Multidimensional assessment method and system for thirst level in hemodialysis patients

Abstract

The present application relates to the technical field of medical monitoring, in particular to a multi-dimensional evaluation method and system for the thirst degree of hemodialysis patients, which collects the saliva osmotic pressure time series signal, the pharyngeal muscle electromyogram signal and the digital analog thirst scale score filled in by the patient from a plurality of sensing terminals; accordingly, a thirst physiological marker graph reflecting physiological thirst sensation driving is established, and the scale score is checked to obtain a patient thirst state benchmark vector. The natural dialogue voice stream and the neck surface video stream of the patient during the dialysis interval are synchronously captured, from which the thirst associated keyword frequency, the voice dryness degree characteristics, and the frequency and amplitude profile characteristics of the swallowing action are analyzed to generate a behavior representation index set. The thirst state benchmark vector and the behavior representation index set are cross-modally associated and mapped to construct a dynamic thirst evaluation network, through processing the continuously input real-time data, an individualized thirst degree trajectory evolving over time is output, and objective, continuous and multi-dimensional thirst state evaluation is realized.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a multidimensional assessment method and system for the thirst level of hemodialysis patients. Background Technology

[0002] Patients undergoing maintenance hemodialysis commonly experience chronic thirst due to strict fluid restriction. Accurate assessment of their thirst levels is of direct clinical significance for adjusting dialysis protocols and implementing drinking behavior interventions. Current assessment techniques primarily rely on patient self-reporting, using various thirst scales for periodic or event-triggered scoring. These methods simply map the complex perception of thirst into a single number or level, and their accuracy is easily affected by patient cognitive function, emotional state, willingness to report, and memory biases, leading to significant subjectivity and lag in clinical judgment.

[0003] Further technological attempts have introduced single physiological parameters as objective supplements, such as monitoring saliva secretion or blood biochemical indicators. However, thirst is a complex state driven by both physiological factors and behavioral-psychological regulation. A single physiological parameter can only reflect a momentary slice of the thirst-driven circuit, failing to characterize its dynamic fluctuation patterns and making it even more difficult to capture the unconscious behavioral regulation that patients use to alleviate discomfort. Existing methods suffer from information fragmentation; physiological signals, subjective reports, and behavioral manifestations are isolated from each other, lacking an effective fusion and verification mechanism. This results in biased assessments, failing to distinguish between the intensity of physiological thirst and the subjective thirst distress that patients actually feel and attempt to alleviate through behavior, and also failing to achieve continuous, uninterrupted monitoring over the long interdialysis intervals. This invention aims to overcome the limitations of relying on subjective reports and construct an objective assessment system that can simultaneously reflect physiological drivers and behavioral expressions. The key lies in establishing a correlation mapping method that connects internal physiological fluctuations with overt behavioral patterns. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies by proposing a multidimensional assessment method and system for the thirst level of hemodialysis patients.

[0005] To achieve the above objectives, the present invention employs the following technical solution: a multidimensional assessment method for the thirst level of hemodialysis patients, comprising:

[0006] The salivary osmotic pressure timing signal, pharyngeal muscle electromyography signal, and self-filled digital simulated thirst scale scores of hemodialysis patients were collected from multiple sensor terminals.

[0007] Based on the fluctuation pattern of salivary osmolality time-series signal and the activation cycle of pharyngeal muscle electromyography signal, a thirst physiological marker atlas reflecting physiological thirst drive was established.

[0008] The thirst physiological marker atlas was hierarchically verified with the digital simulation thirst scale score to obtain a baseline vector of patient thirst status containing verification weights.

[0009] Simultaneously capture the patient's natural conversational speech stream and the neck surface video stream of unconscious swallowing movements during the interdialysis period;

[0010] The frequency and speech dryness features of thirst-related keywords were extracted from natural dialogue speech streams, and the frequency and amplitude contour features of swallowing movements were extracted from neck surface video streams.

[0011] By integrating the frequency of the thirst-related keywords, the speech dryness features, the frequency of the swallowing action, and the amplitude contour features, a set of behavioral representation indicators is generated.

[0012] A dynamic thirst assessment network is constructed by cross-modal association mapping between the patient's thirst status baseline vector and the behavioral representation index set.

[0013] The dynamic thirst assessment network processes continuously input real-time sensor data and behavioral data to generate an individualized thirst trajectory that evolves over time.

[0014] As a further aspect of the present invention, the step of establishing a thirst physiological marker atlas reflecting physiological thirst drive based on the fluctuation pattern of salivary osmotic pressure timing signal and the activation cycle of pharyngeal muscle electromyography signal specifically includes:

[0015] Multiscale fractal analysis was performed on the salivary osmolality time series signal to calculate its Hearst exponent at different time scales, so as to quantify the long-range dependence and self-similarity of the signal and identify specific fluctuation patterns related to body fluid imbalance.

[0016] The electromyography signals of the pharyngeal muscles were decomposed into motor unit action potentials to extract the activation cycle, recruitment sequence, and discharge frequency codes corresponding to the swallowing preparation and execution phases.

[0017] The identified specific fluctuation pattern is analyzed in the time domain with the extracted activation period, recruitment order and discharge frequency code to detect the phase-locked relationship between the specific fluctuation pattern and the activation period, recruitment order and discharge frequency code;

[0018] Based on the phase-locked relationship, a two-dimensional mapping plane is generated. In the two-dimensional mapping plane, the intensity of the specific wave pattern is used as one dimension and the regularity of the activation cycle is used as another dimension to partition the physiological thirst-driven state.

[0019] Each partition is defined as a type of thirst physiological marker, and the set of all partitions constitutes the thirst physiological marker atlas.

[0020] As a further aspect of the present invention, the hierarchical verification of the thirst physiological marker map and the digital simulated thirst scale score to obtain a patient thirst state baseline vector containing verification weights is specifically as follows:

[0021] Obtain the patient's self-reported digital thirst scale score during the same time period when the salivary osmolality timing signal and the pharyngeal muscle electromyography signal were collected;

[0022] Each thirst physiological marker in the thirst physiological marker atlas is paired with the corresponding digital simulated thirst scale score;

[0023] Establish a discriminant model to analyze the consistency and discrepancy between the classification results of thirst physiological markers in each pair of paired data and the numerical simulation thirst scale scores;

[0024] An initial confidence weight is assigned to each type of thirst physiological marker based on the consistency and divergence.

[0025] A time sliding window is introduced to statistically analyze the stability coefficient of the initial confidence weight of the same thirst physiological marker within a continuous time window;

[0026] The initial confidence weights are dynamically adjusted based on the stability coefficients to generate the final verification weights.

[0027] Using each marker category in the thirst physiological marker atlas as a basis, and combining it with the corresponding final verification weight, a weighted synthesis is performed to output the baseline vector of the patient's thirst state.

[0028] As a further aspect of the present invention, the step of parsing the frequency and speech dryness features of thirst-related keywords from natural dialogue speech streams, and extracting the frequency and amplitude contour features of swallowing movements from neck surface video streams, specifically includes:

[0029] The natural dialogue speech stream is automatically recognized and converted into a text sequence. The text sequence is then matched in a preset thirst-related word library, and the frequency of occurrence of thirst-related keywords is counted per unit time.

[0030] Simultaneously, subband spectrum centroid analysis is performed on the original audio signal of the natural dialogue speech stream to calculate the mean and variance of the energy ratio between the high-frequency subband and the low-frequency subband. The mean and variance are jointly defined as the speech dryness feature.

[0031] A dense trajectory detection algorithm is applied to the video stream of the neck surface to track the movement trajectory of the Adam's apple and its surrounding area in three-dimensional spacetime.

[0032] Based on the motion trajectory, periodic motion patterns that conform to the swallowing biomechanical model are identified in the trajectory, and the number of times the periodic motion pattern occurs per unit time is calculated as the frequency of the swallowing action.

[0033] For each identified periodic motion pattern, the maximum displacement amplitude of its motion trajectory and the peak value of its motion velocity are quantified to form an amplitude profile feature describing the intensity of a single swallowing action.

[0034] As a further aspect of the present invention, the step of fusing the frequency of the thirst-related keywords, the speech dryness features, the frequency of the swallowing action, and the amplitude contour features to generate a behavioral representation index set specifically includes:

[0035] The frequency of the thirst-related keywords per unit time is normalized to a value between zero and one to obtain the keyword frequency index.

[0036] The mean and variance of the speech dryness features are transformed into a single speech dryness index through linear combination.

[0037] The frequency of the swallowing action is logarithmically transformed to smooth out extreme values, resulting in a swallowing frequency index.

[0038] Principal component analysis was performed on the maximum displacement amplitude and the peak value of the motion velocity in the amplitude contour features to reduce the dimensionality and obtain the principal component score of swallowing force.

[0039] A behavioral feature fusion space is constructed, and the keyword frequency index, the speech dryness index, the swallowing frequency index, and the principal component score of swallowing intensity are used as four orthogonal bases of the behavioral feature fusion space.

[0040] Calculate the coordinates of the four index values ​​at each sampling time point in the behavioral feature fusion space;

[0041] The sequence of all coordinate points within a continuous time period and its density distribution evolving over time are defined as the behavioral characterization index set.

[0042] As a further aspect of the present invention, the step of performing cross-modal correlation mapping between the patient's thirst state baseline vector and the behavioral representation index set to construct a dynamic thirst assessment network specifically includes:

[0043] Design a dual-channel attention coupling architecture, where one channel is used to process the baseline vector of the patient's thirst state and the other channel is used to process the set of behavioral representation indicators;

[0044] In the channel processing the patient's thirst state baseline vector, a gated recurrent unit network is used to capture the temporal dependencies of different physiological markers in the patient's thirst state baseline vector;

[0045] In the channel processing the behavioral representation index set, a convolutional neural network is used to extract local and global patterns of the coordinate point sequence in the behavioral representation index set;

[0046] A cross-attention mechanism is introduced in the middle layer between the two channels, which enables the physiological channel to focus on the behavioral patterns in the behavioral channel that are most relevant to the current physiological state, while the behavioral channel can focus on the physiological markers in the physiological channel that best match the current behavior.

[0047] The cross-attention mechanism is used to calculate the dynamic correlation weight matrix between physiological features and behavioral features;

[0048] Based on the dynamic correlation weight matrix, the feature representations of the two channel outputs are adaptively weighted and fused to form a unified cross-modal state representation;

[0049] The cross-modal state representation is input into a fully connected layer sequence, and the nonlinear mapping relationship from fused features to thirst scale is learned through the fully connected layer sequence. The entire architecture constitutes the dynamic thirst assessment network.

[0050] As a further aspect of the present invention, the step of processing continuously input real-time sensor data and behavioral data through the dynamic thirst assessment network to generate an individualized thirst level trajectory that evolves over time specifically involves:

[0051] The real-time collected salivary osmolality time-series signal and pharyngeal muscle electromyography signal are input into the thirst physiological marker atlas generation module to update the patient's thirst state baseline vector;

[0052] The real-time acquired natural dialogue speech stream and neck surface video stream are input into the behavior feature analysis module to update the behavior representation index set;

[0053] The updated patient thirst status baseline vector and the updated behavioral representation index set are synchronously input into the trained dynamic thirst assessment network.

[0054] Within the dynamic thirst assessment network, the dynamic association weight matrix is ​​calculated in real time through the dual-channel attention coupling architecture, and cross-modal feature fusion is performed.

[0055] The fully connected layer sequence outputs an estimate of the instantaneous thirst level at the current sampling moment based on the fused cross-modal state representation.

[0056] Record the instantaneous thirst level estimates at a series of consecutive sampling times, and connect them in chronological order to form the individualized thirst level trajectory that reflects the continuous change in thirst level.

[0057] As a further aspect of the present invention, the method for applying a dense trajectory detection algorithm to the video stream of the neck surface to track the motion trajectory of the Adam's apple and its surrounding area in three-dimensional spatiotemporal space is as follows:

[0058] In the initial frame of the video stream of the neck surface, the initial position coordinates of the Adam's apple are located using a key point detection model;

[0059] Using the initial position coordinates of the Adam's apple as the center, a rectangular region of interest is defined that includes the Adam's apple and its surrounding muscle tissue;

[0060] Within the rectangular region of interest, multiple feature points are sampled uniformly at a fixed grid spacing.

[0061] For each frame of the video stream, calculate the optical flow vector of each feature point to obtain the displacement of each feature point between adjacent frames;

[0062] Based on the optical flow vector of each feature point, predict the position coordinates of each feature point in the next frame image;

[0063] For each feature point, its position coordinates in the continuous frame sequence are connected in chronological order to form the motion trajectory of the feature point;

[0064] Remove feature points whose trajectory length is less than a preset trajectory length threshold, and retain feature points whose trajectory length is greater than or equal to the preset trajectory length threshold.

[0065] The set of motion trajectories of all retained feature points is taken as the motion trajectory of the Adam's apple and its surrounding area in three-dimensional spacetime.

[0066] As a further aspect of the present invention, the method for calculating the coordinates of the four index values ​​at each sampling time in the behavioral feature fusion space is as follows:

[0067] Obtain four orthogonal basis vectors in the behavioral feature fusion space, consisting of keyword frequency index, speech dryness index, swallowing frequency index, and swallowing force principal component score;

[0068] For each sampling time, the keyword frequency index, speech dryness index, swallowing frequency index, and swallowing force principal component score are obtained.

[0069] The first vector component is obtained by performing a scalar multiplication operation between the keyword frequency index value and the orthogonal basis vector corresponding to the keyword frequency index.

[0070] The second vector component is obtained by performing a scalar multiplication operation between the spoken language dryness index value and the orthogonal basis vector corresponding to the spoken language dryness index.

[0071] The third vector component is obtained by performing a scalar multiplication operation between the swallowing frequency index value and the orthogonal basis vector corresponding to the swallowing frequency index.

[0072] The fourth vector component is obtained by performing a scalar multiplication operation between the principal component score of swallowing force and the orthogonal basis vector corresponding to the principal component score of swallowing force.

[0073] The first, second, third, and fourth vector components are added together to obtain the coordinates of the sampling time in the behavioral feature fusion space.

[0074] As a further aspect of the present invention, the present invention also includes a multidimensional assessment system for the thirst level of hemodialysis patients, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described multidimensional assessment method for the thirst level of hemodialysis patients.

[0075] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0076] By analyzing the temporal fluctuation patterns of salivary osmolality signals and the periodic activation characteristics of pharyngeal muscle electromyography (EMG), a physiological marker atlas of thirst is constructed by combining the two. This method moves beyond reliance on single instantaneous physiological values ​​or subjective responses, instead characterizing the dynamic synergistic patterns of the core physiological mechanisms driving thirst. Salivary osmolality is directly related to oral mucosal fluid balance, while pharyngeal EMG reflects unconscious swallowing preparation; the temporal correlation patterns between the two can more fundamentally characterize the generation and intensity changes of physiological thirst signals. Based on this atlas and a hierarchical verification with subjective scales, the obtained baseline vector not only contains quantitative information on physiological drives but also calibrates the individualized difference weights between physiological sensations and subjective cognitive reports, giving the assessed physiological dimensions individualized interpretability and higher objectivity.

[0077] Simultaneously capturing patients' natural conversational speech and neck videos of unconscious swallowing, multimodal features are extracted from these two types of non-invasive, readily available behavioral signals. The frequency of mentions of specific semantic keywords and acoustic dryness characteristics are analyzed from the speech stream, while the dynamic contours of swallowing movements are quantified from the video stream. The behavioral representation index set generated by fusing these features essentially captures the unconscious or habitual traces left by patients in their verbal expressions and bodily movements in response to thirst. This behavioral data is collected continuously and passively in the patient's daily environment, avoiding response biases caused by subjective questionnaires and truly reflecting the penetration and influence of thirst in the patient's daily behavior. By performing a cross-modal correlation mapping between the aforementioned physiological baseline vector and this behavioral index set, the constructed dynamic network can learn the complex correspondence between an individual's unique physiological thirst signals and diverse overt behaviors. This mapping enables the system to infer and track the continuous evolution trajectory of the internal thirst state based on continuously input, easily monitored behavioral data, realizing a shift from discrete, active assessment to continuous, non-intrusive, objective monitoring. Attached Figure Description

[0078] Figure 1 The flowchart is a multidimensional assessment method and system for thirst in hemodialysis patients as described in this invention;

[0079] Figure 2 A flowchart for obtaining the baseline vector of a patient's thirst state;

[0080] Figure 3 Contour plot for kernel density estimation of behavioral feature fusion space;

[0081] Figure 4 Physiological marker weighting analysis plot for assessing thirst in hemodialysis patients. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0083] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0084] Please see Figure 1 A multidimensional assessment method for thirst in hemodialysis patients includes the following steps: collecting salivary osmolality time-series signals, pharyngeal muscle electromyography signals, and self-reported digital simulated thirst scale scores from multiple sensor terminals; establishing a thirst physiological marker atlas reflecting physiological thirst drive based on the fluctuation pattern of the salivary osmolality time-series signals and the activation cycle of the pharyngeal muscle electromyography signals; performing hierarchical verification between the thirst physiological marker atlas and the digital simulated thirst scale scores to obtain a baseline vector of patient thirst status including verification weights; and simultaneously capturing the patient's natural conversational speech and unconscious speech during interdialysis intervals. The system extracts video streams of the neck surface during swallowing; it also extracts frequency and speech dryness features of thirst-related keywords from natural conversational speech streams, and frequency and amplitude contour features of swallowing movements from the neck surface video stream; it integrates the frequency of thirst-related keywords, speech dryness features, and frequency and amplitude contour features of swallowing movements to generate a set of behavioral representation indicators; it performs cross-modal association mapping between the patient's thirst state baseline vector and the set of behavioral representation indicators to construct a dynamic thirst assessment network; and it processes continuously input real-time sensor data and behavioral data through the dynamic thirst assessment network to generate an individualized thirst trajectory that evolves over time.

[0085] In one embodiment of the present invention, multi-scale fractal analysis is performed on the salivary osmolality time-series signal to calculate its Hearst exponent at different time scales, so as to quantify the long-range dependence and self-similarity of the signal and identify specific fluctuation patterns related to fluid imbalance; motor unit action potential decomposition is performed on the pharyngeal muscle electromyography signal to extract the activation cycle, recruitment sequence, and discharge frequency encoding corresponding to the swallowing preparation and execution phases; temporal synchronization analysis is performed on the identified specific fluctuation patterns and the extracted activation cycle, recruitment sequence, and discharge frequency encoding to detect the phase-locked relationship between the specific fluctuation patterns and the activation cycle, recruitment sequence, and discharge frequency encoding; based on the phase-locked relationship, a two-dimensional mapping plane is generated, in which the intensity of the specific fluctuation pattern is used as one dimension and the regularity of the activation cycle is used as another dimension to partition the physiological thirst-driven state; each partition is defined as a type of thirst physiological marker, and the set of all partitions constitutes a thirst physiological marker atlas.

[0086] In practice, salivary osmolality time-series signals are continuously acquired at a preset sampling frequency using implanted or wearable osmolality sensors, while pharyngeal muscle electromyography signals are synchronously acquired using a surface electrode array attached to specific anatomical locations in the anterior neck. The multi-scale fractal analysis of the salivary osmolality time-series signals is performed using a detrended fluctuation analysis method. This method divides the salivary osmolality time-series signal into multiple non-overlapping time windows of varying lengths. Within each window, the local trend of the signal is calculated and removed, and then the fluctuation function of the detrended sequence is calculated. The slope of the linear relationship between the fluctuation function and the window length on a logarithmic coordinate system is the Hearst exponent. The formula for calculating the Hearst exponent of the salivary osmolality time-series signal is as follows:

[0087]

[0088] in: Represents the length of the time window. Indicates the window length is The fluctuation function value at time, This indicates that the length that can be divided within the entire length of the signal is... The total number of windows, Indicates the first The integral profile sequence of the salivary osmotic pressure time-series signal within a window. This represents the first value obtained by fitting using the least squares method. Local trend lines within a window. Calculated by multiple different... corresponding And analyze and The linear relationship was obtained to obtain the Hearst exponent, which characterizes the long-range dependence and self-similarity of the salivary osmolality time series signal. The variation patterns of the Hearst exponent at different scales were identified as specific fluctuation patterns related to fluid imbalance.

[0089] In some embodiments, the decomposition of motor unit action potentials in the pharyngeal muscle electromyography (EMG) signal is accomplished by combining a blind source separation algorithm and a template matching algorithm. The blind source separation algorithm separates several independent motor unit action potential sequences from the original pharyngeal muscle EMG signal, while the template matching algorithm clusters and identifies the separated action potential waveforms to extract the activation time point sequence of each independent motor unit. Based on the activation time point sequence, the interval between adjacent activation time points is further calculated to obtain the activation cycle; the sequential relationship of activation time points of different motor units is analyzed to determine the recruitment order; and the number of activations of each motor unit per unit time is counted to obtain the discharge frequency code. These features are clearly corresponding to the preparatory and execution phases of swallowing.

[0090] Optionally, the time-domain synchronization analysis of the identified specific fluctuation patterns with the extracted activation cycles, recruitment sequences, and discharge frequency codes is performed using a phase-locked value calculation method. Phase-locked value calculation first requires converting the specific fluctuation pattern segments of the salivary osmolality time-series signal and the activation event sequences extracted from the pharyngeal muscle electromyography signal into phase sequences. Then, the consistency of the phase difference distribution between the two phase sequences within a specific time window is calculated; a high degree of consistency indicates a phase-locked relationship. It can be understood that a phase-locked value close to 1 indicates strong phase-locking, and close to 0 indicates no phase-locking. A threshold is set to detect whether a significant phase-locked relationship exists between the specific fluctuation patterns and the activation cycles, recruitment sequences, and discharge frequency codes.

[0091] In some embodiments, a two-dimensional mapping plane is generated based on the detected phase-locked relationship. Specifically, the horizontal axis of the two-dimensional mapping plane is defined as the intensity of a specific fluctuation pattern, which can be quantified by the mean or variance of the Hearst exponent over the duration of the pattern; the vertical axis is defined as the regularity of the activation cycle, which can be quantified by the reciprocal of the coefficient of variation of the activation cycle sequence or the first positive peak of the autocorrelation function. Based on historical data or prior clinical knowledge, multiple well-defined regions are delineated on the two-dimensional mapping plane, such as "high fluctuation intensity - high periodicity" regions, "low fluctuation intensity - low periodicity" regions, and other combinations thereof. Each delineated region is defined as a type of thirst physiological marker, used to characterize a specific physiological thirst-driven state; the complete set of all regions constitutes a thirst physiological marker atlas for subsequent steps.

[0092] In one embodiment of the present invention, see [reference] Figure 2 The study acquires the patient's self-reported digital thirst scale score during the same time period of collecting salivary osmolality and pharyngeal muscle electromyography signals. Each thirst physiological marker in the thirst physiological marker atlas is paired with its corresponding digital thirst scale score. A discriminant model is established to analyze the consistency and discrepancy between the classification results of the thirst physiological markers and the digital thirst scale scores in each pair of paired data. An initial confidence weight is assigned to each type of thirst physiological marker based on the consistency and discrepancy. A time sliding window is introduced to statistically analyze the stability coefficient of the initial confidence weight of the same thirst physiological marker within a continuous time window. The initial confidence weight is dynamically adjusted based on the stability coefficient to generate the final verification weight. Using each marker category in the thirst physiological marker atlas as a basis, combined with its corresponding final verification weight, a weighted synthesis is performed to output the patient's thirst state baseline vector.

[0093] In practice, obtaining the patient's self-reported digital thirst scale score during the same time period of collecting salivary osmolality time-series signals and pharyngeal muscle electromyography signals is accomplished by simultaneously displaying the digital thirst scale on the sensor data acquisition interface and recording the patient's submitted score. In practice, pairing each thirst physiological marker in the thirst physiological marker atlas with its corresponding digital thirst scale score is performed according to the timestamp alignment principle. That is, for each thirst physiological marker generated by the joint analysis of salivary osmolality time-series signals and pharyngeal muscle electromyography signals, the digital thirst scale score with the highest overlap in the time window is matched, forming a "marker-score" data pair.

[0094] In some embodiments, a discriminant model is established to analyze the consistency and divergence between the classification results of thirst physiological markers in each pair of paired data and the numerical scores of a digital simulated thirst scale. This discriminant model is implemented using a support vector machine or logistic regression model. The input of the discriminant model is a feature vector representing the thirst physiological marker, and the output is a classification prediction of the degree of thirst. The numerical simulated thirst scale scores are discretized into identical classification labels according to a preset threshold. The classification accuracy is calculated as a consistency measure by comparing the predicted classification of the discriminant model with the actual classification based on the numerical simulated thirst scale scores. Simultaneously, the mean absolute error between the classification probability output by the discriminant model and the normalized numerical scores of the numerical simulated thirst scale is calculated as a divergence measure. The formula for assigning an initial confidence weight to each class of thirst physiological markers based on consistency and divergence is as follows:

[0095]

[0096] in: Representing the Initial confidence weights for thirst-like physiological markers Representative discriminant model in the first Classification accuracy of thirst-like physiological markers across all samples Representing the The mean absolute error between the predicted probability and the normalized score of the thirst-like physiological markers across all samples. It is a preset harmonic coefficient between 0 and 1, used to balance the relative importance of consistency and divergence measures.

[0097] Optionally, a time-sliding window is introduced to statistically determine the stability coefficient of the initial confidence weights of the same thirst physiological marker within consecutive time windows. The length of the time-sliding window is set according to the dialysis cycle or clinical observation interval. The stability coefficient is quantified by calculating the inverse of the standard deviation of the initial confidence weights of the same type of thirst physiological marker within multiple consecutive time-sliding windows, or by calculating the coefficients of its autoregressive model. The initial confidence weights are dynamically adjusted based on the stability coefficient to generate the final validation weights. The dynamic adjustment can be achieved through linear scaling or mapping using a sigmoid function. For example, the stability coefficient can be multiplied by the initial confidence weights as a gain factor. When the stability coefficient is high, the initial confidence weights are enhanced; when the stability coefficient is low, the initial confidence weights are suppressed.

[0098] It can be understood that using each marker category in the thirst physiological marker atlas as a basis, and combining it with its corresponding final verification weights to output a patient thirst state baseline vector, treats each thirst physiological marker as an independent dimension. In some embodiments, the patient thirst state baseline vector is a multi-dimensional vector, the number of which is equal to the total number of marker categories defined in the thirst physiological marker atlas. The component value of the vector in each dimension is obtained by multiplying the binary indicator of whether that thirst physiological marker is activated at the current time (1 for activation, 0 for inactivation) with the final verification weight corresponding to that marker. It can be understood that weighted synthesis is to arrange the component values ​​calculated in all dimensions in order to form a patient thirst state baseline vector that can comprehensively reflect the current physiological thirst state and has been weighted by subjective report verification.

[0099] In one embodiment of the present invention, automatic speech recognition is performed on the natural dialogue speech stream, which is converted into a text sequence. The text sequence is matched in a preset thirst-related word library, and the frequency of occurrence of thirst-related keywords per unit time is counted. At the same time, subband spectral centroid analysis is performed on the original audio signal of the natural dialogue speech stream to calculate the mean and variance of the energy ratio of high-frequency subbands and low-frequency subbands. The mean and variance are jointly defined as speech dryness features. A dense trajectory detection algorithm is applied to the neck surface video stream to track the motion trajectory of the larynx and its surrounding area in three-dimensional spatiotemporal space. Based on the motion trajectory, periodic motion patterns that conform to the swallowing biomechanical model are identified in the trajectory, and the number of occurrences of the periodic motion pattern per unit time is calculated as the frequency of swallowing action. For each identified periodic motion pattern, the maximum displacement amplitude of its motion trajectory and the peak value of its motion velocity are quantified to form an amplitude contour feature describing the intensity of a single swallowing action. In the initial frame of the video stream of the neck surface, the initial position coordinates of the larynx are located using a keypoint detection model. A rectangular region of interest (ROI) encompassing the larynx and its surrounding muscle tissue is defined centered on the initial position coordinates of the larynx. Within this ROI, multiple feature points are uniformly sampled at a fixed grid spacing. For each frame of the video stream, the optical flow vector of each feature point is calculated to obtain the displacement of each feature point between adjacent frames. Based on the optical flow vector of each feature point, the position coordinates of each feature point in the next frame are predicted. For each feature point, its position coordinates in the consecutive frame sequence are connected in chronological order to form its motion trajectory. Feature points with a trajectory length less than a preset trajectory length threshold are discarded, while those with a trajectory length greater than or equal to the preset trajectory length threshold are retained. The set of motion trajectories of all retained feature points is taken as the motion trajectory of the larynx and its surrounding region in three-dimensional spacetime.

[0100] In practice, the automatic speech recognition and conversion of natural dialogue speech streams into text sequences is accomplished by a speech recognition engine deployed locally or in the cloud. The speech recognition engine receives the real-time audio stream and outputs timestamp-aligned text transcription results. In this implementation, a pre-defined thirst-related vocabulary library contains a set of words and phrases directly or indirectly related to the perception of thirst, such as "thirsty," "wanting to drink water," "dry mouth," "dry lips," "dry throat," and words describing the act of drinking. The text sequence generated by automatic speech recognition is pattern-matched with the thirst-related vocabulary library, and the frequency of all matched thirst-related keywords is counted within a set unit time window.

[0101] In some embodiments, subband spectrum centroid analysis is performed on the original audio signal of a natural dialogue speech stream to calculate the mean and variance of the energy ratio between high-frequency and low-frequency subbands. Subband spectrum centroid analysis first decomposes the original audio signal into multiple subband signals using a set of bandpass filters. The mean and variance of the energy ratio between high-frequency and low-frequency subbands are calculated within each analysis frame. Energy of the body The subbands are then classified into high-frequency subband groups based on predefined division thresholds. With low-frequency subband group Energy ratio of high-frequency subband to low-frequency subband The calculation formula is:

[0102]

[0103] in: The set of indices representing all high-frequency subbands. A set of indices representing all low-frequency subbands. and These correspond to the energy of the sub-bands. Within a continuous analysis time window, the energy of all frames is calculated. The arithmetic mean of the values ​​is the average. Calculate all frames value relative to mean The variance is obtained by taking the average of the squared deviations. mean and variance They are collectively defined as speech dryness features.

[0104] It is understandable that applying a dense trajectory detection algorithm to a neck surface video stream to track the motion trajectory of the Adam's apple and its surrounding area in three-dimensional spacetime involves several steps. Specifically, in the initial frame of the neck surface video stream, a keypoint detection model is used to locate the initial coordinates of the Adam's apple. This keypoint detection model employs a deep learning-based human pose estimation network. Centered on the initial coordinates of the Adam's apple, a rectangular region of interest with fixed length and width is defined based on the prior dimensions of the human neck anatomy. This rectangular region of interest covers the surface projection areas of the Adam's apple and surrounding muscles that may participate in swallowing, such as the thyrohyoid and sternohyoid muscles. Within this rectangular region of interest, multiple feature points are uniformly sampled at a fixed grid spacing. The number of feature points is determined by the grid spacing, which is set according to the video resolution and tracking accuracy requirements.

[0105] In some embodiments, for each frame of the video stream, the optical flow vector of each feature point is calculated to obtain the displacement of each feature point between adjacent frames. The optical flow vector is calculated using the Farneback dense optical flow algorithm or the Lucas-Kanade sparse optical flow algorithm. Based on the optical flow vector of each feature point, the position coordinates of each feature point in the next frame are predicted. The prediction formula is the current frame coordinates plus the optical flow vector displacement. For each feature point, its position coordinates in the continuous frame sequence are connected in chronological order to form the feature point's motion trajectory. Feature points whose trajectory length is less than a preset trajectory length threshold are discarded. The preset trajectory length threshold is used to filter out transient false trajectories caused by noise, occlusion, or tracking failure. Feature points whose trajectory length is greater than or equal to the preset trajectory length threshold are retained. The set of motion trajectories of all retained feature points is taken as the motion trajectory of the throat and its surrounding area in three-dimensional spacetime.

[0106] Optionally, identifying periodic motion patterns in the trajectory that conform to the swallowing biomechanical model is achieved by analyzing the periodic peaks of the displacement-time curve in the vertical direction of the motion trajectory. The swallowing biomechanical model defines a typical swallowing action with the Adam's apple rising, moving forward, and then returning to its original position. Candidate swallowing events are identified by matching the waveform characteristics of this pattern. The number of times the identified candidate swallowing events occur per unit time is calculated as the frequency of the swallowing action. For each identified periodic motion pattern, the maximum displacement amplitude and peak velocity of its trajectory are quantified. The maximum displacement amplitude is obtained by calculating the Euclidean distance from the starting point to the highest point of the trajectory in three-dimensional space, and the peak velocity is obtained by calculating the maximum value of the ratio of displacement to time interval between adjacent points in the trajectory point sequence. The maximum displacement amplitude and the peak velocity together form an amplitude profile feature describing the intensity of a single swallowing action.

[0107] In one embodiment of the present invention, the frequency of thirst-related keywords per unit time is normalized to a value between zero and one to obtain a keyword frequency index; the mean and variance in the speech dryness feature are transformed into a single speech dryness index through linear combination; the frequency of swallowing actions is logarithmically transformed to smooth extreme values ​​to obtain a swallowing frequency index; principal component analysis is performed on the maximum displacement amplitude and peak value of movement speed in the amplitude contour feature to reduce dimensionality and obtain the swallowing force principal component score; a behavioral feature fusion space is constructed, and the keyword frequency index, speech dryness index, swallowing frequency index, and swallowing force principal component score are used as four orthogonal bases of the behavioral feature fusion space; the coordinates of the four index values ​​at each sampling time in the behavioral feature fusion space are calculated; the sequence of all coordinate points in continuous time and its density distribution over time are defined as the behavioral representation index set. Four orthogonal basis vectors, consisting of keyword frequency index, speech dryness index, swallowing frequency index, and swallowing strength principal component score, are obtained in the behavioral feature fusion space. For each sampling time, the keyword frequency index, speech dryness index, swallowing frequency index, and swallowing strength principal component score are obtained. The keyword frequency index value is multiplied by the orthogonal basis vector corresponding to the keyword frequency index to obtain the first vector component. The speech dryness index value is multiplied by the orthogonal basis vector corresponding to the speech dryness index to obtain the second vector component. The swallowing frequency index value is multiplied by the orthogonal basis vector corresponding to the swallowing frequency index to obtain the third vector component. The swallowing strength principal component score is multiplied by the orthogonal basis vector corresponding to the swallowing strength principal component score to obtain the fourth vector component. The first, second, third, and fourth vector components are added together to obtain the coordinates of the sampling time in the behavioral feature fusion space.

[0108] In practical implementation, the frequency of thirst-related keywords within a unit of time is normalized to a value between zero and one to obtain the keyword frequency index. The normalization operation is achieved through a min-max scaling method, which involves taking the maximum and minimum frequency values ​​statistically observed within a calibration period, subtracting the minimum frequency value from the real-time frequency value, and then dividing by the difference between the maximum and minimum frequency values, thus mapping the keyword frequency index to the range of zero to one. In practical implementation, the mean and variance of the speech dryness features are transformed into a single speech dryness index through linear combination. The weighting coefficients of the linear combination are pre-set based on the discriminative power of each component of the speech dryness features in distinguishing different oral cavity moisture states. The formula for calculating the speech dryness index is expressed as follows:

[0109]

[0110] in: Represents the level of speech dryness. This represents the average energy ratio of the high-frequency subband to the low-frequency subband. The variance representing the energy ratio of the high-frequency subband to the low-frequency subband. and It is a preset linear combination weighting coefficient used to balance the contribution ratio of the mean and variance components in the formation of the speech dryness index.

[0111] In some embodiments, the frequency of swallowing movements is logarithmically transformed to smooth out extreme values ​​to obtain a swallowing frequency exponent. The logarithmic transformation uses a logarithmic function with the natural constant e as its base, i.e., the swallowing frequency exponent equals... ,in: The original frequency value of the swallowing action identified per unit time is represented by the value of 1. The increment operation is to avoid the undefined logarithmic operation when the frequency is zero. Principal component analysis (PCA) is performed on the maximum displacement amplitude and the peak value of the movement velocity in the amplitude contour feature to obtain the swallowing force principal component score. PCA is trained on a historically collected amplitude contour feature dataset. The first principal component direction of the dataset is calculated. The two-dimensional vector composed of the real-time collected maximum displacement amplitude value and the peak value of the movement velocity is projected onto the first principal component direction. The scalar value obtained after projection is the swallowing force principal component score.

[0112] Optionally, a behavioral feature fusion space can be constructed using keyword frequency index, speech dryness index, swallowing frequency index, and swallowing intensity principal component score as four orthogonal bases. This means that the behavioral feature fusion space is a four-dimensional real space, with the four basis vectors being orthogonal to each other and having a magnitude of 1. Each basis vector corresponds to a standard unit direction of a behavioral indicator. The orthogonal basis vectors of the behavioral feature fusion space can usually be represented as standard unit vectors. For example, in a specific mathematical construction, referring to Table 1, a set of basis vectors as shown in Table 1 can be defined.

[0113] Table 1: Orthogonal basis vector table for behavioral feature fusion space

[0114]

[0115] In some embodiments, the calculation of the coordinates of the four index values ​​in the behavioral feature fusion space at each sampling time is performed based on the principle of linear representation of vector space. Four orthogonal basis vectors, consisting of the keyword frequency index, speech dryness index, swallowing frequency index, and swallowing strength principal component score, are obtained in the behavioral feature fusion space. For each sampling time, the keyword frequency index, speech dryness index, swallowing frequency index, and swallowing strength principal component score are obtained. The keyword frequency index value is multiplied by the orthogonal basis vector corresponding to the keyword frequency index to obtain the first vector component. The speech dryness index value is multiplied by the orthogonal basis vector corresponding to the speech dryness index to obtain the second vector component. The swallowing frequency index value is multiplied by the orthogonal basis vector corresponding to the swallowing frequency index to obtain the third vector component. The swallowing strength principal component score value is multiplied by the orthogonal basis vector corresponding to the swallowing strength principal component score to obtain the fourth vector component. The first, second, third, and fourth vector components are added together to obtain the coordinates of the sampling time in the behavioral feature fusion space. It can be understood that this coordinate point is a four-dimensional vector, and its four components are respectively equal to the keyword frequency index, the speech dryness index, the swallowing frequency index, and the principal component score of swallowing intensity at that moment.

[0116] It can be understood that the sequence of all coordinate points in a continuous time period and its density distribution evolving over time are defined as the behavioral representation index set. The behavioral representation index set not only includes the sequence of coordinate points arranged in chronological order, but also includes the probability density function obtained by estimating the kernel density of the coordinate point sequence in the behavioral feature fusion space. This probability density function describes the clustering pattern and trend of the patient's behavioral features in the fusion space within a specific time period.

[0117] See Figure 3 This is a kernel density estimation contour plot of the behavioral feature fusion space. It projects the four-dimensional behavioral features of hemodialysis patients (keyword frequency index, dry speech index, swallowing frequency index, and principal component score of swallowing force) onto a two-dimensional principal component space, and uses color and contour lines to show the probability density distribution of the patient's behavioral features. The central high-density area can serve as a baseline for assessing whether an individual patient's behavior is "normal." If a patient's behavioral feature points are far from the center, it suggests that their thirst-related behavior may be abnormal and requires further evaluation. The unimodal symmetrical distribution pattern verifies the rationality of the "behavioral feature fusion space" design, indicating that these four behavioral indicators can be effectively aggregated into a clinically meaningful potential dimension. This figure verifies the input feature quality of the dynamic thirst assessment network from the perspective of data distribution, providing important statistical basis for subsequent model training and result interpretation.

[0118] In one embodiment of the present invention, a dual-channel attention coupling architecture is designed, where one channel processes a baseline vector of a patient's thirst state, and the other channel processes a set of behavioral representation indicators. In the channel processing the baseline vector of the patient's thirst state, a gated recurrent unit network is used to capture the temporal dependencies of different physiological markers in the baseline vector of the patient's thirst state. In the channel processing the set of behavioral representation indicators, a convolutional neural network is used to extract local and global patterns of the coordinate point sequences in the set of behavioral representation indicators. A cross-attention mechanism is introduced in the middle layer of the two channels, enabling the physiological channel to focus on the behavioral patterns most relevant to the current physiological state in the behavioral channel, while the behavioral channel can focus on the physiological markers most matching the current behavior in the physiological channel. Through the cross-attention mechanism, a dynamic correlation weight matrix between physiological features and behavioral features is calculated. Based on the dynamic correlation weight matrix, the feature representations output by the two channels are adaptively weighted and fused to form a unified cross-modal state representation. The cross-modal state representation is input into a fully connected layer sequence, and the nonlinear mapping relationship from the fused features to the thirst severity scale is learned through the fully connected layer sequence. The entire architecture constitutes a dynamic thirst assessment network. Real-time collected salivary osmolality time-series signals and pharyngeal muscle electromyography signals are input into the thirst physiological marker atlas generation module to update the patient's thirst state baseline vector. Real-time collected natural dialogue speech streams and neck surface video streams are input into the behavioral feature analysis module to update the behavioral representation index set. The updated patient thirst state baseline vector and the updated behavioral representation index set are synchronously input into the trained dynamic thirst assessment network. Inside the dynamic thirst assessment network, a dynamic correlation weight matrix is ​​calculated in real time through a dual-channel attention coupling architecture, and cross-modal feature fusion is performed. The fully connected layer sequence outputs the instantaneous thirst level estimate at the current sampling moment based on the fused cross-modal state representation. A series of consecutive sampling moments of instantaneous thirst level estimates are recorded and connected in chronological order to form an individualized thirst level trajectory reflecting the continuous change in thirst level.

[0119] In the specific implementation, a dual-channel attention coupling architecture is designed. One channel processes the patient's thirst state baseline vector, and the other channel processes the behavioral representation index set. In the channel processing the patient's thirst state baseline vector, a gated recurrent unit network (GRN) is used to capture the temporal dependencies of different physiological markers in the patient's thirst state baseline vector. The number of neurons in the input layer of the GRN is consistent with the dimension of the patient's thirst state baseline vector, that is, equal to the total number of marker categories in the thirst physiological marker atlas. The GRN uses its update gate and reset gate mechanism to model the state transitions and information filtering of the patient's thirst state baseline vector over time. In the channel processing the behavioral representation index set, a convolutional neural network (CNN) is used to extract the local and global patterns of the coordinate point sequences in the behavioral representation index set. Since the coordinate points in the behavioral representation index set are four-dimensional vectors, the one-dimensional convolutional kernel of the CNN slides along the time dimension, performing convolution operations on the four component channels of the coordinate points to capture local temporal patterns. By stacking multiple convolutional layers and pooling layers, the receptive field is gradually expanded to integrate global contextual information.

[0120] In some embodiments, a cross-attention mechanism is introduced in the intermediate layer between the two channels, enabling the physiological channel to focus on the behavioral pattern most relevant to the current physiological state in the behavioral channel, while the behavioral channel focuses on the physiological marker that best matches the current behavior in the physiological channel. The specific implementation of the cross-attention mechanism involves first denoting the hidden state vector of the physiological channel-gated recurrent unit network at a certain time step as follows: The feature vector flattened from the feature map extracted by the behavior channel convolutional neural network within the corresponding time window is denoted as... Calculate the dynamic correlation weight matrix between physiological and behavioral characteristics. The formula is:

[0121]

[0122] in: and These are trainable linear transformation weight matrices, used to transform the physiological hidden state vector... and behavioral feature vector Projected into the query space and key space. It is the dimension of the key vector, used to scale the dot product result. The function is normalized along the row direction, resulting in a dynamic correlation weight matrix. The sum of the elements in each row is 1.

[0123] Optionally, the feature representations of the two channel outputs can be adaptively weighted and fused based on a dynamic correlation weight matrix to form a unified cross-modal state representation. The fusion process involves using the dynamic correlation weight matrix... After another linear transformation The projected behavioral feature vectors are weighted and summed to obtain the context vector of the behavioral channel to the physiological channel. Simultaneously, the transpose of the dynamic correlation weight matrix is ​​used to perform a weighted summation of the physiological feature value vectors to obtain the context vector of the physiological channel to the behavioral channel. ,in This is another set of trainable weights. Subsequently, the original physiological feature vector is concatenated or added to its corresponding context vector. The original behavioral feature vector is processed similarly, and then the feature vectors from the two channels are further concatenated to form a unified cross-modal state representation. .

[0124] The dynamic thirst assessment network can be understood as follows: Cross-modal state representations are input into a sequence of fully connected layers. This sequence learns a non-linear mapping from fused features to a thirst level scale, forming the overall architecture. The fully connected layer sequence typically contains multiple hidden layers with non-linear activation functions such as ReLU. The final output layer uses a linear activation function or a sigmoid function (when the output needs to be normalized to the 0-1 range), outputting a scalar value as a thirst level estimate. The entire dynamic thirst assessment network undergoes end-to-end supervised training using historically collected labeled data, and the loss function is typically the mean squared error loss.

[0125] In some embodiments, a dynamic thirst assessment network processes continuously input real-time sensor data and behavioral data to generate an individualized thirst trajectory that evolves over time. Real-time acquired salivary osmolality time-series signals and pharyngeal muscle electromyography signals are input to a thirst physiological marker atlas generation module to update the patient's thirst state baseline vector. Real-time acquired natural conversational speech streams and neck surface video streams are input to a behavioral feature analysis module to update the behavioral representation index set. The updated patient thirst state baseline vector and the updated behavioral representation index set are then synchronously input into the trained dynamic thirst assessment network.

[0126] Within the dynamic thirst assessment network, a dynamic correlation weight matrix is ​​calculated in real time using a dual-channel attention coupling architecture, and cross-modal feature fusion is performed. The fully connected layer sequence outputs an estimate of the instantaneous thirst level at the current sampling moment based on the fused cross-modal state representation. A series of consecutive sampling moments' instantaneous thirst level estimates are recorded and connected chronologically to form an individualized thirst level trajectory reflecting continuous changes in thirst level. This individualized thirst level trajectory, presented as a time series, demonstrates the dynamic fluctuations in a patient's thirst level between dialysis sessions or during specific monitoring periods, with each point on the trajectory corresponding to an assessment result at a sampling moment.

[0127] See Figure 4This is a weighted analysis chart of physiological markers for assessing thirst in hemodialysis patients. It shows the weight changes of four core physiological markers at different stages and their correlation with the Patient Self-Reported Thirst Scale (DAS) score. The chart clearly illustrates the complete validation process of "initial confidence weight → stability coefficient → final validation weight," verifying the scientific validity of the physiological marker weight calculation. The high degree of agreement between the DAS score and the final validation weight trends indicates that the physiological markers can effectively reflect the patient's subjective thirst experience, providing empirical evidence for the system's reliability. The parameters of the assessment network can be dynamically adjusted based on the weight performance of each marker; for example, a higher decision weight can be assigned to the "pharyngeal electromyography activation" marker, or the calculation method of the stability coefficient can be optimized for the "swallowing preparation phase" marker.

[0128] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multidimensional assessment method for thirst in hemodialysis patients, characterized in that, Includes the following steps: The salivary osmotic pressure timing signal, pharyngeal muscle electromyography signal, and self-filled digital simulated thirst scale scores of hemodialysis patients were collected from multiple sensor terminals. Based on the fluctuation pattern of salivary osmolality time-series signal and the activation cycle of pharyngeal muscle electromyography signal, a thirst physiological marker atlas reflecting physiological thirst drive is established. Specifically, the establishment of this thirst physiological marker atlas involves: Multiscale fractal analysis was performed on the salivary osmolality time series signal to calculate its Hearst exponent at different time scales, so as to quantify the long-range dependence and self-similarity of the signal and identify specific fluctuation patterns related to body fluid imbalance. The electromyography signals of the pharyngeal muscles were decomposed into motor unit action potentials to extract the activation cycle, recruitment sequence, and discharge frequency codes corresponding to the swallowing preparation and execution phases. The identified specific fluctuation pattern is subjected to time-domain synchronization analysis with the extracted activation period, recruitment order, and discharge frequency code to detect the phase-locked relationship between the specific fluctuation pattern and the activation period, recruitment order, and discharge frequency code. Based on the phase-locked relationship, a two-dimensional mapping plane is generated. In the two-dimensional mapping plane, the intensity of the specific wave pattern is used as one dimension and the regularity of the activation cycle is used as another dimension to partition the physiological thirst-driven state. Each partition is defined as a type of thirst physiological marker, and the set of all partitions constitutes the thirst physiological marker atlas; The thirst physiological marker atlas was hierarchically verified with the digital simulation thirst scale score to obtain a baseline vector of patient thirst status containing verification weights. Simultaneously capture the patient's natural conversational speech stream and the neck surface video stream of unconscious swallowing movements during the interdialysis period; The frequency and speech dryness features of thirst-related keywords were extracted from natural dialogue speech streams, and the frequency and amplitude contour features of swallowing movements were extracted from neck surface video streams. By integrating the frequency of the thirst-related keywords, the speech dryness features, the frequency of the swallowing action, and the amplitude contour features, a set of behavioral representation indicators is generated. A dynamic thirst assessment network is constructed by cross-modal association mapping between the patient's thirst status baseline vector and the behavioral representation index set. The dynamic thirst assessment network processes continuously input real-time sensor data and behavioral data to generate an individualized thirst trajectory that evolves over time.

2. The multidimensional assessment method for thirst in hemodialysis patients according to claim 1, characterized in that, The process of hierarchically verifying the thirst physiological marker atlas with the digital simulated thirst scale score to obtain a baseline vector of patient thirst status containing verification weights is as follows: Obtain the patient's self-reported digital thirst scale score during the same time period when the salivary osmolality timing signal and the pharyngeal muscle electromyography signal were collected; Each thirst physiological marker in the thirst physiological marker atlas is paired with the corresponding digital simulated thirst scale score; Establish a discriminant model to analyze the consistency and discrepancy between the classification results of thirst physiological markers in each pair of paired data and the numerical simulation thirst scale scores; An initial confidence weight is assigned to each type of thirst physiological marker based on the consistency and divergence. A time sliding window is introduced to statistically analyze the stability coefficient of the initial confidence weight of the same thirst physiological marker within a continuous time window; The initial confidence weights are dynamically adjusted based on the stability coefficients to generate the final verification weights. Using each marker category in the thirst physiological marker atlas as a basis, and combining it with the corresponding final verification weight, a weighted synthesis is performed to output the baseline vector of the patient's thirst state.

3. The multidimensional assessment method for thirst in hemodialysis patients according to claim 2, characterized in that, The process involves parsing the frequency and speech dryness features of thirst-related keywords from natural dialogue speech streams, and extracting the frequency and amplitude contour features of swallowing movements from neck surface video streams. Specifically: Automatic speech recognition is performed on the natural dialogue speech stream, which is converted into a text sequence. The text sequence is then matched in a preset thirst-related word library, and the frequency of occurrence of thirst-related keywords is counted per unit time. Simultaneously, subband spectrum centroid analysis is performed on the original audio signal of the natural dialogue speech stream to calculate the mean and variance of the energy ratio between the high-frequency subband and the low-frequency subband. The mean and variance are jointly defined as the speech dryness feature. A dense trajectory detection algorithm is applied to the video stream of the neck surface to track the movement trajectory of the Adam's apple and its surrounding area in three-dimensional spacetime. Based on the motion trajectory, periodic motion patterns that conform to the swallowing biomechanical model are identified in the trajectory, and the number of times the periodic motion pattern occurs per unit time is calculated as the frequency of the swallowing action. For each identified periodic motion pattern, the maximum displacement amplitude of its motion trajectory and the peak value of its motion velocity are quantified to form an amplitude profile feature describing the intensity of a single swallowing action.

4. The multidimensional assessment method for thirst in hemodialysis patients according to claim 3, characterized in that, The method of integrating the frequency of thirst-related keywords, the speech dryness features, the frequency of swallowing actions, and the amplitude contour features to generate a set of behavioral representation indicators, specifically: The frequency of the thirst-related keywords per unit time is normalized to a value between zero and one to obtain the keyword frequency index. The mean and variance of the speech dryness features are transformed into a single speech dryness index through linear combination. The frequency of the swallowing action is logarithmically transformed to smooth out extreme values, resulting in a swallowing frequency index. Principal component analysis was performed on the maximum displacement amplitude and the peak value of the motion velocity in the amplitude contour features to reduce the dimensionality and obtain the principal component score of swallowing force. A behavioral feature fusion space is constructed, and the keyword frequency index, the speech dryness index, the swallowing frequency index, and the principal component score of swallowing intensity are used as four orthogonal bases of the behavioral feature fusion space. Calculate the coordinates of the four index values ​​at each sampling time point in the behavioral feature fusion space; The sequence of all coordinate points within a continuous time period and its density distribution evolving over time are defined as the behavioral characterization index set.

5. The multidimensional assessment method for thirst in hemodialysis patients according to claim 4, characterized in that, The step of constructing a dynamic thirst assessment network by performing a cross-modal association mapping between the patient's thirst state baseline vector and the behavioral representation index set is as follows: Design a dual-channel attention coupling architecture, where one channel is used to process the baseline vector of the patient's thirst state and the other channel is used to process the set of behavioral representation indicators; In the channel processing the patient's thirst state baseline vector, a gated recurrent unit network is used to capture the temporal dependencies of different physiological markers in the patient's thirst state baseline vector; In the channel processing the behavioral representation index set, a convolutional neural network is used to extract local and global patterns of the coordinate point sequence in the behavioral representation index set; A cross-attention mechanism is introduced in the middle layer between the two channels, which enables the physiological channel to focus on the behavioral patterns in the behavioral channel that are most relevant to the current physiological state, while the behavioral channel can focus on the physiological markers in the physiological channel that best match the current behavior. The cross-attention mechanism is used to calculate the dynamic correlation weight matrix between physiological features and behavioral features; Based on the dynamic correlation weight matrix, the feature representations of the two channel outputs are adaptively weighted and fused to form a unified cross-modal state representation; The cross-modal state representation is input into a fully connected layer sequence, and the nonlinear mapping relationship from fused features to thirst scale is learned through the fully connected layer sequence. The entire architecture constitutes the dynamic thirst assessment network.

6. The multidimensional assessment method for thirst in hemodialysis patients according to claim 5, characterized in that, The process of processing continuously input real-time sensor data and behavioral data through the dynamic thirst assessment network to generate an individualized thirst trajectory that evolves over time is specifically as follows: The real-time collected salivary osmolality time-series signal and pharyngeal muscle electromyography signal are input into the thirst physiological marker atlas generation module to update the patient's thirst state baseline vector; The real-time acquired natural dialogue speech stream and neck surface video stream are input into the behavior feature analysis module to update the behavior representation index set; The updated patient thirst status baseline vector and the updated behavioral representation index set are synchronously input into the trained dynamic thirst assessment network. Within the dynamic thirst assessment network, the dynamic association weight matrix is ​​calculated in real time through the dual-channel attention coupling architecture, and cross-modal feature fusion is performed. The fully connected layer sequence outputs an estimate of the instantaneous thirst level at the current sampling moment based on the fused cross-modal state representation. Record the instantaneous thirst level estimates at a series of consecutive sampling times, and connect them in chronological order to form the individualized thirst level trajectory that reflects the continuous change in thirst level.

7. The multidimensional assessment method for thirst in hemodialysis patients according to claim 6, characterized in that, The method for applying a dense trajectory detection algorithm to the video stream of the neck surface to track the motion trajectory of the Adam's apple and its surrounding area in three-dimensional spacetime is as follows: In the initial frame of the video stream of the neck surface, the initial position coordinates of the Adam's apple are located using a key point detection model; Using the initial position coordinates of the Adam's apple as the center, a rectangular region of interest is defined that includes the Adam's apple and its surrounding muscle tissue; Within the rectangular region of interest, multiple feature points are sampled uniformly at a fixed grid spacing. For each frame of the video stream, calculate the optical flow vector of each feature point to obtain the displacement of each feature point between adjacent frames; Based on the optical flow vector of each feature point, predict the position coordinates of each feature point in the next frame image; For each feature point, its position coordinates in the continuous frame sequence are connected in chronological order to form the motion trajectory of the feature point; Remove feature points whose trajectory length is less than a preset trajectory length threshold, and retain feature points whose trajectory length is greater than or equal to the preset trajectory length threshold. The set of motion trajectories of all retained feature points is taken as the motion trajectory of the Adam's apple and its surrounding area in three-dimensional spacetime.

8. The multidimensional assessment method for thirst in hemodialysis patients according to claim 7, characterized in that, The method for calculating the coordinates of the four index values ​​at each sampling time in the behavioral feature fusion space is as follows: Obtain four orthogonal basis vectors in the behavioral feature fusion space, consisting of keyword frequency index, speech dryness index, swallowing frequency index, and swallowing force principal component score; For each sampling time, the keyword frequency index, speech dryness index, swallowing frequency index, and swallowing force principal component score are obtained. The first vector component is obtained by performing a scalar multiplication operation between the keyword frequency index value and the orthogonal basis vector corresponding to the keyword frequency index. The second vector component is obtained by performing a scalar multiplication operation between the spoken language dryness index value and the orthogonal basis vector corresponding to the spoken language dryness index. The third vector component is obtained by performing a scalar multiplication operation between the swallowing frequency index value and the orthogonal basis vector corresponding to the swallowing frequency index. The fourth vector component is obtained by performing a scalar multiplication operation between the principal component score of swallowing force and the orthogonal basis vector corresponding to the principal component score of swallowing force. The first, second, third, and fourth vector components are added together to obtain the coordinates of the sampling time in the behavioral feature fusion space.

9. A multidimensional assessment system for thirst levels in hemodialysis patients, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multidimensional assessment method for the thirst level of hemodialysis patients as described in any one of claims 1 to 8.

Images