A method and system for early warning of dysphagia risk based on multi-signal analysis

Abstract

This invention provides a method and system for risk warning of dysphagia based on multi-signal analysis. The method includes: acquiring surface electromyography (EMG) signals, bioimpedance signals, laryngeal vibration signals, and neck microenvironment temperature and humidity signals from the wearer's neck using a flexible wearable sensing unit; using the respiratory phase determined by the neck microenvironment temperature and humidity signals as a time reference, performing time alignment and preprocessing on the multimodal physiological signals to obtain multi-channel time-series data; extracting and fusing features from the multi-channel time-series data to generate a discriminative fusion feature vector; inputting the vector into a pre-trained classification model for real-time inference to identify swallowing function states, including effective swallowing, ineffective swallowing, and suspected aspiration; and generating a graded risk warning signal in response to the identified abnormal states. This invention solves the problems of motion artifact interference and time alignment through multimodal signal fusion and respiratory rhythm calibration, achieving accurate and continuous monitoring of dysphagia risk.

Description

Technical Field

[0001] This invention relates to the fields of medical rehabilitation monitoring and biomedical signal processing technology, and in particular to a method and system for early warning of swallowing disorders based on multi-signal analysis. Background Technology

[0002] Dysphagia is a common complication in stroke patients, patients with neurodegenerative diseases, and the elderly. If not detected and intervened in a timely manner, it can easily lead to serious consequences such as aspiration pneumonia, malnutrition, and even suffocation. Currently, clinical assessment of swallowing function mainly relies on video swallowing imaging or fiberoptic endoscopy swallowing function examination. Although these imaging examinations can clearly present the anatomical structure and movement of the swallowing organs, they are limited by the large size of the equipment, high examination costs, and the potential for radiation exposure or invasive discomfort. They are usually only used as a phased diagnostic tool and are difficult to apply to daily continuous monitoring in ward care or home rehabilitation scenarios. This means that medical staff can often only obtain the patient's functional status at a specific examination moment, and cannot capture the dynamic swallowing dysfunction that occurs over a long period of time due to fatigue, distraction, or changes in body position.

[0003] To address the challenges of continuous monitoring, various portable monitoring solutions based on non-invasive sensors have emerged in recent years. Most of these solutions attempt to identify swallowing movements by collecting surface electromyography (EMG) signals from the neck or acoustic signals from the larynx. However, swallowing is a complex physiological process involving neuromuscular control, laryngeal mechanical movement, and respiratory airflow coordination. Single-modal physiological signals often fail to fully reflect the dynamic characteristics of this process. For example, simple EMG signals are easily interfered with by speaking, chewing, or head and neck movements, while simple acoustic signals struggle to accurately distinguish between environmental noise and internal biological sounds. More critically, physiological studies have shown that safe swallowing requires strict temporal coordination with the respiratory cycle; swallowing typically occurs during apnea. Existing wearable monitoring technologies largely lack the ability to synchronously sense neck respiratory airflow, making it impossible to accurately determine the relative relationship between swallowing movements and respiratory phases. This makes it difficult to accurately identify the hidden risk of aspiration caused by respiratory-swallowing dyscodynamics in its early stages, resulting in a high false alarm rate or missed detection of critical risks in the monitoring system. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for early warning of swallowing disorders based on multi-signal analysis, so as to solve the problems pointed out in the background art.

[0005] In a first aspect, the present invention provides a method for early warning of swallowing disorder risk based on multi-signal analysis, comprising the following steps:

[0006] S1. Collect multimodal physiological signals from the wearer's neck, including at least surface electromyography signals, bioimpedance signals, laryngeal vibration signals, and neck microenvironment temperature and humidity signals.

[0007] S2. Perform time alignment and preprocessing on the multimodal physiological signals to obtain multi-channel time-series data;

[0008] S3. Perform feature extraction and feature fusion on the multi-channel time-series data to generate a discriminative fusion feature vector;

[0009] S4. Input the discriminative fusion feature vector into the pre-trained swallowing pattern classification model for real-time reasoning to identify the swallowing function state, which includes effective swallowing, ineffective swallowing, and suspected aspiration.

[0010] S5. In response to the identified invalid swallowing or suspected aspiration, generate a corresponding risk warning signal.

[0011] Optionally, in step S1, the multimodal physiological signals are acquired via a flexible wearable sensing unit:

[0012] The surface electromyography signal and the bioimpedance signal were acquired by a composite electrode symmetrically embedded at the horizontal position of the thyroid cartilage;

[0013] The vibration signal of the larynx was acquired by a miniature accelerometer placed above the Adam's apple;

[0014] The temperature and humidity signals of the neck microenvironment are collected by a temperature and humidity sensor located at the mandible.

[0015] Optionally, in step S2, the preprocessing includes auxiliary calibration of the respiratory rhythm using the temperature and humidity signal of the neck microenvironment:

[0016] The periodic fluctuation patterns of the temperature and humidity signals in the neck microenvironment are detected to determine the respiratory phase;

[0017] Identify temperature and humidity plateaus or interruption features during swallowing and use them as a time reference for aligning other multimodal physiological signals.

[0018] Optionally, in step S3, the feature extraction includes:

[0019] The maximum contraction amplitude, contraction duration, integrated electromyography value, and left-right symmetry coefficient of the surface electromyography signal were extracted, as well as the maximum rate of change of impedance, laryngeal elevation amplitude index, and swallowing recovery time of the bioimpedance signal.

[0020] Extract the three-dimensional motion trajectory features and vibration spectrum features of the throat vibration signal;

[0021] The feature fusion includes: concatenating the features of each modality, and then using principal component analysis or linear discriminant analysis algorithms to reduce the dimensionality and generate the discriminative fusion feature vector.

[0022] Optionally, in step S4, the swallowing pattern classification model is constructed using a support vector machine, a convolutional neural network, or a long short-term memory network.

[0023] During the process of identifying the state of swallowing function, the dynamic characteristics of the laryngeal tissue are evaluated by the magnitude change of the bioimpedance signal, and the mechanical strength of the swallowing action is evaluated by the laryngeal vibration signal.

[0024] Optionally, in step S4, the invalid swallowing identification logic is as follows:

[0025] When a swallowing initiation signal is detected, but within a 500ms time window after the swallowing initiation signal is triggered, the root mean square amplitude of the laryngeal vibration signal does not reach the preset effective swallowing vibration energy threshold, and the calculated muscle contraction coordination coefficient... When the value is less than a preset muscle contraction coordination threshold, it is determined to be an invalid swallow; wherein, the muscle contraction coordination coefficient The calculation formula is:

[0026] ;

[0027] in, This is the muscle group contraction coordination coefficient within the current calculation time window, with a value between 0 and 1; This indicates that the operation takes the maximum value of the cross-correlation function within the set sliding time delay range; For the time variable within the current calculation time window; This is a sliding time delay variable used to represent the relative shift of the signals on the left and right sides on the time axis; The integral domain in the formula is the time length of the current detection window. The specific data segment for the calculation is limited; The surface electromyography signals acquired and extracted by the left composite electrode at time 10:00 The envelope amplitude; The surface electromyography (EMG) signals acquired and extracted by the right-side composite electrode after a time delay After translation, the corresponding time is The envelope amplitude; and They represent the times at time 1 and 2 respectively. The squares of the envelope amplitudes of electromyographic signals on the left and right sides; It is a time integral infinitesimal element.

[0028] Optionally, in step S4, the identification logic for suspected aspiration is to satisfy any of the following conditions:

[0029] Condition 1: A temporal disorder in the laryngeal movement sequence is detected, that is, the laryngeal repositioning time indicated by the bioimpedance signal is later than the respiratory recovery time indicated by the temperature and humidity signal.

[0030] Condition 2: Within a 2-second window after the swallowing action, a sudden electromyographic signal characterizing the cough reflex is detected, and the temperature and humidity signal of the neck microenvironment shows an abnormal interruption of the respiratory rhythm.

[0031] Optionally, in step S5, generating the corresponding risk warning signal includes: counting the total number of swallows within a 60-minute sliding time window. and number of ineffective swallows And calculate the swallowing efficiency index. ;

[0032] ;

[0033] in, The swallowing efficiency index is used to characterize the patient's current swallowing function. To count the cumulative number of consecutive invalid swallows within the obtained sliding time window; To count the total number of swallowing events that occur within the same sliding time window; and To convert the calculation results into a constant with percentage dimensions;

[0034] If the swallowing efficiency index If less than 70% of the swallows are swallowed or if more than 5 consecutive ineffective swallows occur, a function deterioration alarm will be triggered. Once a suspected aspiration is identified, an audible and visual alarm will be triggered immediately and the signal waveform data at the current moment will be recorded.

[0035] Optionally, a multi-signal analysis-based method for early warning of swallowing disorder risk also includes:

[0036] Generate a structured rehabilitation report, which includes historical records of swallowing frequency trends, swallowing strength scores, and safety risk levels.

[0037] Secondly, an embodiment of the present invention provides a swallowing disorder risk early warning system based on multi-signal analysis, comprising:

[0038] A flexible wearable sensing unit, configured to be worn around a user's neck, is used to perform the aforementioned signal acquisition operation;

[0039] A signal processing and wireless transmission unit is connected to the flexible wearable sensing unit and is configured to perform preliminary filtering on the collected signal and transmit it wirelessly.

[0040] The intelligent early warning terminal includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the method described in any one of the first aspects.

[0041] The present invention has achieved the following beneficial effects:

[0042] This invention integrates multiple sensors, including surface electromyography, bioimpedance, laryngeal vibration, and neck microenvironment temperature and humidity sensors, into a flexible wearable sensing unit. This constructs a comprehensive monitoring system covering neural drive, tissue deformation, mechanical vibration, and airflow exchange, effectively overcoming the problems of insufficient information and poor anti-interference capabilities of single signal sources in complex daily activities. The invention reconstructs respiratory rhythm using microenvironment temperature and humidity signals sensitive to airflow changes and uses the apnea characteristics during swallowing as a time alignment benchmark for multi-source signals. This solves the problems of physiological signal synchronization difficulties and motion artifact interference in uncontrolled environments, improving the purity of feature extraction. The system combines the timing logic of breathing and swallowing with a graded early warning strategy, objectively distinguishing between functional decline trends caused by ineffective swallowing and immediate safety hazards caused by aspiration. While ensuring monitoring sensitivity, it avoids unnecessary frequent alarms, providing continuous, objective, and pathologically interpretable auxiliary diagnostic evidence for clinical nursing and home rehabilitation.

[0043] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0044] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0046] Figure 1 This is a flowchart illustrating a method for early warning of swallowing disorder risk based on multi-signal analysis in an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of the composition structure of a swallowing disorder risk warning system based on multi-signal analysis in an embodiment of the present invention. Detailed Implementation

[0048] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0049] This application provides a swallowing disorder risk warning system based on multi-signal analysis. The system addresses pain points in hospital clinical nursing, neurological rehabilitation, and geriatric care scenarios, namely, the inherent subjectivity, inability to continuously monitor swallowing, and difficulty in accurately identifying hidden aspiration risks of existing swallowing assessment methods.

[0050] The swallowing disorder risk warning system based on multi-signal analysis has a hardware architecture that mainly includes three core interactive units: a flexible wearable sensing unit worn around the patient's neck, a signal processing and wireless transmission unit responsible for edge computing, and an intelligent warning terminal located at the nurse station or on a mobile device.

[0051] The flexible wearable sensing unit serves as the physical front end for data acquisition, employing a non-invasive conformal design that conforms to the anatomical characteristics of the human neck. Considering the potential skin sensitivity and long-term bedridden needs of target users (such as stroke patients and the elderly), the main structure of the unit utilizes a flexible neckband made of a medical-grade liquid silicone (LSR) blend with a one-way moisture-wicking silver fiber fabric. This neckband possesses excellent elasticity and breathability, adapting to the physiological curvature of the neck of different patients (ranging from thin to obese, with neck circumferences from 30cm to 50cm), ensuring that the sensor array maintains stable physical contact with the skin surface during swallowing, head turning, or speaking, thus suppressing motion artifacts at the source. Multiple high-precision sensor modules are integrated on the inner side of the neckband, strictly following the anatomical structure of the anterior cervical triangle.

[0052] Specifically, the flexible wearable sensing unit integrates the following four types of core sensor modules, and the spatial layout and functional definition of each module are as follows:

[0053] First, a composite acquisition module for surface electromyography and bioimpedance analysis. This module includes two pairs of composite electrode probes symmetrically embedded at both ends of the inner side of the neckband. When the wearer wears the neckband correctly, these two pairs of electrodes are positioned at the level of the thyroid cartilage, extending laterally to cover the area inside the anterior border of the sternocleidomastoid muscle. This area is the core projection area of ​​the submental muscle group (including the anterior belly of the digastric muscle and the mylohyoid muscle) and the infrahyoid muscle group (including the sternohyoid muscle and the thyrohyoid muscle). The composite electrodes are made of a flexible conductive polymer material doped with silver nanowires (AgNWs). Compared with traditional Ag / AgCl gel electrodes, this dry electrode avoids the problems of gel drying and impedance drift caused by prolonged wear. This module has a dual function of time-division multiplexing:

[0054] Firstly, it serves as a high-input-impedance biopotential pickup terminal, used to acquire microvolt-level surface electromyography (sEMG) signals generated by swallowing-related muscle groups under neural drive. This signal reflects the excitability and recruitment capacity of the neuromuscular system when the swallowing reflex is initiated.

[0055] Secondly, it serves as both the excitation and measurement terminals for four-wire bioimpedance measurement. The system injects a 50kHz frequency current into the neck tissue via the outer electrode, with the current intensity strictly controlled. A safe constant current source is used, and the resulting voltage drop is detected through an inner electrode. Based on Ohm's law and the principle of volumetric conductivity of biological tissues, a bioimpedance signal reflecting the movement state of deep larynx tissues is obtained. This signal is extremely sensitive to changes in the cross-sectional area of ​​neck tissues (such as Adam's apple elevation and pharyngeal cavity contraction), and can objectively depict the mechanical displacement of the swallowing organs.

[0056] Second, the laryngeal vibration acquisition module. This module is positioned in the center of the neck strap, precisely above the wearer's Adam's apple (the protrusion of the thyroid cartilage) or the cricothyroid membrane region. Internally, this module encapsulates a high-sensitivity MEMS (Micro-Electro-Mechanical Systems) triaxial accelerometer and a wide-band response piezoelectric thin-film sensor. The triaxial accelerometer is configured to monitor low-frequency mechanical motion in the range of 0Hz to 50Hz, primarily used to capture the three-dimensional motion trajectory of the laryngeal complex during swallowing, including elevation, forward movement, and repositioning. The piezoelectric thin-film sensor is configured to monitor high-frequency acoustic vibration in the range of 50Hz to 2000Hz, used to capture hydrodynamic noise generated as the bolus flows through the pharynx, vibrations caused by vocal cord closure, and chest shock waves caused by coughing or throat clearing.

[0057] Third, the neck microenvironment temperature and humidity acquisition module. This module is not directly attached to the skin to measure body temperature, but is located in a miniature flow-guiding groove on the inside of the neckband near the chin. This location was designed using fluid dynamics simulation, directly facing the diffusion path of the wearer's exhaled airflow from the nose and mouth, and maintaining an air gap of approximately 1mm to 2mm from the skin surface. This module integrates an ultra-miniature digital temperature and humidity sensor with extremely low heat capacity and millisecond-level response time. Its working principle is to utilize the thermodynamic difference between exhaled airflow (typically close to 37°C and saturated with humidity) and ambient air (typically cooler and drier) to construct a non-contact respiratory airflow monitoring field. By monitoring the temperature and relative humidity fluctuations of the microenvironment in real time, the system can accurately reconstruct the patient's respiratory rhythm and use the apnea phenomenon that inevitably occurs during swallowing as a time reference to assess the coordination between breathing and swallowing.

[0058] The signal processing and wireless transmission unit is the edge computing node of the entire wearable system. It is a miniaturized rigid-flexible board encapsulated in a magnetic battery compartment on the side of the neckband. This unit mainly includes:

[0059] Analog Front-End Circuit (AFE): Integrates a multi-channel, high-precision instrumentation amplifier. For surface electromyography (EMG) signals, the AFE is configured with a differential amplification link with a high common-mode rejection ratio (CMRR > 110dB) and connected in series with a fourth-order Butterworth bandpass filter (20Hz-500Hz) to effectively suppress power frequency interference and baseline drift. For bioimpedance signals, the AFE integrates a direct digital synthesis (DDS) waveform generator and a quadrature demodulator, which can accurately separate the real part (resistance R) and imaginary part (reactance X) of the impedance.

[0060] Microcontroller (MCU): Utilizing a low-power, high-performance ARM Cortex-M4F core processor, it is responsible for controlling the synchronous sampling timing of each sensor. To ensure strict alignment of multimodal data, the MCU employs a hardware timer-triggered DMA (Direct Memory Access) for data acquisition.

[0061] Wireless communication module: It adopts the Bluetooth Low Energy (BLE5.0) protocol to transmit pre-processed data packets to external terminals in real time.

[0062] Power management module: It integrates a lithium polymer battery charging management chip, supports fast charging and power monitoring, and ensures continuous 24-hour operation of the system.

[0063] This application further provides a method for early warning of swallowing disorder risk based on multi-signal analysis. Please refer to the appendix. Figure 1 The method includes the following detailed steps:

[0064] Step S1: Collect multimodal physiological signals from the wearer's neck. The multimodal physiological signals include at least surface electromyography signals, bioimpedance signals, laryngeal vibration signals, and neck microenvironment temperature and humidity signals.

[0065] After system startup, a sensor self-test procedure is first executed to confirm that the contact impedance of the composite electrode is below a threshold (e.g., 50kΩ) and that communication between all sensors is normal. To acquire the aforementioned multimodal physiological signals, in this embodiment, the multimodal physiological signals are collected through a flexible wearable sensing unit: the surface electromyography (EMG) signal and the bioimpedance signal are acquired through a composite electrode symmetrically embedded at the level of the thyroid cartilage; the laryngeal vibration signal is acquired through a miniature accelerometer placed above the Adam's apple; and the neck microenvironment temperature and humidity signal is acquired through a temperature and humidity sensor located at the mandible. Subsequently, a global synchronous acquisition mode is entered. To balance signal integrity with system power consumption, differentiated sampling strategies are set for different modalities: for surface EMG signals, a sampling rate of 2000Hz is set. This is because the spectrum of motor unit action potentials is mainly distributed between 20Hz and 500Hz, and a high sampling rate helps retain high-frequency harmonic characteristics and prevent aliasing.

[0066] For bioimpedance signals, the sampling rate was set to 1000 Hz. The excitation signal was a 50 kHz sine wave. The acquired impedance magnitude changes directly reflected the dynamic deformation of the laryngeal tissue.

[0067] For the throat vibration signal, the sampling rate is set to 1000Hz. Acceleration data along the X, Y, and Z axes are collected, where the X-axis corresponds to the left-right direction, the Y-axis corresponds to the up-down direction (the main axis of throat elevation), and the Z-axis corresponds to the front-back direction.

[0068] For the temperature and humidity signals of the neck microenvironment, a sampling rate of 50 Hz was set. Although the environmental parameters change slowly, this sampling rate is necessary to capture transient switching points in respiratory airflow (such as the start of exhalation).

[0069] All data from all channels are stamped with a uniform 64-bit hardware timestamp by the MCU at the acquisition end to ensure strict correspondence on the timeline.

[0070] Step S2: Perform time alignment and preprocessing on the multimodal physiological signals to obtain multi-channel time-series data.

[0071] The acquired raw signals are mixed with environmental noise and motion artifacts, and the response characteristics of each sensor are different, so direct fusion will lead to feature misalignment. Therefore, step S2 mainly performs signal cleaning and time alignment based on physiological characteristics.

[0072] First, perform basic signal filtering and noise reduction:

[0073] For surface electromyography (EMG) signals, a wavelet thresholding denoising algorithm was used to remove background white noise, and an adaptive notch filter was used to filter out 50Hz power frequency interference. Subsequently, full-wave rectification and envelope extraction were performed (using a root mean square sliding window with a window length of 50ms) to obtain the EMG envelope signal reflecting changes in muscle energy.

[0074] For bioimpedance signals, a low-pass filter with a cutoff frequency of 5Hz is used to remove minute impedance fluctuations (impedance graph components) caused by heartbeat, retaining only the large-amplitude swallowing motion component.

[0075] For the throat vibration signal, a high-pass filter (cutoff frequency 0.5Hz) is used to remove the gravitational acceleration component, retaining only the dynamic acceleration.

[0076] For temperature and humidity signals, detrending processing is performed to eliminate baseline drift caused by slow changes in ambient temperature and highlight high-frequency fluctuations caused by respiration.

[0077] Secondly, as the core of preprocessing, the system uses the temperature and humidity signals of the neck microenvironment to assist in the calibration of respiratory rhythm. Since the thermal response of the temperature and humidity sensor has a physical hysteresis (usually lags behind the actual airflow by 0.5s-1.0s), the system first uses the sensor's step response model for deconvolution compensation to recover the actual airflow change time.

[0078] Next, the algorithm detects the periodic fluctuation pattern of the temperature and humidity signal in the neck microenvironment to determine the respiratory phase: peak and trough detection is performed on the compensated temperature and humidity signal to divide the respiratory cycle, defining the rising edge of temperature and humidity as the expiratory phase and the falling edge or low flat period as the inspiratory phase.

[0079] Finally, the system identifies temperature and humidity plateaus or interruption characteristics during swallowing and uses these as a time reference for aligning other multimodal physiological signals.

[0080] Specifically, the system searches for swallowing apnea features: under normal respiratory fluctuations, if a plateau or signal interruption is detected in the temperature and humidity signal, lasting between 0.5s and 2.0s, and the absolute value of the first derivative during this period is lower than the stability threshold, then this time window is marked as an apnea window.

[0081] Using the center point of the apnea window as the reference anchor, the timing of the burst onset of surface electromyography (EMG) signals, the trough of bioimpedance signals, and the peak of laryngeal vibration signals are time-matched. Specifically, the system checks for significant EMG activity and laryngeal vibration within the apnea window. If present, signals from each channel are truncated around this anchor point (e.g., 1.5 seconds before and after), forming strictly aligned multi-channel time-series data slices. This alignment mechanism effectively eliminates artifacts caused by a single breath-hold (no EMG), a single speech (no airflow interruption plateau), or a single neck movement (no EMG burst), greatly improving data purity.

[0082] Step S3: Perform feature extraction and feature fusion on the multi-channel time series data to generate a discriminative fusion feature vector.

[0083] After acquiring synchronized multi-channel time-series data, the system extracts core features characterizing swallowing function from three dimensions: time domain, frequency domain, and nonlinear dynamics. Specifically, the feature extraction includes:

[0084] First, extract the temporal amplitude and slope characteristics of the surface electromyography signal and the bioimpedance signal.

[0085] Based on surface electromyography (EMG) signals, features reflecting neuromuscular drive capacity are extracted: maximum contraction amplitude (quantifying the explosive force of muscle contraction); contraction duration (the duration for which the EMG signal exceeds the resting threshold); integrated EMG value (reflecting the total amount of muscle work done during swallowing); and left-right symmetry coefficient (calculating the cross-correlation coefficient of the signals from the composite electrodes on both sides to assess the coordination of bilateral muscle group contraction).

[0086] Based on bioimpedance signals, features reflecting laryngeal movement dynamics are extracted: maximum rate of change of impedance (calculated by the maximum slope of the falling edge of the impedance magnitude curve); laryngeal elevation amplitude index (characterized by the relative displacement of laryngeal elevation after normalization of the peak value of impedance change); and swallowing recovery time (the time required to recover from the impedance minimum point to 90% of the baseline).

[0087] Second, extract the three-dimensional motion trajectory features and vibration spectrum features of the throat vibration signal.

[0088] Extracting three-dimensional motion trajectory features includes calculating the integral of the vector sum of triaxial accelerations (three-dimensional trajectory energy, reflecting the overall mechanical energy) and using frequency domain entropy to quantify the motion smoothness of the throat trajectory.

[0089] Extracting vibrational spectrum features: i.e., swallowing sound features, extracting the energy percentage of the high-frequency band (>500Hz). Aspiration-induced coughing or moist rales will lead to a significant increase in high-frequency energy.

[0090] Third, based on temperature and humidity signals, respiratory-swallowing coordination features are extracted:

[0091] Swallowing onset phase: Calculate the relative phase angle (0-360 degrees) of the swallowing initiation point within the respiratory cycle. Safe swallowing usually occurs in the mid-to-late expiratory phase.

[0092] Breathing reset delay: Calculate the time difference between the end of the swallowing action and the moment the first normal breathing resumes. An excessively long delay may indicate a reflexive breath-holding due to silent aspiration.

[0093] After completing the single-modal feature extraction described above, the system constructs a high-dimensional original feature vector. To eliminate redundancy between features (such as the high correlation between electromyographic amplitude and impedance change rate) and reduce model complexity, this embodiment employs a feature-level fusion strategy. The feature fusion includes: concatenating the features of each modality, then using principal component analysis or linear discriminant analysis algorithms for dimensionality reduction to generate the discriminative fused feature vector. Specifically:

[0094] First, all features are Z-score standardized to eliminate differences in different physical dimensions.

[0095] Next, the modal features are concatenated and spliced, and dimensionality reduction is performed using principal component analysis or linear discriminant analysis (LDA) algorithms, either supervised or unsupervised. LDA aims to find an optimal projection direction that minimizes the variance of the projected data among similar samples (e.g., effective swallowing) and maximizes the distance between dissimilar samples (e.g., effective swallowing and aspiration).

[0096] Finally, a discriminative fusion feature vector with 10 to 15 dimensions is generated. This vector highly integrates comprehensive information on muscle drive force, mechanical movement amplitude, movement smoothness, and respiratory coordination, and serves as the core basis for accurate judgment by the subsequent classification model.

[0097] Step S4: Input the discriminative fusion feature vector into the pre-trained swallowing pattern classification model for real-time inference to identify the state of swallowing function.

[0098] In step S3, the system has generated a high-dimensional, redundancy-free discriminative fusion feature vector (hereinafter referred to as the "input vector"). This vector highly condenses the wearer's neuromuscular electrophysiological characteristics, laryngeal tissue kinematic characteristics, mechanical vibration acoustic characteristics, and respiratory airflow thermodynamic characteristics within the current time window. To transform these abstract mathematical features into swallowing function states with clear clinical guidance, this embodiment employs a hybrid inference architecture combining statistical learning theory and expert rule systems.

[0099] In this embodiment, the swallowing pattern classification model is not a single black-box model, but a cascaded system consisting of a first-level classifier (initial event screening) and a second-level logic determiner (refined state determination). This architecture design ensures both efficient filtering of non-swallowing noise and medical interpretability of pathological state determination.

[0100] 1. First-level classifier:

[0101] First, the input vector is fed into a support vector machine (SVM) classifier pre-installed in a microcontroller (MCU) or intelligent early warning terminal processor. Considering the computational resource limitations of embedded systems and the nonlinear distribution characteristics of physiological signals in the feature space, the SVM classifier preferably uses a radial basis function (RBF) as its kernel function.

[0102] The core task of this classifier is to perform highly sensitive binary classification, that is, to determine whether the currently acquired signal segment belongs to a swallowing-related event or non-swallowing background noise (such as speaking, coughing, turning the head, or yawning). During the model training phase, a large number of samples labeled with the gold standard (such as video swallowing angiography VFSS) were used for supervised learning, enabling the classifier to find the optimal hyperplane in the high-dimensional feature space.

[0103] During real-time inference, if the classification confidence score of the SVM output is lower than a preset threshold (e.g., 0.75), or is determined to be non-swallowing background noise, the system will automatically discard the current data frame and reset the buffer, waiting for the next trigger. This mechanism effectively reduces the system's false alarm rate and power consumption. If the SVM determines it to be a swallowing-related event, the data stream will be immediately sent to the secondary logic determiner.

[0104] 2. Two-level logic diagnostic unit:

[0105] The secondary decision-maker no longer relies on simple probability statistics, but executes physical constraint logic based on pathophysiological mechanisms to accurately identify three specific states: effective swallowing, ineffective swallowing, and suspected aspiration.

[0106] A valid swallow is determined when the input vector satisfies all of the following physical constraints:

[0107] First, the integrated electromyography (iEMG) value of the surface electromyography signal is within the range of 80% to 120% of the wearer's personalized baseline value, indicating that the nerve drive is normal and the muscle contraction force is moderate.

[0108] Secondly, the laryngeal elevation displacement inverted by the bioimpedance signal is greater than the minimum effective displacement threshold corresponding to the wearer's anatomical structure (usually around 20mm), and the entire elevation-reduction process lasts between 0.6 seconds and 1.5 seconds.

[0109] Third, the spectral center of the throat vibration signal remains in the low-frequency range (<200Hz), with no high-frequency noise;

[0110] Fourth, the temperature and humidity signals of the neck microenvironment showed that within 0.5 to 1.0 seconds after the swallowing action ended, the respiratory airflow returned to normal, and the temperature and humidity readings showed typical expiratory phase characteristics (temperature rises and humidity increases).

[0111] Ineffective swallowing clinically corresponds to the pathological phenomenon of initiating swallowing but failing to deliver the bolus. It is commonly seen in patients with sarcopenia or bulbar palsy, manifesting as neurological intent without mechanical effectiveness. This embodiment concretizes this judgment logic into the following calculation process:

[0112] First, the system detects swallowing initiation. A neural initiation moment is marked when the instantaneous energy of the surface electromyography signal exceeds three standard deviations from the resting baseline. .

[0113] Subsequently, at the neural activation moment... Within the subsequent 500ms time window, the root mean square amplitude (RMS) of the laryngeal vibration signal and the peak value of the bioimpedance change were calculated. .

[0114] Set two key thresholds:

[0115] Effective swallowing vibration energy threshold. This threshold is set at 50% of the maximum swallowing vibration energy measured during the patient calibration phase.

[0116] : Muscle group contraction coordination threshold, which is set to 0.7 in this embodiment.

[0117] The system calculates the current muscle group contraction coordination coefficient. This coefficient is obtained by calculating the maximum value of the cross-correlation function of the electromyography signals on the left and right sides, as shown in the following formula:

[0118] ;

[0119] in, This is the muscle group contraction coordination coefficient within the current calculation time window, with a value between 0 and 1; This indicates that the operation takes the maximum value of the cross-correlation function within the set sliding time delay range; For the time variable within the current calculation time window; This is a sliding time delay variable used to represent the relative shift of the signals on the left and right sides on the time axis; The duration of the current detection window (in this embodiment) The value is taken from the aforementioned neural activation time. (500ms after), the integration domain in the formula The specific data segment for the calculation is limited; The surface electromyography signals acquired and extracted by the left composite electrode at time 10:00 The envelope amplitude; The surface electromyography (EMG) signals acquired and extracted by the right-side composite electrode after a time delay After translation, the corresponding time is The envelope amplitude; and They represent the times at time 1 and 2 respectively. The squares of the envelope amplitudes of electromyographic signals on the left and right sides; It is a time integral infinitesimal element.

[0120] The determination formula is:

[0121] If the conditions are met and ,in, The root mean square amplitude of the laryngeal vibration signal within the current detection window; that is, if the mechanical vibration energy generated by the larynx is too weak and the contraction of the muscles on both sides is severely asynchronous, the system determines that the event is an invalid swallowing.

[0122] The physical significance of this logic is that it eliminates signal loss caused by poor sensor contact (poor contact usually leads to extremely high power frequency interference rather than low energy), and accurately identifies swallowing failure caused by muscle weakness or motor incoordination.

[0123] Suspected aspiration is the highest priority risk monitored by this system. Aspiration often occurs amidst disruptions to fine motor timing or the triggering of the body's defensive reflexes. This embodiment constructs a multi-dimensional cross-validation logic:

[0124] Dimension 1: Detection of temporal disorder in laryngeal movement sequence: Normal swallowing follows a strict physiological sequence: electromyographic bursts larynx elevation Glottal closure (breathing apnea) Laryngeal repositioning Breathing has resumed.

[0125] The system calculates the laryngeal repositioning time indicated by the bioimpedance signal. Respiratory recovery time indicated by temperature and humidity signals .

[0126] If breathing resumes Laryngeal repositioning time This means that the airflow resumes before the throat has fully closed, which means that the airway is open while inhaling, and there is a very high probability that food will be inhaled into the trachea.

[0127] Dimension Two: Capturing the Defensive Cough Reflex:

[0128] During the 2-second high-risk window following the end of a swallowing event, the system continuously scans surface electromyography and vibration signals.

[0129] If a short-duration (duration < 300 ms) pulse train with extremely high amplitude (> 5 times the baseline) is detected in the surface electromyography signal, and the synchronous laryngeal vibration signal shows high-frequency (center frequency > 500 Hz) and high-energy shock wave characteristics, this is labeled by the algorithm as a reflexive cough or violent throat clearing.

[0130] Dimension 3: Verification of Abnormal Disruptions in Respiratory Rhythm:

[0131] If the temperature and humidity signal remains at the environmental baseline level for a long time (>3 seconds) after swallowing (i.e., no exhaled hot airflow), or if there are disordered high-frequency low-amplitude fluctuations (indicating airflow turbulence or wet speech), it is determined to be an abnormal respiratory rhythm.

[0132] Comprehensive Judgment Rules:

[0133] When either (Dimension 1 is true) or (Dimension 2 is true and Dimension 3 is true) is satisfied, the system immediately locks the current state as suspicious accidental aspiration.

[0134] Step S5: In response to the identified invalid swallowing or suspected aspiration, generate a corresponding risk warning signal.

[0135] This embodiment designs a hierarchical and layered closed-loop feedback mechanism to adapt to the complex nursing processes in hospitals and avoid alarm fatigue.

[0136] Considering that a single ineffective swallowing may be caused by accidental factors (such as talking interference or body position adjustment) and does not pose an immediate danger, this system will not trigger an audible or visual alarm.

[0137] The system opens a sliding time window (60 minutes long) in the background processor and counts the total number of swallows within this window in real time. and number of ineffective swallows .

[0138] Calculate the swallowing efficiency index :

[0139] ;

[0140] in, The swallowing efficiency index is used to characterize the patient's current swallowing function. To count the cumulative number of consecutive invalid swallows within the obtained sliding time window (e.g., a length of 60 minutes); To count the total number of all swallowing events (including effective swallowing and ineffective swallowing) that occur within the same sliding time window; and This is a constant used to convert the calculation results into percentage units.

[0141] If swallowing efficiency index If the percentage of ineffective swallowing exceeds 30%, or if there are more than 5 consecutive ineffective swallowings, the system will trigger a "functional decline alarm".

[0142] The alert was sent via Bluetooth to the central monitoring terminal at the nurses' station. A yellow notification popped up on the screen, displaying: "Patient XX in bed has experienced decreased swallowing efficiency; assessment of muscle fatigue is recommended." Simultaneously, the system marked the data for that period as requiring attention for doctors to review.

[0143] Once step S4 outputs a suspected aspiration determination, the system triggers the highest priority alarm procedure:

[0144] A miniature linear motor worn in the patient's neckband immediately executes three short, powerful vibrations (mode: strong-strong-strong), using a biofeedback mechanism to remind the patient to stop eating and try to cough voluntarily.

[0145] The intelligent early warning terminal immediately emits a high-decibel (>65dB) alarm sound (an intermittent sound with a frequency of 2000Hz), and the screen background flashes red throughout, displaying the words "Accidental Inhalation Alarm!".

[0146] The system immediately freezes all raw sensor waveform data from 30 seconds before to 30 seconds after the alarm. This 60-second data is not overwritten by subsequent data but is permanently stored in non-volatile memory and timestamped. This provides objective evidence for doctors to subsequently analyze the cause of aspiration.

[0147] Based on monitoring data throughout the day, the system automatically generates electronic reports in HL7 (Healthcare Information Transmission Standard) format.

[0148] The report includes: a swallowing frequency histogram (showing the diurnal swallowing distribution and assessing saliva management ability), a safety score curve (showing the fluctuation of the aspiration risk index throughout the day), and rehabilitation recommendations (automatically recommending the bolus characteristics for the next stage based on changes in electromyography amplitude).

[0149] This application also provides a swallowing disorder risk early warning system based on multi-signal analysis. Please refer to... Figure 2 The system includes:

[0150] 1. Flexible wearable sensing unit:

[0151] This unit is the front end for data acquisition. Its main body adopts a multi-layer composite fabric structure: the inner layer is a skin-friendly conductive silver fiber knitted fabric, which directly serves as the dry electrode for electromyography and impedance measurement; the middle layer is a highly elastic TPU (thermoplastic polyurethane) film, which is used to encapsulate the miniature accelerometer and temperature and humidity sensor chip and provide waterproof protection; the outer layer is a breathable Lycra fabric.

[0152] Specifically, the temperature and humidity sensor is mounted within a specially designed 3D-printed flexible airflow channel. This channel, located under the chin of the neck strap, has an inverted "V"-shaped opening designed to collect exhaled airflow from the nasal cavity while preventing skin sweat from directly contacting the sensor surface, thereby improving the signal-to-noise ratio of respiratory monitoring.

[0153] 2. Signal processing and wireless transmission unit:

[0154] This unit is a coin-sized rigid-flexible composite plate. It is equipped with:

[0155] Analog Front-End Chip (AFE): Integrates a 4-channel 24-bit high-precision ADC for synchronous acquisition of microvolt-level electromyographic and bioimpedance signals.

[0156] Main control chip (MCU): It adopts a low-power Cortex-M4 processor with DSP instruction set and runs a real-time operating system (RTOS). Its internal firmware implements all the algorithms of step S2 (preprocessing) and step S3 (feature extraction).

[0157] Wireless radio frequency chip: Supports Bluetooth 5.0 (BLE) protocol and is responsible for sending the compressed feature vector to the external terminal.

[0158] 3. Intelligent Early Warning Terminal

[0159] This terminal can be a desktop workstation at the nurses' station or a dedicated PDA carried by medical staff. The terminal's internal memory contains a swallowing intelligent monitoring app. This software includes the following functional modules:

[0160] Communication Guardian Module: Responsible for maintaining Bluetooth connections with multiple neckband devices, handling disconnection and reconnection, and data packet verification.

[0161] Inference Engine Module: Loads the pre-trained SVM model file and performs the classification inference in step S4.

[0162] Alarm Management Module: Executes the hierarchical alarm logic in step S5 and interfaces with the Hospital Information System (HIS).

[0163] To address the issue of poor model generalization ability caused by differences in neck anatomy (such as fat thickness and skin impedance) among different patients, this system designed a standardized baseline parameter calibration procedure. This procedure is enforced when the patient first wears the device and is a prerequisite for ensuring algorithm accuracy. Specifically, it includes the following steps:

[0164] Instruct the patient to remain seated, mouth closed, and breathing calmly for one minute. During this time, the system automatically measures the background noise level for each channel. Calculate the resting root mean square value of surface electromyography (EMG). And the fluctuation range of the temperature and humidity sensor during calm breathing. These values ​​will serve as the zero points for subsequent signals.

[0165] Instruct the patient to perform forceful dry swallowing (saliva swallowing) three times. Record the maximum amplitude of surface electromyography (EMG) during these three attempts. and the maximum change in bioimpedance .

[0166] The system calculates the normalization factor:

[0167]

[0168] in, This is the normalization factor for the surface electromyography signal calculated for this specific wearer; To record the maximum amplitude of surface electromyography signals obtained when guiding patients to perform standard forceful dry swallowing during the calibration phase; The root mean square value of the baseline surface electromyography signal was calculated to guide patients to remain at rest during the calibration phase.

[0169] In subsequent real-time monitoring, all electromyographic features were converted into relative values ​​(%MVC), i.e.:

[0170]

[0171] in, The calculated relative electromyography percentage value at the current moment (i.e., relative maximum voluntary contractile force %MVC); This represents the instantaneous amplitude of the surface electromyography signal acquired at the current time point during real-time monitoring. The root mean square value of the resting surface electromyography signal baseline obtained in the aforementioned calibration steps; This is the normalization factor calculated in the aforementioned calibration steps. It is calculated using this formula. This completely eliminates the interference of differences in subcutaneous fat thickness, electrode contact impedance, and basic muscle strength among different wearers on the fixed threshold of the subsequent classification algorithm.

[0172] Instruct the patient to drink 5ml of warm water. The system records the laryngeal vibration energy and timing characteristics of this standard swallow and stores it as an individual standard template. Subsequent invalid swallows will be judged based on the parameters of this template (for example, vibration energy below 50% of the template is considered invalid).

[0173] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for early warning of swallowing disorder risk based on multi-signal analysis, characterized in that, Includes the following steps: S1. Collect multimodal physiological signals from the wearer's neck, including at least surface electromyography signals of the thyroid cartilage, bioimpedance signals, laryngeal vibration signals, and neck microenvironment temperature and humidity signals. S2. Time alignment and preprocessing of the multimodal physiological signals are performed to obtain multi-channel time-series data. The preprocessing includes auxiliary calibration of respiratory rhythm using the temperature and humidity signals of the neck microenvironment. The periodic fluctuation patterns of the temperature and humidity signals in the neck microenvironment are detected to determine the respiratory phase; Identify the temperature and humidity plateau or interruption characteristics during swallowing and use them as a time reference to align other multimodal physiological signals. Perform time-series matching on the burst start point of surface electromyography signal, the trough point of bioimpedance signal, and the peak point of laryngeal vibration signal. S3. Perform feature extraction and feature fusion on the multi-channel time-series data to generate a discriminative fusion feature vector; S4. Input the discriminative fusion feature vector into the pre-trained swallowing pattern classification model for real-time inference to identify the swallowing function state. The swallowing function state includes effective swallowing, ineffective swallowing, and suspected aspiration. The identification logic for ineffective swallowing is as follows: When a swallowing initiation signal is detected, but within a 500ms time window after the swallowing initiation signal is triggered, the root mean square amplitude of the laryngeal vibration signal does not reach the preset effective swallowing vibration energy threshold, and the calculated muscle contraction coordination coefficient... If the swallowing rate is less than the preset threshold for muscle group contraction coordination, it is determined to be invalid swallowing. S5. In response to the identified invalid swallowing or suspected aspiration, generate a corresponding risk warning signal.

2. The method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 1, characterized in that, In step S1, the multimodal physiological signals are acquired through a flexible wearable sensing unit: The surface electromyography signal and the bioimpedance signal were acquired by a composite electrode symmetrically embedded at the horizontal position of the thyroid cartilage; The vibration signal of the larynx was acquired by a miniature accelerometer placed above the Adam's apple; The temperature and humidity signals of the neck microenvironment are collected by a temperature and humidity sensor located at the mandible.

3. The method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 1, characterized in that, In step S3, the feature extraction includes: The maximum contraction amplitude, contraction duration, integrated electromyography value, and left-right symmetry coefficient of the surface electromyography signal were extracted, as well as the maximum rate of change of impedance, laryngeal elevation amplitude index, and swallowing recovery time of the bioimpedance signal. Extract the three-dimensional motion trajectory features and vibration spectrum features of the throat vibration signal; The feature fusion includes: concatenating the features of each modality, and then using principal component analysis or linear discriminant analysis algorithms to reduce the dimensionality and generate the discriminative fusion feature vector.

4. The method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 1, characterized in that, In step S4, the swallowing pattern classification model is constructed using a support vector machine, a convolutional neural network, or a long short-term memory network. During the process of identifying the state of swallowing function, the dynamic characteristics of the laryngeal tissue are evaluated by the magnitude change of the bioimpedance signal, and the mechanical strength of the swallowing action is evaluated by the laryngeal vibration signal.

5. The method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 1, characterized in that, in, The muscle group contraction coordination coefficient The calculation formula is: ； in, This is the muscle group contraction coordination coefficient within the current calculation time window, with a value between 0 and 1; This indicates that the operation takes the maximum value of the cross-correlation function within the set sliding time delay range; For the time variable within the current calculation time window; This is a sliding time delay variable used to represent the relative shift of the signals on the left and right sides on the time axis; The integral domain in the formula is the time length of the current detection window. The specific data segment for the calculation is limited; The surface electromyography signals acquired and extracted by the left composite electrode at time 10:00 The envelope amplitude; The surface electromyography (EMG) signals acquired and extracted by the right-side composite electrode after a time delay After translation, the corresponding time is The envelope amplitude; and They represent the times respectively. The squares of the envelope amplitudes of electromyographic signals on the left and right sides; It is a time integral infinitesimal element.

6. The method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 1, characterized in that, In step S4, the identification logic for suspected aspiration is based on satisfying any of the following conditions: Condition 1: A temporal disorder in the laryngeal movement sequence is detected, that is, the laryngeal repositioning time indicated by the bioimpedance signal is later than the respiratory recovery time indicated by the temperature and humidity signal. Condition 2: Within a 2-second window after the swallowing action, a sudden electromyographic signal characterizing the cough reflex is detected, and the temperature and humidity signal of the neck microenvironment shows an abnormal interruption of the respiratory rhythm.

7. The method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 1, characterized in that, In step S5, generating the corresponding risk warning signal includes: counting the total number of swallows within a 60-minute sliding time window. and number of ineffective swallows And calculate the swallowing efficiency index. ; ； in, The swallowing efficiency index is used to characterize the patient's current swallowing function. To count the cumulative number of consecutive invalid swallows within the obtained sliding time window; To count the total number of swallowing events that occur within the same sliding time window; and To convert the calculation results into a constant with percentage dimensions; If the swallowing efficiency index If less than 70% of the swallows are swallowed or if more than 5 consecutive ineffective swallows occur, a function deterioration alarm will be triggered. Once a suspected aspiration is identified, an audible and visual alarm will be triggered immediately and the signal waveform data at the current moment will be recorded.

8. A method for early warning of swallowing disorder risk based on multi-signal analysis according to claim 7, characterized in that, Also includes: Generate a structured rehabilitation report, which includes historical records of swallowing frequency trends, swallowing strength scores, and safety risk levels.

9. A swallowing disorder risk early warning system based on multi-signal analysis, characterized in that, include: A flexible wearable sensing unit, configured to be worn around a user's neck, is used to perform signal acquisition operations; A signal processing and wireless transmission unit is connected to the flexible wearable sensing unit and is configured to perform preliminary filtering on the collected signal and transmit it wirelessly. The intelligent early warning terminal includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images