An atrial fibrillation early warning intelligent bracelet based on multi-modal biosignal fusion and an analysis method

Abstract

This invention discloses a smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion and its analysis method. The bracelet includes a ring-shaped main body, adapted to fit the user's wrist, made of biocompatible material. The main body contains cavities and mounting structures to accommodate and fix various functional modules, and exposes the optical window of the PPG signal acquisition module and the electrode contacts of the ECG signal acquisition module. Each functional module further includes a motion sensor module, a signal processing unit, a storage unit, a wireless communication module, and a power management unit. This invention relates to the field of smart bracelet technology. This smart bracelet and analysis method for atrial fibrillation early warning, through the fusion of dual-modal data from PPG and ECG, combined with the artifact removal capabilities of motion sensor data, overcomes the limitations of single signals. The PPG signal provides continuous hemodynamic information and is sensitive to early and subtle changes in heart rate and rhythm.

Description

Technical Field

[0001] This invention relates to the field of smart bracelet technology, specifically to a smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion and its analysis method. Background Technology

[0002] Cardiovascular diseases are a leading cause of death and disability worldwide, posing a serious threat to human health. Among them, atrial fibrillation, with its high incidence and significantly increased risk of stroke and heart failure, has received increasing attention in the medical and health field. Because atrial fibrillation often manifests as paroxysmal or asymptomatic, its early detection and timely intervention face enormous challenges.

[0003] Reference patent publication number "CN202010261627.8" discloses an electronic device and its atrial fibrillation early warning method and medium, including: the electronic device extracting multiple sinus intervals from the photoplethysmography data of the user collected by the electronic device; the electronic device calculating the degree of change of the relative difference between multiple temporally adjacent sinus intervals among the multiple sinus intervals; and the electronic device determining whether to issue an atrial fibrillation early warning to the user based on the calculated degree of change.

[0004] As the technologies described above show, existing active measurement modes mean that users cannot perform continuous, unobtrusive, background real-time monitoring. For atrial fibrillation, a paroxysmal or asymptomatic disease, many early attacks are often brief and irregular, and may occur without the patient's awareness. Traditional on-demand measurement methods are prone to missing these instantaneous events, thus delaying the best time for early warning and intervention. Especially in complex and ever-changing daily environments, how to achieve continuous, unobtrusive, highly accurate, highly sensitive, and forward-looking real-time early warning of atrial fibrillation risk has become a major problem. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion and its analysis method, which solves the problem that existing smart bracelets and analysis methods for atrial fibrillation early warning based on multimodal biosignal fusion have limitations in complex and ever-changing daily environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion, comprising: The main body of the bracelet has a ring-shaped structure, which is adapted to be worn on the user's wrist. Its material is a biocompatible material. The main body of the bracelet has a cavity and mounting structure to accommodate and fix the various functional modules, and exposes the optical window of the PPG signal acquisition module and the electrode contacts of the ECG signal acquisition module. Each functional module also includes: a motion sensor module, a signal processing unit, a storage unit, a wireless communication module, a power management unit, and a user interaction module. The signal processing unit is an embedded microcontroller unit that is connected to the PPG signal acquisition module, the ECG signal acquisition module and the motion sensor module through a high-speed serial interface. It is used to receive and manage the PPG signal, the ECG signal and the motion data, and to perform data preprocessing, multi-dimensional physiological feature extraction, multi-modal feature fusion and atrial fibrillation risk assessment based on a deep learning model.

[0007] Preferably, the PPG signal acquisition module includes at least one light-emitting unit and at least one photoelectric detection unit, and the ECG signal acquisition module includes at least two bioelectrodes and an ECG front-end amplification circuit.

[0008] Preferably, the signal processing unit is an embedded microcontroller unit, which integrates a central processing unit core, random access memory and flash memory. The RAM capacity is not less than 512KB and the flash memory capacity is not less than 4MB. The signal processing unit is specifically used to perform the following functions: data preprocessing, multi-dimensional physiological feature extraction, multi-modal feature fusion, and atrial fibrillation risk assessment.

[0009] Preferably, the deep learning model is a hybrid model combining a convolutional neural network and a long short-term memory network. The convolutional neural network layer is used to automatically extract local pattern features from time series data. The user interaction module includes a vibration motor and an LED indicator. The vibration motor is used to provide a vibration alert when a potential risk of atrial fibrillation is detected.

[0010] The present invention also discloses an analysis method for atrial fibrillation early warning smart bracelet based on multimodal biosignal fusion, which specifically includes the following steps: Step S100: Synchronous acquisition and preprocessing of multimodal biosignals: The signal processing unit continuously receives the digital PPG signal output by the PPG signal acquisition module, the digital ECG signal output by the ECG signal acquisition module, and the acceleration data and angular velocity data output by the motion sensor module; For PPG signals, perform the following preprocessing: Sub-step S110: Bandpass filtering: The digital PPG signal is filtered using a digital infinite impulse response or finite impulse response bandpass filter, with the cutoff frequency set, for example, from 0.5 Hz to 10 Hz, to filter out baseline drift, high-frequency noise, and breathing artifacts. Sub-step S120: Motion artifact elimination: Using the acceleration and angular velocity data provided by the motion sensor module, an adaptive filtering algorithm or independent component analysis method is used to compensate for and eliminate motion artifacts in the PPG signal in real time; Sub-step S130: Signal quality assessment: Based on the signal-to-noise ratio, peak-to-valley amplitude, periodic stability and other indicators, perform real-time quality assessment on the processed PPG signal and mark low-quality signal segments; For ECG signals, perform the following preprocessing: Sub-step S140: Bandpass filtering: The digital ECG signal is filtered using a digital bandpass filter; Sub-step S150: Signal quality assessment: Based on indicators such as the detection rate, amplitude stability, and baseline drift of the QRS group, the processed ECG signal is assessed in real time, and low-quality signal segments are marked. If the ECG signal quality continues to be lower than the preset threshold, the smart bracelet will prompt the user to check the wearing status or electrode contact. Synchronization processing: PPG signals, ECG signals and motion sensor data are synchronized with high precision on the time axis, with an error of no more than 5ms.

[0011] Preferably, S200: Multidimensional physiological feature extraction: The signal processing unit extracts multidimensional physiological feature parameters from the preprocessed multimodal biological signals. The feature parameters are divided into PPG source features, ECG source features, spatiotemporal domain features and composite physiological features. PPG source feature extraction: Sub-step S210: Heart rate and heart rate variability characteristics: Identify the peak point of the pulse wave from the PPG signal, calculate the time interval between adjacent pulse cycles, and calculate the time-domain and frequency-domain heart rate variability characteristics such as heart rate, standard deviation, root mean square continuous difference, percentage of adjacent pulse intervals greater than 50ms, and low-frequency / high-frequency power ratio based on the PPI sequence. Sub-step S220: Pulse wave morphological characteristics: Analyze the morphological characteristics of the PPG pulse wave, including but not limited to: rise time, fall time, peak width, systolic time, diastolic time, enhancement index, reflection index, presence and location of biphasic peaks, etc. These characteristics are obtained through derivative analysis, Fourier transform or wavelet transform of the pulse wave. ECG source feature extraction: Sub-step S230: Heart rate and heart rate variability characteristics: Accurately detect the peak value of the R wave from the ECG signal, calculate the time interval between adjacent R waves, and calculate the heart rate and various ECG heart rate variability characteristics corresponding to the PPG source heart rate variability characteristics based on the time interval sequence between adjacent R waves. ECG heart rate variability characteristics are the gold standard indicator for the diagnosis of atrial fibrillation, and their accuracy is better than that of PPG heart rate variability. Sub-step S240: P wave detection and QRS complex morphology characteristics: accurately identify the P wave, QRS complex and T wave in the ECG signal, extract the duration, amplitude and axial deflection information of the QRS complex, pay special attention to the presence and morphology of the P wave, the relationship between the P wave and the QRS complex, the typical characteristics of atrial fibrillation are the absence of the P wave or its replacement by an irregular f wave, and the absolute irregularity of the RR interval; Spatiotemporal feature extraction: Sub-step S250: Joint time-frequency analysis features of fused signals: Perform joint time-frequency analysis on the synchronized PPG and ECG signals to extract features such as energy distribution and phase relationship at different frequencies and time scales; Sub-step S260: Cardiovascular coupling characteristics: Calculate the time delay or correlation between the PPG pulse wave and the ECG R wave; Preferably, step S300: Multimodal feature fusion and deep learning model construction: The signal processing unit fuses the PPG source features, ECG source features, spatiotemporal features and cardiovascular coupling features extracted in step S200 to form a high-dimensional comprehensive feature vector. The fusion method adopts feature-level fusion, that is, the features of different modalities are spliced ​​or weighted before being input into the deep learning model.

[0012] The deep learning model is pre-stored in the storage unit and runs on the signal processing unit. The deep learning model is a hybrid model that combines convolutional neural networks and long short-term memory networks. The convolutional neural network layer is used to automatically extract local pattern features from time series data, and the long short-term memory network layer is used to capture long-term dependencies and sequence patterns in time series data. The input layer of the deep learning model receives high-dimensional comprehensive feature vectors. The training data of the model includes a large amount of normal heart rhythm data, paroxysmal atrial fibrillation data, persistent atrial fibrillation data, and other common arrhythmia data, especially physiological signal data in the early stage of atrial fibrillation. The model is trained through supervised learning so that it can identify early warning signals and diagnostic features of atrial fibrillation. The output layer of the deep learning model outputs an atrial fibrillation risk score, which represents the likelihood of the user developing atrial fibrillation within the current time window. The score ranges from 0 to 1, where 0 represents extremely low risk and 1 represents extremely high risk.

[0013] Preferred step S400: Atrial fibrillation risk assessment and early warning: The signal processing unit performs atrial fibrillation risk assessment and early warning based on the atrial fibrillation risk score output by the deep learning model, combined with a preset risk threshold and time window mechanism; Sub-step S410: Real-time risk assessment: Compare the average atrial fibrillation risk score of the current time window with multiple preset risk thresholds; Sub-step S420: Continuous judgment: In order to avoid false alarms caused by occasional noise or artifacts, the smart bracelet requires the risk score to be continuously higher than the preset threshold within a specific time period or to show an abnormal pattern within a specified number of heartbeat cycles before triggering the warning.

[0014] Sub-step S430: Motion state correction: Correct the atrial fibrillation risk assessment results using activity data provided by the motion sensor module.

[0015] Sub-step S440: Personalized baseline adjustment. After the first wear or after a period of normal wear, the smart bracelet automatically establishes the user's personalized physiological baseline data. In subsequent risk assessments, the real-time detected features are compared with the personalized baseline to identify subtle changes that deviate from the normal range.

[0016] Preferably, step S500: User feedback and data upload: Once the risk of atrial fibrillation is detected to reach the warning level, the smart bracelet sends multimodal warning information to the user through the user interaction module; Sub-step S510: Tactile warning: The vibration motor of the smart bracelet is activated, vibrating at a specific frequency and duration; Sub-step S520: Visual warning: The LED indicator of the smart bracelet displays the warning level in a specific color or flashing pattern; Sub-step S530: Data Upload: The signal processing unit uploads detailed atrial fibrillation warning event data, relevant raw PPG and ECG signal segments, characteristic parameters, and timestamps to the paired smart terminal device via the wireless communication module. Upon receiving the data, the companion application on the smart terminal device immediately displays detailed warning information to the user and can optionally synchronize the data to a cloud server for user review or remote access by a doctor.

[0017] This invention provides a smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion and its analysis method. Compared with existing technologies, it has the following advantages: 1. This smart bracelet for early warning of atrial fibrillation based on multimodal biosignal fusion and its analysis method overcomes the limitations of single signals by fusing PPG and ECG data in a dual-modal manner and combining the artifact removal capabilities of motion sensor data. PPG signals provide continuous hemodynamic information and are sensitive to early and subtle changes in heart rate and rhythm, while ECG signals provide the gold standard information on cardiac electrical activity and have high specificity for atrial fibrillation diagnostic features such as missing P waves and irregular RR intervals. The synergistic analysis of the two significantly improves the accuracy and detection rate of early warning of atrial fibrillation.

[0018] 2. This smart bracelet for atrial fibrillation early warning and its analysis method based on multimodal biosignal fusion optimizes ECG electrode design and signal acquisition strategy, enabling users to continuously and backgroundly acquire PPG and ECG signals during daily wear without active operation. This avoids the problem that traditional ECG devices may miss transient atrial fibrillation events due to on-demand measurement, and greatly improves the possibility of early detection of paroxysmal or asymptomatic atrial fibrillation.

[0019] 3. This atrial fibrillation early warning smart bracelet and its analysis method based on multimodal biosignal fusion employs a deep learning model to deeply mine multi-dimensional features, enabling it to identify subtle physiological change patterns before the onset of atrial fibrillation. This achieves "early warning" of atrial fibrillation risk rather than just "detection during occurrence." Combined with personalized baseline establishment and a continuous judgment mechanism, it effectively reduces the false alarm rate while improving the ability to capture real risks. The introduction of motion sensor data can intelligently identify and eliminate signal artifacts introduced by daily activities such as movement and posture changes, significantly enhancing the signal quality and algorithm stability of the system in complex and changing environments, and ensuring the reliability of early warning in different scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the smart bracelet of the present invention; Figure 2 This is a schematic diagram of the data flow of the smart bracelet system of the present invention; Figure 3 This is a flowchart illustrating the atrial fibrillation early warning analysis method based on multimodal biological signal fusion according to the present invention. Detailed Implementation

[0021] The technical solutions of 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.

[0022] Please see Figure 1 and Figure 2 This smart bracelet for atrial fibrillation early warning, based on multimodal biosignal fusion, provides a technical solution: It includes a wristband body, a photoplethysmography (PPG) signal acquisition module, an electrocardiogram (ECG) signal acquisition module, a motion sensor module, a signal processing unit, a storage unit, a wireless communication module, a power management unit, and a user interaction module.

[0023] The main body of the bracelet has a ring structure, and its design fully considers the principles of ergonomics to ensure the comfort and stability of users during long-term daily wear. The main body of the bracelet is made of medical-grade silicone or thermoplastic polyurethane (TPU), which has excellent biocompatibility, sweat corrosion resistance and moderate elasticity.

[0024] The main body of the bracelet is meticulously designed with multiple independent and isolated cavities and precision mounting structures. These cavities and structures are used to house and secure the PPG signal acquisition module, ECG signal acquisition module, motion sensor module, signal processing unit, storage unit, wireless communication module, power management unit, and user interaction module, ensuring the precise position of each component inside the bracelet and avoiding mutual interference. In addition, the outer and inner surfaces of the bracelet body are precisely perforated or made transparent corresponding to the optical window of the PPG signal acquisition module and the electrode contact area of ​​the ECG signal acquisition module to achieve effective signal acquisition while maintaining the overall waterproof and dustproof performance of the bracelet, with a protection level of IP68.

[0025] The PPG signal acquisition module is precisely positioned on the inside of the wristband body to ensure close contact with the user's wrist skin surface when worn, thereby continuously and non-invasively acquiring optical signals caused by changes in the volume of the wrist arteries. The core components of the PPG signal acquisition module include at least one light-emitting unit and at least one photoelectric detection unit.

[0026] The ECG signal acquisition module is strategically positioned at specific locations on the inner and outer sides of the main body of the bracelet. Its design goal is to achieve seamless and continuous single-lead electrocardiogram signal acquisition when the user wears it daily. The core components of the ECG signal acquisition module include at least two bioelectrodes and an ECG front-end amplifier circuit.

[0027] The motion sensor module integrates a high-performance multi-axis accelerometer and a multi-axis gyroscope, employing a six-axis inertial measurement unit (IMU) chip, such as STMicroelectronics' LSM6DS3. The three-axis accelerometer has a measurement range of ±16g and a resolution of 16 bits, enabling precise capture of the band's linear acceleration changes in the X, Y, and Z directions. The three-axis gyroscope has a measurement range of ±2000 dps (degrees per second) and a resolution of 16 bits, allowing real-time monitoring of the band's three-dimensional angular velocity and attitude changes. The motion sensor module synchronously transmits acceleration and angular velocity data to the signal processing unit at a sampling rate of at least 100Hz (e.g., 104Hz). This high-precision motion data serves as crucial auxiliary information, primarily used for motion artifact recognition, quantitative assessment, and advanced elimination in PPG and ECG signals. Simultaneously, it provides real-time assessment of the user's current activity status (resting, walking, running, etc.), offering contextual correction for atrial fibrillation risk assessment algorithms.

[0028] The signal processing unit uses a low-power, high-performance embedded microcontroller unit (MCU). In a preferred embodiment, the MCU is STMicroelectronics' STM32L4 series, which integrates an ARM Cortex-M4F core and operates at a frequency of not less than 120MHz. The MCU's internal integrated random access memory (RAM) has a capacity of not less than 512KB (e.g., 512KB SRAM) for efficiently storing real-time acquired biosignal data streams, complex intermediate processing results, temporary variables during deep learning model runtime, and algorithm execution stacks. The flash memory has a capacity of not less than 4MB (e.g., 4MB onboard Flash) for securely storing the embedded operating system (e.g., FreeRTOS or μC / OS), all complex signal processing algorithms (including filtering, feature extraction, artifact removal, etc.), pre-trained deep learning model parameters, personalized baseline data, and short-term historical event records. The signal processing unit interacts with the PPG signal acquisition module, ECG signal acquisition module, and motion sensor module via a high-speed serial peripheral interface (e.g., SPI or I2C) to ensure real-time and efficient transmission of raw data. The signal processing unit is responsible for the overall management of all acquired raw multimodal biological data and sequentially executes the following core functions: First, data preprocessing, including filtering, denoising, and motion artifact removal of the raw signals; second, multi-dimensional physiological feature extraction, extracting rich time-domain, frequency-domain, and morphological features from PPG and ECG signals; third, multimodal feature fusion, organically integrating features from different modalities; fourth, atrial fibrillation risk assessment based on a deep learning model; and finally, result output, triggering alerts and managing data uploads. The MCU also features advanced power management capabilities, maximizing battery life while maintaining computing power through Dynamic Voltage-Frequency Scaling (DVFS) and multi-level sleep modes.

[0029] The storage unit is a component used for long-term, non-volatile storage of large amounts of user physiological data and event records. The wireless communication module uses a Bluetooth Low Energy (BLE) module, which is mainly used for wireless data transmission and command interaction between the smart bracelet and external smart terminal devices (such as smartphones or tablets).

[0030] The power management unit includes a high-energy-density rechargeable lithium-ion battery. The user interaction module mainly includes a miniature vibration motor and a multi-color LED indicator. The vibration motor is a linear resonant actuator (LRA) with fast response speed and programmable vibration mode, which can provide clear tactile feedback to the user. The LED indicator 192 is an RGB tri-color LED that can intuitively display the device's working status, battery power status, and different levels of atrial fibrillation warning information through different colors (e.g., green, yellow, red) and flashing modes (e.g., constant light, slow flashing, fast flashing).

[0031] Please see Figure 3 The present invention also discloses an analysis method for atrial fibrillation early warning based on multimodal biosignal fusion. The method is executed on the signal processing unit of a smart bracelet, and its detailed steps are as follows: Step S100: Synchronous acquisition and preprocessing of multimodal biological signals.

[0032] This step is the cornerstone of the entire analysis method, aiming to acquire high-quality, artifact-free raw physiological signals. The signal processing unit continuously receives and buffers the digital PPG signal output from the PPG signal acquisition module, the digital ECG signal output from the ECG signal acquisition module, and the triaxial acceleration and triaxial angular velocity data output from the motion sensor module. To ensure the accuracy of subsequent multimodal fusion analysis, all signals are synchronized with high precision on the time axis, with the synchronization error strictly controlled within ±2ms. This is achieved through a shared high-precision clock source and timestamp synchronization mechanism.

[0033] For PPG signals, the following preprocessing procedure is performed: Sub-step S110: Bandpass filtering. A digital IIR (Infinite Impulse Response) or FIR (Finite Impulse Response) bandpass filter is applied to the acquired raw digital PPG signal. In a typical implementation, a fourth-order Butterworth bandpass filter is used, with its cutoff frequency set between 0.5 Hz and 10 Hz. The main function of this filter is to effectively filter out low-frequency artifacts caused by respiratory movements and baseline drift (caused by changes in skin contact), as well as interference caused by high-frequency environmental noise and muscle tremors, thereby separating the core pulse wave component.

[0034] Sub-step S120: Motion Artifact Removal. PPG signals are highly susceptible to motion artifacts. This sub-step utilizes the triaxial acceleration and triaxial angular velocity data provided by the motion sensor module 14, employing an advanced adaptive filtering algorithm to compensate for and eliminate motion artifacts in the PPG signal in real time. Specifically, an adaptive noise canceller based on the Recursive Least Squares (RLS) or Normalized Least Mean Squares (NLMS) algorithm can be used, with motion data as the reference input and the PPG signal as the main input. Motion-induced artifacts are estimated and subtracted from the PPG signal by dynamically adjusting the filter coefficients. Alternatively, Empirical Mode Decomposition (EMD) or Independent Component Analysis (ICA) methods can be combined to decompose the PPG signal into several independent components, identify and remove components containing motion artifacts, thereby significantly improving the purity of the PPG signal and making it more suitable for subsequent feature extraction.

[0035] Sub-step S130: Signal Quality Assessment. Real-time quality assessment is performed on the PPG signal after filtering and motion artifact removal. Assessment metrics include, but are not limited to: signal-to-noise ratio (SNR, e.g., SNR required to be no less than 20dB), peak-to-valley amplitude stability of the pulse wave (peak-to-valley amplitude fluctuation rate between adjacent cycles less than 15%), periodic stability (standard deviation of PPI between adjacent pulse cycles less than 10%), and the presence of significant saturation or truncation. If the PPG signal quality consistently falls below a preset threshold (e.g., a PPG signal quality score below 0.7 for 30 consecutive seconds, out of 1.0), the signal segment is marked as low-quality data and will be considered or discarded in subsequent feature extraction, or the user will be prompted to adjust the wearing position.

[0036] For ECG signals, perform the following preprocessing steps: Sub-step S140: Bandpass Filtering. A digital bandpass filter is applied to the acquired raw digital ECG signal. In one implementation, a second- or fourth-order Butterworth bandpass filter with zero phase shift is used, with a cutoff frequency set between 0.05 Hz and 40 Hz. This filter is used to remove baseline drift (low-frequency artifacts) caused by breathing and body movement, as well as high-frequency electromyography interference and equipment noise. For power frequency interference (50 Hz or 60 Hz, depending on the local power grid standard), a narrowband notch filter (e.g., a Butterworth notch filter with a Q factor of 30) or a higher-order adaptive notch filtering algorithm (e.g., a DFT-based adaptive filter) is used for precise filtering to avoid distortion of key ECG waveforms, particularly the P and T waves.

[0037] Sub-step S150: Signal Quality Assessment. Real-time quality assessment is performed on the preprocessed ECG signal. Assessment metrics primarily include: QRS complex detection rate (requiring a QRS complex detection rate of no less than 95% within a specified time window), R-wave amplitude stability, baseline drift degree (requiring the baseline drift amplitude to be less than 10% of the QRS complex peak amplitude), and the discriminability of the P and T waves. If the ECG signal quality remains below a preset threshold (e.g., an ECG signal quality score below 0.6 for two consecutive minutes), the smart bracelet 10 will issue a wearing status check or electrode contact malfunction warning to the user via the user interaction module 19, ensuring the accuracy of subsequent diagnoses.

[0038] Step S200: Extraction of multi-dimensional physiological features.

[0039] The signal processing unit extracts multidimensional and complementary physiological feature parameters from the preprocessed multimodal biological signals. These feature parameters are divided into PPG source features, ECG source features, spatiotemporal domain features, and composite physiological features to comprehensively characterize the electrical activity and hemodynamic state of the cardiovascular system.

[0040] PPG source feature extraction: Sub-step S210: Heart Rate (HR) and Heart Rate Variability (HRV) Features. The peak point of each pulse wave is accurately identified from the high-purity PPG signal. Real-time heart rate (HR, beats / minute) is calculated based on the time interval between adjacent pulse wave peaks (pulse interval, PPI). Simultaneously, time-domain and frequency-domain HRV features are extracted from the continuous PPI sequence. Time-domain HRV features include: the standard deviation of the PPI sequence (SDNN), the root mean square of the difference between adjacent pulse intervals (RMSSD), and the percentage of adjacent pulse intervals greater than 50 ms (pNN50). Frequency-domain HRV features are obtained by performing a Fast Fourier Transform (FFT) or wavelet transform on the PPI sequence, including the power spectral density and its ratio (LF / HF) for very low frequencies (VLF, <0.04 Hz), low frequencies (LF, 0.04–0.15 Hz), and high frequencies (HF, 0.15–0.4 Hz). These indicators reflect the regulation of cardiac activity by the autonomic nervous system.

[0041] Sub-step S220: Pulse Wave Morphology Features. A thorough morphological analysis of the PPG pulse wave is performed to extract a series of features reflecting changes in vascular elasticity and hemodynamics. This includes, but is not limited to: pulse wave rise time (time from trough to peak), fall time (time from peak to the next trough), peak width (half-width at half maximum at peak), systolic time (time from trough to dicrotic notch or sub-peak), and diastolic time (time from dicrotic notch or sub-peak to the next trough). Furthermore, by performing first and second derivative analysis on the pulse wave, the Augmentation Index (AIx) and Reflection Index (RI), which are correlated with arterial stiffness, can be extracted. Simultaneously, the presence and location of the dicrotic notch in the descending limb are detected and quantified, which is diagnostically significant in specific cardiovascular diseases. The extraction of these morphological features typically involves waveform normalization, feature point detection (e.g., based on thresholding or curve fitting), and time / amplitude measurements.

[0042] ECG source feature extraction: Sub-step S230: Heart Rate (HR) and Heart Rate Variability (HRV) Features. From high-quality ECG signals, advanced QRS complex detection algorithms (e.g., Pan-Tompkins algorithm or wavelet transform methods) are used to accurately identify R-wave peaks. Real-time heart rate (HR) is calculated based on the time interval between adjacent R-wave peaks (RR interval, RRi). Similar to PPG HRV features, various ECG HRV features, including SDNN, RMSSD, pNN50, and LF / HF, are accurately calculated from continuous RRi sequences. Because ECG signals directly reflect cardiac electrical activity, their HRV features are considered the gold standard in atrial fibrillation diagnosis, with significantly higher accuracy than HRV features based solely on PPG.

[0043] Sub-step S240: P-wave detection and QRS complex morphology. This sub-step accurately identifies and quantifies the P wave (atrial depolarization wave), QRS complex (ventricular depolarization wave), and T wave (ventricular repolarization wave) in the ECG signal. It focuses on extracting the duration, amplitude (R-wave peak value), and axial deflection information of the QRS complex (determined by relative amplitude in different leads). More importantly, it accurately determines the presence of the P wave, its morphology (e.g., clarity, flatness, fibrillation), and the relationship between the P wave and the QRS complex (PR interval, normal range 120-200ms). One of the typical diagnostic features of atrial fibrillation is the absence of the P wave or its replacement by an irregular, low-amplitude f wave (fibrillation wave), and the absolute irregularity of the RR interval. In addition, it analyzes possible indicators of myocardial ischemia, such as J-point deviation and ST-segment changes, to help assess overall cardiovascular health.

[0044] Spatiotemporal feature extraction: Sub-step S250: Joint time-frequency analysis of the fused signals. Joint time-frequency analysis is performed on the highly synchronized PPG and ECG signals, for example, using Short Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT) methods. This analysis extracts features such as energy distribution, instantaneous frequency, instantaneous phase, and intermodal correlations at different frequencies and time scales. These features can capture subtle changes and potential signal correlations that are difficult to detect with single-mode signals, such as the synchronicity or deviation between the PPG pulse waveform and the ECG waveform at specific frequency components, providing richer dynamic information for identifying early atrial fibrillation precursors.

[0045] Sub-step S260: Cardiovascular Coupling Characteristics. Calculate the time delay (pulse conduction time, PTT) between the peak value of the PPG pulse wave and the peak value of the ECG R wave. PTT is the time from the onset of ventricular systole to the arrival of the pulse wave in peripheral vessels, and its variability (PTTV) is highly sensitive to cardiovascular autonomic activity, changes in vascular elasticity, and arrhythmias. By tracking the absolute value and variability of PTT, additional dimensions for atrial fibrillation risk assessment can be provided. For example, atrial fibrillation may cause abnormal fluctuations in PTT or changes in the mean PTT.

[0046] Step S300: Multimodal feature fusion and deep learning model construction.

[0047] The signal processing unit 15 fuses the PPG source features (e.g., HR, SDNN_PPG, RMSSD_PPG, pNN50_PPG, LF_PPG, HF_PPG, AIx, RI, and presence of Digital Notch) extracted in step S200, ECG source features (e.g., HR, SDNN_ECG, RMSSD_ECG, pNN50_ECG, LF_ECG, HF_ECG, presence of P waves, RRi irregularity, and QRS duration), spatiotemporal features (e.g., PPG-ECG joint time-frequency energy ratio), and cardiovascular coupling features (e.g., PTT mean and PTT standard deviation) to form a high-dimensional (e.g., containing more than 50 feature dimensions) comprehensive feature vector. The fusion method employs feature-level fusion, which involves concatenating or weighting features from different modalities before inputting them into the deep learning model.

[0048] The deep learning model is pre-trained offline and stored in the smart bracelet's storage unit, then runs on the signal processing unit for inference. The deep learning model is a hybrid model combining a convolutional neural network (CNN) and a long short-term memory network (LSTM), specifically designed for processing time-series physiological data. The CNN layers (e.g., containing 3-5 one-dimensional convolutional layers, each with 32-128 kernels and a kernel size of 3-7) are used to automatically extract local pattern features from the input feature vector, such as capturing microstructural changes, local irregularities, or specific rhythmic patterns in PPG and ECG waveforms in continuous heart rate cycles. The LSTM layers (e.g., containing 1-2 LSTM layers, each with 64-128 units) follow the CNN layers and are used to capture long-term dependencies and sequence patterns in the time-series data, such as the persistence, evolution trend, and deviation from normal rhythms of arrhythmias. The model training data includes large-scale, diverse clinical physiological signal data, including: normal sinus rhythm data, various types of paroxysmal atrial fibrillation data, persistent atrial fibrillation data, and other common arrhythmia data (such as ventricular premature beats and supraventricular premature beats). Crucially, the training dataset contains a large amount of precisely labeled physiological signal data from the pre-atrial fibrillation phase. Supervised learning and the cross-entropy loss function are used to train the model, enabling it to accurately identify early warning signals and diagnostic features of atrial fibrillation. The training process employs the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 200 training epochs.

[0049] The output layer of the deep learning model (e.g., a fully connected layer followed by a sigmoid activation function) outputs an atrial fibrillation risk score, representing the probability that the user will develop atrial fibrillation within the current time window (e.g., a sliding time window of 5 consecutive minutes). The score ranges from 0 to 1, where 0 represents very low risk and 1 represents very high risk. This score is a continuous variable, providing a more refined risk quantification than a simple binary classification (normal / abnormal).

[0050] Step S400: Atrial fibrillation risk assessment and early warning.

[0051] The signal processing unit performs a final assessment and early warning triggering of atrial fibrillation risk based on the atrial fibrillation risk score output by the deep learning model, combined with multiple preset risk thresholds and time window mechanisms.

[0052] Sub-step S410: Real-time risk assessment. The average atrial fibrillation risk score for the current time window (e.g., 5 minutes of continuously processed signal data, with the assessment result updated every 30 seconds) is compared with multiple preset risk thresholds. This invention sets three tiered thresholds: T1 (low-risk threshold, e.g., 0.3), T2 (medium-risk threshold, e.g., 0.6), and T3 (high-risk threshold, e.g., 0.85). When the average risk score first exceeds T1, the system enters observation mode; when it exceeds T2, a preliminary warning is issued; and when it exceeds T3, an emergency warning is issued.

[0053] Sub-step S420: Continuity Assessment. To minimize false alarms caused by sporadic noise, transient artifacts, or arrhythmias, the smart bracelet requires that the atrial fibrillation risk score remain consistently above a preset threshold for a specific time period or exhibit an abnormal pattern within a specified number of heartbeat cycles before triggering any alert. For example, for a medium-risk alert, the average risk score per minute must be higher than T2 for 10 consecutive minutes; for a high-risk alert, the average risk score per minute must be higher than T3 for 5 consecutive minutes. Furthermore, the system will also count whether irregularities in the RR intervals (e.g., the proportion of adjacent RR interval differences greater than 50ms) persist for at least 300 consecutive heartbeat cycles within these time windows.

[0054] Sub-step S430: Motion State Correction. The atrial fibrillation risk assessment results are contextually corrected using real-time activity data provided by the motion sensor module. For example, when the user is in a state of vigorous exercise (determined by accelerometer data, e.g., an RMS acceleration value exceeding 5g for more than 1 minute), the system will appropriately raise the warning threshold (e.g., T2 and T3 are increased by 0.05 each), or perform more stringent motion artifact suppression on the signal processing algorithm to reduce the risk of physiological fluctuations caused by exercise being misdiagnosed as atrial fibrillation. Conversely, when the user is in a resting or sleeping state (determined by accelerometer and gyroscope data, e.g., an RMS acceleration value below 0.5g and an angular velocity below 10 dps for more than 5 minutes), the system will appropriately lower the warning threshold (e.g., T2 and T3 are decreased by 0.03 each) to improve sensitivity and early detection capability for occult, low-symptom atrial fibrillation episodes. This dynamic threshold adjustment mechanism significantly enhances the system's robustness and environmental adaptability.

[0055] Sub-step S440: Personalized Baseline Adjustment. After initial wear or at least 7 days of normal wear, the smart bracelet automatically establishes personalized physiological baseline data for the user. This baseline includes the user's average heart rate, normal heart rate variability range (normal fluctuation ranges of SDNN, RMSSD, and pNN50), typical PPG waveform characteristics (e.g., average values ​​and fluctuation ranges of AIx and RI), and ECG characteristics (P wave morphology, mean PR interval). In subsequent real-time risk assessments, the detected characteristics are dynamically compared with the personalized baseline to identify subtle changes deviating from the user's normal range, thereby improving the personalization and accuracy of alerts and reducing false alarm rates caused by individual differences. The personalized baseline is automatically updated every month to adapt to long-term changes in the user's physiological state.

[0056] Step S500: User feedback and data upload.

[0057] Once the risk of atrial fibrillation is detected to reach the warning level, the smart bracelet will send a clear multimodal warning message to the user through the user interaction module and upload relevant data simultaneously.

[0058] Sub-step S510: Tactile Alert. The smart bracelet's vibration motor activates, vibrating at a specific frequency and duration to provide the user with intuitive tactile feedback. For example, a low-risk alert is indicated by two short vibrations (100ms each, 100ms interval); a medium-risk alert is indicated by continuous periodic vibrations (200ms each, 300ms interval, lasting 5 seconds); and a high-risk alert is indicated by continuous strong vibrations (lasting 10 seconds), which can be repeated.

[0059] Sub-step S520: Visual warning. The LED indicator on the smart bracelet displays the warning level using a specific color or flashing pattern. For example, a solid green light indicates a normal state; a slow yellow flash (once per second) indicates a medium-risk warning; and a fast red flash (three times per second) indicates a high-risk warning, which may be accompanied by a buzzer (if the bracelet is equipped with one) emitting an alarm sound.

[0060] Sub-step S530: Data Upload. The signal processing unit immediately uploads detailed atrial fibrillation warning event data, relevant raw PPG and ECG signal segments (e.g., data from 5 minutes before and 1 minute after the warning trigger), key physiological characteristic parameters, risk scores, and precise timestamps to the paired smart terminal device (e.g., a smartphone) via the wireless communication module (BLE). Upon receiving the data, the companion application on the smart terminal device immediately displays detailed warning information to the user through pop-ups, push notifications, or voice prompts. The user can choose to synchronize the data to a cloud server via the companion application for review, long-term health trend analysis, or authorized remote access and diagnosis by medical professionals. TLS / SSL encryption is used during data upload to ensure data transmission security.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion, characterized in that: include: The main body of the bracelet has a ring-shaped structure, which is adapted to be worn on the user's wrist. Its material is a biocompatible material. The main body of the bracelet has a cavity and mounting structure to accommodate and fix the various functional modules, and exposes the optical window of the PPG signal acquisition module and the electrode contacts of the ECG signal acquisition module. Each functional module also includes: a motion sensor module, a signal processing unit, a storage unit, a wireless communication module, a power management unit, and a user interaction module. The signal processing unit is an embedded microcontroller unit that is connected to the PPG signal acquisition module, the ECG signal acquisition module and the motion sensor module through a high-speed serial interface. It is used to receive and manage the PPG signal, the ECG signal and the motion data, and to perform data preprocessing, multi-dimensional physiological feature extraction, multi-modal feature fusion and atrial fibrillation risk assessment based on a deep learning model.

2. The atrial fibrillation early warning smart bracelet based on multimodal biosignal fusion according to claim 1, characterized in that: The PPG signal acquisition module includes at least one light-emitting unit and at least one photoelectric detection unit, and the ECG signal acquisition module includes at least two bioelectrodes and an ECG front-end amplifier circuit.

3. The atrial fibrillation early warning smart bracelet based on multimodal biosignal fusion according to claim 1, characterized in that: The signal processing unit is an embedded microcontroller unit, which integrates a central processing unit core, random access memory and flash memory. The RAM capacity is not less than 512KB and the flash memory capacity is not less than 4MB. The signal processing unit is specifically used to perform the following functions: data preprocessing, multi-dimensional physiological feature extraction, multi-modal feature fusion, and atrial fibrillation risk assessment.

4. The atrial fibrillation early warning smart bracelet based on multimodal biosignal fusion according to claim 3, characterized in that: The deep learning model is a hybrid model combining a convolutional neural network and a long short-term memory network. The convolutional neural network layer is used to automatically extract local pattern features from time series data. The user interaction module includes a vibration motor and an LED indicator. The vibration motor is used to provide a vibration alert when a potential risk of atrial fibrillation is detected.

5. An analytical method for atrial fibrillation early warning based on multimodal biosignal fusion, implemented using a smart bracelet for atrial fibrillation early warning based on multimodal biosignal fusion as described in any one of claims 1-4, characterized in that: Step S100: Synchronous acquisition and preprocessing of multimodal biological signals: The signal processing unit continuously receives the digital PPG signal output by the PPG signal acquisition module, the digital ECG signal output by the ECG signal acquisition module, and the acceleration and angular velocity data output by the motion sensor module; For PPG signals, perform the following preprocessing: Sub-step S110: Bandpass filtering: The digital PPG signal is filtered using a digital infinite impulse response or finite impulse response bandpass filter; Sub-step S120: Motion artifact elimination: Using the acceleration and angular velocity data provided by the motion sensor module, an adaptive filtering algorithm or independent component analysis method is used to compensate for and eliminate motion artifacts in the PPG signal in real time; Sub-step S130: Signal quality assessment: Based on the signal-to-noise ratio, peak-to-valley amplitude, and periodic stability index, perform real-time quality assessment on the processed PPG signal and mark low-quality signal segments; For ECG signals, perform the following preprocessing: Sub-step S140: Bandpass filtering: The digital ECG signal is filtered using a digital bandpass filter; Sub-step S150: Signal quality assessment: Based on the detection rate, amplitude stability, and baseline drift index of the QRS group, the processed ECG signal is assessed in real time, and low-quality signal segments are marked. If the ECG signal quality continues to be lower than the preset threshold, the smart bracelet will prompt the user to check the wearing status or electrode contact. Synchronization processing: PPG signals, ECG signals and motion sensor data are synchronized with high precision on the time axis.

6. The analytical method for atrial fibrillation early warning based on multimodal biosignal fusion according to claim 5, characterized in that: Step S200: Multidimensional physiological feature extraction: The signal processing unit extracts multidimensional physiological feature parameters from the preprocessed multimodal biological signals. The feature parameters are divided into PPG source features, ECG source features, spatiotemporal domain features, and composite physiological features. PPG source feature extraction: Sub-step S210: Heart rate and heart rate variability characteristics: Identify the peak point of the pulse wave from the PPG signal, calculate the time interval between adjacent pulse cycles, and calculate the heart rate, standard deviation, root mean square continuous difference, percentage of adjacent pulse intervals greater than 50ms, and time-domain and frequency-domain heart rate variability characteristics based on the PPI sequence. Sub-step S220: Pulse wave morphological characteristics: Analyze the morphological characteristics of the PPG pulse wave, including but not limited to: rise time, fall time, peak width, systolic time, diastolic time, enhancement index, reflection index, presence and location of biphasic peaks. These characteristics are obtained through derivative analysis, Fourier transform or wavelet transform of the pulse wave. ECG source feature extraction: Sub-step S230: Heart rate and heart rate variability characteristics: Accurately detect the peak value of the R wave from the ECG signal, calculate the time interval between adjacent R waves, and calculate the heart rate and various ECG heart rate variability characteristics corresponding to the PPG source heart rate variability characteristics based on the time interval sequence between adjacent R waves. ECG heart rate variability characteristics are the gold standard indicator for the diagnosis of atrial fibrillation, and their accuracy is better than that of PPG heart rate variability. Sub-step S240: P wave detection and QRS complex morphology characteristics: accurately identify the P wave, QRS complex and T wave in the ECG signal, extract the duration, amplitude and axial deflection information of the QRS complex, pay special attention to the presence and morphology of the P wave, the relationship between the P wave and the QRS complex, the typical characteristics of atrial fibrillation are the absence of the P wave or its replacement by an irregular f wave, and the absolute irregularity of the RR interval; Spatiotemporal feature extraction: Sub-step S250: Joint time-frequency analysis characteristics of fused signals: Perform joint time-frequency analysis on the synchronized PPG and ECG signals to extract energy distribution and phase relationship characteristics at different frequencies and time scales; Sub-step S260: Cardiovascular coupling characteristics: Calculate the time delay or correlation between the PPG pulse wave and the ECG R wave.

7. The analytical method for atrial fibrillation early warning based on multimodal biosignal fusion according to claim 5, characterized in that: Step S300: Multimodal feature fusion and deep learning model construction: The signal processing unit fuses the PPG source features, ECG source features, spatiotemporal features and cardiovascular coupling features extracted in step S200 to form a high-dimensional comprehensive feature vector. The fusion method adopts feature-level fusion, that is, the features of different modalities are concatenated or weighted before being input into the deep learning model. The deep learning model is pre-stored in the storage unit and runs on the signal processing unit. The deep learning model is a hybrid model that combines convolutional neural networks and long short-term memory networks. The convolutional neural network layer is used to automatically extract local pattern features from time series data, and the long short-term memory network layer is used to capture long-term dependencies and sequence patterns in time series data. The input layer of the deep learning model receives high-dimensional comprehensive feature vectors. The training data of the model includes a large amount of normal heart rhythm data, paroxysmal atrial fibrillation data, persistent atrial fibrillation data, and other common arrhythmia data, especially physiological signal data in the early stage of atrial fibrillation. The model is trained through supervised learning so that it can identify early warning signals and diagnostic features of atrial fibrillation. The output layer of the deep learning model outputs an atrial fibrillation risk score, which represents the likelihood of the user developing atrial fibrillation within the current time window. The score ranges from 0 to 1, where 0 represents extremely low risk and 1 represents extremely high risk.

8. The atrial fibrillation early warning analysis method based on multimodal biosignal fusion according to claim 5, characterized in that: Step S400: Atrial fibrillation risk assessment and early warning: The signal processing unit performs atrial fibrillation risk assessment and early warning based on the atrial fibrillation risk score output by the deep learning model, combined with a preset risk threshold and time window mechanism; Sub-step S410: Real-time risk assessment: Compare the average atrial fibrillation risk score of the current time window with multiple preset risk thresholds; Sub-step S420: Continuous judgment: In order to avoid false alarms caused by occasional noise or artifacts, before triggering an alarm, the smart bracelet requires the risk score to be continuously higher than the preset threshold within a specific time period, or to show an abnormal pattern within a specified number of heartbeat cycles. Sub-step S430: Motion state correction: Correct the atrial fibrillation risk assessment results using activity data provided by the motion sensor module; Sub-step S440: Personalized baseline adjustment. After the first wear or after a period of normal wear, the smart bracelet automatically establishes the user's personalized physiological baseline data. In subsequent risk assessments, the real-time detected features are compared with the personalized baseline to identify subtle changes that deviate from the normal range.

9. The analytical method for atrial fibrillation early warning based on multimodal biosignal fusion according to claim 5, characterized in that: Step S500: User Feedback and Data Upload: Once the risk of atrial fibrillation is detected to reach the warning level, the smart bracelet sends multimodal warning information to the user through the user interaction module; Sub-step S510: Tactile warning: The vibration motor of the smart bracelet is activated, vibrating at a specific frequency and duration; Sub-step S520: Visual warning: The LED indicator of the smart bracelet displays the warning level in a specific color or flashing pattern; Sub-step S530: Data Upload: The signal processing unit uploads detailed atrial fibrillation warning event data, relevant raw PPG and ECG signal segments, characteristic parameters, and timestamps to the paired smart terminal device via the wireless communication module. Upon receiving the data, the companion application on the smart terminal device immediately displays detailed warning information to the user and can optionally synchronize the data to a cloud server for user review or remote access by a doctor.

Images