative of heart rate variability power spectral density in relation to baseline sleep stage information; creating a model using machine learning based on the vital sign features with reference to the baseline sleep stage information, the model including a convolution filter; A method comprising:

23. 23. The method of claim 22, wherein the vital sign features are derived from physiological signals for 3 minute intervals within each 5 minute interval.

24. The method of claim 22 or 23, wherein the vital sign features are normalized.

25. The method of any one of claims 22 to 24, wherein the machine learning comprises a convolutional neural network.

26. 1. A method of delivering a medical medication, comprising: receiving sleep stage information for a first period of time; receiving sleep stage information of a second period of time for comparison with the sleep stage information of the first period of time; providing a medical medication based on a result of said comparing; A method comprising:

27. 1. A method for assessing responsiveness to a medical medication, comprising: receiving sleep stage information for a first period of time; receiving sleep stage information of a second period of time for comparison with the sleep stage information of the first period of time; assessing responsiveness to medical administration administered during the first and second time periods based on results of the comparison; and A method comprising:

28. 28. The method of claim 26 or 27, wherein the normative medical medication relates to at least one of an anticonvulsant, a benzodiazepine, and emazepam.

29. A method comprising the steps of: A computer-readable medium.

30. 30. The computer-readable medium of claim 29, wherein the instructions are configured to be implemented in firmware.

31. A processor; Instructions for causing the processor to carry out the method according to any one of claims 1 to 28. Storage devices including A computing device comprising:

32. The computing device of claim 31 , wherein the storage device is a firmware chip.

33. Including instructions for causing a processor to carry out the method according to any one of claims 1 to 28. Storage Devices A wearable device comprising:

34. 34. The wearable device of claim 33, wherein the storage device is a firmware chip.

35. 35. A wearable device as claimed in claim 33 or 34, in the form of a wristwatch.

Images