A Smart Early Warning System and Method for Sleep Apnea Based on Multi-Sensor Fusion in Mobile Phones

Abstract

This invention discloses a smart sleep apnea early warning system and method based on multi-sensor fusion in a mobile phone. The system includes: a data acquisition module for collaboratively acquiring multimodal data from the phone's built-in microphone, accelerometer, and optional wearable device heart rate sensor in low-power mode; a data processing module, including a data preprocessing unit and an improved CNN-LSTM early warning model, for noise reduction, feature extraction, and identification of apnea events in the acquired data; and an output module for triggering a graded early warning and generating a health report when an apnea event meeting preset conditions (single apnea ≥ 10 seconds and confidence ≥ 85%) is identified. This invention achieves high-precision (accuracy ≥ 90%) and low-power (nighttime monitoring power consumption ≤ 5mA) non-contact sleep monitoring through multi-sensor fusion and lightweight model design, effectively solving the problems of poor convenience, missing early warnings, and high power consumption in existing technologies, and is particularly suitable for home health management scenarios.

Description

Technical Field

[0001] This invention relates to the fields of mobile terminals and digital health technology, and in particular to a system and method for low-cost, non-contact, and intelligent early warning of sleep apnea syndrome based on data fusion using built-in sensors in smartphones. Background Technology

[0002] Sleep apnea syndrome is a common sleep disorder characterized by repeated pauses in breathing (usually defined as pauses lasting ≥10 seconds) and hypoventilation during sleep. Long-term suffering from this syndrome can significantly increase the risk of developing chronic diseases such as hypertension, coronary heart disease, stroke, and type II diabetes, posing a serious threat to the patient's quality of life and life safety.

[0003] Currently, the "gold standard" for diagnosing sleep apnea in clinical practice is polysomnography (PSG). PSG requires connecting multiple sensors to the patient's body and recording more than ten physiological parameters throughout the night, including electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), electrocardiography (ECG), nasal and oral airflow, blood oxygen saturation, and chest and abdominal respiratory movements. Although PSG results are accurate and reliable, it has significant drawbacks: the equipment is expensive and bulky, must be operated by technicians in a professional sleep laboratory, patients need to fall asleep in an unfamiliar environment, which may affect the authenticity of sleep, and the waiting time for appointments is long. It is only suitable for clinical diagnosis, and the professional equipment depends on medical settings, making it unsuitable for routine screening and long-term monitoring.

[0004] With the increasing popularity of wearable devices, smartwatches, smart bracelets, and other devices have integrated heart rate monitoring functions and attempted to indirectly infer respiratory events by analyzing the periodic changes in heart rate variability. However, these wearable devices also have certain limitations: First, they require users to wear additional devices, which may cause discomfort or allergies, resulting in poor compatibility. Furthermore, the devices may shift or fall off during sleep, causing monitoring interruptions. Second, relying solely on indirect monitoring through a heart rate sensor makes it susceptible to motion artifacts, making it difficult to distinguish between physiological heart rate fluctuations and changes caused by apnea. This is especially true during the "normal silent sleep" and "apnea" phases, where heart rate change patterns may be similar, leading to misjudgments or missed detections.

[0005] In recent years, some health monitoring apps based on smartphones have emerged, such as those that use the phone's microphone to record snoring. While these solutions are convenient, they are usually limited in function, only able to collect snoring sounds and record the intensity of snoring, and cannot accurately identify silent sleep apnea events. In addition, microphones are highly susceptible to interference from environmental noise (such as car noise outside the window, partner's snoring, and air conditioning noise), resulting in a very high false alarm rate. More importantly, most existing apps remain at the data recording level and lack effective, real-time or near-real-time early warning mechanisms based on intelligent algorithms, making it impossible to intervene in a timely manner when dangerous events occur. At the same time, continuous high-frequency audio collection will quickly drain the phone's battery, making it impossible to support monitoring all night.

[0006] In summary, existing sleep apnea syndrome monitoring methods suffer from drawbacks such as high usage barriers and insufficient convenience. There is an urgent need in this field for a sleep apnea monitoring solution that can balance accuracy, convenience, low power consumption, and effective early warning. As a highly ubiquitous mobile terminal, smartphones, with their built-in high-precision microphones, accelerometers, and other sensors, provide the hardware foundation for developing low-cost, non-invasive daily monitoring. The core starting point of this invention is how to fuse multi-sensor information through innovative algorithm models to achieve high-precision respiratory event recognition and intelligent early warning under low power consumption constraints. Summary of the Invention

[0007] This invention aims to provide a smart sleep apnea early warning system and method based on multi-sensor fusion of mobile phones. Its main purpose is to solve the problems of poor convenience, low accuracy, lack of effective early warning and high power consumption of existing monitoring solutions, and to realize a home sleep health management tool that requires no additional equipment, is non-contact, highly accurate, low power consumption and can provide timely early warning.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A smart sleep apnea early warning system based on multi-sensor fusion of mobile phones is based on a three-layer architecture that integrates data acquisition, intelligent analysis, and decision-making early warning. The system is mainly deployed on users' smartphones and can be selectively linked with wearable devices.

[0009] Users simply place their phones on the mattress next to their pillow before going to sleep. When the user turns over in their sleep, their body movement causes the mattress to shake slightly. This shaking is transmitted through the mattress to the phone, causing it to tilt, wobble, or even shift slightly. Even if the user only turns their body, it may cause uneven pressure on the mattress and a slight change in angle. Once the phone's posture (angle) changes relative to the direction of gravity, the components of the gravitational acceleration vector on the phone's three axes will change accordingly. The accelerometer in the phone can sensitively capture this continuous change in component value. The processor reads the three-axis data from the accelerometer in real time and calculates the phone's current posture angle (such as pitch angle and roll angle) using a specific posture calculation algorithm (e.g., complementary filtering based on changes in the gravity vector or simple tilt angle calculation).

[0010] In practical applications, instead of directly measuring the swaying angle or linear displacement in centimeters of the phone on the mattress, the integral result of the acceleration data or the significant change in the attitude angle is equivalent to the lateral displacement (e.g., more than 3 centimeters) of the user's torso center. This is an inference and calibration based on a physical model. For foldable screen phones and curved screen phones, the compatibility of different models can be ensured by optimizing the sensor acquisition angle and algorithm parameters.

[0011] By using a highly accurate accelerometer to detect minute changes in posture caused by a user's turning over, which are transmitted to the phone via the mattress, the system then uses an algorithm model to convert this posture change signal into a reliable inference of the user's body rotation (such as turning over). This is an equivalent result after algorithmic conversion, thus achieving contactless sleep behavior monitoring. For example, when the accelerometer detects a continuous change signal with an amplitude exceeding a preset threshold, and it matches the time characteristics of a turning over action (rather than a short tap), the system determines that "a turning over event has occurred" and triggers the "state switching" logic.

[0012] Specifically, at the data acquisition layer, the system enters a low-power monitoring mode after the phone screen is off (stops at sunrise or when manually turned off), coordinating the microphone and accelerometer to collect data. The microphone focuses on the 50-500Hz breathing-related frequency band at an 8kHz sampling rate, storing one frame of data every 10 seconds to reduce storage pressure. When the accelerometer detects shaking or displacement of the phone at a 10Hz sampling rate, equivalent to detecting a rolling motion with an amplitude greater than 3cm, it marks "state switch" and pauses audio acquisition to avoid invalid data. At the same time, if the user is also wearing a smart bracelet / watch on their wrist and it is paired with the phone, the system can intermittently (every 30 seconds) synchronize the heart rate data of the smart bracelet / watch's heart rate sensor through Bluetooth Low Energy technology as an auxiliary judgment feature.

[0013] Specifically, at the data processing layer, the collected raw data undergoes wavelet threshold denoising, feature extraction, and standardization in the preprocessing unit to generate a unified feature vector. Subsequently, this feature vector is input into an improved CNN-LSTM hybrid model, which uses a convolutional neural network to capture spatial local patterns in breathing sounds and motion signals, and analyzes respiratory rate, pause duration, and heart rate variability (e.g., heart rate variability ≥15 bpm is an anomalous feature). Then, a long short-term memory network is used to learn the dynamic evolution of these patterns over time. Furthermore, the model is trained using the Focal Loss function to address the class imbalance problem where apnea events (positive samples) are far fewer than normal breathing events (negative samples), thereby improving the ability to identify the minority class.

[0014] Specifically, at the output layer, the system makes decisions based on the model's output. When the model determines that a sleep apnea event has occurred, and the event lasts for ≥10 seconds and the model confidence level is ≥85%, a graded warning mechanism is triggered. For example, a weak-to-strong phone vibration combined with a very low-brightness screen light (graded vibration + light warning) is used to alert the user without waking them. The next morning, the system automatically generates a detailed and quantitative sleep health report, including the number of sleep apneas, the longest duration, the respiratory rate curve, etc., and performs risk classification, supporting local encrypted storage and secure sharing.

[0015] Compared with existing technologies, this invention requires no additional equipment, provides non-contact monitoring, and boasts an early warning accuracy rate of ≥90%. It effectively addresses the core pain points of existing technologies, making it highly suitable for daily home health management scenarios, and possesses the following significant advantages: 1. High convenience and contactless operation: It makes full use of the user's existing smartphone, without the need to purchase or wear any additional professional equipment, achieving truly "unobtrusive" monitoring and greatly reducing the threshold for use.

[0016] 2. High accuracy and low false positive rate: By fusing multi-source information from a microphone (directly monitoring breathing sounds), an accelerometer (assisting in judging body position and interference), and an optional heart rate sensor (adding physiological evidence), it pioneered the collaborative acquisition of mobile phone microphone, accelerometer and optional wearable heart rate sensor to achieve "no additional equipment + non-contact" monitoring, solving the problem of interference from a single sensor in wearable devices; and combined with the deep learning CNN-LSTM model, it can effectively distinguish normal breathing, snoring, environmental noise and sleep apnea, taking into account spatial and temporal features, reducing false positives caused by sample imbalance, and improving the warning accuracy rate to over 90%.

[0017] 3. Effective intelligent early warning: A dual-threshold early warning mechanism based on duration and confidence level is introduced, and a gentle graded early warning method is designed in combination with sleep scenarios. With the addition of quantitative health reports, a closed loop of "monitoring-early warning-intervention" is constructed, realizing the leap from "passive recording" to "active intervention".

[0018] 4. Excellent low-power performance: By optimizing the sampling rate (strategy), data storage rules (strategy) and lightweight model design, the extra power consumption of monitoring throughout the night (8 hours) is controlled within 3%, achieving low-power operation with nighttime monitoring power consumption ≤5mA, balancing data integrity and battery life, and ensuring the practicality and sustainability of the solution.

[0019] 5. Data security and privacy protection: Core data processing and early warning are all completed locally on the mobile phone, and sensitive physiological data does not need to be uploaded to the cloud, effectively protecting user privacy. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall hardware and logic architecture of the system provided in an embodiment of the present invention; Figure 2 This is a flowchart of an intelligent sleep apnea early warning method provided in an embodiment of the present invention; Figure 3 This is a detailed structural diagram of an improved CNN-LSTM early warning model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of multi-sensor data time-series collaborative acquisition provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the warning interface and health report user interface provided in an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will now be described in detail and completely with reference to the accompanying drawings. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] ( Example 1 This embodiment provides a typical physical architecture and logical composition of a smart sleep apnea early warning system based on multi-sensor fusion in a mobile phone. (See reference...) Figure 1The document illustrates the hardware environment and logical module composition of the system 100 of the present invention. The system 100 mainly runs on a smartphone 101. Logically, the system 100 is divided into three main modules: a data acquisition module 110, a data processing module 120, and an output module 130. The modules interact with each other through a system bus to form a complete closed-loop system. The smartphone 101 serves as the core hardware platform and internally includes at least one processor 102, a memory 103, a microphone 104, an accelerometer 105, a display screen 106, a vibration motor 107, and a Bluetooth module 108. Optionally, the system 100 can communicate with wearable devices such as a smart bracelet or smartwatch 109 through the Bluetooth module 108 to obtain heart rate data.

[0023] Specifically, the data acquisition module 110 is responsible for scheduling the heart rate data of the microphone (104), accelerometer (105) and wearable device (109) acquired through Bluetooth module (108); during sleep at night, the data acquisition module 110 coordinates various sensors to collect data in a low-power manner; after the user turns on the "sleep monitoring" mode through the APP interface, the smartphone 101 will prompt the user to place the phone (such as next to the pillow), and then the screen will turn off and the system will enter a low-power operation state; the power control unit in the data acquisition module 110 will dynamically manage the power supply and sampling frequency of each sensor.

[0024] Specifically, the data processing module 120 is the "brain" of the system 100. It receives raw data from the data acquisition module 110 and performs calculations and analysis. First, the data is cleaned and features are extracted in the data preprocessing unit 121. Then, the standardized (e.g., normalized to [0,1] and concatenated into a 64-dimensional (with heart rate) / 48-dimensional (without heart rate) vector) feature vector is sent to the core improved CNN-LSTM warning model 122 for inference calculation. The CNN-LSTM warning model 122 has been trained on massive amounts of labeled data and can identify sleep apnea events. Optionally, the CNN-LSTM warning model 122 supports OTA upgrades of model parameters and functional modules to adapt to emerging sensor technologies (such as flexible sensors) and new research results in sleep health, thus extending the technology life cycle.

[0025] Specifically, the output module 130 drives the vibration motor 107 and the display screen 106 to perform early warning and report generation based on the analysis and inference results of the CNN-LSTM early warning model 122. If an event that meets the early warning conditions is identified, the early warning execution unit 131 of the output module 130 will immediately start a graded early warning. After the entire monitoring cycle (such as from falling asleep to waking up the next day) ends, the report generation unit 132 of the output module 130 will also integrate the data from the whole night to generate a visualized sleep health report and display it on the display screen 106.

[0026] ( Example 2 This embodiment provides a smart sleep apnea early warning method based on multi-sensor fusion of mobile phones. (See attached document.) Figure 2 This demonstrates the complete workflow of the method of the present invention: after the user launches the relevant application and enters the sleep monitoring mode, the process begins (step S201); the system 100 first initializes each sensor and activates the low-power acquisition strategy through the power consumption control unit (step S202); combined with Figure 4 As shown, the subsequent data acquisition process involves multi-sensor time-coordinated operation: 1. On the audio acquisition channel 401, the microphone 104 continuously records at a sampling rate of 8kHz, but only stores one frame of valid data after 50-500Hz bandpass filtering every 10 seconds (corresponding to step S203). 2. On the motion acquisition channel 402, the accelerometer 105 samples at a frequency of 10Hz and calculates the displacement in real time. When the displacement exceeds the preset threshold (e.g., 3cm), it is marked as a rolling event 403. At this time, the system will trigger a short audio acquisition pause window 404 (e.g., 10 seconds) to avoid recording invalid noise data (corresponding to step S204). 3. If the wearable device 109 is connected to the heart rate acquisition channel 405, the system 100 will synchronize the heart rate data every 30 seconds via the Bluetooth module 108 (corresponding to step S205).

[0027] This collaborative strategy ensures data validity while keeping the average operating current of system 100 below 5mA. After preprocessing (e.g., using wavelet thresholding to remove environmental noise) and feature extraction (e.g., respiratory frequency / pause duration / intensity variance, movement turning frequency / stability, (optional) average heart rate / HRV) (step S206), the collected data is fed into the improved CNN-LSTM warning model 122 for inference (step S207). If the CNN-LSTM warning model 122 determines that the warning conditions are met (step S208: pause ≥ 10 seconds and confidence ≥ 85%), it executes a graded warning (step S209). System 100 continuously monitors in a loop until the preset end time (e.g., sunrise) is reached, and then generates the final sleep health report (step S210).

[0028] ( Example 3 This embodiment details the internal structure of the improved CNN-LSTM early warning model 122, which effectively integrates spatial and temporal features, a key factor in achieving high accuracy; for example... Figure 3As shown, the preprocessed feature vector sequence 301 is first input into the feature learning layer 302, which consists of three alternating one-dimensional convolutional layers (302a, 302b, and 302c) and two pooling layers (302d and 302e), with a kernel size of 3, used to extract the local spatial patterns of the features (i.e., extract the spatial correlation between local features), such as the relationship between the periodic intensity changes of breathing sounds and brief silence intervals. The feature sequence abstracted by the feature learning layer 302 is then fed into the temporal analysis layer 303, which consists of two stacked LSTM network layers (303a and 303b), each containing 128 neurons, used to learn the long-term dependencies of features over time. The LSTM network, due to its gating mechanism, It excels at learning dependencies in long-term time series and can memorize long-term contextual information of breathing patterns, such as the gradual change in respiratory rate before and after a sleep apnea event. Then, the output of the last time step of the time series analysis layer 303 is sent to the classification output layer 304. This classification output layer 304 is a fully connected layer. Through a fully connected network and using the Sigmoid activation function, it outputs the classification probability 304a of the sleep apnea event and the confidence level 304b of the current prediction, that is, it outputs a scalar between 0 and 1. The closer the value of this scalar is to 1, the higher the probability that the warning model 122 believes that a sleep apnea event will occur within the current time window. At the same time, the warning model 122 also outputs a confidence score, which reflects the degree of certainty of the warning model 122 in making this judgment.

[0029] For example, the early warning model 122 was trained using a database containing 500 subjects and a total of 100,000 labeled data points. The data was divided into training, validation, and test sets in a 7:2:1 ratio. During training, the Focal Loss function (α_t=0.75, γ=2) was used to focus on samples that were difficult to classify, thus alleviating the class imbalance problem. After testing, the early warning model 122 achieved an accuracy of 92%, a recall of 89%, and a single inference time of less than 15 milliseconds, fully meeting the real-time requirements.

[0030] ( Example 4 This embodiment provides the specific process for triggering early warnings and generating reports, such as... Figure 1As shown, when the CNN-LSTM warning model 122 outputs an indication of a breathing apnea event (e.g., output value > 0.5), and the system timer 100 confirms that the duration of the apnea has exceeded 10 seconds, and the confidence level of the warning model 122 is higher than 85%, the warning execution unit 131 is triggered. First, the vibration motor 107 is controlled to generate a gentle short vibration. If the apnea event is detected again within the following 30 seconds, a second, slightly stronger vibration is triggered, and a local area of ​​the display screen 106 is lit up at the same time. The brightness of the display screen 106 is strictly controlled to be below 5 cd / m², emitting a dim light prompt. This tiered warning method aims to try to awaken the user's spontaneous breathing with minimal interference and avoid excessive disturbance.

[0031] After monitoring, the report generation unit 132 will compile key indicators for the entire night, including the total number of apneas, the longest apnea duration, the average respiratory rate, and the number of times the user turns over, and generate a health report. According to the guidelines of the American Academy of Sleep Medicine (AASM), the sleep apnea-hypopnea index (approximate estimate) is divided into low risk (≤5 times / hour), medium risk (6-15 times / hour), and high risk (≥16 times / hour). The health report clearly presents this information in chart form and provides brief suggestions. All data is encrypted and stored locally on the user's phone by default, and users can choose to share the report with family or doctors via a secure interface. It can collaborate with health management apps and physical examination institutions, providing data interface access and supporting doctors to remotely view monitoring reports, adapting to chronic disease management scenarios.

[0032] ( Example 5 This embodiment provides two key interfaces for system 100 to interact with the user, such as... Figure 5 As shown, when the warning conditions are triggered, the system 100 will activate the warning interface 501; as shown in the interface diagram, most of the mobile phone screen 106 remains dark, and only the edges or designated areas display a dim light prompt 502 with the brightness strictly controlled below 5 cd / m²; at the same time, the vibration motor 107 will emit a tactile vibration that gradually increases in intensity; this design is intended to gently remind the user and avoid disturbing sleep.

[0033] After the monitoring period ends, the system 100 will generate and display a health report interface 503, which includes the following core elements: overall risk assessment level 504 (such as "medium risk"), statistical charts of sleep apnea events 505 (such as a bar chart of the distribution of the number of apneas per hour), a list of key indicators 506 (such as total number of apneas, longest apnea duration, average respiratory rate, etc.), and data sharing options 507; this health report can provide users with an intuitive and quantitative assessment of sleep quality.

[0034] The intelligent sleep apnea early warning system and method based on multi-sensor fusion of mobile phones provided by this invention can be widely used in consumer electronic devices such as smartphones, tablets, and smart speakers. It can be deployed through software updates or pre-installed applications and seamlessly integrated with the existing mobile ecosystem, providing ordinary users with professional-grade sleep health screening and management tools. It has broad market prospects and significant industrial applicability.

[0035] It should be understood that the above description is only a preferred embodiment of the present invention and is not sufficient to limit the technical solution of the present invention. For those skilled in the art, within the spirit and principles of the present invention, additions, subtractions, substitutions, transformations, or improvements can be made based on the above description. For example, a barometer (i.e., a barometric pressure sensor) can be added to monitor minute pressure changes caused by breathing (accuracy ≥ 92%), or a microphone can be replaced with an acoustic sensor (such as a bone conduction sensor), or a lightweight Transformer model can be used to replace the CNN-LSTM warning model (40% reduction in volume and 25% increase in speed), or an RNN can be used to replace LSTM, or a transfer learning module can be added to the model, or the application APP can be extended to terminals such as tablets and smart speakers (the core architecture remains unchanged, only the hardware interface is adapted); or cloud collaborative analysis can be implemented (i.e., "local collection + cloud collaboration", historical data is stored in the cloud and trend analysis is performed, and real-time warnings are retained locally), or correlation analysis can be performed during sleep, or a dynamic sampling rate can be used to replace a fixed sampling rate, or a ringtone can be used to replace vibration for warning, or multi-user data storage can be implemented, etc. All such technical solutions after additions, subtractions, substitutions, transformations, or improvements should fall within the protection scope of the appended claims of the present invention.

Claims

1. A smart sleep apnea early warning system based on multi-sensor fusion in a mobile phone, the system being deployed on a smartphone terminal and communicating with an optional wearable device, characterized in that, The system includes: The data acquisition module is configured to collaboratively acquire raw data from multiple sensors using a low-power strategy after the device enters sleep monitoring mode, wherein the sensors include at least the microphone and accelerometer built into the mobile phone. The data processing module is communicatively connected to the data acquisition module and is configured to preprocess and extract features from the raw data, and to analyze the extracted features using a trained respiratory event recognition model to identify sleep apnea events. The output module is communicatively connected to the data processing module and is configured to trigger an early warning operation when the respiratory event recognition model determines that a sleep apnea event that meets the preset early warning conditions has occurred, and to generate a sleep health report after the monitoring period ends.

2. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 1, characterized in that, The data acquisition module further includes: An audio acquisition unit is configured to control the microphone to acquire ambient audio data at a first sampling rate, filter the data based on a preset frequency band range, and store audio frames in a first cycle. The motion acquisition unit is configured to control the accelerometer to acquire device motion data at a second sampling rate, and to determine the user's rolling over action based on a displacement threshold, and to pause the acquisition of some data when the rolling over action is detected. An optional heart rate acquisition unit is configured to connect to wearable devices via Bluetooth Low Energy protocol to synchronously acquire the user's heart rate data and heart rate variability data at a third sampling rate; The power consumption control unit is configured to activate the sleep monitoring mode after the device screen is turned off, and dynamically manage the sampling rate and data storage frequency of each acquisition unit so that the overall system operating current does not exceed 5mA.

3. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 2, characterized in that: The first sampling rate is 8kHz, the preset frequency band range is 50Hz to 500Hz, and the first period is 10 seconds; the second sampling rate is 10Hz, the displacement threshold is 3 cm, and the third sampling rate is 30 seconds.

4. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 1, characterized in that, The data processing module further includes: The data preprocessing unit is configured to perform noise reduction and standardization on the raw data and extract multidimensional feature vectors related to apnea, wherein the features include at least respiratory rate, apnea duration, respiratory intensity variance, turning frequency, exercise stability, average heart rate and heart rate variability. The respiratory event recognition model employs an improved convolutional neural network-long short-term memory hybrid model. Its input is the multidimensional feature vector, and its output is the classification result of the sleep apnea event and the corresponding confidence score.

5. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 4, characterized in that, The respiratory event recognition model includes, in sequence: The feature learning layer, consisting of at least three convolutional layers and pooling layers, is used to extract the spatial local correlations of the input features. The time series analysis layer, consisting of at least two LSTM network layers, is used to learn the evolution of the spatial features over time. The classification output layer uses the Sigmoid activation function to output a binary classification result and its confidence score.

6. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 5, characterized in that: The breathing event recognition model uses Focal Loss as the loss function during training, with the adjustment factor α_t set to 0.75 and the focusing parameter γ set to 2, to address the imbalance in the number of samples between sleep apnea events and normal breathing events.

7. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 1, characterized in that: The preset warning conditions are: the duration of a single apnea event is greater than or equal to 10 seconds, and the corresponding confidence level output by the respiratory event recognition model is greater than or equal to 85%.

8. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 1, characterized in that, The output module further includes: The warning execution unit is configured to trigger a graded warning when the preset warning conditions are met. The graded warning includes tactile vibration with progressive intensity and a dim light prompt with a brightness of less than 5 cd / m². The report generation unit is configured to, at the end of the monitoring period, count the total number of apneas, the longest apnea duration, the average respiratory rate, and the frequency of turning over during the monitoring period, perform risk assessment and classification according to preset rules, and generate a data report containing visual charts.

9. The intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion according to claim 8, characterized in that: The report generation unit is also configured to encrypt and store the generated data report locally, and provide a secure data sharing interface to support sharing the report to a designated health management application or medical platform.

10. A method for intelligent early warning of sleep apnea based on multi-sensor fusion in mobile phones, characterized in that, The method, applied to any one of the following claims, in the intelligent sleep apnea early warning system based on mobile phone multi-sensor fusion, comprises the following steps: Activate sleep monitoring mode and control the phone to enter a low-power operation state; It collaboratively acquires multimodal data from the phone's microphone, accelerometer, and optional heart rate sensor; The collected multimodal data are preprocessed to extract multidimensional feature vectors representing respiratory states; The multidimensional feature vector is input into a pre-trained improved CNN-LSTM hybrid model for inference to obtain the judgment result of whether a sleep apnea event exists and its confidence level; If the judgment result indicates the existence of a breathing apnea event with a duration of ≥10 seconds and a confidence level of ≥85%, then a graded warning operation will be performed; At the end of the monitoring period, a sleep health report containing statistical data and risk assessment is automatically generated.

11. The intelligent sleep apnea early warning method based on mobile phone multi-sensor fusion according to claim 10, characterized in that, The steps for preprocessing multimodal data include: Wavelet thresholding noise reduction algorithm is applied to the audio data captured by the microphone to suppress ambient noise; The respiratory audio spectrum features are calculated from the noise-reduced audio data, including estimating the respiratory rate and pause interval using a peak detection algorithm; The accelerometer data is analyzed to calculate the vector amplitude of the triaxial acceleration, and the rolling event is detected by sliding window variance. When heart rate data is available, the coefficient of variation of adjacent heartbeat intervals is calculated as an indicator of heart rate variability. All extracted features are normalized and concatenated into a fixed-dimensional feature vector.

12. The intelligent sleep apnea early warning method based on mobile phone multi-sensor fusion according to claim 10, characterized in that, The training steps for the improved CNN-LSTM hybrid model include: Collect a large-scale labeled sleep dataset, which contains data collected synchronously by multiple sensors and their corresponding respiratory event annotations; The dataset is cleaned and augmented, and then divided into training, validation, and test sets proportionally. Using the training set, the model is iteratively trained with Focal Loss as the loss function, and the model performance is monitored on the validation set to prevent overfitting. The trained model is then evaluated using the test set to ensure that its accuracy is no less than 90% and its recall is no less than 88%.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent sleep apnea warning method based on mobile phone multi-sensor fusion as described in any one of claims 10 to 12.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the intelligent sleep apnea warning method based on mobile phone multi-sensor fusion as described in any one of claims 10 to 12.

Images