Method and device for identifying the risk of mis-swallowing for a person with disabilities

Abstract

This invention relates to the field of nursing technology for disabled individuals, and particularly to a method and device for identifying the risk of aspiration in disabled individuals. The method includes: establishing a multimodal, non-invasive data acquisition system to simultaneously collect temporal information on laryngeal vibration, respiratory status, and facial movements; preprocessing to extract effective signals; constructing an individual-specific physiological baseline model using historical normal physiological samples and binding it to the individual's identity; segmenting and processing real-time signals to construct a three-dimensional feature collaboration matrix, calculating mutual information entropy, and dynamically iterating the associated baseline; temporally aligning swallowing and respiratory abnormality anchor points, marking key intervals, and screening suspected risk events; and outputting the final aspiration risk identification result after environmental interference filtering, feature weight adaptation, and verification. This invention enables accurate, real-time identification and early warning of aspiration risk without requiring any wearable devices and without affecting the daily care and rest of disabled individuals.

Description

Technical Field

[0001] This invention relates to the technical field of care for disabled persons, and in particular to a method and device for identifying the risk of aspiration in disabled persons. Background Technology

[0002] Individuals with disabilities have a higher probability of accidental swallowing due to limited independent movement and weaker physiological expression abilities. Once accidental swallowing occurs, it can easily lead to serious safety accidents such as suffocation. Therefore, effectively identifying the risk of accidental swallowing in individuals with disabilities is an important guarantee for improving their care safety.

[0003] Currently, the identification of aspiration risk in disabled individuals mainly relies on regular rounds and observations by caregivers, manual review of video surveillance, or wearable monitoring devices based on sensors. Manual monitoring is highly dependent on continuous observation by caregivers, is limited by their energy levels, and is prone to blind spots, making it difficult to achieve real-time, continuous risk identification. Wearable monitoring devices require disabled individuals to wear related sensors, which are not only easily torn off unconsciously but also interfere with daily care procedures such as feeding and turning over. None of these methods can achieve accurate, real-time identification of aspiration risk without affecting the daily care and rest of disabled individuals, failing to meet the actual needs of care scenarios for disabled individuals. Summary of the Invention

[0004] This invention provides a method and apparatus for identifying the risk of aspiration in disabled persons, which can effectively solve the problems in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying the risk of aspiration in disabled individuals includes: A multimodal, non-intrusive data acquisition system was established to simultaneously collect the laryngeal vibration timing signal, respiratory status timing data, and facial movement timing information of disabled individuals. The collected raw data was preprocessed to extract effective signals. Based on historical normal physiological signal samples, we constructed an individual physiological baseline model specific to disabled individuals and completed the binding of the data collection system with the identity information of disabled individuals. The effective signal stream acquired in real time is segmented and processed to construct a three-dimensional feature coordination matrix of laryngeal vibration, respiratory state and facial movements. The mutual information entropy value of the three-dimensional features is calculated, and dynamic iteration of the associated baseline is achieved by combining the individual physiological baseline model. Based on the dynamic iteration results of the associated baseline, swallowing action anchor points and breathing abnormality anchor points are extracted and time-aligned, key monitoring intervals are marked, and suspected risk events of accidental swallowing are screened out. Environmental interference filtering and feature weight adaptive adaptation are performed on suspected risk events to retain valid suspected events; Verify valid suspected events, determine and confirm the risk of accidental swallowing, and output the identification results.

[0006] Furthermore, the multimodal non-contact acquisition system includes a non-contact throat vibration sensor, a non-contact breathing sensor, and a high-definition camera.

[0007] Furthermore, the preprocessing includes: The original timing signal of throat vibration is filtered and denoised to remove environmental vibration and transmission noise. Baseline calibration and interference removal were performed on the raw time series data of respiratory status to remove baseline drift and non-respiratory pressure interference. The raw image data of facial movement time sequence is processed by grayscale conversion, noise reduction and region of interest extraction to focus on the facial areas related to accidental swallowing.

[0008] Furthermore, constructing individual physiological baseline models specific to disabled individuals includes: Based on continuous signal acquisition for no less than a preset duration, signal samples corresponding to normal physiological states are selected. The vibration frequency, amplitude, and duration parameters of the laryngeal vibration signal, the respiratory frequency, depth, and rhythm parameters of the respiratory state signal, and the mouth closure angle, mandibular position, and laryngeal movement amplitude parameters of the facial movement signal were extracted respectively. Weights are assigned based on the correlation between the three types of signals and the risk of aspiration, and the weighted fused feature parameters and normal fluctuation range are used to form an individual physiological baseline model.

[0009] Furthermore, screening for potential accidental swallowing risk events includes: The swallowing action anchor point is located based on the peak value of laryngeal vibration, and the breathing abnormality anchor point is located based on respiratory mutation. The two types of anchor points are time-series aligned, and if the time interval is within a preset threshold, they are determined to be associated anchor points. Key monitoring intervals are defined by associated anchor points, and suspected risk events with abnormal three-dimensional features within the intervals are screened out.

[0010] Furthermore, environmental interference filtering includes: Construct an environmental disturbance feature library containing characteristics of common nursing disturbance events; Calculate the matching degree between the features of suspected risk events and the feature database, remove interfering events based on the matching degree threshold, and retain valid suspected events.

[0011] Furthermore, the adaptive adaptation of feature weights includes: Real-time assessment of environmental disturbance intensity levels; When interference is strong, reduce the weight of laryngeal vibration features and increase the weight of respiratory status and facial movement features.

[0012] Furthermore, the verification includes forward verification and reverse verification. Only when both verifications pass is it determined to be a confirmed risk event of accidental swallowing.

[0013] Furthermore, the output recognition results include: Early warning levels are classified according to the degree of anomalies in the characteristics, and differentiated early warning responses are implemented accordingly; Generate a time-series analysis map that includes full-time signal trends and risk event markers.

[0014] On the other hand, the present invention also provides a device for identifying the risk of aspiration in disabled persons, comprising: The multimodal acquisition module is used to build a multimodal non-sensory acquisition system, which simultaneously acquires the laryngeal vibration timing signal, respiratory status timing data and facial movement timing information of disabled persons, and preprocesses the acquired raw data to extract effective signals. The baseline construction module is used to build an individual physiological baseline model specific to disabled persons based on historical normal physiological signal samples, and to complete the binding of the data collection system with the identity information of disabled persons; The dynamic iteration module is used to segment the real-time acquired effective signal stream, construct a three-dimensional feature coordination matrix of laryngeal vibration, respiratory state and facial movements, calculate the mutual information entropy value of the three-dimensional features, and realize the dynamic iteration of the associated baseline by combining the individual physiological baseline model. The temporal labeling module is used to extract swallowing action anchors and breathing abnormality anchors based on the dynamic iteration results of the associated baseline, align them temporally, mark key monitoring intervals, and screen out suspected risk events of accidental swallowing. The event filtering module is used to filter out environmental interference and adaptively adapt feature weights for suspected risk events, retaining valid suspected events. The event determination module is used to verify valid suspected events, determine and confirm the risk of accidental swallowing, and output the identification results.

[0015] The technical solution of this invention can achieve the following technical effects: This invention establishes a non-contact, multimodal, and sensorless data acquisition system. By combining synchronously acquired three-dimensional temporal signals of laryngeal vibration, respiratory status, and facial movements, and utilizing an individual-specific physiological baseline model and dynamic mutual information entropy calculation, it achieves real-time and accurate identification of the risk of aspiration in disabled individuals. It eliminates the need for personnel to wear any equipment, avoiding the drawbacks of wearable sensors that are prone to falling off and interfering with daily care procedures. Furthermore, through adaptive temporal segmentation, multi-dimensional feature temporal alignment, and environmental interference filtering mechanisms, it addresses the limitations of human monitoring in terms of energy levels and blind spots, reducing misjudgment and missed detection rates. This improves care safety and response efficiency without affecting the rest and comfort of disabled individuals.

[0016] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method for identifying the risk of aspiration in disabled individuals in an embodiment of the present invention. Figure 2 This is a structural block diagram of an aspiration risk identification device for disabled persons in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] like Figure 1 As shown, the present invention provides a method for identifying the risk of aspiration in disabled individuals, which specifically includes the following steps: Step S1: Build a multimodal non-sensory acquisition system to simultaneously acquire the laryngeal vibration timing signal, respiratory status timing data and facial movement timing information of disabled persons, and preprocess the acquired raw data to extract effective signals. Step S2: Based on historical normal physiological signal samples, construct an individual physiological baseline model specific to disabled persons and complete the binding of the data collection system with the identity information of disabled persons; Step S3: Segment the real-time acquired effective signal stream, construct a three-dimensional feature collaboration matrix of laryngeal vibration, respiratory state and facial movements, calculate the mutual information entropy value of the three-dimensional features, and combine it with the individual physiological baseline model to realize the dynamic iteration of the associated baseline. Step S4: Based on the dynamic iteration results of the associated baseline, extract the swallowing action anchor point and the breathing abnormality anchor point and align them in time sequence, mark the key monitoring interval, and screen out suspected risk events of accidental swallowing. Step S5: Perform environmental interference filtering and feature weight adaptive adaptation on suspected risk events, and retain valid suspected events; Step S6: Verify the valid suspected events, determine and confirm the risk of accidental swallowing, and output the identification results.

[0022] In this embodiment, by combining multimodal non-invasive data acquisition, individual physiological baseline modeling, and three-dimensional feature collaboration matrix with temporal anchor point alignment, synchronous monitoring and joint judgment of laryngeal vibration, respiratory status, and facial movements of disabled individuals can be achieved without the need for wearable devices or interference with daily care and rest. This overcomes the monitoring blind spots and inability to identify in real time at all times that exist in manual monitoring, and avoids the defects of wearable devices that are easy to fall off and interfere with nursing operations. At the same time, it improves the accuracy of risk identification by relying on individual-specific baselines and dynamic mutual information entropy calculation, and reduces the false judgment rate through temporal alignment and interference filtering. Finally, it completes the real-time and accurate identification of aspiration risk under non-invasive conditions, thereby improving the safety of care for disabled individuals.

[0023] In a specific implementation, as one example, given the core shortcomings of existing methods for identifying aspiration risk—namely, the inability of manual monitoring to achieve real-time monitoring around the clock, and the ease with which wearable data collection devices can fall off and interfere with daily care—it is necessary to achieve accurate acquisition and effective extraction of multi-dimensional temporal signals without contacting the disabled person's body or interfering with their daily care and rest. This example achieves this by building a multimodal non-intrusive acquisition system, simultaneously acquiring target signals, and performing targeted preprocessing, as detailed below: Step S11: Establish a multimodal non-contact acquisition system and determine the signal acquisition equipment, deployment method and acquisition parameters for each dimension. The construction process follows the principles of non-contact installation, non-interference deployment and multi-signal collaborative acquisition. Considering the nursing scenario where disabled persons are mostly bedridden and have weak independent activity ability, all acquisition points are deployed around the bedridden area of ​​disabled persons, without direct contact with the body, to avoid contact interference and the risk of unconscious tearing. The timing signal acquisition of laryngeal vibration uses a non-contact vibration sensor, deployed to the side of the disabled person's bedside. The distance is determined based on the typical distance between the laryngeal region and the head of the bed when the disabled person is lying down, balancing signal capture accuracy with equipment interference avoidance. This avoids interference caused by being too close and signal attenuation caused by being too far away. The laryngeal region produces characteristic vibrations when the disabled person swallows or accidentally swallows. The non-contact sensor can capture this signal without contacting the skin, without needing to be placed against the laryngeal region, and does not interfere with nursing operations such as feeding or turning over. It will also not interrupt signal acquisition due to the disabled person's unconscious movements. The sensor sampling frequency must match the frequency range of laryngeal vibration. A specification that can fully cover the vibration frequency range corresponding to swallowing actions is selected, such as 50Hz. This frequency is determined based on the fact that laryngeal vibration during human swallowing is mostly in the low to mid-frequency range. This avoids the loss of vibration characteristics due to the sampling frequency being too low and the generation of redundant data due to the sampling frequency being too high, thus reducing the processing burden. The respiratory status time-series data acquisition uses a non-contact respiratory sensor deployed under the mattress of the disabled person's bed, close to the corresponding area of ​​the chest. This location can accurately capture the subtle pressure changes transmitted by the chest respiratory movements when the disabled person is lying down, and then convert them into respiratory status time-series data. Deploying it under the mattress completely avoids contact with the body and does not affect daily care such as turning over and bathing, while enabling 24-hour uninterrupted data collection, solving the problem that manual monitoring cannot monitor at all times. The sensor acquisition parameters are adapted to the respiratory characteristics of disabled persons. The respiratory rate of disabled persons is usually lower than that of healthy adults and the respiratory depth fluctuates. The acquisition parameters need to be able to capture subtle changes in respiratory rate and the range of fluctuations in respiratory depth to ensure that abnormal respiratory signals can be effectively captured. The parameters are determined by pre-collecting respiratory data of disabled persons in different physical states and statistically analyzing their respiratory rate and depth within a normal range. Based on this, the acquisition threshold and sampling interval are set to ensure that the acquired data covers both normal and abnormal respiratory states, while avoiding invalid data redundancy. Facial movement timing information is collected using a high-definition camera, deployed directly above the head of the disabled person's bed. The lens angle is adjusted to fully cover the facial area, including the mouth, throat, and jaw. The deployment angle and height must avoid obstruction by caregivers, bedding, etc., while ensuring clear capture of subtle facial movements. In cases of aspiration, disabled individuals typically exhibit abnormal mouth closure, slight jaw twitching, and abnormal throat movement. This deployment method accurately captures these characteristic movements, and the non-contact deployment does not cause visual disturbance or psychological stress to the disabled person, nor does it interfere with daily care. The camera's acquisition parameters balance image clarity and data volume, with resolution... Select a specification that can clearly identify subtle facial movements, and match the frame rate to the speed of facial movement changes, such as 15 frames per second. This frame rate is determined based on the characteristic that the facial movements of disabled people are slower than those of healthy adults. It can completely capture the temporal changes of every subtle movement, avoiding the loss of movement features due to too low a frame rate and the image data redundancy caused by too high a frame rate, thus reducing the computational load of subsequent image processing. At the same time, the infrared supplementary light function is turned on, and the supplementary light intensity is automatically adjusted according to changes in indoor light to ensure clear acquisition at night or in low-light environments. The supplementary light intensity adjustment range is determined through prior testing to avoid excessive stimulation of the eyes and image blurring due to too weak a light intensity, so as not to affect the rest of disabled people. Step S12: Achieve multi-dimensional signal synchronous acquisition to ensure the consistency of timing of various types of time-series signals. This can be achieved by setting up a unified multimodal signal synchronous calibration unit. This unit establishes communication connections with the laryngeal vibration sensor, respiratory sensor, and high-definition camera, and sets a unified acquisition start time and sampling timestamp to ensure that the acquisition start time of the laryngeal vibration timing signal, respiratory state timing data, and facial movement timing information is consistent and the sampling interval is synchronized, avoiding subsequent feature association errors caused by asynchronous acquisition. When swallowing occurs, the changes in laryngeal vibration, respiratory state, and facial movements are temporally correlated. Asynchronous acquisition will lead to misalignment of the three-dimensional signal timing, which cannot accurately reflect the coordinated change relationship of the three. The multimodal signal synchronous calibration unit ensures that each set of synchronously acquired data corresponds to the same time node by calibrating the time of each acquisition device in real time. Step S13: Perform targeted preprocessing on the collected raw data to extract effective signals: For the original throat vibration timing signal, the main interferences are environmental vibration and signal transmission noise. Preprocessing involves filtering and signal denoising. The filtering method is selected to filter low-frequency environmental vibration and high-frequency noise. The filtering range is determined based on the effective frequency range of the throat vibration signal pre-acquired. Signals outside this range are considered interference signals and are removed. For example, low-frequency environmental vibration below 10Hz and high-frequency noise above 200Hz are filtered to ensure that only effective signals related to throat vibration are retained. The signal denoising method is selected to remove random noise. The reason is that random noise in the original signal will blur the vibration characteristics. This processing can remove random noise and restore the true characteristics of the throat vibration signal. For the raw time-series data of respiratory status, the main interferences are pressure interference caused by the disabled person turning over and the movement of bedding, as well as baseline drift of the sensor itself. Preprocessing adopts baseline calibration and interference removal. Baseline calibration is carried out by collecting respiratory data of the disabled person in a quiet resting state to determine the baseline of the respiratory signal. The collected raw data is compared with this baseline, and the deviation data caused by baseline drift is removed. The baseline calibration frequency is set to once per hour, based on the fact that the disabled person's physical condition may change slowly. Regular calibration can ensure that the baseline matches the current physical condition and avoid the error in respiratory status judgment caused by baseline deviation. Interference removal is achieved by setting a pressure change threshold. In the early stage, the range of pressure change caused by the disabled person turning over and the movement of bedding is collected and statistically analyzed. Pressure signals exceeding the range are regarded as interference signals and removed to ensure that the pressure change signals generated only by respiratory movement are retained, which accurately reflects the true changes in respiratory rate and respiratory depth. For raw image data of facial movement time sequence, the main interferences are changes in lighting, background clutter, and image noise. Preprocessing includes image grayscale conversion, denoising, and region of interest (ROI) extraction. Color images contain redundant color information, which increases the computational load of subsequent processing. Grayscale conversion can preserve image contours and movement features while simplifying the processing flow. Image denoising can be performed by removing random noise and preserving detailed features. Image noise can blur facial movement details and affect subsequent movement anchor point positioning. This processing can remove noise and restore the true contours of facial movements. ROI extraction uses a preset facial contour recognition algorithm to extract facial regions from the image, focusing on preserving areas related to accidental swallowing such as the mouth, throat, and jaw. Background clutter and irrelevant areas are removed. Focusing on ROI can reduce the computational load of subsequent processing and improve the accuracy of feature extraction.

[0024] In this embodiment, the shortcomings of existing acquisition methods are avoided by using a non-contact multimodal acquisition system. The synchronously acquired three-dimensional time-series signals can provide high-quality and effective signals after targeted preprocessing. Compared with the prior art, this embodiment does not require contact with the disabled person's body and does not interfere with daily care. Multi-dimensional synchronous acquisition can comprehensively capture the characteristics related to accidental swallowing, and the preprocessing method can effectively eliminate interference, ensuring the effectiveness and accuracy of the signal, and adapting to the actual use needs of the care scenario for disabled persons.

[0025] In some embodiments of the present invention, existing physiological feature analysis for identifying the risk of aspiration typically uses a general physiological baseline, failing to consider the inherent differences in physiological characteristics among different individuals, nor taking into account the dynamic changes in physiological characteristics of the same disabled person under different physical conditions. This easily leads to misjudgments and omissions in risk identification, and the data collection system is not bound to the identity information of disabled persons, making it impossible to achieve corresponding monitoring in scenarios with multiple disabled persons. Based on the above problems, this embodiment provides a judgment criterion for risk identification by constructing an individual physiological baseline model specific to disabled persons, while simultaneously binding the data collection system to the identity information of disabled persons, achieving correspondence between the baseline, data collection, and identity. Specifically, the following operations are performed: Step S21: Based on the multimodal non-sensory acquisition system, continuously acquire signals from the target disabled person for no less than a preset duration. The preset duration is determined according to the stability of the disabled person's physiological state, such as 72 hours, to ensure that the acquired signals can cover various physiological fluctuations under normal conditions. The acquisition process is accompanied by professional nursing staff throughout, who record the disabled person's physiological state, mark the signals within the corresponding time period, and use the signals marked as corresponding to the normal physiological state as screening samples. The characteristic ranges of three types of signals in the sample, namely laryngeal vibration timing signal, respiratory state timing data, and facial movement timing information, are statistically analyzed. Based on this, a screening threshold is set to remove abnormal fluctuation signals. Step S22: Based on the filtered signals, feature parameters of the three types of signals are extracted respectively, and a complete individual physiological baseline model is formed after weighted fusion. When extracting characteristic parameters of laryngeal vibration signals, three parameters are extracted: vibration frequency, vibration amplitude, and vibration duration. These three parameters can accurately reflect the normal physiological state of the larynx. The extraction method is as follows: the screened laryngeal vibration time sequence signal is segmented for analysis. The duration of each segment is determined according to the duration of a normal swallowing action, for example, 0.5 seconds. The frequency, amplitude, and duration of each segment are statistically analyzed, and the average value is taken as the baseline value of the parameter. The fluctuation range is determined based on the maximum fluctuation value in the pre-collected normal samples to cover normal physiological fluctuations and avoid misjudgment. When extracting respiratory status signal feature parameters, three parameters are extracted: respiratory rate, respiratory depth, and respiratory rhythm. These three parameters can reflect the normal physiological state of breathing. The extraction method is as follows: the time series data of the filtered respiratory status are continuously statistically analyzed, and a statistical period is set, such as 1 minute. The average respiratory rate and respiratory depth within each statistical period are calculated to determine the pattern of respiratory rhythm. The average value and rhythm pattern are used as the baseline value of respiratory characteristics. At the same time, a normal fluctuation range is set. The fluctuation range is determined by statistical analysis of pre-collected normal samples to take into account the individual differences and physiological fluctuations of the respiratory status of disabled persons. When extracting facial motion signal feature parameters, three parameters are extracted: mouth closure angle, mandibular position, and laryngeal undulation amplitude. These three parameters can accurately reflect normal facial motion features. The extraction method is as follows: frame analysis is performed on the selected facial motion time sequence images, feature points of the mouth, mandible, and laryngeal region are extracted in each frame, the corresponding parameter values ​​are calculated, and the average value and fluctuation range of the parameters in all frames are statistically analyzed as the baseline value of facial motion features. The fluctuation range is determined based on the fluctuation of the parameter in the pre-collected normal samples to ensure that it is adapted to individual facial motion characteristics. The weighting was based on the correlation between the three types of signals and the risk of aspiration. Among them, the laryngeal vibration signal was directly related to the swallowing action and had the highest weight. The respiratory state signal and facial movement signal were used as auxiliary verification and had a relatively lower weight. The weight values ​​were determined by comparing normal and abnormal samples collected in the early stage to ensure that the fused baseline model could comprehensively and accurately reflect the normal physiological state of disabled persons. The individual physiological baseline model formed after fusion includes the baseline values ​​and normal fluctuation range of the three types of signals. Step S23: Reuse the acquisition device that collects facial motion timing information to capture facial features. The facial recognition parameters are adapted in coordination with the facial motion acquisition parameters, and the resolution and frame rate are kept consistent. Nursing staff enter the basic information of the disabled person, including the disabled person's name, age, degree of disability, past medical history, etc., and generate a unique identity feature template through facial recognition to establish a binding relationship between identity and facial features. The correspondence between identity and facial features is entered into the acquisition system. When the acquisition system is started, it captures the disabled person's facial features in real time through facial recognition and matches them with the preset identity feature template. If the match is successful, the acquisition system calls the disabled person's unique individual physiological baseline model. If the match fails, a prompt signal is issued, and the nursing staff confirms the identity and rematches to avoid identity confusion.

[0026] In this embodiment, the individual-specific baseline model fits the physiological characteristics of disabled persons and adapts to the differences between individuals; the identity binding adopts a non-contact method, which does not increase the operational burden of disabled persons, does not interfere with daily care, and at the same time achieves accurate correspondence in scenarios with multiple disabled persons, avoiding identity confusion; the data collection system and identity recognition share hardware devices, saving equipment costs, avoiding interference caused by the deployment of multiple devices, and adapting to the actual needs of disabled persons' care scenarios.

[0027] In a specific implementation, as one example, existing effective signal analysis for risk identification typically uses a fixed-length window to segment the signal. The window size cannot adapt to individual differences and dynamic changes, easily leading to overly fine segmentation resulting in redundant data, or overly coarse segmentation losing key features. Based on these problems, this embodiment uses an adaptive temporal window to segment the effective signal, constructs a three-dimensional feature coordination matrix, and combines it with an individual physiological baseline model to achieve dynamic iteration of the associated baseline. The specific implementation steps are as follows: Step S31: Use an adaptive timing window to segment the real-time acquired effective signals and determine the dynamic adaptation rules of the adaptive timing window; the size of the adaptive timing window is dynamically adapted according to the daily swallowing frequency of the disabled person, and the adaptation rules are determined through pre-collection and statistics: when the swallowing frequency increases, the window is reduced to ensure that the complete features of each swallowing action are captured; when the swallowing frequency decreases, the window is increased to reduce redundant data and reduce computational pressure. Segmentation is performed synchronously for three types of valid signals to ensure that the time range of the three types of signals is consistent within the same time window, and to avoid feature association errors caused by segmentation asynchrony. During the segmentation process, a continuous signal interception method is used to avoid destroying the temporal continuity of the signals. A reasonable overlap area is set between each two adjacent windows. The size of the overlap area is determined according to the duration of the swallowing action to avoid losing key features at the window connection. Step S32: Based on the segmented effective signal, extract the feature parameters of the three types of signals within each time window and construct a three-dimensional feature coordination matrix; the feature parameters are consistent with the parameters extracted when constructing the individual physiological baseline model, and the extraction method follows the same logic to ensure consistent and comparable results; during the extraction process, abnormal feature points that exceed the normal fluctuation range of the individual physiological baseline model are removed. The three-dimensional feature coordination matrix uses time-series windows as units, and takes the nine types of feature parameters in each window as matrix elements to form a matrix structure of time-series windows and feature parameters. It presents the correlation between the three types of features in different time-series windows, which is consistent with the characteristics of abnormal coordination of the three types of features when aspiration occurs. Step S33: Calculate the mutual information entropy values ​​of the three types of features in the three-dimensional feature coordination matrix, and determine the calculation logic and judgment criteria for the mutual information entropy values. The mutual information entropy value is used to measure the coordination strength of the three types of features. Under normal physiological conditions, the coordination strength is stable and the entropy value is within a fixed range. When swallowed accidentally, the coordination strength changes abruptly and the entropy value deviates from the normal range. The mutual information entropy value is calculated separately for each time window. The logic is as follows: calculate the pairwise mutual information values ​​of laryngeal vibration and respiratory state, laryngeal vibration and facial movements, and respiratory state and facial movements respectively, and then calculate the overall mutual information entropy value. The normal range of entropy values ​​is determined based on the individual physiological baseline model. By statistically collecting the entropy values ​​of normal physiological signals corresponding to each time window, the average value and fluctuation range are taken as the normal range. Step S34: Determine the iteration cycle based on the rate of change in the disabled person's physical condition. After each iteration cycle, extract the feature parameters and mutual information entropy values ​​of all time windows within that cycle and compare them with the normal range of feature parameters and mutual information entropy values ​​in the individual's physiological baseline model. If the feature parameters and mutual information entropy values ​​are within the normal range and show a stable trend in multiple consecutive iteration cycles, appropriately reduce the fluctuation range of the association strength threshold to improve the accuracy of anomaly judgment. If the feature parameters and mutual information entropy values ​​show slight fluctuations but do not exceed the normal range, appropriately expand the fluctuation range of the association strength threshold to avoid misjudging normal physiological fluctuations as abnormalities. If feature parameters are abnormal but do not reach the risk standard for aspiration, adjust the association strength threshold to adapt to the changes in the current physical condition. The mutual information entropy value calculated for each time window is compared with the preset range of the dynamic baseline after the current iteration. When the mutual information entropy value exceeds the preset range, it is determined that there is a suspected risk of accidental swallowing and a preliminary warning is triggered. When the mutual information entropy value is within the preset range, it is determined to be a normal physiological state and no warning is triggered.

[0028] In this embodiment, the adaptive temporal window adapts to the dynamic changes in swallowing frequency, avoiding the unreasonable segmentation problem of fixed windows, and taking into account both the integrity of feature capture and data redundancy control; the three-dimensional feature collaboration matrix integrates three types of signal features to comprehensively reflect the multi-dimensional changes in misswallowing; the mutual information entropy value quantifies the correlation strength, the judgment standard is clear, and subjective bias is avoided; the correlation baseline dynamically iterates to adapt to changes in physical state, reducing the probability of misjudgment and omission, and the overall approach is in line with the individual differences of disabled persons and the needs of nursing scenarios.

[0029] In a specific implementation, as one example, to accurately extract key temporal nodes related to aspiration, clarify feature correlations, and screen out suspected aspiration risk events, it is necessary to extract swallowing action anchor points and respiratory abnormality anchor points based on the dynamic iteration results of the correlation baseline, align them temporally, and mark key monitoring intervals. The specific implementation steps are as follows: Step S41: Based on the dynamic iteration results of the associated baseline, extract the swallowing action anchor point and clarify the positioning rules and judgment criteria; the swallowing action anchor point is located according to the peak signal of the laryngeal vibration characteristics. When a disabled person swallows, the laryngeal muscles contract regularly, which drives the laryngeal vibration to produce obvious peaks. The peak time is highly synchronized with the swallowing action, which can accurately characterize the swallowing sequence. The positioning rules are determined through preliminary data collection and testing: based on the pre-collected normal physiological signals, the peak characteristics of laryngeal vibration corresponding to normal swallowing are statistically analyzed. Combined with the associated baseline after dynamic iteration, the judgment thresholds for peak amplitude and duration are set to distinguish between normal and abnormal swallowing peaks and exclude false peaks. The anchor point extraction process is as follows: traverse the laryngeal vibration time-series signal, identify the peak signal that meets the judgment threshold in each time-series window, take the peak start time as the swallowing action anchor point time-series position, and record the corresponding feature parameters and time-series window; remove false peaks with peak amplitude below the threshold, duration exceeding the normal range, or excessive deviation from the normal features of the associated baseline to ensure the accuracy of the anchor points; Step S42: Based on the dynamic iteration results of the associated baseline, extract the respiratory abnormality anchor points and clarify the localization rules and judgment criteria. The respiratory abnormality anchor points are key signal nodes that characterize respiratory abnormalities. When swallowed accidentally, airway obstruction will cause sudden changes in respiratory rate and respiratory depth. These two types of parameter mutations are highly identifiable and directly related to swallowing. The localization rules are determined through preliminary data collection and testing: based on the pre-collected normal physiological signals, the frequency and depth range of normal breathing are statistically analyzed, and combined with the correlation baseline after dynamic iteration, the thresholds for judging the amplitude and duration of mutations are set. The anchor point extraction process is as follows: traverse the segmented respiratory time series data, monitor the changes in respiratory frequency and depth within each time series window, and when the change amplitude exceeds the mutation amplitude threshold and the duration reaches the threshold, take the change start time as the respiratory abnormality anchor point time series position, and record the corresponding mutation parameters and time series window; remove false abnormal anchor points that do not reach the threshold, have insufficient duration, or whose overall characteristics are within the normal range when the mutation occurs, to ensure the accuracy of the anchor points; Step S43: Using the unified timestamp of all time windows as a reference, convert the temporal positions of the two types of anchor points into unified time coordinates; set a temporal association threshold, i.e., the maximum time interval between the two types of anchor points during accidental swallowing, which is determined by pre-collecting accidental swallowing samples; if the time interval between the two types of anchor points is ≤ the threshold, it is determined that there is a temporal association and they are regarded as a set of associated anchor points; if the swallowing action anchor point has no corresponding respiratory abnormality anchor point or the interval exceeds the threshold, it is determined to be normal swallowing; if the respiratory abnormality anchor point has no corresponding swallowing action anchor point or the interval exceeds the threshold, it is determined to be an abnormality not caused by accidental swallowing and is not included in the suspected risk screening scope. Step S44: Using the start time of the swallowing action anchor point in each set of associated anchor points as the starting point of the interval and the end time of the breathing abnormality anchor point as the ending point of the interval, extend the interval forward and backward by a preset time to cover the signal precursors before aspiration and the breathing recovery process afterward; after marking, extract three types of effective signal features within the interval and compare them with the preset range of the associated baseline. If all three types of features are abnormal and exceed the range, they are judged as suspected aspiration risk events; if a single dimension is abnormal or the abnormality does not meet the standard, it is judged as a normal physiological or minor interference event and is not included in the suspected range; during the screening process, record the associated anchor point time sequence, monitoring interval range, and signal abnormality parameters for each suspected event, and sort them by occurrence time.

[0030] In this embodiment, the swallowing action anchor point is based on sufficient evidence and has high recognition, which can accurately distinguish between normal and abnormal swallowing; the respiratory abnormality anchor point is directly related to abortion and can accurately capture respiratory abnormalities caused by airway obstruction; the temporal alignment clarifies the relationship between the two types of anchor points and excludes irrelevant abnormalities; the key monitoring interval fully covers the temporal process of abortion, and combined with secondary verification of the associated baseline, it ensures the accuracy of suspected events. All technical features are supported by pre-collection and testing and are adapted to the needs of nursing scenarios.

[0031] In a specific implementation, as one example, given that the screened suspected swallowing risk events may be mixed with environmental interference events unrelated to swallowing, such as nursing procedures and environmental noise, and that the reliability of the three types of signal features differs under different environmental interference intensities, it is necessary to filter out interference from suspected risk events and dynamically adapt feature weights to ensure the accuracy of the retained valid suspected events. This embodiment retains valid suspected events by filtering out environmental interference from suspected risk events and adaptively adapting feature weights. The specific implementation steps are as follows: Step S51: Preset environmental interference feature library, determine the construction rules, contents and update logic of the interference feature library; this feature library is used to provide a reference for interference event features, quickly distinguish interference from real suspected events, and avoid subsequent misjudgment caused by interference; the environmental interference feature library contains feature parameters of four common interference events: nursing staff turning over, feeding collisions, environmental noise, and slight vibration of equipment. The signals generated by these four types of events are easily confused with signals related to accidental swallowing. The characteristic parameters of each type of interference event were determined through prior pre-collection: various interference events were simulated in nursing scenarios, three types of signals were collected and preprocessed simultaneously, and the corresponding characteristic parameters were extracted as basic data in the database; among them, the turning over of nursing staff corresponds to irregular fluctuations in laryngeal vibration, brief fluctuations in breathing, and irregular facial displacement; feeding collision corresponds to instantaneous impact of laryngeal vibration and instantaneous facial displacement; environmental noise corresponds to high-frequency noise of laryngeal vibration and irregular fluctuations in breathing; slight vibration of equipment corresponds to periodic fluctuations of laryngeal vibration and slight and stable fluctuations in breathing. The characteristics of each type are significantly different from those of aspiration signals. The update logic for the environmental interference feature library is to periodically collect feature parameters of newly added interference events and add them to the feature library. The update cycle is determined based on the changes in interference types in the nursing scenario. New interference types may appear in the nursing scenario. Regular updates can ensure that the feature library can cover all common interference events and avoid incomplete filtering due to omission of interference types. At the same time, after each update, the original feature parameters need to be verified to remove outdated or inaccurate feature parameters, ensuring the validity and accuracy of all parameters in the feature library. Step S52: Preset high and low thresholds for matching degree, determined through preliminary testing: Collect a large number of interference events and real suspected event signals, and calculate the lowest value of the matching degree of each type of interference event as the high threshold, and the highest matching degree of the real suspected event with the features in the database as the low threshold, clearly distinguishing between the two types of events; extract three types of signal features within the monitoring interval of each suspected event, match them with each type of interference feature in the database, and calculate the matching degree; if the matching degree is ≥ the high threshold, it is determined to be an interference event and is removed; if the matching degree is ≤ the low threshold, it is determined to be a valid suspected event and is retained; when the matching degree is between the two thresholds, extract the core parameters of the suspected event and perform a secondary matching with the core parameters of the corresponding interference type; if the matching degree of the core parameters is ≥ the high threshold, it is removed, otherwise it is retained, avoiding misjudgment caused by fuzzy matching; Step S53: For the retained valid suspected events, adaptively adapt the feature weights; the three types of signals have different anti-interference capabilities under different interference intensities: the throat vibration signal is easily affected by environmental vibration and noise, while the breathing state and facial movement signals have stronger anti-interference capabilities. Dynamically adjusting the weights can avoid interference effects. The determination of environmental interference intensity is based on the following criteria: real-time extraction of effective signals within the current time period, separation of environmental interference signals from physiological signals, calculation of the proportion of interference signal amplitude to total signal amplitude, and classification of environmental interference intensity levels according to this proportion into two categories: strong interference and quiet state. The classification threshold is determined through prior testing. The feature weights are initially set to balanced weights and remain balanced in the quiet state. When the interference is strong, the weights of breathing state and facial movement features are increased, while the weights of laryngeal vibration features are decreased. The adjustment range is determined through prior testing to ensure recognition accuracy in the interference environment, while avoiding excessive weight adjustment that could lead to feature omission. The weight adaptation is real-time. After each interference filtering is completed, the current interference intensity is determined and the weight is adjusted to adapt to the dynamic changes in the interference environment, ensuring that valid suspected events can be accurately identified in different scenarios.

[0032] In this embodiment, the environmental interference feature library covers common interference types, and the feature parameters are based on sufficient evidence to accurately identify interference events; the matching degree threshold is clearly defined, and secondary matching verification further improves the filtering accuracy; the feature weights are adaptively adapted to different interference scenarios to ensure accurate screening of effective suspected events, providing reliable support for subsequent confirmation of aspiration risk, and meeting the actual needs of nursing scenarios.

[0033] In some embodiments of the present invention, given that retained valid suspected events may still have misjudgments, a single-dimensional verification method cannot ensure the accuracy of the determination of accidental swallowing risk events, and it is necessary to implement differentiated early warning for confirmed accidental swallowing risk events and retain risk data for all time periods for traceability; based on this, this embodiment performs three-dimensional collaborative bidirectional verification on valid suspected events, determines and confirms accidental swallowing risk events and outputs identification results, classifies early warning levels to achieve differentiated responses, and generates a full-time risk time series analysis map, specifically implementing the following steps: Step S61: For the retained valid suspected events, perform three-dimensional collaborative bidirectional verification to confirm the aspiration risk event. The three-dimensional collaborative bidirectional verification uses the individual physiological baseline model and the dynamically iterated associated baseline as references. The verification logic is divided into forward verification and reverse verification. The bidirectional verification corroborates each other to ensure the accuracy of the judgment result. Forward verification involves extracting three types of signal feature parameters within the key monitoring interval of the valid suspected event and comparing them with the preset abnormal range of the associated baseline to verify whether the abnormality of each type of feature parameter conforms to the characteristic pattern of aspiration risk events. Reverse verification involves comparing the three types of signal feature parameters with the normal feature range in the individual physiological baseline model to verify the various feature parameters. The degree of deviation must meet the criteria for determining the risk of accidental swallowing. Only after both bidirectional verifications pass can an event be considered a confirmed risk of accidental swallowing. In the forward verification, all three types of signal characteristic parameters must fall within the abnormal range of the associated baseline, and the abnormal characteristics must be consistent with the characteristic patterns of the risk of accidental swallowing. In the reverse verification, the deviation values ​​of all three types of signal characteristic parameters must reach the preset deviation threshold, and the deviation trend must be consistent with the characteristic changes caused by accidental swallowing. If both conditions are met, the valid suspected event is determined to be a confirmed risk of accidental swallowing. If only one condition is met, it is determined to be a suspected misjudgment event, and environmental interference filtering and weight adaptation must be performed again. If neither condition is met, it is determined to be a non-accidental swallowing event and is removed. During the verification process, bidirectional verification data for each valid suspected event must be recorded, including outliers, deviations, verification results, and judgment criteria for various characteristic parameters. This type of data can be used for subsequent early warning level classification and time series analysis graph generation, and also facilitates subsequent traceability of the verification process to ensure the verifiability of the verification results. Step S62: Based on the degree of abnormality of the confirmed aspiration risk events, divide them into multiple warning levels to achieve differentiated warning responses. The purpose of dividing the warning levels is to adopt different warning response methods according to the severity of the aspiration risk. The division criteria are determined by the aspiration samples collected in advance. Aspiration events of different severity are collected in advance, and the range of characteristic abnormal parameters of various events is statistically analyzed. Based on this, at least three warning levels are divided, such as mild warning, moderate warning, and severe warning. The threshold for dividing the warning levels and the response methods can be adjusted according to the individual differences of disabled persons. Step S63: Generate a time-series analysis map of the risk of aspiration in disabled persons throughout the entire period, and complete the non-intrusive identification of the risk of aspiration in disabled persons. The map is generated by combining time-series curves and event markers. The curves show the changing trends of three types of signal characteristics throughout the entire period. The event markers indicate the event type and warning level at the corresponding time nodes, which can intuitively present the risk of aspiration in disabled persons throughout the entire period, making it easier for nursing staff to trace the pattern of risk changes and assess the swallowing status of disabled persons.

[0034] In this embodiment, three-dimensional collaborative two-way verification ensures accurate risk event judgment, early warning level classification matches the severity of risk, and differentiated response improves nursing efficiency; the all-time time sequence analysis map is intuitive and clear, facilitating traceability and analysis, and is suitable for monitoring scenarios with multiple disabled persons; the entire process achieves seamless identification, fully adapting to the actual needs of disabled persons' care scenarios.

[0035] Based on the same inventive concept as the aspiration risk identification method for disabled persons in the foregoing embodiments, the present invention also provides an aspiration risk identification device for disabled persons, such as... Figure 2 As shown, it includes: The multimodal acquisition module is used to build a multimodal non-sensory acquisition system, which simultaneously acquires the laryngeal vibration timing signal, respiratory status timing data and facial movement timing information of disabled persons, and preprocesses the acquired raw data to extract effective signals. The baseline construction module is used to build an individual physiological baseline model specific to disabled persons based on historical normal physiological signal samples, and to complete the binding of the data collection system with the identity information of disabled persons; The dynamic iteration module is used to segment the real-time acquired effective signal stream, construct a three-dimensional feature coordination matrix of laryngeal vibration, respiratory state and facial movements, calculate the mutual information entropy value of the three-dimensional features, and realize the dynamic iteration of the associated baseline by combining the individual physiological baseline model. The temporal labeling module is used to extract swallowing action anchors and breathing abnormality anchors based on the dynamic iteration results of the associated baseline, align them temporally, mark key monitoring intervals, and screen out suspected risk events of accidental swallowing. The event filtering module is used to filter out environmental interference and adaptively adapt feature weights for suspected risk events, retaining valid suspected events. The event determination module is used to verify valid suspected events, determine and confirm the risk of accidental swallowing, and output the identification results.

[0036] The device described above in this invention can effectively realize the method for identifying the risk of accidental swallowing for disabled persons, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0037] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A mis-swallowing risk identification method for a disabled person, characterized by, include: A multimodal, non-intrusive data acquisition system was established to simultaneously collect the laryngeal vibration timing signal, respiratory status timing data, and facial movement timing information of disabled individuals. The collected raw data was preprocessed to extract effective signals. Based on historical normal physiological signal samples, we constructed an individual physiological baseline model specific to disabled individuals and completed the binding of the data collection system with the identity information of disabled individuals. The effective signal stream acquired in real time is segmented and processed to construct a three-dimensional feature coordination matrix of laryngeal vibration, respiratory state and facial movements. The mutual information entropy value of the three-dimensional features is calculated, and the dynamic iteration of the associated baseline is realized by combining the individual physiological baseline model. Based on the dynamic iteration results of the associated baseline, swallowing action anchor points and breathing abnormality anchor points are extracted and time-aligned, key monitoring intervals are marked, and suspected risk events of accidental swallowing are screened out. Environmental interference filtering and feature weight adaptive adaptation are performed on the suspected risk events to retain valid suspected events; The valid suspected events are verified to determine and confirm the risk of accidental swallowing and output the identification results.

2. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The multimodal non-contact acquisition system includes a non-contact throat vibration sensor, a non-contact breathing sensor, and a high-definition camera.

3. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The preprocessing includes: The original timing signal of throat vibration is filtered and denoised to remove environmental vibration and transmission noise. Baseline calibration and interference removal were performed on the raw time series data of respiratory status to remove baseline drift and non-respiratory pressure interference. The raw image data of facial movement time sequence is processed by grayscale conversion, noise reduction and region of interest extraction to focus on the facial areas related to accidental swallowing.

4. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The construction of an individual physiological baseline model specific to disabled individuals includes: Based on continuous signal acquisition for no less than a preset duration, signal samples corresponding to normal physiological states are selected. The vibration frequency, amplitude, and duration parameters of the laryngeal vibration signal, the respiratory frequency, depth, and rhythm parameters of the respiratory state signal, and the mouth closure angle, mandibular position, and laryngeal movement amplitude parameters of the facial movement signal were extracted respectively. Weights are assigned based on the correlation between the three types of signals and the risk of aspiration, and the weighted fused feature parameters and normal fluctuation range are used to form an individual physiological baseline model.

5. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The screening for suspected accidental swallowing risk events includes: The swallowing action anchor point is located based on the peak value of laryngeal vibration, and the breathing abnormality anchor point is located based on respiratory mutation. The two types of anchor points are time-series aligned, and if the time interval is within a preset threshold, they are determined to be associated anchor points. Key monitoring intervals are defined by associated anchor points, and suspected risk events with abnormal three-dimensional features within the intervals are screened out.

6. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The environmental interference filtering includes: Construct an environmental disturbance feature library containing characteristics of common nursing disturbance events; Calculate the matching degree between the features of suspected risk events and the feature library, remove interfering events based on the matching degree threshold, and retain valid suspected events.

7. The method for identifying the risk of aspiration in disabled persons according to claim 6, characterized in that, The adaptive adaptation of feature weights includes: Real-time assessment of environmental disturbance intensity levels; When interference is strong, reduce the weight of laryngeal vibration features and increase the weight of respiratory status and facial movement features.

8. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The verification includes forward verification and reverse verification. Only when both bidirectional verifications pass is it determined to be a confirmed risk event of accidental swallowing.

9. The method for identifying the risk of aspiration in disabled persons according to claim 1, characterized in that, The output recognition results include: Early warning levels are classified according to the degree of anomalies in the characteristics, and differentiated early warning responses are implemented accordingly; Generate a time-series analysis map that includes full-time signal trends and risk event markers.

10. A device for identifying the risk of aspiration in disabled persons, characterized in that, include: The multimodal acquisition module is used to build a multimodal non-sensory acquisition system, which simultaneously acquires the laryngeal vibration timing signal, respiratory status timing data and facial movement timing information of disabled persons, and preprocesses the acquired raw data to extract effective signals. The baseline construction module is used to build an individual physiological baseline model specific to disabled persons based on historical normal physiological signal samples, and to complete the binding of the data collection system with the identity information of disabled persons; The dynamic iteration module is used to segment the real-time acquired effective signal stream, construct a three-dimensional feature coordination matrix of laryngeal vibration, respiratory state and facial movements, calculate the mutual information entropy value of the three-dimensional features, and realize the dynamic iteration of the associated baseline in combination with the individual physiological baseline model. The temporal marking module is used to extract swallowing action anchor points and breathing abnormality anchor points based on the dynamic iteration results of the associated baseline, perform temporal alignment, mark key monitoring intervals, and screen out suspected risk events of accidental swallowing. The event filtering module is used to filter out environmental interference and adaptively match feature weights for the suspected risk events, and retain valid suspected events. The event determination module is used to verify the valid suspected events, determine and confirm the risk of accidental swallowing, and output the identification results.

Images