Method and system for assessing swallowing function of a psychiatric patient based on multi-dimensional physiological signals

Abstract

This application relates to the field of physiological function assessment technology, and in particular to a method and system for assessing swallowing function in psychiatric patients based on multidimensional physiological signals. The method includes: S1, acquiring basic information and medication information of the psychiatric patient to be assessed, and collecting multidimensional physiological signals; S2, extracting swallowing feature vectors based on these signals; S3, identifying behavioral risks in psychiatry based on swallowing event sequences and inertial measurement signals, obtaining behavioral risk labels and their confidence levels; S4, extracting extrapyramidal response representation vectors based on multidimensional physiological signals, and outputting the probability of swallowing difficulties related to extrapyramidal responses through a pre-set classification model; S5, mapping different abnormal indicators to item scores for nursing risk assessment items, obtaining a total score; S6, generating an assessment report containing item scores, total score, risk level, and key physiological evidence. This application solves the problem that existing assessments rely on subjective observation and lack objective evidence support.

Description

Technical Field

[0001] This application relates to the field of physiological function assessment technology, and in particular to a method and system for assessing the swallowing function of mental patients based on multidimensional physiological signals. Background Technology

[0002] Psychiatric patients are at risk of swallowing coordination disorders due to factors such as antipsychotic medications, abnormal emotional and behavioral states, and poor eating habits. In clinical practice, healthcare professionals use an itemized scale to assess and manage patients' swallowing safety. This scale includes multiple assessment items such as swallowing difficulties caused by drug-induced extrapyramidal reactions, overeating, extreme excitement, food grabbing, eating too quickly, eating while lying down, and diminished swallowing reflexes such as choking. By statistically analyzing the scores of each item and classifying the risk into three levels—Level I, Level II, and Level III—according to preset thresholds, it provides a reference for clinical nursing. This assessment method has been widely used in psychiatric clinical nursing and has become an important tool for assessing patients' swallowing safety risks.

[0003] However, in the current technology, there is a lack of objective quantitative means to assess the swallowing function of psychiatric patients, making it difficult to conduct continuous monitoring and risk assessment when patients are agitated or uncooperative. Summary of the Invention

[0004] This application provides a method and system for assessing the swallowing function of mental patients based on multidimensional physiological signals, in order to solve the above-mentioned problems.

[0005] In a first aspect, this application provides a method for assessing the swallowing function of mental patients based on multidimensional physiological signals. The method includes: S1, acquiring basic information and medication information of the mental patient to be assessed, and deploying and simultaneously collecting multidimensional physiological signals for the mental patient, including surface electromyography signals of the anterior cervical region, cervical inertial measurement signals, and swallowing acoustic signals; S2, detecting swallowing events based on the multidimensional physiological signals to obtain a swallowing event sequence, and extracting a swallowing feature vector for each swallowing event; S3, identifying psychiatric behavioral risks based on the swallowing event sequence and the inertial measurement signals to obtain behavioral risk labels and their confidence levels; 4. Extract the extrapyramidal response characterization vector based on the multidimensional physiological signals, and input the extrapyramidal response characterization vector and the swallowing feature vector into a preset classification model to output the probability of swallowing difficulties related to extrapyramidal response; S5. Map the behavioral risk label, the swallowing difficulties related to extrapyramidal response, and the swallowing reflex abnormality index corresponding to the swallowing feature vector to the item scores of nursing risk assessment items, and sum the item scores to obtain the total score; S6. Output the aspiration / choking risk level according to the total score and the preset grading threshold, and generate an assessment report containing the item scores, the total score, the risk level, and key physiological evidence.

[0006] Through the above technical solution, a multidimensional physiological evidence chain is constructed by simultaneously collecting surface electromyography signals, neck inertial measurement signals, and swallowing acoustic signals in the anterior cervical region. Objective characterization of swallowing movements is achieved through swallowing event detection and feature extraction. Coupled discrimination of typical behavioral risks in psychiatry is completed based on swallowing rhythm analysis and body position recognition. Specific identification of drug-induced neuromuscular disorders is achieved by combining extrapyramidal reaction characterization and swallowing feature joint modeling. Finally, various risk indicators are mapped to clinically compatible item scores, forming a complete closed loop from raw signals → physiological characteristics → behavioral labels → risk probability → nursing items → graded reporting. This closed loop not only solves the problem of existing assessments relying on subjective observation and lacking objective evidence support, but also achieves the mechanism separation and synergistic quantification of EPS-related swallowing difficulties and behavioral risks, making aspiration / choking risk assessment repeatable, interpretable, and clinically operable.

[0007] Optionally, the swallowing event detection employs a multimodal consistency confirmation mechanism. This mechanism includes: if, within a preset time tolerance, the surface electromyography (SEMG) envelope peak appears simultaneously, the laryngeal elevation characteristic peak of the neck inertial measurement signal appears simultaneously, and the swallowing acoustic energy exceeds a threshold, then the swallowing event is confirmed as valid. The swallowing event detection is achieved through two-channel signal consistency confirmation. This confirmation includes: the SEMG envelope peak and the laryngeal elevation peak of the neck inertial measurement signal appear simultaneously within a preset time tolerance, and the swallowing acoustic signal meets the swallowing sound energy threshold within the same time window, thus confirming the swallowing event. The preset time tolerance is 10ms to 80ms, and the multidimensional physiological signals undergo synchronous calibration before entering the swallowing event detection process, ensuring that the multimodal time alignment error is no greater than 20ms.

[0008] Through the above technical solution, by confirming the multimodal consistency of surface electromyography envelope peak, laryngeal elevation characteristic peak of neck inertial measurement signal, and swallowing acoustic energy within a preset time tolerance, combined with a strict synchronous calibration mechanism, high specificity identification of real swallowing events in psychiatric patients is achieved. With the flexible confirmation strategy of two-way signal dominance and third-way corroboration, the robustness of detection is ensured while effectively suppressing interference from non-swallowing acoustic artifacts such as verbal agitation and shouting. The ≤20ms time alignment accuracy ensured by synchronous calibration provides the prerequisite for time coupling analysis of multimodal response, making the swallowing event sequence have a repeatable and traceable physiological basis, thereby supporting the accurate extraction of swallowing feature vectors in step S2, as well as risk identification and quantitative assessment in subsequent steps S3 to S6.

[0009] Optionally, the swallowing feature vector includes: swallowing initiation delay, peak amplitude, duration and integral area obtained based on surface electromyography signals; laryngeal elevation amplitude, elevation speed and number of swallowing episodes obtained based on neck inertial measurement signals; and the swallowing feature vector further includes statistical features of swallowing event intervals to characterize the eating rhythm.

[0010] Through the above technical solution, the neuromuscular activation characteristics represented by surface electromyography signals, the pharyngeal motor biomechanics characteristics represented by neck inertial measurement signals, and the dynamic characteristics of eating rhythm represented by swallowing event intervals are structurally fused to form a swallowing feature vector covering three dimensions: single swallow quality, pharyngeal motor efficiency, and eating behavior pattern. This vector not only supports the swallowing event feature extraction in step S2 of this application, but also provides a calculable, distinguishable, and traceable underlying feature foundation for the identification of eating too fast / grabbing, the correlation analysis of eating in a supine position, and the coupled modeling of EPS and swallowing disorders in this application. In particular, by introducing the statistical characteristics of swallowing event intervals, the limitations of traditional swallowing assessment focusing on single events are overcome, enabling the system to objectively identify typical rhythmic eating abnormalities in psychiatric patients, thereby improving the clinical relevance and early warning capability of aspiration / choking risk assessment.

[0011] Optionally, the psychiatric behavioral risk identification includes identifying eating too quickly / grabbing food. This identification includes: calculating the mean and coefficient of variation of a swallowing event interval sequence within a preset time window; outputting a risk index for eating too quickly / grabbing food when the mean is less than a first threshold or the coefficient of variation is greater than a second threshold; wherein the preset time window is 30s to 180s; the identification of eating too quickly / grabbing food further incorporates a chewing adequacy constraint, which includes: calculating the number of chews, chewing energy, or chewing-to-swallowing ratio based on chewing-related electromyography or oromandibular movement signals; increasing the confidence level of the risk index for eating too quickly / grabbing food when the chewing-to-swallowing ratio is lower than a preset threshold and simultaneously meets the mean or coefficient of variation criteria.

[0012] By combining the statistical characteristics (mean and coefficient of variation) of swallowing event intervals with the chewing-swallowing ratio, a physiological coordination indicator, the above-mentioned technical solution enables the stratified identification of typical impulsive eating behaviors in psychiatry. The mean and coefficient of variation jointly characterize the abnormality of swallowing rhythm, constituting the basis for initial risk screening. The chewing-swallowing ratio serves as a constraint on the adequacy of chewing, excluding physiological rapid eating due to accelerated metabolism and high energy consumption, and retaining only the truly risky wolfing-down behavior. The synergistic effect of the two makes the behavioral risk identification both sensitive (capturing rhythmic mutations) and specific (filtering out non-dangerous fast eating), thereby supporting the feasibility and clinical credibility of the behavioral risk identification function in psychiatry required by step S3 of this application.

[0013] Optionally, the psychiatric behavioral risk identification also includes supine eating identification, which includes: calculating the trunk tilt angle from the trunk inertial measurement signal and classifying the body position; when the trunk tilt angle is within a preset supine angle range and overlaps with the swallowing event time window, outputting a supine eating risk index; wherein the supine angle range is 30° to 80°.

[0014] By using the above technical solution, trunk inertial measurement signals are analyzed into clinically significant trunk tilt angles, and their strict overlap with swallowing events in the time dimension is defined, achieving imperceptible, objective, and verifiable identification of supine eating, a typical psychiatric maladaptive eating behavior. This identification does not rely on video monitoring or manual observation, protecting patient privacy. The set angle range of 30° to 80° is consistent with the physiological safety threshold for swallowing, ensuring that the risk assessment criteria are medically reasonable. This technical feature, together with the swallowing event sequence, inertial measurement signal acquisition, and behavioral risk identification framework defined in this application, forms an organic synergy, jointly supporting the complete generation of psychiatric behavioral risk labels in step S3.

[0015] Optionally, the extrapyramidal response characterization vector includes: 4Hz–7Hz tremor spectrum peak energy characteristics obtained from the spectrum of inertial measurement signals; baseline elevation characteristics and co-contraction ratio characteristics obtained from surface electromyography signals; motor initiation delay characteristics obtained from motion-related signals; and the classification model outputs the probability of swallowing difficulties related to extrapyramidal response, which is used to form an EPS-related swallowing risk index.

[0016] By integrating the tremor peak energy characteristics from inertial measurement signals, baseline elevation and co-contraction ratio characteristics from surface electromyography signals, and motor initiation delay characteristics from movement-related signals, a multidimensional physiological representation vector covering the three core clinical manifestations of EPS (resting tremor, rigidity, and bradykinesia) is constructed. Through joint modeling of this vector and the swallowing feature vector, the classification model can distinguish between drug-induced swallowing disorders and risks caused by behavioral feeding or reflex abnormalities at the physiological mechanism level. This allows for a stable output of EPS-related swallowing difficulties in step S4, providing an interpretable, verifiable, and traceable objective basis for subsequent item mapping and comprehensive risk grading.

[0017] Optionally, the classification model is a classification model obtained based on supervised learning training. The input of the classification model includes a concatenated vector of the extrapyramidal response representation vector and the swallowing feature vector. During the training phase, information on changes in drug dosage is introduced as a conditional feature or sample stratification feature to improve the stability of identifying drug-induced EPS-related swallowing disorders.

[0018] By incorporating information on changes in drug dosage as a conditional feature or a basis for sample stratification into the training process of the subtyping model, the model not only relies on static physiological characteristics but can also perceive the dynamic context of drug intervention. Based on this, by combining the joint representation of extrapyramidal reaction representation vectors and swallowing feature vectors, personalized and time-sensitive identification of drug-induced EPS-related swallowing disorders is achieved. This supports the stable output of the probability of swallowing difficulties related to extrapyramidal reactions in S4, providing a reliable, interpretable, and reproducible quantitative basis for subsequent nursing risk assessment items (such as EPS-related swallowing risk indicators).

[0019] Optionally, the swallowing reflex abnormality index is obtained from airway protection abnormality detection, which includes any of the following: detection of cough event clusters after a swallowing event; detection of abnormal respiratory phase switching after a swallowing event; detection of a decrease in blood oxygen saturation exceeding a threshold after a swallowing event; detection of wet sound / residual features in swallowing acoustic signals exceeding a threshold; and fusing the detection results to form a reflex / airway protection insufficiency risk index; in the swallowing event detection and / or the airway protection abnormality detection, anti-artifact processing is performed on artifacts caused by psychiatric patients' speaking, shouting, or agitation, which includes: when an acoustic event meets the characteristics of speech bands but lacks swallowing consistency characteristics of surface electromyography and inertial measurement, suppressing the contribution of the acoustic event to the swallowing event sequence and risk index.

[0020] Through the above technical solution, a multidimensional airway protection abnormality criterion system is constructed by using four complementary pathways: cough event cluster detection, respiratory phase switching abnormality detection, blood oxygen saturation decrease detection, and wet sound / residual feature detection. With the help of anti-artifact processing mechanism, based on the physiological principle that true swallowing must have multimodal consistency, common interferences in psychiatry such as speech and shouting are effectively eliminated. Finally, the results of the four tests are integrated to generate a continuous reflex / airway protection insufficiency risk index, realizing an objective, quantitative, and interference-resistant mapping of swallowing reflex reduction / choking, significantly improving the reliability and clinical applicability of swallowing function assessment in psychiatric patients.

[0021] Optionally, the item-level output corresponding to the nursing risk assessment item is an item risk indicator rather than a fixed score. The item risk indicators include EPS-related swallowing risk indicators, behavioral feeding risk indicators, supine eating risk indicators, and reflex / airway protection inadequate risk indicators. Furthermore, the total score or comprehensive risk value is a comprehensive risk value obtained by normalizing and fusing the item risk indicators. The fusing includes any one of weighted summation, logistic regression fusing, or Bayesian fusing, wherein the weights or fusing parameters are learned from training data or adaptively updated from individual baselines. The comprehensive risk value is used to output the aspiration / choking risk level, which includes Level I, Level II, and Level III. The assessment report further outputs the dominant risk factors that lead to the risk level. The dominant risk factors are determined by ranking the sensitivity contribution or confidence contribution of each item risk indicator.

[0022] By upgrading the item-level output from discrete fixed scores to continuous item-based risk indicators, and combining this with a learnable and adaptive normalization fusion mechanism, a technological leap from static scoring to dynamic risk modeling is achieved. Different approaches, such as weighted summation, logistic regression, or Bayesian fusion, ensure a balance between the general applicability of the comprehensive risk value and individual specificity. Furthermore, through sensitivity- or confidence-driven identification of dominant risk factors, the assessment results not only have the ability to output risk levels but also the function of etiological tracing and intervention guidance. Ultimately, this solution significantly improves the objectivity, sensitivity, and clinical operability of risk identification, providing repeatable, verifiable, and traceable technical support for psychiatric swallowing safety management.

[0023] Secondly, this application provides a swallowing function assessment system for mental patients based on multidimensional physiological signals. The system includes: a data collection module for acquiring basic information and medication information of the mental patient to be assessed, and deploying and synchronously collecting multidimensional physiological signals for the patient, including surface electromyography signals in the anterior cervical region, cervical inertial measurement signals, and swallowing acoustic signals; a vector extraction module for detecting swallowing events based on the multidimensional physiological signals, obtaining a swallowing event sequence, and extracting a swallowing feature vector for each swallowing event; and a risk identification module for identifying psychiatric behavioral risks based on the swallowing event sequence and the inertial measurement signals, obtaining behavioral risk labels and their confidence levels; and swallowing... The analysis module is used to extract extrapyramidal response characterization vectors based on the multidimensional physiological signals, and input the extrapyramidal response characterization vectors and swallowing feature vectors into a preset classification model to output the probability of swallowing difficulties related to extrapyramidal responses. The swallowing abnormality mapping module is used to map the behavioral risk labels, the swallowing difficulties related to extrapyramidal responses, and the swallowing reflex abnormality indicators corresponding to the swallowing feature vectors into item scores for nursing risk assessment items, and sum the item scores to obtain a total score. The swallowing assessment module is used to output the aspiration / choking risk level based on the total score and a preset grading threshold, and generate an assessment report containing the item scores, the total score, the risk level, and key physiological evidence. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a method for assessing the swallowing function of mental patients based on multidimensional physiological signals, provided as an embodiment of this application;

[0026] Figure 2 This is a structural diagram of a swallowing function assessment system for mental patients based on multidimensional physiological signals, provided in one embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0028] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0029] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0030] Example 1: Figure 1 A flowchart illustrating a method for assessing swallowing function in mental patients based on multidimensional physiological signals, as provided in one embodiment of this application, is shown below. Figure 1 As shown, the method includes:

[0031] S1: Obtain basic information and medication information of the mental patients to be evaluated, and set up and collect multidimensional physiological signals for the mental patients simultaneously. The multidimensional physiological signals include surface electromyography signals of the anterior cervical region, neck inertial measurement signals, and swallowing acoustic signals.

[0032] The basic information may include at least one of the following: patient age, gender, history of choking, history of comorbid neuromuscular diseases, and cognitive function rating; medication information may include the type, dosage, frequency of administration, and medication adjustment records for the past 7 days of the currently used antipsychotic drugs; surface electromyography (sEMG) in the anterior cervical region is a muscle electrical activity signal acquired by bipolar electrodes attached to the suprahyoid or thyrohyoid muscle regions, used to characterize the timing of swallowing initiation and the intensity of muscle mobilization; inertial measurement unit (IMU) in the neck is a three-dimensional acceleration and angular velocity signal acquired by a miniature inertial sensor worn at the level of the cricoid cartilage, used to characterize the trajectory of laryngeal movement, elevation amplitude, and changes in body position; swallowing acoustic signal (SAS) is an acoustic response induced by pharyngeal vibration and airflow disturbance, acquired by a bone conduction microphone or attached acoustic sensor, used to reflect the energy distribution and temporal structure of the swallowing action.

[0033] This application, for example, automatically retrieves basic and medication information based on patient wristband scanning or electronic medical record systems; it also involves registering medication information by scanning drug packaging barcodes with a mobile terminal and combining this with voice input; furthermore, it integrates with hospital HIS systems to synchronize structured medication data in real time. This application obtains complete and traceable basic and medication information based on any of the above methods, serving as the data foundation for EPS risk prior modeling and individualized threshold setting.

[0034] For example, before the assessment begins, the nurse uses a handheld terminal to read the patient's wristband RFID information, automatically loading the patient's age, history of choking, and current medication list from their electronic medical record. Subsequently, after cleaning the patient's skin in the anterior neck area, an sEMG electrode (placed 1 cm above the hyoid bone), an IMU sensor (fixed directly in front of the cricoid cartilage), and a bone conduction microphone (closely attached to the lateral side of the thyroid cartilage) are attached in sequence. All sensors are connected to the central acquisition unit via a wireless module, and a timestamp synchronization protocol is initiated to ensure that the time alignment error of the multi-source signals is no more than 20 ms. During the acquisition process, the patient drinks 50 mL of water according to the standard procedure, and the system enters passive monitoring mode, requiring no active cooperation or instruction feedback throughout the process.

[0035] S2: Swallowing event detection is performed based on multidimensional physiological signals to obtain a swallowing event sequence, and swallowing feature vectors are extracted for each swallowing event;

[0036] Here, the swallowing event sequence can refer to the set of discrete time points arranged in chronological order, each confirmed as a valid swallowing action, denoted as . Swallowing Feature Vector (SFV) is a feature vector for the first digit of a digit. A swallowing event An ordered array of quantifiable physiological parameters is extracted and used to characterize the quality, timing characteristics, and coordination of the swallowing action.

[0037] This application confirms swallowing events, for example, based on the joint satisfaction relationship of three signals—sEMG envelope peak, IMU laryngeal elevation characteristic peak, and SAS energy threshold—within a preset time tolerance window. Another application employs a two-signal consistency confirmation mechanism, whereby the sEMG envelope peak and the IMU laryngeal elevation peak appear synchronously within the time tolerance, and the SAS meets the swallowing sound energy threshold within the same time window. Furthermore, this application uses a deep learning model to perform end-to-end event localization on the original multidimensional signal, outputting the start and end times of swallowing and confidence levels. This application obtains highly robust swallowing event sequences based on any of the above methods, supporting subsequent refined feature extraction.

[0038] For example, the synchronously acquired sEMG signal is bandpass filtered (10–500Hz) and rectified-low-pass envelope extracted to obtain sEMG_envelope(t); the vertical acceleration integral of the larynx is calculated from the IMU signal to identify the peak moment of larynx elevation. Perform a short-time Fourier transform on the SAS signal to extract the average energy in the 200–1500 Hz frequency band. When this energy continuously exceeds a dynamic threshold for 300 ms, it is marked as the acoustically active window. (peak time of sEMG envelope) , (When the swallowing sound energy threshold is reached) If the time difference between any two of the three is ≤50ms, then a swallowing event is confirmed. For each confirmed event, a 500ms signal window is extracted before and after the event, and the following data are extracted: sEMG start-up delay, peak amplitude, duration and integral area; IMU laryngeal elevation amplitude, elevation speed and single swallow count; SAS swallowing sound energy, duration and wet sound characteristics, forming a swallowing feature vector with a dimension of 12.

[0039] S3: Identify behavioral risks in psychiatry based on swallowing event sequences and inertial measurement signals, and obtain behavioral risk labels and their confidence levels;

[0040] Among them, behavioral risk identification in psychiatry can refer to the identification of typical high-risk behavioral patterns exhibited by psychiatric patients during eating by using swallowing rhythm, changes in body position and auxiliary physiological signals; behavioral risk labels are one or more of the following: eating too fast / grabbing food, eating in a supine position, and extreme excitement accompanied by swallowing disorders; confidence score (CS) is a normalized index of the probability value or rule criterion satisfied by the model output, used to characterize the reliability of the risk judgment.

[0041] This application identifies abnormal eating rhythms, for example, by combining the statistical distribution characteristics of swallowing event interval sequences with a preset time window to calculate the mean and coefficient of variation; it also determines whether supine eating behavior exists by analyzing the trend of trunk IMU tilt changes and the overlap of swallowing event time windows; furthermore, this application integrates decreased heart rate variability (HRV) and increased electrical skin activity (EDA) as auxiliary criteria for extreme arousal states, improving the specificity of risk identification. Based on any of the above methods, this application obtains clinically interpretable behavioral risk labels and corresponding confidence levels to support graded nursing responses.

[0042] For example, a sequence of swallowing events Based on this, calculate the interval between adjacent events. Within a sliding time window of 60 seconds, statistics were collected. mean With coefficient of variation ;when s or When the risk label of eating too fast / grabbing food is triggered, the confidence level is set to [value missing]. Simultaneously, the trunk tilt angle is calculated using the chest IMU. ,like If the overlap with any swallowing event time window lasts for ≥2 seconds, then the risk label for eating in a supine position is triggered, with a confidence level set to [value missing]. The two labels are output independently and do not inhibit each other.

[0043] S4: Extract extrapyramidal response representation vectors based on multidimensional physiological signals, and input the extrapyramidal response representation vectors and swallowing feature vectors into a preset classification model to output the probability of swallowing difficulties related to extrapyramidal responses;

[0044] The Extrapyramidal Symptom Representation Vector (ESRV) is a set of quantitative indicators reflecting typical physiological manifestations of EPS, including resting tremor peak energy, abnormal muscle tone-related sEMG baseline elevation and co-contraction ratio, and bradykinesia-related movement initiation delay. The EPS-Related Dysphagia Classification Model (EDCM) is a classifier trained based on supervised learning. Its input is a concatenated vector of ESRV and SFV, and its output is an estimate of the probability of EPS-related dysphagia. The EPS-Associated Dysphagia Probability (EDP) is a scalar value output by the model, used to quantify the degree of influence of drug-induced neuromuscular control disorders on swallowing function.

[0045] This application characterizes the intensity of resting tremor, for example, based on the peak power spectral density energy of the neck IMU signal in the 4–7 Hz frequency band; it characterizes the degree of muscle rigidity, for example, based on the increase ratio of the sEMG signal's resting baseline amplitude relative to a healthy reference value, and the overlap of sEMG activation during the swallowing preparation and execution phases, calculating the co-contraction ratio; further, this application characterizes the degree of bradykinesia based on the difference between the sEMG initiation delay and the IMU laryngeal movement initiation delay. This application obtains the ESRV based on any of the above methods, splices it with the SFV, and inputs it into the EDCM to achieve objective classification and discrimination of EPS-related swallowing difficulties.

[0046] For example, power spectrum analysis was performed on the neck IMU signal to extract the maximum spectral peak energy in the 4–7 Hz frequency band. ; Calculate the root mean square amplitude of sEMG at rest. and compared with reference values ​​for healthy people of the same age. The comparison yields the baseline elevation ratio. Calculate the sEMG activation energy ratio during the swallowing preparation phase (t–500ms to t–100ms) and the execution phase (t–100ms to t+400ms), and define it as the co-contraction ratio. ;Will , , The SFV obtained above is concatenated to form a 15-dimensional vector, which is then input into a lightweight random forest model trained with clinical data, and the EDP value is output. When EDP ≥ 0.65, it is determined to be EPS-related dysphagia.

[0047] S5: Map the behavioral risk labels, extrapyramidal reaction-related dysphagia, and swallowing reflex abnormality indicators corresponding to the swallowing feature vectors to the item scores of the nursing risk assessment items, and sum the item scores to obtain the total score;

[0048] Among them, the Swallowing Reflex Abnormality Index (SRAI) is a quantifiable physiological evidence derived from impaired airway protection, including at least one of the following: post-swallowing cough clusters, abnormal respiratory phase switching, decrease in blood oxygen saturation, and wet / residual swallowing acoustic features; nursing risk assessment items are independent items with clear clinical significance in the current psychiatric swallowing risk scale, including EPS-related swallowing difficulties, behavioral feeding / eating too quickly, eating in a supine position, and decreased swallowing reflex / choking; item score (IS) is a deterministic score assigned based on whether objective evidence meets the pre-set criteria, or a continuous score obtained by normalizing the risk indicators.

[0049] This application, for example, maps EDP values ​​to discrete scores using a threshold criterion: EDP ≥ 0.65 → 4 points; this application, for example, linearly maps the confidence level of behavioral risk labels to continuous scores within the range of 0–3 points; furthermore, this application maps multiple test results in SRAI to a unified score after logical fusion (such as OR relationship triggering). This application completes item-level objective scoring based on any of the above methods, ensuring semantic consistency and numerical compatibility with clinical scales.

[0050] For example, in this application: when EDP≥0.65, Item_3=4 points are output; when the confidence level of the eating too fast / grabbing label CS_fast≥0.7, Item_7=3 points are output; when the confidence level of the lying-down eating label CS_bed≥0.6, Item_8=3 points are output; when any sub-item in SRAI is true (such as cough cluster C_cough≥2 times / 30s, or ΔSpO2≥3%, or wet sound feature R_wet≥threshold), Item_10=3 points are output; the remaining items (such as age≥70 years, history of choking) are directly assigned values ​​by the above information; the final total score Score=ΣIS_i, with an upper limit of 22 points.

[0051] S6: Output the risk level of aspiration / choking based on the total score and the preset grading threshold, and generate an assessment report that includes item scores, total score, risk level and key physiological evidence;

[0052] The Aspiration / Choking Risk Level (ACRL) is a three-tiered classification system fully aligned with current clinical practice, including Level I (low risk), Level II (medium risk), and Level III (high risk). The Predefined Grading Threshold (PGT) is an empirical cutoff point set based on clinical validation data. Key Physiological Evidence (KPE) consists of raw signal fragments, characteristic values, or visualizations that support the score determination for each item, including swallowing event timestamps, laryngeal elevation amplitude, supine angle, EDP value, and ΔSpO2 decrease curve.

[0053] This application compares the total score with a preset threshold, for example: Score ∈ [1,6] → ACRL = Level I; Score ∈ [7,14] → ACRL = Level II; Score ≥ 15 → ACRL = Level III. This application also introduces a dynamic threshold mechanism to adaptively shift the PGT based on the patient's age and underlying disease status. Furthermore, this application ranks the scores based on their sensitivity contribution, identifies dominant risk factors, and highlights them in the report. This application generates a structured, traceable, and auditable assessment report based on any of the above methods, seamlessly integrating it into the clinical nursing workflow.

[0054] For example, this application is as follows: After completing all steps, the system automatically generates a PDF report. The first page displays the risk level (e.g., Level II (medium risk)), total score (e.g., 11 points), and dominant risk factors (e.g., behavioral feeding / eating too fast is dominant, EPS-related swallowing difficulties are secondary); the second page lists the detailed scores of each item and corresponding KPE screenshots, such as Item_7 with a note indicating a mean swallowing interval of 2.1s (threshold 3s) and a confidence level of 0.82; the third page includes an overview of the original signals, including synchronous waveforms from sEMG, IMU, and SAS channels, and automatically labeled swallowing event locations; a nursing suggestion link is embedded at the end of the report, which can be clicked to jump to standard operating documents such as dedicated seating and care, and feeding speed observation forms.

[0055] Example 2: In yet another embodiment, this application also provides a multimodal consistency confirmation mechanism for swallowing event detection, including:

[0056] Step 1: If, within a preset time tolerance, the surface electromyography envelope peak appears, the laryngeal elevation characteristic peak of the neck inertial measurement signal appears, and the swallowing acoustic energy exceeds the threshold, then it is confirmed as a valid swallowing event.

[0057] Among them, the surface electromyography (SEMG) envelope peak can refer to the local maximum point on the energy envelope curve obtained after rectification, low-pass filtering, and envelope extraction of the original anterior cervical SEMG signal. It reflects the contraction intensity and initiation timing of the suprahyoid or thyrohyoid muscle groups during the swallowing initiation phase. The laryngeal elevation characteristic peak of the cervical inertial measurement signal can refer to the extreme point of the rising segment in the vertical displacement trajectory of the laryngeal protuberance calculated based on the triaxial acceleration and angular velocity signals of the cervical IMU after coordinate transformation, gravity compensation, and motion integration. It characterizes the amplitude and timing of the active elevation of the larynx during the swallowing phase. Swallowing acoustic energy exceeding a threshold (can...) This refers to calculating the short-time energy of acoustic signals acquired by bone conduction / attached microphones within the swallowing characteristic frequency band (e.g., 500Hz–2kHz), and determining effective activation of swallowing sounds when this energy continuously exceeds an adaptively set dynamic threshold (e.g., based on the average energy of the previous 3-second resting period plus twice the standard deviation); the preset time tolerance is the maximum permissible time offset used to determine whether the multimodal signal response has physiological consistency, with a value range of 10ms to 80ms. This range covers the neuromuscular conduction delay and motor coordination variations that may occur in healthy adults and mental patients under the influence of drugs;

[0058] This application, for example, can use the surface electromyography (sEMG) envelope peak as the swallowing initiation trigger source, and expand a preset time tolerance window forward and backward with the peak time as the center. Within this window, it searches for the IMU laryngeal elevation feature peak and the swallowing acoustic energy exceeding the limit range, and determines whether the three overlap. Alternatively, this application can use the cervical IMU laryngeal elevation feature peak as the main temporal anchor point, and verify whether the sEMG envelope peak and acoustic energy are synchronously activated within preset time tolerances before and after it. Furthermore, this application can use the swallowing acoustic energy exceeding the limit range as the first response window, and reversely search within this range to see if the sEMG envelope peak and the IMU laryngeal elevation peak fall within the same time tolerance constraint range. Based on any of the above methods, this application obtains high-confidence recognition results for real swallowing events, thereby suppressing acoustic artifact interference in non-swallowing contexts.

[0059] For example, in this application, during a single meal, the system can acquire a segment of anterior cervical sEMG signal in real time. After envelope extraction, an envelope peak with an amplitude of 1.8mV at t=2.341s is detected. Simultaneously, the neck IMU signal acquired is processed by motion analysis, and a characteristic peak with a laryngeal elevation amplitude of 4.2mm is identified at t=2.345s. At the same time, the acoustic sensor detects that the energy in the 500Hz–2kHz frequency band is continuously higher than the dynamic threshold by 12dB within a time window of t=2.343s±15ms. The time deviation of the three is within 5ms, which is less than the preset time tolerance lower limit of 10ms. Therefore, they are jointly confirmed as a valid swallowing event. Another event at t=5.112s, where only the acoustic energy exceeds the limit but there is no corresponding sEMG envelope peak and IMU elevation peak, is excluded.

[0060] Step 2: Swallowing event detection is achieved through consistency confirmation of two signals. Consistency confirmation includes: the peak value of surface electromyography envelope and the peak value of laryngeal elevation of neck inertial measurement signal appear simultaneously within a preset time tolerance, and the swallowing acoustic signal meets the swallowing sound energy threshold within the same time window to confirm the swallowing event.

[0061] Among them, the consistency confirmation of two signals can mean that it is not mandatory for all three modes to be satisfied, but rather that any two signals can reach a consistent response within the time tolerance to constitute a preliminary confirmation condition. The third signal participates in the confidence weighting as an enhanced verification item. This mechanism takes into account both detection robustness and clinical applicability. For example, when the patient wears the device loosely, causing the IMU signal quality to decline, or when the signal-to-noise ratio of sEMG is reduced due to sweat interference, the basic discrimination ability can still be maintained through the other two reliable signals. The same time window can refer to a unified judgment window defined according to a preset time tolerance, centered on the peak time of any confirmed signal. All signals participating in the confirmation must complete their respective characteristic responses within this window to ensure that the time logic is unified rather than independently judged.

[0062] For example, this application can output a bimodal confirmed + acoustically corroborated swallowing event label and assign a high confidence score when the time difference between the sEMG envelope peak and the IMU laryngeal elevation peak is ≤80ms, and the swallowing acoustic energy within both time windows meets the threshold. Alternatively, this application can output an EMG-acoustic dominated + motor corroborated swallowing event label when the deviation between the sEMG envelope peak and the center time of the acoustic energy excess interval is ≤60ms, and the rate of change of the IMU laryngeal elevation amplitude within this interval exceeds a preset gradient threshold. Further, this application can output a motor-acoustic dominated + EMG corroborated swallowing event label when the overlap between the IMU laryngeal elevation peak and the acoustic energy excess interval is ≥70%, and the sEMG envelope energy slope within this period is greater than 0.5mV / ms. This application obtains a graded confidence output for swallowing events based on any of the above methods, supporting the confidence-weighted fusion of subsequent risk indicators.

[0063] Step 3: The preset time tolerance is 10ms to 80ms, and the multidimensional physiological signals are synchronously calibrated before entering the swallowing event detection, so that the multimodal time alignment error is no more than 20ms.

[0064] The preset time tolerance of 10ms to 80ms is a parameter range set based on the time scale of the swallowing physiological process: 10ms corresponds to the minimum neuromuscular response synchronization deviation in healthy adults under optimal coordination; 80ms covers the cumulative effect of bradykinesia, increased muscle tone, and delayed response caused by dopamine receptor blockade in typical EPS patients. This range avoids both being too narrow, which would lead to the false rejection of normal physiological variations, and being too wide, which would introduce non-physiological coupling noise. Synchronization calibration can refer to the calibration performed during the signal acquisition or preprocessing stage, using hardware trigger pulses (such as TTL synchronization signals) or software timestamps. Alignment (such as PTP protocol or NTP time synchronization) or offline interpolation resampling and other methods are used to normalize the time axes of the three raw data streams of sEMG, IMU and acoustic to the same reference clock, eliminating systematic offsets caused by sampling rate differences, transmission delays, buffer accumulation, etc. The target accuracy index for synchronous calibration is that the multimodal time alignment error is no more than 20ms, which is stricter than the preset time tolerance lower limit (10ms). This ensures that when consistency judgment is made based on the calibrated signal, the response captured within the time tolerance window is indeed from the same swallowing action, rather than a false overlap caused by residual calibration error.

[0065] For example, after sensor deployment, the system sends a nanosecond-precision hardware synchronization pulse to each acquisition module, records the actual response timestamp of each channel to the pulse, constructs a time offset mapping table between channels, and performs real-time interpolation compensation in subsequent data streams. Another example is guiding the patient to perform three standard swallowing actions before each assessment, using the sEMG envelope peak as the gold standard timing reference, calculating the average delay of the IMU and acoustic channels relative to the sEMG, and uniformly subtracting this delay value from all subsequent data. Further, this application can perform sliding time window cross-correlation analysis on the continuously acquired raw signals, estimating the optimal hysteresis between sEMG and IMU, and between sEMG and acoustic channels within each 5-second window, and performing dynamic resampling after denoising the hysteresis sequence using median filtering. Based on any of the above methods, this application obtains multidimensional physiological signals with a time alignment error ≤20ms, providing a reliable timing basis for multimodal consistency confirmation.

[0066] Example 3: In yet another embodiment, this application also provides a swallowing feature vector, including:

[0067] Step 1: Swallowing initiation delay, peak amplitude, duration, and integral area obtained from surface electromyography signals;

[0068] The swallowing initiation delay refers to the time interval from the triggering of swallowing intention (e.g., the end of chewing or the start of liquid flow) to the first time the sEMG signal envelope exceeds a preset baseline threshold. It characterizes the efficiency of neural conduction and motor preparation for the swallowing reflex. In this embodiment, this delay serves as input for subsequent behavioral risk identification and EPS-related swallowing difficulty discrimination; a prolonged delay suggests potential central initiation disorders or peripheral muscle bradykinesia. The peak amplitude refers to the maximum amplitude of the sEMG envelope curve during a single swallow, reflecting the maximum activation intensity of relevant muscle groups during that swallow. In this embodiment, this amplitude is positively correlated with swallowing strength and airway protection ability. The correlation is significant, which may correspond to muscle weakness or drug-induced abnormal muscle tone; the duration can refer to the length of time it takes for the sEMG envelope signal to rise from the initial threshold and fall back below the threshold, used to characterize the ability to maintain muscle contraction; in this embodiment, the shortening of this duration may indicate hasty swallowing or decreased coordination, often co-occurring with food-grabbing behavior or bradykinesia caused by EPS; the integral area can refer to the integral value below the sEMG envelope curve within the swallowing event time window, used to comprehensively reflect the total amount of muscle work done during this swallowing process; in this embodiment, the reduction of this area is often associated with swallowing weakness, repetitive swallowing tendency, or fatigue-related swallowing disorders.

[0069] This application determines the swallowing reflex nerve preparation time by extracting the swallowing initiation delay from the sEMG signal envelope; it also determines the swallowing reflex nerve preparation time by detecting the swallowing initiation delay using the time-frequency energy distribution of the sEMG signal combined with the motion initiation marker; further, it determines the swallowing reflex nerve preparation time by jointly aligning the sEMG signal with the motion initiation point of the synchronously acquired masticatory electromyography or laryngeal IMU. Based on any of the above methods, this application obtains the swallowing initiation delay, peak amplitude, duration, and integral area, which are used to construct a subset of basic features reflecting the neuromuscular functional state of a single swallow.

[0070] Step 2: Based on the neck inertial measurement signals, obtain the laryngeal elevation amplitude, elevation speed, and number of swallowing episodes;

[0071] Among them, the laryngeal elevation amplitude refers to the maximum vertical displacement of the Adam's apple during swallowing, reflecting the mechanical ability of the pharynx to elevate and close the airway and pull the esophageal inlet. In this embodiment, a decrease in this amplitude is commonly seen in muscle rigidity caused by EPS or age-related swallowing muscle atrophy, which is an important physiological indicator of insufficient airway protection. The elevation speed refers to the maximum first derivative value of the laryngeal elevation displacement curve, reflecting the explosive force and coordination of pharyngeal movement. In this embodiment, a decrease in this speed is often associated with extrapyramidal bradykinesia or motor control disorder in a state of mental excitement. The number of swallowing events refers to the total number of effective swallowing events detected in a single eating / drinking action, reflecting the swallowing load required for a single feeding task. In this embodiment, an increase in this number often indicates decreased swallowing efficiency, incomplete bolus propulsion, or increased residue sensation, which is one of the typical manifestations of dysphagia.

[0072] This application determines the pharyngeal airway closure capacity by extracting the laryngeal elevation amplitude from the three-dimensional trajectory of the larynx calculated by an IMU; it also determines the pharyngeal airway closure capacity by estimating the laryngeal elevation amplitude by integrating the peak acceleration of the IMU and mapping it to the displacement domain; further, it determines the pharyngeal airway closure capacity by jointly inverting the laryngeal elevation amplitude based on the rate of change of IMU angular velocity and the neck flexion angle. Based on any of the above methods, this application obtains the laryngeal elevation amplitude, elevation velocity, and number of swallowing episodes, which are used to construct a key subset reflecting the biomechanical characteristics of the pharynx.

[0073] Step 3: The swallowing feature vector further includes statistical features of swallowing event intervals to characterize the eating rhythm;

[0074] The swallowing event interval can refer to the time difference between two adjacent valid swallowing events, denoted as Δt_sw(k)=t_k–t_(k–1), where t_k is the center timestamp of the k-th swallowing event. The statistical features of the swallowing event interval include the mean, standard deviation, coefficient of variation, skewness, kurtosis, and histogram distribution of swallowing frequency within the sliding window of the swallowing interval sequence, which are used to quantify the stability, regularity, and abnormal tendency of the eating rhythm. In this embodiment, the statistical features do not depend on the quality of a single swallowing event, but rather characterize the eating behavior patterns unique to mental patients from the perspective of dynamic process. For example, food grabbing is characterized by a short mean and low variability, binge eating is characterized by high frequency and high variability, while EPS-related bradyphaly is characterized by a long mean and high volatility. This feature is directly used as the input basis for identifying eating too fast / food grabbing in this application, and together with the static features of sEMG and IMU, it constitutes a multi-scale swallowing function representation.

[0075] This application characterizes the eating rhythm by, for example, calculating the mean and coefficient of variation of swallowing intervals based on the timestamps of swallowing events within a fixed-length sliding time window (30 s); it also characterizes the eating rhythm by, for example, identifying intra-cluster and inter-cluster intervals of consecutive swallowing events using an adaptive swallowing event clustering strategy and statistically analyzing their distribution characteristics; further, it characterizes the eating rhythm by modeling the rhythmic state transition law based on the Markov transition probability matrix of the swallowing interval sequence. This application obtains statistical characteristics of swallowing event intervals based on any of the above methods, which are used to construct a dynamic feature subset reflecting the rhythmic characteristics of eating behavior.

[0076] Example 4: In yet another optional embodiment, this application also provides psychiatric behavioral risk identification, including identification of eating too quickly / grabbing food. Identification of eating too quickly / grabbing food includes:

[0077] Step 1: Calculate the mean and coefficient of variation of the swallowing event interval sequence within a preset time window. When the mean is less than the first threshold or the coefficient of variation is greater than the second threshold, output the risk index of eating too fast / grabbing food. The preset time window is 30s to 180s.

[0078] The swallowing event interval sequence can be a sequence consisting of the time difference between the occurrence times of two adjacent swallowing events, denoted as Δt_sw(k)=t_k–t_(k–1), where t_k represents the start timestamp of the k-th swallowing event; this sequence reflects the stability of the patient's swallowing rhythm during continuous eating.

[0079] The mean can refer to the arithmetic mean of the sequence within the current time window, used to characterize the overall swallowing speed trend; the smaller the mean, the higher the swallowing frequency per unit time, indicating a tendency to eat too quickly;

[0080] The coefficient of variation can be the ratio of the standard deviation to the mean of the sequence, used to quantify the dispersion of swallowing intervals; the larger the coefficient of variation, the more irregular the swallowing rhythm, and the more impulsive behaviors such as snatching food or sudden swallowing exist.

[0081] The first threshold is a fixed value or an individualized dynamic value within the range of 2s to 4s, for example, 3s; when the mean is lower than this value, it indicates that the patient's average swallowing interval within this window is too short, which is consistent with the clinically defined pattern of eating too fast.

[0082] The second threshold is a dimensionless value in the range of 0.4 to 0.7, for example, 0.55. When the coefficient of variation exceeds this value, it indicates that the distribution of swallowing events is highly discrete, suggesting that the patient has irregular swallowing behavior driven by emotions, such as grabbing food or sudden swallowing between shouts.

[0083] The preset time window is a sliding time window with a length of 30s to 180s, for example, 60s. This window length takes into account both clinical observation habits (a single eating action usually lasts 1 to 3 minutes) and the real-time requirements of signal processing, ensuring that effective behavioral segments can still be captured even when the patient is not fully cooperative.

[0084] This application calculates the time difference between adjacent swallowing events based on the timestamp sequence and constructs a sliding window. Within the window, the mean and coefficient of variation statistics are updated in real time. When any statistic exceeds the limit, a preliminary risk marker is triggered. Based on the structured event stream output by the swallowing event detection module, this application caches the timestamps of the most recent N events (N is determined by the sampling rate and window length), and calculates the statistics in a rolling manner after time alignment. Furthermore, this application maps the swallowing event interval sequence to a histogram or probability density function, and judges whether it deviates from the normal eating rhythm distribution by fitting the distribution shape. Based on any of the above methods, this application obtains a preliminary risk indicator of eating too fast / grabbing food.

[0085] For example, this application initiates a 60-second sliding time window after the patient begins drinking water. The system continuously receives the swallowing event timestamp sequence output by the S2 swallowing event detection module. The mean value of the interval Δt_sw(k) of all swallowing events within the current window is 2.1s, and the coefficient of variation is 0.62. Since the mean value is <3s and the coefficient of variation is >0.55, the system outputs a preliminary label indicating the risk of eating too quickly / rushing to eat, and enters the chewing adequacy verification stage.

[0086] Step 2: The identification of eating too fast / grabbing food further combines the chewing adequacy constraint, which includes: calculating the number of chews, chewing energy or chewing-swallowing ratio based on chewing-related electromyography or oromandibular movement signals. When the chewing-swallowing ratio is lower than the preset threshold and simultaneously meets the mean or coefficient of variation criterion, the confidence level of the eating too fast / grabbing food risk indicator is increased.

[0087] Among them, the chewing-related electromyographic signals can refer to electromyographic signals collected from the surface of the masseter, temporalis, or medial pterygoid muscles, used to reflect the initiation, intensity, and duration of chewing movements; this signal has been explicitly mentioned as an optional sensor type in the previous section and belongs to the extended implementation path of multidimensional physiological signals in this application; the jaw movement signals can refer to the mandibular opening and closing trajectory, displacement amplitude, or angular velocity signals obtained by the neck / mandibular IMU, optical markers, or ultrasonic displacement sensors, used to characterize the spatial dynamic characteristics of chewing movements; the number of chewing cycles can refer to the number of chewing movement cycles within a unit time window, each cycle containing one mandibular descent-ascent or one electromyographic burst-resting cycle; the chewing energy can refer to the chewing electromyographic energy in a specific frequency band (such as 20Hz~150Hz). The integral energy within the displacement-velocity phase space, or the area of ​​the orbital envelope of the jaw movement signal; the chewing-swallowing ratio can be the ratio of the number of chewing actions to the number of swallowing events within the same time window, for example, 3:1 to 6:1; this ratio reflects the coordination between chewing and swallowing during eating; the lower the ratio, the fewer chewing actions per unit of swallowing, indicating a wolfing-down behavior of swallowing without sufficient chewing; the preset threshold is a value in the range of 2 to 4, for example, 3; when the chewing-swallowing ratio is lower than this value, and either the mean or the coefficient of variation criterion mentioned above is met, the system will raise the confidence level of the original risk indicator from the basic level (e.g., 0.6) to the enhanced level (e.g., 0.85) to support the priority ranking of subsequent nursing intervention decisions;

[0088] This application detects the starting point of the chewing action based on the peak value of the masticatory electromyography envelope, simultaneously records the timestamp of the swallowing event, counts the number of both within a unit window, and calculates the ratio. This application identifies the chewing cycle based on the zero-crossing point of the mandibular IMU angular velocity signal and counts it after temporal proximity matching with the swallowing event. Furthermore, this application jointly models the masticatory electromyography and swallowing acoustic signals, and uses a temporal attention mechanism to estimate the chewing-swallowing coupling strength, replacing the explicit ratio calculation. This application obtains the chewing adequacy state based on any of the above methods, and adjusts the confidence level of the risk index of eating too fast / grabbing food accordingly.

[0089] For example, in the aforementioned 60s window, the system simultaneously collected sEMG signals of the masseter muscle, detecting a total of 92 chewing movements and 18 swallowing events, with a chewing-swallowing ratio of 5.1. This value is higher than the threshold of 3, so it does not trigger confidence enhancement. However, in another assessment, the same patient exhibited only 25 chewing movements and 15 swallowing events in a state of agitation, with a ratio of 1.7, which is lower than the threshold of 3. Moreover, the mean of the above values ​​was 1.9s and the coefficient of variation was 0.68, satisfying both conditions. The system then increased the confidence level of the risk of eating too quickly / grabbing food from 0.65 to 0.88 and marked the high-confidence grabbing behavior in the assessment report, recommending immediate intervention.

[0090] Example 5: In one possible implementation, this application also provides that psychiatric behavioral risk identification further includes recognizing eating while in a supine position, including:

[0091] Step 1: Calculate the trunk tilt angle and classify body positions based on the trunk inertial measurement signal;

[0092] Among them, trunk inertial measurement signals can refer to triaxial acceleration and triaxial angular velocity signals collected by an inertial measurement unit fixed to the subject's T6–T10 thoracic vertebrae or L2–L4 lumbar vertebrae.

[0093] The signal can be a steady-state attitude signal preprocessed by low-pass filtering (cutoff frequency ≤ 10Hz) and zero-phase digital filtering;

[0094] In this embodiment, the trunk tilt angle can refer to the static tilt angle of the trunk in the sagittal plane relative to the direction of gravity. Its calculation is based on the projection of the gravity component of the acceleration signal in a static or quasi-static state, using the arctangent function relationship: ,in, , , These are the instantaneous acceleration components measured by the IMU in the front-back (x), left-right (y), and head-to-foot (z) directions, respectively.

[0095] This tilt angle can be the mean of a sliding window that is updated in real time (window length is 1–3s) to suppress short-term disturbances;

[0096] Position classification can be achieved by mapping a continuous tilt sequence to discrete position labels, including sitting (0°–30°), semi-recumbent (30°–60°), supine (60°–90°), and upright (–15°–15°), where 30°–80° covers the clinically defined safe range for supine positions.

[0097] This application can determine the torso tilt angle, for example, based on the spatial projection relationship of the torso acceleration signal in the gravitational field; it can also determine the torso tilt angle using a complementary filtering method that combines the integral of the torso angular velocity with acceleration compensation; further, it can estimate the torso tilt angle by fusing acceleration and angular velocity signals using an extended Kalman filter. This application obtains a stable, low-latency estimate of the torso tilt angle based on any of the above methods, which can be used to support subsequent posture-swallowing coupling judgment.

[0098] Step 2: When the trunk tilt angle is within the preset supine angle range and overlaps with the swallowing event time window, output the supine eating risk index;

[0099] Among them, the preset lying position angle range (30°~80°) can refer to the trunk tilt angle range that has been proven in clinical swallowing physiology to significantly reduce airway protection and increase the probability of aspiration. This range covers all high-risk body positions from semi-recumbent position (30°) to near supine position (80°).

[0100] The swallowing event time window can be defined as the time interval [t_k-0.5s, t_k+0.5s] centered on the moment t_k of each confirmed swallowing event, extending 500ms forward and backward. This time window is used to encompass the time distribution of physiological processes such as the initiation of swallowing action, the peak of larynx elevation, and glottal closure.

[0101] Overlap determination can refer to the existence of at least one continuous time period with a duration of not less than 200ms, during which the trunk tilt angle is maintained within the range of 30° to 80° and falls completely within the time window of a swallowing event.

[0102] The risk indicator for eating in a supine position can be a Boolean flag (True / False) or a floating-point value (0-1) with confidence level. Its confidence level can be calculated by weighting the duration of the tilt angle within the time window, the standard deviation of the angle stability, or the time alignment accuracy with the peak swallowing event.

[0103] This application may generate a risk index for eating in a supine position based on the continuous proportion of trunk tilt angle within the swallowing event time window; this application may also generate a risk index for eating in a supine position based on joint modeling of the time offset and angle deviation of trunk tilt angle and the center time of the swallowing event; furthermore, this application may also generate a risk index for eating in a supine position based on a multi-window sliding matching strategy, searching for the longest continuous supine position period within 1 second before and after the swallowing event, and jointly assigning a score to its length and the mean angle.

[0104] For example, this application could involve continuously acquiring acceleration signals from a trunk IMU during the evaluation of a schizophrenic patient treated with olanzapine, and the system calculating the acceleration signal in real time. And it remained stable for 1.2 seconds; at the same time, the system detected a valid swallowing event, the time window of which was Upon comparison, the 47° tilt angle lasted from 12:03:45.320 to 12:03:45.650, completely falling within the time window and lasting for 330ms. Based on this, the risk indicator for eating in a supine position was output as True with a confidence level of 0.86, triggering the marking of eating in a supine position (3 points) in the nursing report. At the same time, the key evidence recorded was the trunk tilt angle of 47°, lasting for 330ms, which highly overlapped with the swallowing event.

[0105] Example 6: In an optional embodiment, this application also provides an extrapyramidal response characterization vector, including:

[0106] Step 1: Energy characteristics of the 4Hz–7Hz flutter spectrum peaks obtained from the spectrum of the inertial measurement signal;

[0107] Among them, the inertial measurement signal can refer to the three-dimensional acceleration and angular velocity signals collected by an inertial sensor worn on the wrist, forehead or neck; the technical feature can refer to the local energy peak and its normalized energy value detected in the 4Hz to 7Hz frequency band after performing short-time Fourier transform (STFT) or wavelet transform on the neck or upper limb IMU signal in a resting or light activity state.

[0108] In this embodiment, this feature is used to characterize resting tremor induced by antipsychotic drugs. Its energy intensity is positively correlated with the tremor amplitude and is used as an objective indicator of early neuromuscular abnormalities in EPS, which is input into the classification model to provide frequency domain physiological basis for distinguishing drug-induced dysphagia.

[0109] Step 2: Baseline elevation characteristics and common contraction ratio characteristics obtained from surface electromyography signals;

[0110] Among them, the baseline elevation feature can refer to the relative increase of the mean value of the envelope obtained by low-pass filtering after rectification of the sEMG signal relative to the healthy baseline level during the resting period without active swallowing tasks; the co-contraction ratio feature can refer to the sEMG energy ratio of antagonistic muscle groups (such as the geniohyoid muscle and the infrahyoid muscle) in the same time window, which is used to quantify the decline in muscle synergistic control ability.

[0111] In this embodiment, the two features mentioned above together reflect the core pathological manifestation of EPS: elevated baseline indicates abnormal recruitment of muscle fibers at rest, and abnormal co-contraction ratio indicates an imbalance in the coordination of larynx elevation and descent during swallowing. The two features are combined and input into the classification model to enhance the robustness of identifying EPS-induced swallowing initiation and execution phase disorders.

[0112] Step 3: Motion initiation delay characteristics obtained from motion-related signals;

[0113] Among them, the action-related signals can refer to physiological signals used to characterize the swallowing preparation period or the initiation process of voluntary movements, including but not limited to: the moment of the initial value jump of the acceleration of the neck IMU, the moment when the sEMG first exceeds the baseline threshold, and the time interval from the termination of chewing sounds to the start of swallowing sounds in the acoustic signals; the motor initiation delay feature can refer to the time difference between the start of the instruction prompt (such as swallowing) or the triggering of the eating action (such as the spoon touching the lips) and the appearance of a clear initiation response of any of the above action-related signals; in this embodiment, this feature is used to quantify the performance of bradykinesia in the swallowing preparation stage. The longer the delay, the lower the efficiency of the extrapyramidal pathway in the transmission and execution of the cortical descending instructions, which in turn affects the timeliness and rhythm of the swallowing reflex; this feature, as a key indicator of swallowing timing regulation disorder, complements the aforementioned tremor and muscle tone features, and constitutes a multidimensional dynamic representation of EPS.

[0114] This application, for example, uses cervical IMU spectral analysis to determine the energy characteristics of the 4Hz–7Hz tremor spectrum peak; it also uses sEMG resting-phase envelope statistics to determine baseline elevation and co-contraction ratio characteristics; further, it uses command-response time alignment to determine motor initiation delay characteristics. Based on any of the above methods, this application obtains a multidimensional physiological characterization of extrapyramidal responses, supporting the classification model's objective discrimination of EPS-related dysphagia.

[0115] For example, this application involves: simultaneously acquiring cervical IMU and anterior cervical sEMG signals while the patient is in a seated and relaxed state; performing a 0.5s sliding window STFT on the IMU signal, extracting the maximum energy value in the 4Hz-7Hz frequency band within each window and smoothing it to obtain a tremor spectrum peak energy sequence; calculating the mean of the rectified low-pass envelope within 30s of resting on the sEMG signal, which is recorded as the baseline elevation value; simultaneously extracting sEMG signals of bilateral antagonistic muscles, calculating the median of their energy ratios within a 1s sliding window, which is recorded as the co-contraction ratio; after issuing a swallowing voice prompt, recording the time difference between the first significant activation of the sEMG and the prompt time as the motor initiation delay; concatenating the three types of features into a fixed-dimensional vector, inputting it into a pre-trained lightweight random forest classifier, and outputting the EPS-related swallowing difficulty probability p_eps; when p_eps ≥ 0.65, it is determined that there is a moderate to high EPS-related swallowing risk.

[0116] Example 7: In an optional implementation, this application also provides a classification model based on supervised learning training. The input of the classification model includes a concatenated vector of extrapyramidal response representation vector and swallowing feature vector. During the training phase, information on changes in drug dosage is introduced as a conditional feature or sample stratification feature to improve the stability of identifying drug-induced EPS-related swallowing disorders.

[0117] Step 1: The classification model is a classification model trained based on supervised learning. The input of the classification model includes a concatenated vector of the extrapyramidal response representation vector and the swallowing feature vector.

[0118] Among them, the classification model can refer to a supervised machine learning model used to map input features to the probability of swallowing difficulties related to extrapyramidal reactions; the model trained based on supervised learning can refer to the model completing parameter optimization by minimizing prediction error on a labeled sample set, and its label is the positive / negative result of EPS-related swallowing difficulties jointly determined by clinical experts based on multidimensional physiological evidence and medication history; the extrapyramidal reaction characterization vector can be a vector defined in Example 6, consisting of the peak energy characteristics of the 4Hz-7Hz tremor spectrum of the inertial measurement signal, the baseline elevation characteristics and common contraction ratio characteristics of the surface electromyography signal, and the motion initiation delay characteristics of the action-related signal; the swallowing feature vector can be a vector defined in Example 3, consisting of the swallowing initiation delay, peak amplitude, duration and integral area of ​​the surface electromyography signal, the laryngeal elevation amplitude, elevation speed and number of swallowing sessions of the neck inertial measurement signal, and the statistical characteristics of the swallowing event interval; the spliced ​​vector can be a joint feature vector formed by directly connecting the above two vectors in dimensional order, the dimension of which is equal to the sum of the dimensions of the two vectors, without introducing cross terms or nonlinear transformations.

[0119] This application can employ a logistic regression model to perform binary classification modeling on the concatenated vectors to output the probability of EPS-related swallowing difficulties; it can also employ a random forest model to perform ensemble learning modeling on the concatenated vectors, determining the probability distribution through voting among multiple decision trees; furthermore, it can employ a lightweight fully connected neural network to perform end-to-end mapping on the concatenated vectors, outputting normalized probability values. Based on any of the above methods, this application obtains a stable ability to discriminate EPS-related swallowing difficulties.

[0120] Step 2: During the training phase, information on changes in drug dosage is introduced as a conditional feature or sample stratification feature to improve the stability of identifying drug-induced EPS-related dysphagia.

[0121] Among them, the information on changes in drug dosage can refer to the relative change rate of the daily dose of antipsychotic drugs in the past 7 days, which is calculated as (current daily dose – daily dose 7 days ago) / daily dose 7 days ago × 100%. This information reflects the dynamic evolution trend of the intensity of drug intervention. As a conditional feature, this value can be directly added to the concatenation vector as an additional input dimension, so that the model can adaptively adjust the weight allocation according to the current medication change status during inference. As a sample stratification feature, the training samples can be divided into a drug-increasing group (change rate ≥ +20%), a stable drug group (change rate ∈ [-10%, +10%]), and a drug-reducing group (change rate ≤ -20%) based on this value, and sub-models can be trained in each layer or differential loss function weights can be set to enhance the model's adaptability to different medication stages. Identification stability can refer to the predictive consistency and robustness of the model under conditions of cross-patient, cross-medication cycle, and cross-device data collection, avoiding misjudgment due to short-term fluctuations of a single patient.

[0122] This application can standardize the drug dosage change rate and input it along with the concatenated vector into a logistic regression model. The model coefficients learn that this feature is positively correlated with the positive probability of EPS, thereby improving the sensitivity of risk warning at the initial stage of drug administration. Furthermore, this application can assign higher sample weights (e.g., 1.5 times) to the drug-administered group samples during random forest training, making them more likely to be preferentially covered during node splitting, thus improving the model's response to drug-induced effects. Further, this application can construct a multi-task learning framework, adding an auxiliary task (drug change stage classification) in addition to the main task (EPS swallowing difficulty identification), sharing underlying feature representations, and enhancing the model's ability to perceive drug context. This application achieves a highly stable identification capability for drug-induced EPS-related swallowing disorders based on any of the above methods.

[0123] Example 8: In another optional embodiment, this application also provides that the swallowing reflex abnormality index is obtained from airway protection abnormality detection, which includes any of the following: detection of cough event clusters after a swallowing event; detection of abnormal respiratory phase switching after a swallowing event; detection of a decrease in blood oxygen saturation exceeding a threshold after a swallowing event; detection of wet sound / residual features in swallowing acoustic signals exceeding a threshold; and fusing the detection results to form a reflex / airway protection inadequacy risk index; in the swallowing event detection and / or airway protection abnormality detection, anti-artifact processing is performed on artifacts caused by the psychiatric patient's speaking, shouting, or agitation, and the anti-artifact processing includes:

[0124] Step 1: Detection of cough event clusters following swallowing events;

[0125] The cough event cluster refers to ≥2 consecutive cough events occurring within a preset time window after a single swallowing event, with the time interval between adjacent cough events not exceeding 1.5 seconds. This time window begins at the end of the swallowing event and lasts for 3 to 10 seconds. The cough event itself is identified by the energy surge and spectral morphology of the chest wall acceleration signal or airflow acoustic signal. In this embodiment, the cough event cluster is used to characterize the decline in airway clearance capacity and the lag in protective reflexes. As indirect evidence of airway reopening failure after swallowing, it is directly associated with the swallowing reflex reduction / choking care item. When at least one cough event cluster is detected, the positive output of this detection path is triggered and it serves as one of the inputs to the risk indicators of insufficient reflex / airway protection.

[0126] This application can identify single cough events based on the instantaneous energy amplitude of the chest wall acceleration signal exceeding the dynamic baseline threshold and lasting for ≥100ms, while the main spectral energy is concentrated in the 100Hz~400Hz frequency band. It can also identify single cough events based on the short-time energy peak density of nasal airflow or perioral acoustic signals being >3 / second, and the peak envelope attenuation rate conforming to the cough acoustic attenuation model. Furthermore, this application can identify single cough events based on multi-channel signal joint verification, requiring the chest wall acceleration signal and acoustic signal to synchronously meet their respective criteria within a time tolerance ≤50ms. This application obtains cough event cluster detection results based on any of the above methods to support the generation of inadequate reflex / airway protection risk indicators.

[0127] Step 2: Detection of abnormal respiratory phase switching after swallowing events;

[0128] Abnormal respiratory phase switching can refer to the failure to complete the normal transition from the expiratory phase to the inspiratory phase within the expected time after a swallowing event, or phenomena such as prolonged apnea, delayed inspiratory initiation, and disordered swallowing-breathing coupling rhythm during swallowing. This abnormality reflects impaired coordination between the medullary respiratory center and the swallowing center, and is a typical manifestation of airway protection reflex pathway disorder. In this embodiment, abnormal respiratory phase switching does not rely on manually labeled respiratory cycles, but is automatically determined based on the temporal morphological changes of nasal airflow signals, chest and abdominal impedance signals, or respiratory band signals. When a respiratory phase switching delay >1.2 seconds after a swallowing event is detected, or an apnea time >2.5 seconds during swallowing, or a respiratory phase mismatch rate >60% in three consecutive swallows, it is determined to be abnormal respiratory phase switching, and serves as another input source for the risk indicator of insufficient reflex / airway protection.

[0129] This application can identify post-swallowing switching delay by judging the respiratory phase using a combination of the direction of the zero-crossing jump of the nasal airflow signal and the rate of change of the amplitude slope; this application can also identify prolonged apnea by comparing the polarity reversal time of the differential waveform of the chest and abdominal impedance signal with the timestamp of the swallowing event; furthermore, this application can also identify rhythm disturbances based on the phase lock value PLV < 0.35 in the respiratory-swallowing coupling phase diagram; this application obtains respiratory phase switching abnormality detection results based on any of the above methods to support the generation of reflex / airway protection inadequate risk indicators.

[0130] Step 3: Detection of blood oxygen saturation drop exceeding a threshold after a swallowing event;

[0131] The decrease in blood oxygen saturation (ΔSpO2) can be defined as the absolute value of the difference between the lowest SpO2 value measured by a pulse oximeter within a specified time window after the swallowing event and the mean steady-state SpO2 value 1 second before swallowing. This indicator reflects the physiological consequences of gas exchange impairment caused by aspiration and serves as downstream confirmatory evidence of airway protection failure. In this embodiment, the specified time window is 20 to 60 seconds from the end of the swallowing event, and the signal-to-noise ratio of the SpO2 signal is required to be ≥15dB to exclude motion artifacts. When ΔSpO2 ≥2% and the duration is ≥5 seconds, or ΔSpO2 ≥4% and the duration is ≥2 seconds, it is determined that the decrease in blood oxygen saturation exceeds the standard. This result serves as the third input source for the risk indicator of insufficient reflex / airway protection.

[0132] This application can determine ΔSpO2 by using a method that obtains the SpO2 trend from the photoplethysmography (PPG) signal of the fingertip or earlobe through bandpass filtering (0.5Hz~5Hz), AC / DC separation, and ratio calculation; this application can also determine ΔSpO2 by using a table to map SpO2 based on the ratio of the AC components of dual-wavelength (660nm / 940nm) PPG signals; furthermore, this application can also improve the robustness of the criterion by combining the confidence weighting of SpO2 decrease with the low-frequency power suppression state of heart rate variability (HRV); this application obtains the blood oxygen saturation decrease detection results based on any of the above methods to support the generation of risk indicators for insufficient reflex / airway protection.

[0133] Step 4: Detection of wet sound / residual features in the swallowing acoustic signal exceeding the threshold;

[0134] The wet sound / residual feature (R_wet) can refer to specific acoustic patterns in the swallowing acoustic signal that reflect residual fluid in the larynx or increased airway secretions, including but not limited to: increased energy proportion in the 200Hz-600Hz frequency band, enhanced high-frequency jitter components during glottal closure, prolonged attenuation time of swallowing sounds, and increased dynamic differences in multi-frame Mel-frequency cepstral coefficients (MFCC). This feature indicates the presence of fluid retention in the airway after swallowing, which is an early acoustic marker of airway protection mechanism decompensation. In this embodiment, the wet sound / residual feature does not rely on a speech recognition model, but extracts the time-frequency energy distribution through a fixed window length (256 points) short-time Fourier transform and compares it with a pre-built healthy swallowing template using cosine similarity. When the similarity is <0.65, or the energy ratio in the 200Hz-600Hz range is > the template mean + 1.5 times the standard deviation, it is determined that the wet sound / residual feature exceeds the standard. This result serves as the fourth input source for the risk index of insufficient reflection / airway protection.

[0135] This application can detect wet sound characteristics by means of the ratio of energy in the 200Hz-600Hz subband to the full-band energy in the short-time spectrum of the swallowing acoustic signal exceeding the dynamic threshold; this application can also detect residual characteristics by means of the acoustic signal envelope decay time constant τ>120ms and accompanied by high-frequency jitter (700Hz-1200Hz) energy enhancement; furthermore, this application can also detect glottal movement incoordination by means of the increase of the first-order differential entropy value of multi-frame MFCC; this application obtains wet sound / residual characteristic detection results based on any of the above methods to support the generation of reflection / airway protection inadequate risk indicators.

[0136] Step 5: Merge the test results to form a risk index for insufficient reflection / airway protection;

[0137] Fusion can refer to the normalization and weighted combination of the aforementioned four test results (cough event clusters, abnormal respiratory phase switching, excessive blood oxygen saturation, and excessive moist rales / residual features) to generate a single continuous risk indicator R_reflex∈[0,1]. Each individual result is first mapped to a score in the 0–1 interval: 1 for positive and 0 for negative. Then, clinical weight coefficients w1 to w4 (ranging from 0.2 to 0.4, with a total of 1) are introduced, and confidence correction factors (such as the number of cough clusters, ΔSpO2 amplitude, and number of moist rales) can be optionally added. Finally, R_reflex=Σ(w i ×v i ×c i ), where v i For the i-th binary output, c i The corresponding confidence level is set to (0.7–1.0). This index is directly mapped to the objective score of swallowing reflex reduction / choking. When R_reflex≥0.65, a score of 3 is assigned; otherwise, a score of 0 is assigned.

[0138] This application can fuse the four test results using a weighted summation method, with the cough event cluster having the highest weight (w1=0.35) due to its strongest specificity. This application can also use a logistic regression model to nonlinearly fuse the four binary inputs and their confidence scores, with model parameters obtained through training on multicenter psychiatric patient data. Furthermore, this application can employ a Bayesian fusion framework, using each test result as likelihood evidence and a history of aspiration as prior information, to iteratively update R_reflex. This application obtains a reflex / airway protection inadequacy risk index based on any of the above methods, which supports item-level scoring and total score calculation.

[0139] Step Six: In swallowing event detection and / or airway protection abnormality detection, anti-artifact processing is performed on artifacts caused by speech, shouting or agitation in psychiatric patients. Anti-artifact processing includes: when an acoustic event meets the characteristics of speech bands but lacks swallowing consistency characteristics of surface electromyography and inertial measurements, the contribution of the acoustic event to the swallowing event sequence and risk indicators is suppressed.

[0140] Among them, the speech band characteristics can refer to the acoustic signal exhibiting a periodic harmonic structure in the 500Hz-4000Hz frequency band, stable fundamental frequency (85Hz-180Hz for males, 165Hz-255Hz for females), and obvious glottal pulse modulation characteristics; this characteristic can be distinguished by short-time autocorrelation function peak detection or cepstral peak-to-valley ratio; the swallowing consistency characteristics can refer to the presence of a suprahyoid muscle group envelope peak in the surface electromyography signal and a laryngeal elevation characteristic peak in the neck inertial measurement signal within the same time window, and the time deviation between the two is ≤80ms; in this embodiment, the suppression contribution does not delete the original acoustic data, but sets the confidence of the acoustic event to zero and prohibits it from triggering any swallowing event confirmation, cough cluster count, wet sound score, or fusion input; this mechanism ensures that it is recognized as a real swallowing-related physiological activity only when the multimodal signal responds collaboratively, thereby avoiding false positive risk indicators caused by speech interference.

[0141] This application can identify speech artifacts and suppress their contribution based on the conditions that the matching degree between the first peak position of the acoustic signal cepstral domain and the estimated fundamental frequency is >0.85, and the surface electromyography envelope peak is missing or has a time offset >100ms. This application can also identify shouting artifacts and suppress their contribution based on the continuity of the F1 / F2 formant trajectory in the Mel spectrogram of the acoustic signal, combined with the condition that the IMU laryngeal movement amplitude is <0.5g. Furthermore, this application can also automatically downweight concurrent acoustic events to 0.05 based on the joint confidence score of sEMG and IMU <0.4, so that they cannot dominate any risk judgment criteria. This application implements anti-artifact processing based on any of the above methods to ensure the data purity of swallowing event sequences and reflex risk indicators.

[0142] For example, this application could involve a patient suddenly shouting during the assessment process, with acoustic signals showing strong energy concentrated between 800Hz and 2500Hz, and the first peak of the cepstral spectrum located at 125Hz, consistent with male fundamental frequency characteristics; however, synchronous sEMG did not show a significant increase in the envelope, IMU did not detect laryngeal elevation movement, and consistency characteristics within the time window were missing; the system determines this event as a speech artifact, and its acoustic energy is not included in cough cluster detection, wet sound scoring, or confirmation of swallowing events, nor is it included in any risk indicator fusion path.

[0143] Example 9: In another optional embodiment, this application further provides that the item-level output corresponding to the nursing risk assessment item is an item risk indicator rather than a fixed score. The item risk indicators include EPS-related swallowing risk indicators, behavioral feeding risk indicators, supine eating risk indicators, and reflex / airway protection inadequacy risk indicators. Furthermore, the total score or comprehensive risk value is a comprehensive risk value obtained by normalizing and fusing the item risk indicators. The fusing includes any one of weighted summation, logistic regression fusing, or Bayesian fusing, wherein the weights or fusing parameters are learned from training data or adaptively updated from the individual baseline. The comprehensive risk value is used to output the aspiration / choking risk level, which includes Level I, Level II, and Level III. The assessment report further outputs the dominant risk factors that lead to the risk level. The dominant risk factors are determined by ranking the sensitivity contribution or confidence contribution of each item risk indicator, including:

[0144] Step 1: The item-level output corresponding to the nursing risk assessment item is the item risk indicator rather than a fixed score. The item risk indicators include EPS-related swallowing risk indicators, behavioral feeding risk indicators, supine eating risk indicators, and reflex / airway protection inadequate risk indicators.

[0145] Among them, the EPS-related swallowing risk index can refer to the probability value p_eps output after the extrapyramidal response characterization vector and swallowing feature vector defined in Example 6 are jointly input into the subtyping model. This probability value represents the degree of possibility that the current swallowing difficulty is caused by EPS induced by antipsychotic drugs. Its value range is [0,1]. The higher the value, the more significant the impact of EPS neuromuscular control disorder on swallowing function.

[0146] Behavioral feeding risk indicators can refer to the risk score generated based on the eating too fast / fighting for food identification results as defined in Example 4. This score can be a continuous value of binary discriminant output (0 or 1) after confidence weighting, or it can be a standardized risk score calculated by combining the coefficient of variation of swallowing interval, the degree of deviation of chewing-swallowing ratio from baseline, and auxiliary evidence of excitation state (decreased HRV, increased EDA). It is used to quantify the degree to which patients' eating rhythm is disrupted due to mental and behavioral abnormalities, thereby increasing the risk of aspiration.

[0147] The risk index for eating in a supine position can be a risk measure obtained by the analysis of the overlap between the trunk tilt angle and the swallowing event time window as defined in Example 5. It can be a continuous index that is a Boolean judgment result (whether or not one is in a supine position for eating) after being weighted by the duration of the position, the frequency of swallowing and the degree of attenuation of the laryngeal movement amplitude. It is used to reflect the tendency of aspiration caused by the weakening of airway protection due to improper position.

[0148] The risk index for insufficient reflex / airway protection can refer to the risk index obtained by fusing the abnormal airway protection detection results as defined in Example 8. It is a weighted synthesis of sub-items such as the intensity of post-swallowing cough event clusters, abnormal respiratory phase switching, the magnitude of decrease in blood oxygen saturation, and swallowing acoustic moist sound characteristics according to preset rules. It is used to comprehensively characterize the susceptibility to aspiration caused by sluggish swallowing reflex, incomplete laryngeal closure, or decreased airway clearance ability. The above four types of risk indicators are all expressed in the form of dimensionless continuous numerical values. They are calculated independently and are not coupled with each other. Each indicator has a clear physiological basis, a reproducible calculation path, and a clinically interpretable semantic boundary, thereby replacing the fixed scoring mode of present / absent and yes / no in traditional scales and realizing a gradient characterization of risk level.

[0149] Step 2: The total score or comprehensive risk value is the comprehensive risk value obtained by normalizing and fusing the item risk indicators. The fusion includes any one of weighted summation, logistic regression fusion or Bayesian fusion, where the weights or fusion parameters are learned from the training data or adaptively updated from the individual baseline.

[0150] The weighted summation method can refer to multiplying the four risk indicators by their corresponding weights and then summing them. The weights reflect the relative contribution of each risk dimension to the overall aspiration mechanism. For example: w1×EPS-related swallowing risk indicators + w2×behavioral feeding risk indicators + w3×feeding risk indicators in supine position + w4×inadequate reflex / airway protection risk indicators, where the weights satisfy w1+w2+w3+w4=1, and w1, w2, w3, w4∈[0,1].

[0151] The logistic regression fusion method can refer to constructing a generalized linear model: logit(P_risk)=β0+β1·EPS index+β2·fighting index+β3·lying position index+β4·reflex index, where β0 is the bias term, β1~β4 are the regression coefficients of each risk index, and P_risk is the output comprehensive aspiration risk probability. The parameters of this model are obtained by training with historical cohort data labeled with real aspiration events.

[0152] Bayesian fusion methods can refer to treating each risk indicator as independent observational evidence and combining it with prior knowledge (such as the high probability of EPS at different medication stages and the baseline level of reflex decline in elderly patients) to construct a posterior risk distribution, i.e., P(Comprehensive Risk|EPS Indicator, Food Competition Indicator, Lying Position Indicator, Reflex Indicator) ∝ P(EPS Indicator|Comprehensive Risk)·P(Food Competition Indicator|Comprehensive Risk)·P(Lying Position Indicator|Comprehensive Risk)·P(Reflex Indicator|Comprehensive Risk)·P(Comprehensive Risk), where each likelihood term and prior term are obtained by fitting clinical expert experience or historical data;

[0153] This application can obtain a comprehensive risk value using a weighted summation method, where weights w1, w2, w3, and w4 are trained from a multi-center clinical dataset, and stratified weight templates are set for different age groups, medication types, and disease stages. Alternatively, this application can obtain a comprehensive risk value using logistic regression fusion. This model has undergone cross-validation before deployment, achieving an AUC ≥ 0.85, and supports online incremental learning to adapt to newly enrolled patient data. Furthermore, this application can obtain a comprehensive risk value using Bayesian fusion. This method uses a population prior at the initial patient assessment and gradually updates the individual posterior distribution during subsequent follow-ups, making the comprehensive risk value more dynamically adaptable. This application obtains an interpretable, traceable, and updatable comprehensive risk value based on any of the above methods, serving as the foundational input for subsequent risk stratification and intervention decisions.

[0154] Step 3: The comprehensive risk value is used to output the risk level of aspiration / choking. The risk levels include Level I, Level II and Level III. The assessment report further outputs the dominant risk factors that cause the risk level. The dominant risk factors are determined by ranking the sensitivity contribution or confidence contribution of each item's risk indicator.

[0155] Among them, the risk levels of aspiration / choking, Level I, Level II, and Level III, correspond to low-risk, medium-risk, and high-risk clinical treatment thresholds, respectively. The classification is based on the preset intervals in which the comprehensive risk value falls: Level I is [0, 0.35), Level II is [0.35, 0.7), and Level III is [0.7, 1]. This three-level classification maintains a consistent statistical distribution with the current nursing scale's Level I (1–6 points), Level II (7–14 points), and Level III (≥15 points), ensuring that clinicians can understand the output results without changing their existing cognitive habits.

[0156] The dominant risk factor can be a key risk indicator that has the greatest impact on the current composite risk value and best explains the cause of the risk level. It can be determined in two ways: Sensitivity contribution analysis: using Locally Interpretable Artificial Intelligence (LIME) or Shapley Additive Explanations (SHAP) methods to calculate the marginal contribution of each risk indicator to the composite risk value output, and selecting the one with the largest absolute value as the dominant risk factor; Confidence contribution analysis: within a Bayesian fusion framework, comparing the posterior likelihood ratio or evidence weight of each risk indicator, and selecting the indicator that has the most significant impact on the shape of the posterior risk distribution as the dominant risk factor.

[0157] This application can determine the dominant risk factors by ranking the SHAP values. The EPS-related swallowing risk index has a SHAP value of 0.42, the behavioral feeding risk index has a SHAP value of 0.28, the supine eating risk index has a SHAP value of 0.15, and the reflex / airway protection inadequate risk index has a SHAP value of 0.15. Therefore, the EPS-related swallowing risk index is identified as the dominant risk factor. This application can also determine the dominant risk factor using a Bayesian evidence weighting method. When the likelihood ratio of the reflex / airway protection inadequate risk index reaches 3.2, significantly higher than the other three (all <1.5), it is identified as the dominant risk factor. Furthermore, this application can simultaneously output the top two dominant risk factors and their contribution percentages in the case of multiple labels, for example: dominant risk factor: EPS-related swallowing risk (contribution rate 62%), secondary risk factor: reflex / airway protection inadequate risk (contribution rate 23%), so that medical staff can quickly identify the key intervention areas. This application obtains the dominant risk factor identification results based on any of the above methods, supporting the formulation of precise and differentiated nursing measures.

[0158] Figure 2 This application provides a schematic diagram of the structure of a multidimensional physiological signal-based swallowing function assessment system for mental patients, as shown in one embodiment. Figure 2 As shown, the swallowing function assessment system 200 for mental patients based on multidimensional physiological signals in this embodiment includes: a data collection module 201, a vector extraction module 202, a risk identification module 203, a swallowing analysis module 204, a swallowing abnormality mapping module 205, and a swallowing assessment module 206.

[0159] The data collection module 201 is used to acquire basic information and medication information of the mentally ill patient to be evaluated, and to deploy and synchronously collect multidimensional physiological signals for the patient, including surface electromyography signals of the anterior cervical region, cervical inertial measurement signals, and swallowing acoustic signals; the vector extraction module 202 is used to detect swallowing events based on the multidimensional physiological signals, obtain a swallowing event sequence, and extract swallowing feature vectors for each swallowing event; the risk identification module 203 is used to identify psychiatric behavioral risks based on the swallowing event sequence and the inertial measurement signals, and obtain behavioral risk labels and their confidence levels; the swallowing analysis module 204 is used to extract swallowing features based on the multidimensional physiological signals. The extrapyramidal response characterization vector is obtained, and the extrapyramidal response characterization vector and the swallowing feature vector are jointly input into a preset classification model to output the probability of swallowing difficulties related to extrapyramidal response; the swallowing abnormality mapping module 205 is used to map the behavioral risk label, the swallowing difficulties related to extrapyramidal response and the swallowing reflex abnormality index corresponding to the swallowing feature vector to the item scores of nursing risk assessment items, and sum the item scores to obtain the total score; the swallowing assessment module 206 is used to output the aspiration / choking risk level according to the total score and the preset grading threshold, and generate an assessment report containing the item scores, the total score, the risk level and key physiological evidence.

[0160] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A method for assessing swallowing function in psychiatric patients based on multidimensional physiological signals, characterized in that, include: S1. Obtain basic information and medication information of the mentally ill patient to be evaluated, and deploy and synchronously collect multidimensional physiological signals for the mentally ill patient. The multidimensional physiological signals include surface electromyography signals of the anterior cervical region, neck inertial measurement signals, and swallowing acoustic signals. S2. Based on the multidimensional physiological signals, swallowing events are detected to obtain a swallowing event sequence, and swallowing feature vectors are extracted for each swallowing event. S3. Based on the swallowing event sequence and the inertial measurement signal, identify psychiatric behavioral risks and obtain behavioral risk labels and their confidence levels; S4. Extract the extrapyramidal response characterization vector based on the multidimensional physiological signal, and input the extrapyramidal response characterization vector and the swallowing feature vector into a preset classification model to output the probability of swallowing difficulties related to extrapyramidal response. S5. Map the behavioral risk labels, the extrapyramidal response-related dysphagia, and the swallowing reflex abnormality index corresponding to the swallowing feature vector to the item scores of the nursing risk assessment items, and sum the item scores to obtain the total score. S6. Output the aspiration / choking risk level based on the total score and the preset grading threshold, and generate an assessment report that includes the item score, the total score, the risk level, and key physiological evidence.

2. The method according to claim 1, characterized in that, The swallowing event detection employs a multimodal consistency confirmation mechanism, which includes: A swallowing event is considered valid if, within a preset time tolerance, the surface electromyography envelope peak appears, the laryngeal elevation characteristic peak of the neck inertial measurement signal appears, and the swallowing acoustic energy exceeds a threshold. The swallowing event detection is achieved through consistency confirmation of two signals. The consistency confirmation includes: the peak value of the surface electromyography envelope and the peak value of the laryngeal elevation of the neck inertial measurement signal appear simultaneously within a preset time tolerance, and the swallowing acoustic signal meets the swallowing sound energy threshold within the same time window, so as to confirm the swallowing event. The preset time tolerance is 10ms to 80ms, and the multidimensional physiological signals are synchronously calibrated before entering the swallowing event detection, so that the multimodal time alignment error is no more than 20ms.

3. The method according to claim 1, characterized in that, The swallowing feature vector includes: Swallowing initiation delay, peak amplitude, duration, and integral area obtained from surface electromyography signals; The laryngeal elevation amplitude, elevation speed, and number of swallowing episodes were obtained based on neck inertial measurement signals. Furthermore, the swallowing feature vector further includes statistical features of swallowing event intervals to characterize the eating rhythm.

4. The method according to claim 1, characterized in that, The aforementioned behavioral risk identification in psychiatry includes the identification of eating too quickly / grabbing food, which includes: The mean and coefficient of variation of the swallowing event interval sequence are calculated within a preset time window. When the mean is less than a first threshold or the coefficient of variation is greater than a second threshold, the risk index of eating too fast / grabbing food is output. The preset time window is 30s to 180s. The rapid eating / grabbing recognition is further combined with chewing adequacy constraints, which include: The number of chews, chewing energy, or chewing-to-swallowing ratio are calculated based on chewing-related electromyography or oromandibular movement signals. When the chewing-to-swallowing ratio is lower than a preset threshold and simultaneously meets the mean or the coefficient of variation criterion, the confidence level of the risk index of eating too fast / grabbing food is increased.

5. The method according to claim 1, characterized in that, The aforementioned behavioral risk identification in psychiatry also includes identification of eating while in a supine position, which includes: The trunk tilt angle is calculated from the trunk inertial measurement signal and the body position is classified. When the trunk tilt angle is within the preset lying angle range and overlaps with the swallowing event time window, the lying position eating risk index is output. The range of the supine angle is 30° to 80°.

6. The method according to claim 1, characterized in that, The extrapyramidal response characterization vector includes: Energy characteristics of the 4Hz–7Hz flutter spectrum peaks obtained from the spectrum of the inertial measurement signal; Baseline elevation characteristics and common contraction ratio characteristics obtained from surface electromyography signals; Motion initiation delay characteristics obtained from motion-related signals; Furthermore, the classification model outputs the probability of swallowing difficulties related to extrapyramidal reactions, which is used to form an EPS-related swallowing risk index.

7. The method according to claim 1, characterized in that, The classification model is a classification model obtained based on supervised learning training. The input of the classification model includes a concatenated vector of the extrapyramidal response representation vector and the swallowing feature vector. During the training phase, information on changes in drug dosage is introduced as a conditional feature or sample stratification feature to improve the stability of identifying drug-induced EPS-related swallowing disorders.

8. The method according to claim 1, characterized in that, The swallowing reflex abnormality index is obtained from the airway protection abnormality detection, which includes any of the following: Cluster detection of cough events following swallowing events; Detection of abnormal respiratory phase switching after swallowing events; Detection of a drop in blood oxygen saturation exceeding a threshold after a swallowing event; Detection of wet sounds / residual features in swallowing acoustic signals exceeding a threshold; The test results are then integrated to form a risk index for insufficient reflection / airway protection. In the swallowing event detection and / or the airway protection abnormality detection, anti-artifact processing is performed on artifacts caused by psychiatric patients speaking, shouting, or agitation. The anti-artifact processing includes: when an acoustic event meets the characteristics of speech bands but lacks swallowing consistency characteristics of surface electromyography and inertial measurements, the contribution of the acoustic event to the swallowing event sequence and risk indicators is suppressed.

9. The method according to claim 1, characterized in that, The item-level output corresponding to the nursing risk assessment item is the item risk index rather than a fixed score. The item risk index includes EPS-related swallowing risk index, behavioral feeding risk index, supine eating risk index, and reflex / airway protection inadequate risk index. Furthermore, the total score or comprehensive risk value is a comprehensive risk value obtained by normalizing and fusing the risk indicators of the items. The fusion includes any one of weighted summation, logistic regression fusion, or Bayesian fusion, wherein the weights or fusion parameters are learned from training data or adaptively updated from individual baselines. The comprehensive risk value is used to output the risk level of aspiration / choking, which includes Level I, Level II and Level III. The assessment report further outputs the dominant risk factors that lead to the risk level. The dominant risk factors are determined by ranking the sensitivity contribution or confidence contribution of each risk indicator.

10. A swallowing function assessment system for mental patients based on multidimensional physiological signals, characterized in that, The method applied to any one of claims 1-9 includes: The data collection module is used to acquire basic information and medication information of the mental patients to be evaluated, and to deploy and synchronously collect multidimensional physiological signals for the mental patients. The multidimensional physiological signals include surface electromyography signals of the anterior cervical region, cervical inertial measurement signals, and swallowing acoustic signals. The vector extraction module is used to detect swallowing events based on the multidimensional physiological signals, obtain a swallowing event sequence, and extract swallowing feature vectors for each swallowing event. The risk identification module is used to identify psychiatric behavioral risks based on the swallowing event sequence and the inertial measurement signal, and to obtain behavioral risk labels and their confidence levels. The swallowing analysis module is used to extract extrapyramidal response characterization vectors based on the multidimensional physiological signals, and input the extrapyramidal response characterization vectors and swallowing feature vectors into a preset subtyping model to output the probability of swallowing difficulties related to extrapyramidal responses. The swallowing abnormality mapping module is used to map the behavioral risk label, the extrapyramidal reaction-related swallowing dysphagia, and the swallowing reflex abnormality index corresponding to the swallowing feature vector into item scores of nursing risk assessment items, and sum the item scores to obtain the total score; The swallowing assessment module is used to output the risk level of aspiration / choking based on the total score and a preset grading threshold, and generate an assessment report that includes the item score, the total score, the risk level and key physiological evidence.

Images