A sleep monitoring system and method based on a smart watch

Abstract

The present application relates to the technical field of intelligent wearable devices, and particularly relates to a sleep monitoring system and method based on a smart watch, which comprises a collection and evaluation module that collects PPG, acceleration and voice data, extracts physiological parameters and respiratory sound envelopes and discards original audio, and calculates a signal quality index; a phase analysis module that extracts arrhythmia waveforms and calculates their phase synchronization characteristics with the respiratory sound envelope; a precursor identification module that generates an airway collapse precursor identifier when multiple conditions are met; a cascaded inference module that constructs two sets of features, generates a fusion confidence, and completes awakening determination and sleep staging in three stages; and a sliding self-adaptive module that updates dynamic thresholds and triggers a health warning based on a precursor cumulative amount. In the present application, the phase synchronization characteristics of respiratory sound and arrhythmia waveforms are extracted, and when decoupling, multiple indicators are combined to generate an airway collapse precursor warning, thus solving the defects of traditional monitoring, which only determines apnea by blood oxygen, does not analyze cardiopulmonary coupling, and cannot identify airway obstruction risks in advance.

Description

Technical Field

[0001] This invention relates to the field of smart wearable device technology, and in particular to a sleep monitoring system and method based on a smartwatch. Background Technology

[0002] With the development of wearable devices and physiological signal acquisition technologies, sleep monitoring technology based on smartwatches is gradually being applied to daily health management. Existing smartwatches typically integrate multiple sensors such as photoplethysmography (PPG) sensors, accelerometers, and microphones. By collecting pulse wave signals, body movement signals, and ambient sound signals during the user's sleep process, they analyze heart rate, sleep activity status, and respiratory information, and combine this with signal processing algorithms to identify different states during sleep. With the development of multi-source physiological information fusion technology, some sleep monitoring solutions further combine heart rate changes, body movement changes, respiratory characteristics, and autonomic nervous system activity characteristics to comprehensively analyze sleep stage division and sleep state assessment.

[0003] Most existing sleep monitoring methods only detect sleep apnea by detecting a drop in blood oxygen saturation. Due to the lack of analysis on the coupling relationship between respiration and heart rate variability, the risk cannot be proactively detected before airway obstruction occurs. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a sleep monitoring system and method based on a smartwatch, aiming to improve the problem that most existing sleep monitoring systems only detect apnea by detecting a drop in blood oxygen saturation, which easily leads to the inability to proactively detect risks before airway obstruction occurs.

[0005] This invention provides the following technical solution: a sleep monitoring system based on a smartwatch, comprising:

[0006] The data acquisition and evaluation module controls the PPG sensor, accelerometer, and microphone triggered by the low-power voice activity detection unit to acquire data; it extracts the heart interval and heart rate from the pulse wave signal, extracts the body motion intensity from the acceleration signal, filters the audio signal to extract the respiratory sound envelope and discards the original audio data, and calculates the PPG signal quality index, body motion signal quality index, and audio signal quality index.

[0007] The phase analysis module extracts the respiratory sinus arrhythmia waveform from the heartbeat interval and calculates the phase synchronization characteristics between the waveform and the respiratory sound envelope;

[0008] The warning sign recognition module generates an upper airway collapse warning sign when the phase synchronization feature is continuously decoupled and the synchronization satisfies the following conditions: the amplitude of the respiratory sound envelope decays, the low-frequency to high-frequency energy ratio extracted from the heartbeat interval exceeds the dynamic autonomic nervous system regulation threshold, and the body movement intensity is lower than the dynamic body movement threshold.

[0009] The cascaded inference module generates a fusion confidence score based on the signal quality index and the phase synchronization features, and constructs a first sleep feature set and a second sleep feature set; it performs a first-level arousal determination by comparing body movement intensity and heart rate with a dynamic arousal threshold; when not awake, it performs a second-level binary sleep staging by comparing the fusion confidence score with the first sleep feature set; when the second-level confidence score is insufficient, it performs a third-level pentatonic sleep staging by comparing the second sleep feature set with the upper airway collapse precursor indicators.

[0010] The sliding adaptive module updates the dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold based on historical sleep monitoring data, and triggers a health warning based on the cumulative results of the upper airway collapse precursor indicators.

[0011] By employing the above technical solution, the phase synchronization features of the breath sound envelope and the waveform of respiratory sinus arrhythmia are extracted. When this feature is continuously decoupled, the breath sound envelope attenuation, the low-frequency to high-frequency energy ratio, and the intensity of body movement are combined to generate an early warning sign of upper airway collapse, thereby achieving early warning. This improves upon the fact that most existing sleep monitoring methods only detect apnea by detecting a decrease in blood oxygen saturation. Due to the lack of analysis on the coupling relationship between respiration and heart rate variability, the risk cannot be proactively detected before airway obstruction occurs.

[0012] Furthermore, the control of the PPG sensor, accelerometer, and microphone data collected by the low-power voice activity detection unit includes:

[0013] The sensing and acquisition unit controls the PPG sensor and the accelerometer to acquire signals respectively, and outputs pulse wave signals and acceleration signals.

[0014] The voice detection unit detects the energy of ambient sound and outputs a voice trigger signal.

[0015] The audio acquisition unit receives the voice trigger signal and controls the microphone to acquire audio signals;

[0016] The time synchronization unit receives the pulse wave signal, the acceleration signal, and the audio signal, aligns the three signals in time, and generates synchronized acquisition data.

[0017] Furthermore, the step of filtering the audio signal to extract the respiratory sound envelope and discarding the original audio data, and calculating the PPG signal quality index, body motion signal quality index, and audio signal quality index includes:

[0018] The audio processing unit filters the audio signal collected by the microphone and extracts the breathing sound envelope. After extraction, the original audio data is discarded.

[0019] The quality calculation unit calculates the PPG signal quality index, body motion signal quality index, and audio signal quality index based on the pulse wave signal, acceleration signal, and respiratory sound envelope, and outputs the PPG signal quality index, body motion signal quality index, and audio signal quality index.

[0020] Furthermore, the calculation of the phase synchronization characteristics between the waveform and the breath sound envelope includes:

[0021] The instantaneous phase extraction unit extracts the instantaneous phase sequences corresponding to the respiratory sinus arrhythmia waveform and the respiratory sound envelope, respectively;

[0022] The phase difference calculation unit calculates the phase difference data between two instantaneous phase sequences;

[0023] The synchronization feature generation unit performs window statistical processing on the phase difference data to generate phase synchronization features.

[0024] Furthermore, the step of generating a fusion confidence score based on the signal quality index and the phase synchronization feature, and constructing a first sleep feature set and a second sleep feature set includes:

[0025] The signal normalization unit quantizes the PPG signal quality index, the body motion signal quality index, the audio signal quality index, and the phase synchronization feature to generate signal reliability parameters.

[0026] The confidence calculation unit generates fusion weights based on the proportional relationship between the reliability parameters of each signal, and generates a fusion confidence score after weighted fusion calculation.

[0027] The first feature construction unit consists of the heartbeat interval, the heart rate, the body movement intensity, and the respiratory sound envelope, forming a first sleep feature set.

[0028] The second feature construction unit generates a second sleep feature set by combining the cardiac interval, body movement intensity, respiratory sound envelope, phase synchronization feature, and upper airway collapse precursor identifier.

[0029] Further, the step of performing a Level 1 arousal determination by comparing body movement intensity and heart rate with a dynamic arousal threshold; performing a Level 2 binary sleep staging by comparing the fusion confidence score with the first sleep feature set when not awake; and performing a Level 3 pentatonic sleep staging by comparing the second sleep feature set with the upper airway collapse precursor markers when the Level 2 confidence score is insufficient, includes:

[0030] The first-level determination unit compares the body movement intensity and the heart rate with the dynamic arousal threshold respectively to generate an arousal determination result;

[0031] The secondary classification unit, when the awakening determination result is an unawakened state, uses the fusion confidence and the first sleep feature set to perform binary sleep staging and generate a binary sleep state result.

[0032] When the confidence level of the binary sleep stage is insufficient, the third-level classification unit uses the second sleep feature set and the upper airway collapse precursor marker to perform a five-category sleep stage and generate a five-category sleep state result.

[0033] The temporal smoothing unit smooths the sleep state categories of adjacent time slices and outputs sleep staging data.

[0034] Furthermore, updating the dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold based on historical sleep monitoring data includes:

[0035] The historical data processing unit filters historical sleep monitoring data, removes data periods that lack heart rate interval features, body movement intensity features, or sleep state features, and extracts heart rate interval features, body movement intensity features, and sleep state features to generate historical sleep feature data.

[0036] The sliding baseline calculation unit establishes a time sliding window based on the historical sleep characteristic data, performs statistical calculations on the features within the window, and generates autonomic nervous regulation baseline parameters, body movement baseline parameters, and wakefulness state baseline parameters.

[0037] The threshold update unit updates the dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold respectively based on the autonomic nervous system regulation baseline parameters, the body movement baseline parameters, and the arousal state baseline parameters, and outputs them to the aura recognition module and the cascaded inference module.

[0038] Secondly, the present invention provides the following technical solution: a sleep monitoring method based on a smartwatch, the method comprising the following steps:

[0039] It collects pulse wave signals and acceleration signals, and performs low-power energy detection on ambient sound. When the ambient sound energy exceeds the wake-up threshold, it activates the microphone to collect audio signals.

[0040] The heart rate and heart interval are extracted from the pulse wave signal, the body motion intensity is extracted from the acceleration signal, the respiratory sound envelope is extracted from the audio signal after filtering and the original audio data is discarded, and the PPG signal quality index, body motion signal quality index and audio signal quality index are calculated.

[0041] The respiratory sinus arrhythmia waveform is extracted from the intercardiac interval, and the phase synchronization characteristics between the waveform and the respiratory sound envelope are calculated.

[0042] When the phase synchronization feature is continuously decoupled, and the synchronization satisfies the following conditions: the amplitude of the respiratory sound envelope decays, the ratio of low-frequency to high-frequency energy extracted from the heartbeat interval exceeds the dynamic autonomic nervous system regulation threshold, and the body movement intensity is lower than the dynamic body movement threshold, an upper airway collapse precursor indicator is generated.

[0043] A fusion confidence score is generated based on the signal quality index and the phase synchronization features, and a first sleep feature set and a second sleep feature set are constructed. A first-level arousal determination is performed using body movement intensity and heart rate as a reference to the dynamic arousal threshold. When not awake, a second-level binary sleep stage is performed using the fusion confidence score and the first sleep feature set. When the second-level confidence score is insufficient, a third-level pentatonic sleep stage is performed using the second sleep feature set and the upper airway collapse precursor indicators.

[0044] The dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold are updated based on historical sleep monitoring data, and a health warning is triggered based on the cumulative results of the upper airway collapse precursor indicators.

[0045] The present invention has the following beneficial effects:

[0046] 1. In this invention, the phase synchronization features of the respiratory sound envelope and the waveform of respiratory sinus arrhythmia are extracted. When the features are continuously decoupled, respiratory sound envelope attenuation, low-frequency to high-frequency energy ratio and body movement intensity are combined to generate an upper airway collapse precursor indicator, thereby achieving early warning. This improves the problem that most existing sleep monitoring methods only detect apnea by a decrease in blood oxygen saturation. Due to the lack of analysis on the coupling relationship between respiration and heart rate variability, the risk cannot be proactively detected before airway obstruction occurs.

[0047] 2. In this invention, a fusion confidence score is generated based on the signal quality index and phase synchronization features, and a differentiated sleep feature set is constructed. Then, the arousal judgment is performed by comparing the body movement intensity and heart rate with the dynamic arousal threshold. The sleep stages are classified into two categories and five categories according to the confidence score. This improves the problem that most existing wearable sleep monitoring devices use fixed thresholds and complex models throughout the day, which cannot adapt to real-time signal quality fluctuations and thus cause high power consumption throughout the night.

[0048] 3. In this invention, the microphone is triggered by a low-power voice activity detection unit to collect audio, and the original audio data is discarded after extracting the breathing sound envelope, thereby providing a breathing sound waveform with the original audio removed for subsequent analysis. This improves the problem that most existing smartwatches use microphones to continuously record sleep data, which causes a high risk of user privacy leakage because the original audio data is stored on the device side. Attached Figure Description

[0049] Figure 1This is a schematic diagram of the overall module architecture of a smartwatch-based sleep monitoring system according to the present invention;

[0050] Figure 2 This is a schematic diagram of the data acquisition and feature extraction process of a smartwatch-based sleep monitoring system according to the present invention;

[0051] Figure 3 This is a schematic diagram of the synchronous analysis process of the waveform and respiratory sound envelope phase of a sleep monitoring system based on a smartwatch in this invention.

[0052] Figure 4 This is a schematic diagram of the process for generating an upper airway collapse precursor indicator in a smartwatch-based sleep monitoring system according to the present invention.

[0053] Figure 5 This is a schematic diagram of the cascaded sleep inference process of a smartwatch-based sleep monitoring system in this invention.

[0054] Figure 6 This is a schematic diagram of the overall process of a sleep monitoring method based on a smartwatch in this invention. Detailed Implementation

[0055] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1: In the first embodiment of the present invention, the present invention provides a sleep monitoring system based on a smartwatch, such as... Figure 1 As shown, it includes:

[0057] The data acquisition and evaluation module controls the PPG sensor, accelerometer, and microphone triggered by the low-power voice activity detection unit to acquire data; it extracts the heart interval and heart rate from the pulse wave signal, extracts the body motion intensity from the acceleration signal, filters the audio signal to extract the respiratory sound envelope and discards the original audio data, and calculates the PPG signal quality index, body motion signal quality index, and audio signal quality index.

[0058] like Figure 2 As shown, further, controlling the PPG sensor, accelerometer, and microphone data acquisition triggered by the low-power voice activity detection unit includes:

[0059] The sensing and acquisition unit controls the PPG sensor and the accelerometer to acquire signals respectively, and outputs pulse wave signals and acceleration signals.

[0060] The voice detection unit detects the energy of ambient sound and outputs a voice trigger signal.

[0061] The audio acquisition unit receives voice trigger signals and controls the microphone to acquire audio signals.

[0062] The time synchronization unit receives pulse wave signals, acceleration signals, and audio signals, aligns the three signals in time, and generates synchronized acquisition data.

[0063] Specifically, the sensor acquisition unit controls the PPG sensor to acquire green light signals at a sampling rate of 50Hz and the accelerometer to acquire acceleration signals at a sampling rate of 25Hz. The pulse wave signal and acceleration signal are output after analog-to-digital conversion. The pulse wave signal is bandpass filtered to remove baseline drift and high-frequency noise, and the peak point of the filtered signal is detected. The time interval between adjacent peak points is used as the heartbeat interval, and the heart rate is calculated based on the number of heartbeats per unit time. The voice detection unit performs energy detection on real-time ambient sound. When the ambient sound energy value exceeds 40dB, a voice trigger signal is output. The microphone remains in sleep mode while the ambient sound energy value is continuously below this threshold. After receiving the voice trigger signal, the audio acquisition unit controls the microphone to acquire ambient audio signals at a sampling rate of 16kHz and generates audio signals. The time synchronization unit timestamps the pulse wave signal, acceleration signal, and audio signal based on the MCU clock, and generates synchronized acquisition data by aligning the timestamps.

[0064] like Figure 2 As shown, further, after filtering the audio signal to extract the respiratory sound envelope, the original audio data is discarded, and the PPG signal quality index, body motion signal quality index, and audio signal quality index are calculated, including:

[0065] The audio processing unit filters the audio signal collected by the microphone and extracts the breathing sound envelope. After extraction, the original audio data is discarded.

[0066] The quality calculation unit calculates the PPG signal quality index, body motion signal quality index, and audio signal quality index based on the pulse wave signal, acceleration signal, and breath sound envelope, and outputs the PPG signal quality index, body motion signal quality index, and audio signal quality index.

[0067] Specifically, the audio processing unit employs a 5-second segmented buffering mechanism for the audio signal. Every 5 seconds, an 80Hz high-pass filter and a 200-1000Hz band-pass filter are sequentially applied to remove low-frequency wind noise and focus on the main spectral components of the breathing sounds. The filtered signal undergoes Hilbert transform to extract the envelope, followed by median filtering and baseline drift removal to obtain the breathing sound envelope reflecting the intensity of the respiratory airflow. After feature extraction, the original 5-second segment of audio data is immediately discarded without local storage or external transmission. The quality calculation unit calculates the PPG signal quality index based on the pulse waveform template matching degree and the motion artifact energy ratio of the pulse wave signal. , The mass motion signal quality index is calculated for acceleration signals based on signal saturation and the proportion of high-frequency noise, and is a continuous value from 0 to 1. The audio signal quality index is estimated based on the energy ratio of the audible and silent segments of the breath sound envelope. It outputs the three quality indices mentioned above.

[0068] like Figure 3 As shown, the phase analysis module extracts the waveform of respiratory sinus arrhythmia from the heartbeat interval and calculates the phase synchronization characteristics between the waveform and the breath sound envelope.

[0069] Furthermore, the phase synchronization characteristics between this waveform and the breath sound envelope are calculated as follows:

[0070] The instantaneous phase extraction unit extracts the instantaneous phase sequences corresponding to the waveform of respiratory sinus arrhythmia and the envelope of respiratory sounds, respectively;

[0071] The phase difference calculation unit calculates the phase difference data between two instantaneous phase sequences;

[0072] The synchronization feature generation unit performs window statistical processing on the phase difference data to generate phase synchronization features.

[0073] Specifically, the intercardiac interval sequence is resampled at 4Hz, and then... Bandpass filtering yields the high-frequency components, which are the waveforms of respiratory sinus arrhythmia. The instantaneous phase extraction unit performs a Hilbert transform on the respiratory sinus arrhythmia waveform, taking the arctangent of the ratio of the imaginary to the real parts to obtain the first instantaneous phase sequence. A Hilbert transform is then performed on the respiratory sound envelope waveform, and the arctangent of the ratio of the imaginary to the real parts is taken to obtain the second instantaneous phase sequence. The phase difference calculation unit calculates the difference between the first and second instantaneous phase sequences at each moment, generating phase difference data. The synchronization feature generation unit calculates the phase lock value within a 2-minute window with a 30-second step size. As a phase synchronization feature, among which For instantaneous phase difference, The imaginary unit, This indicates the average value within the window. When the short-term energy of the breath sound envelope is lower than the preset ambient background noise threshold, this period is not included in the phase lock value calculation, and the corresponding phase lock value is set to invalid.

[0074] like Figure 4 As shown, the precursor identification module generates an upper airway collapse precursor identifier when the phase synchronization feature is continuously decoupled and the synchronization satisfies the following conditions: the amplitude of the respiratory sound envelope decays, the ratio of low-frequency to high-frequency energy extracted from the heartbeat interval exceeds the dynamic autonomic nervous system regulation threshold, and the body movement intensity is lower than the dynamic body movement threshold.

[0075] Specifically, the precursor recognition module uses a 5-minute continuous time window for judgment, and when the phase lock value... A value consistently below 0.5 indicates a decoupling state. Simultaneously, the average amplitude of the breath sound envelope every 30 seconds is checked to see if it exhibits a monotonically decreasing trend and if the slope of the linear fit is less than [value missing]. From the intercardiac interval sequence low frequency band and The analysis of the high-frequency power spectrum determines whether the low-frequency to high-frequency energy ratio exceeds the dynamic autonomic nervous system regulation threshold and whether the body movement intensity is lower than the dynamic body movement threshold. When both conditions are met, an upper airway collapse precursor indicator is generated. Preferably, when the blood oxygen saturation estimated by wrist PPG decreases by more than 1.5% within 5 consecutive minutes, the effectiveness of the precursor pattern is further confirmed. The dynamic autonomic nervous system regulation threshold is composed of the sum of the general baseline of 1.2 and 1.5 times the standard deviation of 0.4 during the cold start phase of the first three nights. From the fourth night onwards, it is updated by the sliding baseline calculation unit based on personalized data from the past 7 nights.

[0076] like Figure 5 As shown, the cascaded inference module generates a fusion confidence score based on the signal quality index and phase synchronization features, and constructs a first sleep feature set and a second sleep feature set; it performs a first-level arousal determination by comparing body movement intensity and heart rate with the dynamic arousal threshold; when not awake, it performs a second-level binary sleep staging by using the fusion confidence score and the first sleep feature set; when the second-level confidence score is insufficient, it performs a third-level pentatonic sleep staging by using the second sleep feature set and upper airway collapse precursor indicators.

[0077] Furthermore, a fusion confidence score is generated based on the signal quality index and phase synchronization features, and a first sleep feature set and a second sleep feature set are constructed, including:

[0078] The signal normalization unit quantizes the PPG signal quality index, body motion signal quality index, audio signal quality index, and phase synchronization characteristics to generate signal reliability parameters.

[0079] The confidence calculation unit generates fusion weights based on the proportional relationship between the reliability parameters of each signal, and generates a fusion confidence score after weighted fusion calculation.

[0080] The first feature construction unit consists of the first sleep feature set, which is composed of the heartbeat interval, heart rate, body movement intensity, and respiratory sound envelope.

[0081] The second feature building unit is generated by combining the heartbeat interval, body movement intensity, respiratory sound envelope, phase synchronization features, and upper airway collapse precursor indicators to form the second sleep feature set.

[0082] Specifically, the signal normalization unit will normalize the PPG signal quality index. Body motion signal quality index Audio signal quality index and phase lock value By normalizing and mapping to the 0-1 interval, the signal reliability parameters corresponding to each channel are obtained, including the audio signal-to-noise ratio. The original value range is 0~60dB. The Sigmoid function is used to map it to the 0~1 interval to obtain the normalized value. When a phase lock value is invalidated for a certain period due to the short-term energy of the breath sound envelope being lower than the noise threshold, the weighting coefficient corresponding to that period is automatically set to zero; the confidence calculation unit is set according to... The adaptive weights of each channel are calculated in the following way: Take in sequence and And then according to Generate fusion confidence , To normalize environmental noise energy from Approximately obtained; the first feature construction unit takes the mean and standard deviation of the heartbeat interval, the mean heart rate, the mean body movement intensity, and the mean short-term energy of the breath sound envelope, and... Together, they constitute the first sleep feature set; the second feature construction unit takes the mean and standard deviation of the heartbeat interval, the time-domain indices RMSSD and SDNN of heart rate variability, the frequency-domain indices LF / HF of heart rate variability, the mean and variance of body movement intensity, the short-time energy and zero-crossing rate of the breath sound envelope, and the phase lock value. Combined with signs of upper airway collapse, a second sleep feature set is generated.

[0083] Furthermore, Level 1 arousal assessment was performed using body movement intensity and heart rate as a reference to dynamic arousal thresholds; during non-awake periods, Level 2 binary sleep staging was performed using fusion confidence and the first sleep feature set; when Level 2 confidence was insufficient, Level 3 pentatonic sleep staging was performed using the second sleep feature set and upper airway collapse precursor indicators, including:

[0084] The first-level judgment unit compares the body movement intensity and heart rate with the dynamic arousal threshold respectively to generate an arousal judgment result;

[0085] The secondary classification unit, when the awakening determination result is an in-awake state, uses the fusion confidence and the first sleep feature set to perform binary sleep staging and generate a binary sleep state result;

[0086] The three-level classification unit, when the confidence of the binary sleep stage is insufficient, uses the second sleep feature set and the upper airway collapse precursor marker to perform five-level sleep stage and generate a five-level sleep state result.

[0087] The temporal smoothing unit smooths the sleep state categories of adjacent time slices and outputs sleep staging data.

[0088] Specifically, the first-level decision unit calculates the average resultant acceleration within the current time window. and mean instantaneous heart rate ,when The awakening judgment result is output at the time, where The preferred value is 2.5. This represents the individual's resting average physical activity level. For individual resting heart rate, and Provided by the sliding baseline calculation unit, within the first 2 hours after falling asleep The value was temporarily lowered to 2.25; the secondary classification unit was in a non-awake state according to the arousal determination result and the fusion confidence score was [missing information]. When the value is greater than 0.6, the first sleep feature set is input into a miniature decision tree model with a depth of 3 for binary classification of light sleep or REM sleep and deep sleep. When the maximum classification probability is higher than 0.8 and Output binary classification sleep state results when the confidence level is greater than 0.7; otherwise, mark it as insufficient confidence. When the energy level is below 0.3 and periodic snoring energy is detected in the audio signal, accompanied by a disordered respiratory sinus arrhythmia waveform, it is determined to be snoring interference from the bedmate. A penalty factor of 0.2 is applied to achieve soft inhibition. When the confidence of binary classification is insufficient, the third-level classification unit inputs the second sleep feature set and the upper airway collapse precursor label into the LightGBM model with a maximum depth of 5, 31 leaves, and 80 trees for five-class classification, and outputs the classification results of wakefulness, REM, light sleep, deep sleep, or disturbance as the five-class sleep state results. The temporal smoothing unit uses a weighted moving average to smooth the sleep state categories of adjacent time slices and outputs the final sleep staging data.

[0089] The sliding adaptive module updates the dynamic autonomic nervous regulation threshold, dynamic body movement threshold, and dynamic arousal threshold based on historical sleep monitoring data, and triggers health warnings based on the cumulative results of upper airway collapse precursor indicators.

[0090] Furthermore, the dynamic autonomic nervous system regulation threshold, dynamic body movement threshold, and dynamic arousal threshold are updated based on historical sleep monitoring data, including:

[0091] The historical data processing unit filters historical sleep monitoring data, removes data periods that lack heart rate interval features, body movement intensity features, or sleep state features, and extracts heart rate interval features, body movement intensity features, and sleep state features to generate historical sleep feature data.

[0092] The sliding baseline calculation unit establishes a time sliding window based on historical sleep characteristic data, performs statistical calculations on the features within the window, and generates autonomic nervous regulation baseline parameters, body movement baseline parameters, and wakefulness state baseline parameters.

[0093] The threshold update unit updates the dynamic autonomic regulation threshold, dynamic body movement threshold, and dynamic arousal threshold based on the autonomic regulation baseline parameters, body movement baseline parameters, and arousal state baseline parameters, and outputs them to the aura recognition module and the cascaded inference module.

[0094] Specifically, the historical data processing unit sorts the sleep monitoring data from the past 7 nights by time and automatically identifies weekdays to distinguish between weekdays and rest days. It checks each cycle for missing features such as heart rate intervals, body movement intensity, or sleep state characteristics, discarding any data cycles with missing features. From the remaining complete data cycles, it extracts the mean heart rate interval, mean body movement intensity, and the percentage of each sleep state stage, establishing historical sleep characteristic data for weekdays and rest days respectively. The sliding baseline calculation unit applies a 7-day sliding window to the historical sleep characteristic data for both weekdays and rest days, sliding forward each day by... The system automatically forgets data from 7 days prior. It uses an exponentially weighted moving average to statistically calculate the mean heart rate interval, mean body movement intensity, and mean body movement during wakefulness within the window. This generates two sets of baseline parameters for autonomic nervous system regulation, body movement, and wakefulness: one for weekdays and one for rest days. The threshold update unit selects the corresponding set of baseline parameters based on the current date type. It multiplies the autonomic nervous system regulation baseline parameters by a preset coefficient to obtain the dynamic autonomic nervous system regulation threshold, multiplies the body movement baseline parameters by a preset coefficient to obtain the dynamic body movement threshold, and multiplies the wakefulness baseline parameters by a preset coefficient to obtain the dynamic wakefulness threshold. These parameters are then output to the aura recognition module and the cascaded inference module.

[0095] Example 2: A middle-aged user has a long-standing problem with nighttime snoring and morning dry mouth, but is unsure whether medical attention is necessary. The user wears a smartwatch equipped with a PPG sensor, accelerometer, and microphone before bed. The watch automatically activates its sleep monitoring function after the user falls asleep, aiming to proactively detect early signs of upper airway collapse before a significant drop in blood oxygen saturation. To address this issue, a sleep monitoring method based on a smartwatch, as provided in this invention, is employed. Figure 6 As shown. The specific implementation process of this method is as follows:

[0096] First, the watch collects pulse wave signals and acceleration signals. Meanwhile, the low-power voice activity detection unit only triggers the microphone to collect audio when the ambient sound energy exceeds the wake-up threshold, thus avoiding the power waste and privacy risks of continuous recording at all times.

[0097] The heart rate and heart interval are extracted from the pulse wave signal, the body motion intensity is extracted from the acceleration signal, the original audio data is discarded after the respiratory sound envelope is extracted by filtering the audio signal, and the signal quality index of each channel is calculated to provide a reliable reference for subsequent fusion inference.

[0098] Next, the respiratory sinus arrhythmia waveform was extracted from the intercardiac interval, and the phase lock value between the waveform and the breath sound envelope was calculated as the phase synchronization feature, thereby quantifying the coupling strength between central respiratory drive and upper airway ventilation.

[0099] Subsequently, when the phase lock value is in a decoupled state for a continuous period of time, and simultaneously meets the following conditions: the amplitude of the breath sound envelope decays, the energy ratio of low frequency to high frequency exceeds the dynamic autonomic nervous system regulation threshold, and the body movement intensity is lower than the dynamic body movement threshold, an upper airway collapse precursor indicator is generated, thereby actively capturing the precursor characteristics of the gradual narrowing of the upper airway before the blood oxygen saturation drops significantly.

[0100] Then, based on the signal quality index and phase lock value of each channel, a fusion confidence score is generated and a differentiated sleep feature set is constructed. First, the body movement intensity and heart rate are compared with the dynamic arousal threshold to perform a first-level arousal judgment to filter obvious activity periods. When not awake, a second-level binary classification is performed using the fusion confidence score and the first sleep feature set. When the second-level confidence score is insufficient, a third-level five-level classification is performed using the second sleep feature set and the upper airway collapse precursor indicator. Thus, the whole night's sleep is staged using low-power cascaded inference.

[0101] Finally, based on historical sleep monitoring data, each dynamic threshold is updated using a sliding window, and a health warning is triggered based on the accumulated upper airway collapse precursor indicators, thus completing the entire process of sleep health management from early risk identification to proactive alerts without the user's awareness during the sleep stage.

[0102] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A sleep monitoring system based on a smartwatch, characterized in that, include: The data acquisition and evaluation module controls the PPG sensor, accelerometer, and microphone triggered by the low-power voice activity detection unit to acquire data. The heart rate and heart interval were extracted from the pulse wave signal, the body motion intensity was extracted from the acceleration signal, the respiratory sound envelope was extracted from the audio signal after filtering and the original audio data was discarded, and the PPG signal quality index, body motion signal quality index and audio signal quality index were calculated. The phase analysis module extracts the respiratory sinus arrhythmia waveform from the heartbeat interval and calculates the phase synchronization characteristics between the waveform and the respiratory sound envelope; The warning sign recognition module generates an upper airway collapse warning sign when the phase synchronization feature is continuously decoupled and the synchronization satisfies the following conditions: the amplitude of the respiratory sound envelope decays, the low-frequency to high-frequency energy ratio extracted from the heartbeat interval exceeds the dynamic autonomic nervous system regulation threshold, and the body movement intensity is lower than the dynamic body movement threshold. The cascaded inference module generates a fusion confidence score based on the signal quality index and the phase synchronization features, and constructs a first sleep feature set and a second sleep feature set; it performs a first-level arousal determination by comparing body movement intensity and heart rate with a dynamic arousal threshold; when not awake, it performs a second-level binary sleep staging by comparing the fusion confidence score with the first sleep feature set; when the second-level confidence score is insufficient, it performs a third-level pentatonic sleep staging by comparing the second sleep feature set with the upper airway collapse precursor indicators. The sliding adaptive module updates the dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold based on historical sleep monitoring data, and triggers a health warning based on the cumulative results of the upper airway collapse precursor indicators.

2. The sleep monitoring system based on a smartwatch according to claim 1, characterized in that, The data collected by the control PPG sensor, accelerometer, and microphone triggered by the low-power voice activity detection unit includes: The sensing and acquisition unit controls the PPG sensor and the accelerometer to acquire signals respectively, and outputs pulse wave signals and acceleration signals. The voice detection unit detects the energy of ambient sound and outputs a voice trigger signal. The audio acquisition unit receives the voice trigger signal and controls the microphone to acquire audio signals; The time synchronization unit receives the pulse wave signal, the acceleration signal, and the audio signal, aligns the three signals in time, and generates synchronized acquisition data.

3. The sleep monitoring system based on a smartwatch according to claim 1, characterized in that, The process of filtering the audio signal to extract the respiratory sound envelope, discarding the original audio data, and calculating the PPG signal quality index, body motion signal quality index, and audio signal quality index includes: The audio processing unit filters the audio signal collected by the microphone and extracts the breathing sound envelope. After extraction, the original audio data is discarded. The quality calculation unit calculates the PPG signal quality index, body motion signal quality index, and audio signal quality index based on the pulse wave signal, acceleration signal, and respiratory sound envelope, and outputs the PPG signal quality index, body motion signal quality index, and audio signal quality index.

4. A sleep monitoring system based on a smartwatch according to claim 1, characterized in that, The calculation of the phase synchronization characteristics between the waveform and the breath sound envelope includes: The instantaneous phase extraction unit extracts the instantaneous phase sequences corresponding to the respiratory sinus arrhythmia waveform and the respiratory sound envelope, respectively; The phase difference calculation unit calculates the phase difference data between two instantaneous phase sequences; The synchronization feature generation unit performs window statistical processing on the phase difference data to generate phase synchronization features.

5. A sleep monitoring system based on a smartwatch according to claim 1, characterized in that, The process of generating a fusion confidence score based on the signal quality index and the phase synchronization features, and constructing a first sleep feature set and a second sleep feature set, includes: The signal normalization unit quantizes the PPG signal quality index, the body motion signal quality index, the audio signal quality index, and the phase synchronization feature to generate signal reliability parameters. The confidence calculation unit generates fusion weights based on the proportional relationship between the reliability parameters of each signal, and generates a fusion confidence score after weighted fusion calculation. The first feature construction unit consists of the heartbeat interval, the heart rate, the body movement intensity, and the respiratory sound envelope, forming a first sleep feature set. The second feature construction unit generates a second sleep feature set by combining the cardiac interval, body movement intensity, respiratory sound envelope, phase synchronization feature, and upper airway collapse precursor identifier.

6. A sleep monitoring system based on a smartwatch according to claim 1, characterized in that, The process of determining primary arousal by comparing body movement intensity and heart rate with a dynamic arousal threshold; performing secondary binary sleep staging by using the fusion confidence level and the first sleep feature set when not awake; and performing tertiary pentatonic sleep staging by using the second sleep feature set and the upper airway collapse precursor markers when the secondary confidence level is insufficient, includes: The first-level determination unit compares the body movement intensity and the heart rate with the dynamic arousal threshold respectively to generate an arousal determination result; The secondary classification unit, when the awakening determination result is an unawakened state, uses the fusion confidence and the first sleep feature set to perform binary sleep staging and generate a binary sleep state result. When the confidence level of the binary sleep stage is insufficient, the third-level classification unit uses the second sleep feature set and the upper airway collapse precursor marker to perform a five-category sleep stage and generate a five-category sleep state result. The temporal smoothing unit smooths the sleep state categories of adjacent time slices and outputs sleep staging data.

7. A sleep monitoring system based on a smartwatch according to claim 1, characterized in that, The step of updating the dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold based on historical sleep monitoring data includes: The historical data processing unit filters historical sleep monitoring data, removes data periods that lack heart rate interval features, body movement intensity features, or sleep state features, and extracts heart rate interval features, body movement intensity features, and sleep state features to generate historical sleep feature data. The sliding baseline calculation unit establishes a time sliding window based on the historical sleep characteristic data, performs statistical calculations on the features within the window, and generates autonomic nervous regulation baseline parameters, body movement baseline parameters, and wakefulness state baseline parameters. The threshold update unit updates the dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold respectively based on the autonomic nervous system regulation baseline parameters, the body movement baseline parameters, and the arousal state baseline parameters, and outputs them to the aura recognition module and the cascaded inference module.

8. A sleep monitoring method based on a smartwatch, characterized in that, The method for a smartwatch-based sleep monitoring system according to any one of claims 1-7 includes the following steps: It collects pulse wave signals and acceleration signals, and performs low-power energy detection on ambient sound. When the ambient sound energy exceeds the wake-up threshold, it activates the microphone to collect audio signals. The heart rate and heart interval are extracted from the pulse wave signal, the body motion intensity is extracted from the acceleration signal, the respiratory sound envelope is extracted from the audio signal after filtering and the original audio data is discarded, and the PPG signal quality index, body motion signal quality index and audio signal quality index are calculated. The respiratory sinus arrhythmia waveform is extracted from the intercardiac interval, and the phase synchronization characteristics between the waveform and the respiratory sound envelope are calculated. When the phase synchronization feature is continuously decoupled, and the synchronization satisfies the following conditions: the amplitude of the respiratory sound envelope decays, the ratio of low-frequency to high-frequency energy extracted from the heartbeat interval exceeds the dynamic autonomic nervous system regulation threshold, and the body movement intensity is lower than the dynamic body movement threshold, an upper airway collapse precursor indicator is generated. A fusion confidence score is generated based on the signal quality index and the phase synchronization features, and a first sleep feature set and a second sleep feature set are constructed. A first-level arousal determination is performed using body movement intensity and heart rate as a reference to the dynamic arousal threshold. When not awake, a second-level binary sleep stage is performed using the fusion confidence score and the first sleep feature set. When the second-level confidence score is insufficient, a third-level pentatonic sleep stage is performed using the second sleep feature set and the upper airway collapse precursor indicators. The dynamic autonomic nervous system regulation threshold, the dynamic body movement threshold, and the dynamic arousal threshold are updated based on historical sleep monitoring data, and a health warning is triggered based on the cumulative results of the upper airway collapse precursor indicators.

Images