Emergency triage method and system based on time-axis audio translation and clinical decision support

Abstract

This invention relates to the field of artificial intelligence technology, specifically to an emergency triage method and system based on time-axis audio-visual translation and clinical decision support. The method includes the following steps: acquiring the chief complaint signal, extracting semantic vocabulary of the chief complaint symptoms, aligning multidimensional physiological objective indicators and a comprehensive disease score to generate a multimodal time-series dataset; regularizing the time axis to generate a gradient vector of the comprehensive disease score and a vector of physiological sign fluctuation amplitudes; calculating the Pearson correlation coefficient based on the dual vectors to mark high-risk coupled events; segmenting coupled events to project risk quadrants to generate a state matrix for graded assessment. In this invention, millisecond-level multimodal alignment and time-domain cross-analysis logic deeply mine the deep coupling patterns of the dynamic evolution of physiological indicators, breaking down the data fragmentation barriers of traditional static decision-making to accurately identify multiple early-stage, hidden severe disease deterioration trends. The feature space mapping mechanism significantly improves the accuracy of risk grading of extremely dangerous coupled events and effectively avoids delays in conventional assessments that could miss the golden period for rescue.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an emergency triage method and system based on time-axis audio-visual translation and clinical decision support. Background Technology

[0002] The field of artificial intelligence technology encompasses a system of disciplines that simulate human cognition and judgment processes through programming logic and computational architecture. Its core focus is on building machine learning algorithm architectures, natural language text parsing, and clinical knowledge graph reasoning. Overall, it relies on tensor operations of the underlying processor to perform feature vector space mapping on the input multimodal data in order to establish the mathematical relationship between the underlying data features and the target classification task.

[0003] The traditional emergency triage method based on time-axis audio translation and clinical decision support refers to the continuous recording of the patient's verbal symptoms and the determination of the severity of the condition during the emergency visit. It uses electret microphones placed at the triage station to continuously collect the audio stream of the conversation between the nurse and the patient. The audio stream is decoded frame by frame using a hidden Markov model to output a text sequence with timestamps. Then, the conditional random field algorithm is used to extract symptom description words and vital sign values ​​from the text sequence. The extracted words and values ​​are input into the root node of a pre-set emergency condition classification decision tree. The attribute determination and downward traversal are performed layer by layer until the leaf node is reached to output the corresponding triage level number.

[0004] Traditional methods rely on microphones to collect audio streams and decode text sequences, then use conditional random field algorithms to extract vocabulary and vital signs, inputting them into a pre-defined decision tree to output triage numbers. Based on simple vocabulary matching and a fixed decision tree operation mode, it is difficult to capture the potential correlations of the dynamic evolution of patients' physiological indicators over time. The single-sign extraction fragments the deep coupling mechanism of multi-dimensional data, making it impossible to accurately identify early hidden deterioration trends when dealing with complex acute and critical illnesses. Relying on traversing surface static data layer by layer can easily lead to lags in disease assessment and deviations in triage level determination, thus delaying the treatment of critically ill patients. Summary of the Invention

[0005] To address the technical problems of traditional methods that rely on microphone-based audio stream acquisition and text sequence decoding, combined with conditional random field algorithms to extract vocabulary and vital signs, inputting them into a pre-defined decision tree to output triage labels, and whose simple vocabulary matching and fixed decision tree operation mode make it difficult to capture the potential correlations of the dynamic evolution of patients' physiological indicators over time, whose singular vital sign extraction fragments the deep coupling mechanism of multidimensional data, and whose inability to accurately identify early hidden deterioration trends when dealing with complex acute and critical illnesses, and whose reliance on surface static data to traverse layer by layer easily leads to lag in disease assessment and deviation in triage level determination, thus delaying the treatment of critically ill patients, this invention provides an emergency triage method based on time-axis audio translation and clinical auxiliary decision-making.

[0006] To achieve the above objectives, this invention employs an emergency triage method based on time-axis audio-visual interpretation and clinical decision support, comprising the following steps: S1: Collect the patient's voice complaint signal and input it into the pre-trained acoustic-language joint model to extract the semantic vocabulary and timestamps of the complaint symptoms, obtain the semantic feature sequence of the complaint, obtain the comprehensive disease score and multidimensional physiological objective indicators, align them with the timestamps of the semantic feature sequence of the complaint at the millisecond level, and generate a multimodal physiological time series dataset; S2: Input the multimodal physiological time series dataset into the dynamic time warping algorithm for time axis calibration, calculate the rate of change of the comprehensive disease score, extract the extreme difference of abnormal fluctuations of vital signs in multidimensional physiological objective indicators, and generate the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. S3: Perform time-domain cross-correlation analysis based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological signs to calculate the Pearson correlation coefficient. When the correlation coefficient exceeds the preset threshold and the peak time difference is less than the set duration, it is marked as a high-risk coupling event. S4: The high-risk coupling events are processed in segments with fixed durations. The ratio of the time integral to the time length of the comprehensive disease score is calculated. The crossover rate of the waveforms of multidimensional physiological objective indicators is statistically analyzed. The ratio of the comprehensive disease score is mapped to the feature space using the support vector machine algorithm to generate a patient state risk matrix.

[0007] As a further aspect of the present invention, the multimodal physiological time-series dataset includes speech fundamental frequency features, MEWS modified early warning score, and vital sign cycle diagram; the comprehensive disease score gradient vector and physiological sign fluctuation amplitude vector include the first derivative of the score, the score range, the duration of abnormal heart rhythm, and the hemodynamic index drift; the high-risk coupling events include paroxysmal tachycardia segments, sympathetic nerve excitation periods, and hemodynamic abnormalities; and the patient status risk matrix includes complication rate, disease deterioration probability, and clinical prognosis index.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect the patient's voice complaint signal, perform inner product operation according to the language transformation matrix, extract the semantic words and timestamps of the complaint symptoms, construct a discrete vector array by mapping the semantic words of the complaint symptoms to the space, and concatenate the discrete vector arrays according to the timestamps to generate the semantic feature sequence of the complaint. S102: Obtain the comprehensive disease score and multidimensional physiological signals through physiological monitoring equipment, extract the periodic sequence of physiological signals, call the semantic feature sequence of the chief complaint to calculate the dimension and truncate the periodic sequence to construct the truncated electrocardiogram sequence, aggregate the comprehensive disease score and truncated sign sequence, and generate a set of basic physiological indicators. S103: Construct a clock baseline based on the timestamp of the chief complaint semantic feature sequence, extract collection nodes for the set of basic physiological indicators, compare the deviation between the collection nodes and the timestamp, establish a key-value pair mapping relationship when the deviation is less than the tolerance baseline value and write it into the multimodal time series feature matrix to generate a multimodal physiological time series dataset.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Extract multidimensional timestamps from the multimodal physiological time series dataset, calculate spatial Euclidean distance to construct a two-dimensional distance matrix, traverse the two-dimensional distance matrix to calculate the cumulative path cost of adjacent nodes, select the sequence with the minimum cumulative path cost as the time series normalization path and resample the values ​​to generate a calibrated physiological feature sequence. S202: Call the calibration physiological feature sequence to extract the comprehensive disease score data, read the score values ​​of adjacent nodes in the comprehensive disease score data to calculate the absolute difference, calculate the first derivative based on the absolute difference and the corresponding time interval, arrange the first derivative in ascending order according to the timestamp, and obtain the gradient vector of the comprehensive disease score. S203: Based on the calibrated physiological feature sequence, retrieve the physiological sign dimension features, search for local maxima within the physiological sign dimension features to determine the peak nodes, quantify the absolute difference of amplitude between adjacent nodes to construct an amplitude feature array, perform dimension alignment processing according to the time axis coordinates, and obtain the physiological sign fluctuation amplitude vector.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations, extract the time-domain segmented sequence, quantify the dispersion and co-variation trend of the node distribution within the time-domain sequence, and perform time-domain cross-correlation analysis based on the dispersion and co-variation trend to obtain the Pearson correlation coefficient. S302: Call the gradient vector of the comprehensive disease score to traverse the time window values, compare the values ​​of adjacent nodes to filter local maxima points and extract the gradient peak timestamp parameter, extract the amplitude peak timestamp parameter for the amplitude vector of physiological sign fluctuations, construct a pairing combination of gradient peak timestamp parameter and amplitude peak timestamp parameter, calculate the absolute value of the difference in the horizontal axis coordinate, and generate the peak time difference; S303: Based on the Pearson correlation coefficient and peak time difference triggering dual judgment logic, the Pearson correlation coefficient is compared with the preset correlation threshold, the peak time difference is compared with the set duration, the segmented feature node allocation status identifier that satisfies the dual judgment logic is extracted, and a high-risk coupling event is established.

[0011] As a further aspect of the present invention, the preset association threshold is set based on: retrieving the normal disease score benchmark vector and the normal sign amplitude benchmark vector from the historical sample database, performing Pearson association calculation on the normal disease score benchmark vector and the normal sign amplitude benchmark vector to generate a benchmark association coefficient sequence, extracting the upper limit edge value of the benchmark association coefficient sequence in the pre-configured confidence interval, and configuring the upper limit edge value as the preset association threshold. The set duration is based on the following: retrieve the set of peak timestamp parameters of stress symptoms and the set of peak timestamp parameters of stress signs from the historical sample database, calculate the absolute value of the difference between the elements in the set of peak timestamp parameters of stress symptoms and the corresponding elements in the set of peak timestamp parameters of stress signs, generate a baseline time difference sequence, extract the maximum distribution limit parameter of the baseline time difference sequence, and configure the maximum distribution limit parameter as the set duration. The condition that the dual-judgment logic is satisfied is that the Pearson correlation coefficient is greater than the preset correlation threshold and the peak time difference is less than the set duration.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: The high-risk coupling event is divided into fixed-length analysis windows at equal intervals along the time axis. The time series parameters of the comprehensive disease score inside the fixed-length analysis window are retrieved. The time series parameters of the comprehensive disease score are summed to obtain the cumulative area. The average value of the time span of the fixed-length analysis window is then processed to obtain the ratio of the comprehensive disease score. S402: Call the high-risk coupling event to read the physiological waveform parameters, configure the upper and lower limits of the safety threshold baseline to traverse the physiological waveform parameters, detect out-of-bounds flip nodes, count the total number of out-of-bounds flip nodes inside the fixed-length analysis window, and calculate the ratio with the time span value to obtain the waveform out-of-bounds crossover rate. S403: Construct a joint feature vector based on the comprehensive score ratio of the condition and the waveform crossover rate, call the preset kernel function to calculate the dot product of the joint feature vector and the sample set, aggregate the dot product and the Lagrange multiplier to obtain the hyperplane decision distance, project the joint feature vector to the risk quadrant, and establish the patient status risk matrix.

[0013] As a further aspect of the present invention, the method further includes step S5: S5: Call the patient status risk matrix classification results to assess the patient's current health status, convert it into the corresponding priority code according to the emergency four-level triage standard, and generate a dynamic triage level identifier; The dynamic triage level identifier includes emergency response time limit, triage level code, and priority order of treatment.

[0014] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Call the patient status risk matrix to extract diagonal values ​​to construct a feature dimensionality reduction sequence, obtain the health baseline distribution function of the preset sample, calculate the absolute deviation between the feature dimensionality reduction sequence and the mean of the health baseline distribution function, extract the deviation out-of-bounds sequence to locate high-risk vital signs parameters, and establish a status assessment vector. S502: Based on the state assessment vector, retrieve the emergency triage standard mapping table, extract the corresponding upper and lower limit boundary conditions in the emergency triage standard mapping table, detect the hierarchical position of the high-risk vital sign parameters falling into the interval of the upper and lower limit boundary conditions, allocate discrete integer variables according to the hierarchical position, and obtain priority codes; S503: Read the timestamp of the clock generator's operating cycle, concatenate the priority code and the timestamp bit by bit to construct a basic code segment, perform modulo-2 division on the basic code segment to extract the remainder parameter as an additional check segment, combine the basic code segment and the additional check segment to generate a dynamic triage level identifier.

[0015] An emergency triage system based on timeline interpretation and clinical decision support includes: The physiological time series acquisition module collects the patient's voice complaint signal and inputs it into a pre-trained acoustic-language joint model to extract the semantic vocabulary and timestamps of the complaint symptoms, obtains the semantic feature sequence of the complaint, obtains the comprehensive disease score and multidimensional physiological objective indicators, aligns them with the timestamps of the semantic feature sequence of the complaint at the millisecond level, and generates a multimodal physiological time series dataset. The dynamic feature warping module inputs the multimodal physiological time series dataset into the dynamic time warping algorithm for time axis calibration, calculates the rate of change of the comprehensive disease score, extracts the extreme difference of abnormal fluctuations of vital signs in multidimensional physiological objective indicators, and generates the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. The time-domain cross-correlation analysis module calculates the Pearson correlation coefficient based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. When the correlation coefficient exceeds the preset correlation threshold and the peak time difference is less than the set duration, it is marked as a high-risk coupling event. The risk space mapping module processes the high-risk coupled events in segments with fixed durations, calculates the ratio of the time integral to the time length of the comprehensive disease score, statistically analyzes the crossover rate of the waveforms of multidimensional physiological objective indicators, and performs feature space mapping on the ratio of the comprehensive disease score using a support vector machine algorithm to generate a patient state risk matrix. The emergency triage decision module calls the patient status risk matrix grading results to assess the patient's current health status, converts it into the corresponding priority code according to the emergency four-level triage standard, and generates a dynamic triage level identifier.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the main complaint signal is collected, semantic vocabulary of the main complaint symptom is extracted to construct a semantic feature sequence, and a multimodal physiological time series dataset is generated by combining the comprehensive disease score and multidimensional objective physiological indicators. The time axis is calibrated to calculate the score gradient vector and the amplitude vector of physiological sign fluctuation. Based on vector time domain cross-analysis, high-risk coupled events are extracted and processed by fixed-length segmentation to extract the integral ratio and waveform cross-boundary rate to construct a state risk matrix. The millisecond-level alignment and time domain cross-analysis logic deeply mine the deep coupling law of the dynamic evolution of physiological indicators, break through the data fragmentation barrier of static decision tree, accurately identify early hidden deterioration trends, and the feature mapping mechanism improves the accuracy of risk classification of high-risk events to avoid assessment lag and delay in rescue. Attached Figure Description

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

[0018] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This invention provides an emergency triage method based on time-axis audio-visual translation and clinical decision support, comprising the following steps: S1: Collect the patient's voice complaint signal and input it into the pre-trained acoustic-language joint model to extract the semantic vocabulary and timestamps of the complaint symptoms, obtain the semantic feature sequence of the complaint, obtain the comprehensive disease score and multidimensional physiological objective indicators, align them with the timestamps of the semantic feature sequence of the complaint at the millisecond level, and generate a multimodal physiological time series dataset; S2: Input the multimodal physiological time series dataset into the dynamic time warping algorithm for time axis calibration, calculate the rate of change of the comprehensive disease score, extract the extreme difference of abnormal fluctuations of vital signs in multidimensional physiological objective indicators, and generate the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. S3: Calculate the Pearson correlation coefficient by performing time-domain cross-correlation analysis based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. When the correlation coefficient exceeds the preset threshold and the peak time difference is less than the set duration, it is marked as a high-risk coupling event. S4: High-risk coupled events are processed in segments with fixed durations. The ratio of the time integral to the time length of the comprehensive disease score is calculated. The crossover rate of the waveforms of multidimensional physiological objective indicators is statistically analyzed. The ratio of the comprehensive disease score is mapped to the feature space through the support vector machine algorithm to generate a patient state risk matrix. S5: Call the patient status risk matrix classification results, assess the patient's current health status, convert it into the corresponding priority code according to the emergency four-level triage standard, and generate a dynamic triage level identifier; The multimodal physiological time-series dataset includes speech fundamental frequency features, MEWS modified early warning score, and vital sign cycle diagram. The gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations include the first derivative of the score, the score range, the duration of abnormal heart rhythm, and the drift of hemodynamic indicators. High-risk coupled events include paroxysmal tachycardia segments, sympathetic nerve excitation periods, and hemodynamic abnormalities. The patient status risk matrix includes the incidence of complications, the probability of disease deterioration, and the clinical prognosis index. The dynamic triage level identifier includes emergency response time limit, triage level code, and priority of medical treatment.

[0022] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect the patient's voice complaint signal, perform inner product operation according to the language transformation matrix, extract the semantic words and timestamps of the complaint symptoms, construct a discrete vector array by mapping the semantic words of the complaint symptoms to the space, and concatenate the discrete vector arrays according to the timestamps to generate the semantic feature sequence of the complaint. The omnidirectional microphone array hardware interface installed at the emergency triage station was used to collect patients' voice complaints. The sampling rate was set to 16000 Hz and the sampling depth to 16 bits to obtain continuous analog speech signals, which were then converted into a 1D digital audio sequence using an analog-to-digital converter. Preprocessing was performed on the 1D digital audio sequence, using a Hamming window with a window length of 25 milliseconds and a step size of 10 milliseconds for frame processing. The 80-dimensional Mel-frequency cepstral coefficients of each frame of audio data were extracted to construct an initial acoustic feature matrix. The initial acoustic feature matrix is ​​input into a pre-defined semantic extraction deep neural network. This network consists of one input layer, three hidden layers, and one output layer. The number of neurons in the input layer is aligned with the 80-dimensional dimension of the Mel-frequency cepstral coefficient feature. All three hidden layers employ a bidirectional long short-term memory (LSTM) network architecture, with 512 memory units per hidden layer. Features are transferred between layers via fully connected connections, and a linear rectified function is used as the activation function for each layer to filter out negative features less than 0. The output layer contains 10,000 neurons, strictly corresponding to the size of a pre-defined medical dictionary vocabulary. A normalized exponential function is used in the output layer to convert the values ​​into a sequence of word probability distributions. The semantic extraction deep neural network outputs a text sequence containing timestamp information. Part-of-speech tags are retrieved from the text sequence, and adjective and noun combinations representing the subjective symptoms are extracted, such as chest tightness, dizziness, and tearing pain. Simultaneously, the start and end timestamps of the corresponding words in the speech signal are extracted as their respective timestamp attributes. A pre-configured language transformation matrix is ​​established. This matrix is ​​a floating-point matrix with 10,000 rows and 300 columns, generated iteratively from a massive amount of medical consultation text through a word vector training algorithm. One-hot encoded vectors of the semantic terms of the chief complaint symptoms are extracted. These vectors have a total dimension of 10,000, with only the values ​​at the corresponding word index positions set to 1, and the values ​​at the remaining 9,999 positions set to 0. An inner product operation is performed based on the language transformation matrix. Specifically, each element of the one-hot encoded vector is multiplied by the element of the corresponding row in the language transformation matrix, and the product is summed column-wise to extract the 300-dimensional row vectors of the corresponding words in the language transformation matrix. This maps the semantic terms of the chief complaint symptoms to a 300-dimensional continuous semantic space, constructing a discrete vector array containing 300 floating-point elements. Multiple discrete vector arrays extracted from the same patient's chief complaint are concatenated according to their row dimensions, following the chronological order of their timestamps, to obtain the chief complaint semantic feature sequence matrix.For example, if a patient's chief complaint contains the words "chest tightness" and "angina," the 50th position of the one-hot encoding vector for "chest tightness" is 1, and the 120th position of the one-hot encoding vector for "angina" is 1. These two one-hot encoding vectors are then subjected to the aforementioned inner product and summation logical operations with the language transformation matrix, extracting the 50th and 120th rows of the language transformation matrix respectively, resulting in two discrete vector arrays containing 300 floating-point numbers. These arrays are then concatenated in the order of the timestamps of "chest tightness" (1.5 seconds) and "angina" (4.2 seconds), ultimately generating a 2-row, 300-column chief complaint semantic feature sequence matrix. The advantage of this operational logic is that by transforming the high-dimensional, sparse, unstructured chief complaint text into a low-dimensional, dense numerical matrix, the diverse ambiguities of spoken expression are eliminated, allowing subsequent processes to directly quantify the patient's objective feelings using numerical features.

[0023] S102: Obtain the comprehensive disease score and multidimensional physiological signals through physiological monitoring equipment, extract the periodic sequence of physiological signals, call the semantic feature sequence of the chief complaint to calculate the dimension and truncate the periodic sequence to construct the truncated electrocardiogram sequence, aggregate the comprehensive disease score and truncated sign sequence, and generate a set of basic physiological indicators. The serial communication interface of a wearable multi-parameter physiological monitor is used to acquire dynamic comprehensive patient condition scores and multi-dimensional physiological signals (including heart rate, respiration, blood oxygen saturation, and blood pressure) data. The comprehensive patient condition score data is derived from discrete integer values ​​between 0 and 10 entered by the patient using a handheld electronic scoring board or the emergency system. The physiological signal data is derived from continuous electrical signals collected by multimodal sensors (such as photoplethysmography, pressure sensors, etc.), with a sampling frequency set to 100 Hz. Baseline drift removal and high-frequency noise filtering are performed on the acquired physiological signals using a bidirectional filtering process with a high-pass filter (cutoff frequency of 0.5 Hz) and a low-pass filter (cutoff frequency of 40 Hz). Local maxima in the filtered physiological signals are extracted to locate the physiological fluctuation cycle, and the time interval between adjacent peaks is recorded to construct a periodic sequence of the physiological signals. The chief complaint semantic feature sequence matrix generated in the previous steps is called, and the number of rows in the matrix is ​​read as the sequence dimension value, which is used as the truncation length benchmark. Based on the start timestamp of the chief complaint semantic feature sequence, the corresponding start node is located in the physiological signal cycle sequence. Then, starting from this start node, continuous physiological feature data is extracted along the positive direction of the time axis according to the calculated truncation length benchmark. Excess cycle nodes exceeding the length benchmark are discarded, constructing a truncated sign sequence with a length strictly consistent with the dimension of the chief complaint semantic feature sequence. The set of discrete disease comprehensive score values ​​collected within the same time window and the truncated sign sequence are combined according to the corresponding time axis nodes. At the data structure level, a multi-dimensional array is used for aggregation, with the disease comprehensive score value as the first dimension feature of the array and the cycle value of the truncated sign sequence as the second dimension feature of the array, generating a set of basic physiological indicators. For example, if the chief complaint semantic feature sequence matrix contains 5 rows of data, it means that 5 semantic feature nodes have been extracted. The extracted comprehensive disease score value is 7. The physiological signal is extracted from the corresponding timestamp and the duration values ​​of 5 complete cycles are extracted, such as 0.8 seconds, 0.82 seconds, 0.79 seconds, 0.81 seconds, and 0.80 seconds. The single comprehensive disease score of 7 is combined with these 5 cycle values ​​by column to generate a basic physiological indicator set array containing 1 row and 6 columns or feature levels side by side.

[0024] S103: Construct a clock baseline based on the timestamp of the chief complaint semantic feature sequence, extract collection nodes for the basic physiological indicator set, compare the deviation between the collection nodes and the timestamp, and establish a key-value pair mapping relationship when it is less than the tolerance benchmark value and write it into the multimodal time series feature matrix to generate a multimodal physiological time series dataset. The system iterates through the timestamp data attached to the semantic feature sequence matrix of the patient's complaints, setting the starting timestamp of the first uttered word as the relative zero point. It then calculates the time difference between subsequent timestamps and this zero point, constructing a one-dimensional clock baseline by arranging the timestamps in ascending order. For the generated set of basic physiological indicators, it scans multiple acquisition nodes that record changes in physiological parameters, extracting the timestamp corresponding to each acquisition node. The absolute time deviation between the timestamp of each acquisition node and the corresponding timestamps on the clock baseline is calculated. A preset tolerance baseline value is read, set to 0.2 seconds. This tolerance baseline value is based on statistical analysis of 500 historical emergency room patients' complaints and synchronized physiological monitoring audio and video recordings, which revealed that the average natural delay time from speech utterance to physiological feature manifestation ranges from 0.15 seconds to 0.25 seconds; the median value was then set to 0.2 seconds. The absolute time deviation value is compared with a tolerance benchmark of 0.2 seconds. When the absolute time deviation value is less than 0.2 seconds, the physiological indicator data of the current acquisition node is considered to be highly synchronized with the corresponding subjective semantic features in the time dimension. For data nodes that meet the tolerance condition, a key-value pair mapping relationship is established. The timestamp on the clock baseline is used as the unique key name of the dictionary structure, and the corresponding comprehensive disease score and multidimensional physiological signs in the basic physiological indicator set are used as the values ​​of the dictionary structure. The data dictionary with the established key-value pair mapping relationship is written into the corresponding column of the multimodal time series feature matrix to complete the alignment and splicing of multi-source heterogeneous data and generate a multimodal physiological time series dataset. For example, the time difference between the first word timestamp and zero is 1.5 seconds. The corresponding time label of the first acquisition node of the basic physiological indicator is converted to 1.6 seconds. Subtracting the two gives an absolute time deviation of 0.1 seconds. Since 0.1 seconds is less than the preset tolerance benchmark of 0.2 seconds, the timestamp of 1.5 seconds is used as the key, and the acquired comprehensive disease score of 7 and the physiological fluctuation cycle of 0.8 seconds are used as the value to form a key-value pair and write it into the matrix to complete the dataset generation operation. The advantage of this operation logic is that, through strict tolerance threshold comparison and key-value pair mapping, it eliminates the asynchronous error of the hardware clock between the voice acquisition device and the physiological monitoring device.

[0025] Please see Figure 3 The specific steps of S2 are as follows: S201: Extract multidimensional timestamps from multimodal physiological time-series datasets, calculate spatial Euclidean distance to construct a two-dimensional distance matrix, traverse the two-dimensional distance matrix to calculate the cumulative path cost of adjacent nodes, select the sequence with the minimum cumulative path cost as the time-series normalized path and resample the values ​​to generate calibrated physiological feature sequences; The generated multimodal physiological time-series dataset is read, and the time-label arrays of feature sequences in different dimensions within the dataset are scanned. The speech timestamps of the chief complaint dimension and the timestamps of the physiological objective indicators dimension are extracted to form a multidimensional timestamp set. For nodes between any two different sequences in the multidimensional timestamp set, the spatial Euclidean distance in the feature space is calculated. The specific Euclidean distance calculation logic is as follows: obtain the feature value of a node in the first sequence and the feature value of the corresponding node in the second sequence, subtract the two to calculate the feature difference, multiply the feature difference by itself to obtain the square value of the feature difference, sum the square values ​​of the feature differences of each dimension, and finally perform a square root operation on the summation to obtain the Euclidean distance between the two nodes. The Euclidean distances of the node combinations are arranged according to the cross-index of rows and columns to construct a 2D distance matrix. A dynamic time warping algorithm is employed to traverse the 2D distance matrix, starting from the top-left corner and searching towards the bottom-right corner. During the search, for each node, the cumulative path cost values ​​of its left-hand neighbor, top-hand neighbor, and top-left diagonal neighbor are extracted. The minimum value is selected and summed with the node's Euclidean distance to obtain the latest cumulative path cost. This process is repeated until the end of the matrix. The entire 2D distance matrix is ​​backtracked, and the node sequence with the minimum cumulative path cost from the start to the end is selected as the time-warped path. Based on the node mapping coordinates recorded by this time-warped path, the physiological index values ​​in the original multimodal physiological time-series dataset are resampled. For nodes with one-to-many mappings, the values ​​are averaged; for nodes with many-to-one mappings, the values ​​are copied, generating calibrated physiological feature sequences with strictly consistent length and phase alignment. For example, to calculate the Euclidean distance between a node with a value of 5 and a node with a value of 2, the difference between the two is 3, and the square of the difference is 9. If there is only one dimension, the square root of the Euclidean distance is 3. If the current node has an Euclidean distance of 3, its cumulative cost to the left is 10, to the top is 15, and to the upper left is 12. Then, the minimum value of 10 is selected, and the sum of 3 and 10 is used to obtain the current cumulative path cost of 13. By finding this minimum cost path and resampling, non-linear alignment of the time axis is achieved.

[0026] S202: Call the calibration physiological feature sequence to extract the comprehensive disease score data, read the score values ​​of adjacent nodes in the comprehensive disease score data to calculate the absolute difference, calculate the first derivative based on the absolute difference and the corresponding time interval, arrange the first derivative in ascending order of timestamp, and obtain the gradient vector of the comprehensive disease score. The output calibrated physiological characteristic sequence is retrieved, and a subset of the comprehensive disease score data is read according to the key index of the data dictionary to obtain a list of continuous comprehensive disease score values ​​arranged in chronological order. The score values ​​of two adjacent nodes in this list are read, and the comprehensive disease score value of the later node is subtracted from that of the earlier node. The difference is calculated by taking the absolute value of this difference to eliminate the negative impact of score decline. Simultaneously, the timestamp values ​​corresponding to these two nodes are extracted, and the time interval between them is obtained by subtracting the earlier timestamp value from the later one. The calculated absolute difference is divided by the corresponding time interval to calculate the rate of change of the comprehensive disease score within that time interval, and this rate of change is defined as the first derivative of the comprehensive disease score. The above subtraction, absolute value taking, and division operations are repeated for adjacent node combinations in the sequence to obtain a set of first derivatives for each time period. Based on the chronological order of the timestamps generated by the first derivatives, the numerical elements within the set of first derivatives are sorted in ascending order and combined into a continuous 1D floating-point array to obtain the gradient vector of the comprehensive disease score. For example, if the comprehensive disease score values ​​of three consecutive nodes in the calibrated physiological feature sequence are 4, 7, and 6, with corresponding timestamps of 1.2 seconds, 2.2 seconds, and 3.2 seconds, first, the first node 4 and the second node 7 are read. The absolute difference between them is 3, and the time interval is 2.2 seconds minus 1.2 seconds, which equals 1.0 seconds. Dividing 3 by 1.0 gives the first first derivative value of 3.0. Then, the second node 7 and the third node 6 are read. The absolute difference is 1, and the time interval is 1.0 seconds. Dividing 1 by 1.0 gives the second first derivative value of 1.0. Finally, these values ​​are concatenated according to their chronological order to obtain the gradient vector of the comprehensive disease score containing the values ​​3.0 and 1.0.

[0027] S203: Based on the calibration of physiological feature sequences, retrieve the dimensional features of physiological signs, search for local maxima within the dimensional features of physiological signs to determine the peak nodes, quantify the absolute difference of amplitude between adjacent nodes to construct an amplitude feature array, and perform dimensional alignment processing according to the time axis coordinates to obtain the amplitude vector of physiological sign fluctuations. Based on the generated calibrated physiological feature sequence, the column index of the physiological objective indicator dimension feature in the sequence multidimensional array is retrieved, and all continuous values ​​in that column are extracted to form a 1D feature array. The numerical fluctuations within this feature array are searched. A sliding search window containing three consecutive data nodes is set, and the relationship between the value of the middle node and the values ​​of its two adjacent nodes is compared. If the value of the middle node is simultaneously greater than both the preceding and following node values, the middle node is determined to be a local maximum and marked as a peak node. The index position of the selected peak node in the sequence is recorded. The physiological parameter values ​​corresponding to the peak nodes are read as amplitude values. The absolute difference between the amplitude values ​​of two adjacent peak nodes in chronological order is calculated to indicate the severity of pathological fluctuations caused by physiological dysfunction or compensation. The absolute differences calculated for adjacent peak nodes are combined sequentially to construct an amplitude feature array. The midpoint timestamp coordinates corresponding to each absolute difference element in the amplitude feature array are extracted. These midpoint timestamp coordinates are then numerically compared and interpolated with the time axis coordinates of the comprehensive disease score gradient vector for dimensional alignment. This ensures that both vectors have corresponding feature values ​​at the same timestamp position. After alignment, a 1-dimensional physiological sign fluctuation amplitude vector is output. For example, if the values ​​of five consecutive nodes in the physiological objective indicator dimensional feature array are 75, 82, 78, 85, and 80, a sliding window comparison is used. Since 82 is greater than 75 and 78, 82 is determined to be the first peak node; and since 85 is greater than 78 and 80, 85 is determined to be the second peak node. The absolute difference between the amplitude values ​​82 and 85 of these two adjacent peak nodes is then calculated, yielding a value of 3. This value 3 is used as the fluctuation amplitude feature and written into the array. Finally, its time coordinates are aligned to a unified time axis system, completing the extraction of the physiological sign fluctuation amplitude vector.

[0028] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations, the time-domain segmented sequence is extracted, the dispersion and covariance trend of the node distribution within the time-domain sequence are quantified, and the time-domain cross-correlation analysis is performed based on the dispersion and covariance trend to obtain the Pearson correlation coefficient. The gradient vector of the comprehensive disease score and the amplitude vector of physiological indicator fluctuations are used as parallel inputs. A fixed time-domain truncation window of 5 seconds is set, and two sets of time-domain segmented sequences are generated synchronously along the time axis. The mean of the sequences is calculated by summing the values ​​in the two sets and dividing by the total number of nodes. The deviation is obtained by subtracting the mean from the value of each node in the sequence. The sum of the squares of the deviations is then divided by the total number of nodes and the square root is taken to obtain the standard deviation, which quantifies the dispersion of the node distribution in the sequence. The deviation of the gradient vector of the comprehensive disease score and the deviation of the amplitude of physiological indicator fluctuations at the same time node are multiplied. The multiplications of the nodes are summed to obtain the covariance, which quantifies the covariance trend. The calculated covariance is then divided by the product of the standard deviations of the two sequences to finally calculate the Pearson correlation coefficient. For example, the gradient vector values ​​of 5 nodes are 1, 2, 3, 4, and 5, with an average value of 3. The magnitude vector values ​​are 2, 4, 6, 8, and 10, with an average value of 6. The gradient vector deviations are calculated as -2, -1, 0, 1, and 2, and the magnitude vector deviations are calculated as -4, -2, 0, 2, and 4. Multiplying the deviations of the same node and summing them gives a covariance of 20. The standard deviation of the gradient is approximately 1.414, and the standard deviation of the magnitude is approximately 2.828. The product is 4.0. Dividing the covariance 20 by the number of nodes 5 gives 4. Dividing this by 4.0 gives a Pearson correlation coefficient of 1.0, indicating that the two are completely positively correlated.

[0029] S302: Call the gradient vector of the comprehensive disease score to traverse the time window values, compare the values ​​of adjacent nodes to filter local maxima points and extract the gradient peak timestamp parameter, extract the amplitude peak timestamp parameter for the amplitude vector of physiological sign fluctuations, construct a pairing combination of gradient peak timestamp parameter and amplitude peak timestamp parameter, calculate the absolute value of the difference in the horizontal axis coordinate, and generate the peak time difference. The gradient vector of the comprehensive disease assessment is invoked, and a sliding time window with a width of 10 seconds is used to traverse continuous gradient values. The values ​​of three adjacent nodes are compared; if the value of the middle node is greater than the values ​​of the two adjacent nodes, it is determined to be a local maximum, and the exact time coordinate of its occurrence is recorded and extracted as the gradient peak timestamp parameter. The same operation is performed on the amplitude vector of physiological sign fluctuations, extracting the time coordinate of the moment of most intense fluctuation as the amplitude peak timestamp parameter. The gradient peak timestamp parameter and the amplitude peak timestamp parameter within the same observation period are paired and combined. The horizontal axis coordinate values ​​of the two time node elements are extracted, a subtraction operation is performed, and the absolute value of the result is taken to calculate the peak time difference. For example, in a 10-second traversal window, the gradient vector of the comprehensive disease score is selected to have a local maximum of 5.2 at 3.5 seconds, and 3.5 seconds is extracted as the gradient peak timestamp parameter. The amplitude vector of physiological sign fluctuations is selected to have a local maximum of 12.5 at 4.1 seconds, and 4.1 seconds is extracted as the amplitude peak timestamp parameter. After pairing the two, 4.1 minus 3.5 gives a difference of 0.6. The absolute value is then used to obtain the final peak time difference of 0.6 seconds.

[0030] S303: Based on the Pearson correlation coefficient and peak time difference, trigger the dual judgment logic, compare the Pearson correlation coefficient with the preset correlation threshold, compare the peak time difference with the set duration, extract the segmented feature nodes that meet the dual judgment logic, assign status identifiers, and establish high-risk coupling events. The Pearson correlation coefficient is compared with a preset correlation threshold. This preset threshold is calculated by retrieving the baseline correlation coefficient sequence from the comprehensive score benchmark vector and the baseline amplitude benchmark vector of normal physical signs from 1000 cases of normal condition in the cloud historical sample database, performing the same logic as above, and then configuring the upper edge value of 0.45 for the distribution ratio reaching 95% confidence interval. The peak time difference is compared with a set duration. This set duration is calculated by retrieving the peak timestamp parameters of stress condition and peak timestamp parameters of stress physical signs from 500 cases of severe stress patients in the historical sample database, subtracting them and taking the absolute value to generate a baseline time difference sequence, and then configuring the maximum distribution boundary parameter of 2.5 seconds covering 90% of the samples. A dual logic condition judgment is performed. When the Pearson correlation coefficient is greater than 0.45 and the peak time difference is less than 2.5 seconds, the data segment is determined to meet the dual judgment logic. Risk identifiers with a state of 1 are assigned to the feature nodes within the segment sequence that meet this logic, and a high-risk coupled event data packet is established accordingly. For example, if the current Pearson correlation coefficient is 0.72 and the peak time difference is 0.6 seconds, since 0.72 is greater than 0.45 and 0.6 is less than 2.5, the conditions are all true, triggering the establishment of a high-risk coupling event.

[0031] Table 1: Example Table of Feature Extraction and Threshold Determination Parameters

[0032] Table 1 lists the numerical comparison of key parameters in the dual-decision logic and the execution results that ultimately trigger high-risk coupling events.

[0033] Please see Figure 5 The specific steps of S4 are as follows: S401: Extract fixed-length analysis windows by equidistantly dividing high-risk coupled events along the time axis, retrieve time-series parameters of comprehensive disease scores within the fixed-length analysis window, perform cumulative summation on the time-series parameters of comprehensive disease scores to obtain the cumulative area, perform mean processing on the time span values ​​of the fixed-length analysis window, and obtain the ratio of comprehensive disease score to the total score. The system receives high-risk coupled event data packets and, starting from the initial timestamp, equally divides the timeline along the positive direction to extract a fixed-length analysis window containing continuous 30-second data. It retrieves the time-series parameter array of the comprehensive disease score within the fixed-length analysis window. For each score element, it uses a loop structure to read and add it to an accumulator variable, performing a summation operation to obtain the cumulative area. This cumulative area is then divided by the 30-second time span of the fixed-length analysis window to perform an average calculation, yielding the comprehensive disease score ratio parameter. For example, if there are 30 sampling points within the 30-second fixed-length analysis window, and each sampling point's score remains at 8, the cumulative area is 240 (30 eights). Dividing 240 by the 30-second time span yields a comprehensive disease score ratio of 8.0, reflecting the severity of the persistent disease.

[0034] S402: Call the high-risk coupling event to read physiological waveform parameters, configure the upper and lower limits of the safety threshold baseline, traverse the physiological waveform parameters, detect out-of-bounds flip nodes, count the total number of out-of-bounds flip nodes inside the fixed-length analysis window, calculate the ratio with the time span value, and obtain the waveform out-of-bounds crossover rate. The high-risk coupled event data packet containing high-dimensional physiological features is invoked again. Based on the data dimension index, the parameter sequence of objective physiological indicators (such as ECG waveform, respiratory waveform, and arterial pressure waveform) within the data packet is read. This sequence records millivolt-level voltage values ​​that fluctuate in the microsecond range over time. To eliminate errors caused by device baseline drift, the average value of the voltage values ​​in the entire sequence is extracted and configured as the zero-potential baseline (i.e., the safety threshold baseline). A loop detection algorithm is used to traverse each voltage value node in the ECG waveform parameter sequence from left to right. The voltage value of the current node is subtracted from the zero-potential baseline value. If the difference is positive, the polarity is marked as positive; if the difference is negative, the polarity is marked as negative. The polarity marking states of two adjacent time nodes are compared. If the previous node is positive and the next node is negative, or vice versa, it is determined that a signal penetration of the baseline has occurred between these two nodes. This event is detected and counted as one out-of-bounds flip node. The algorithm continues to iterate through the entire 30-second fixed-length analysis window, accumulating the count of each detected flip event to calculate the total number of out-of-bounds flip nodes within the window. This total number is then divided by the 30-second time span of the analysis window to obtain the frequency of waveform crossing the baseline per unit time, defined as the waveform crossover rate parameter. For example, if during the iteration, an ECG waveform is detected to abruptly change from below to above the baseline (one flip), and then subsequently drops below the baseline again (a second flip), a total of 45 such polarity switching events are detected within 30 seconds. Dividing this total of 45 by the 30-second time span yields a waveform crossover rate of 1.5 times per second. The advantage of this logic is that by statistically analyzing the crossover rate, high-frequency abnormal fluctuations in the patient's circulatory or respiratory system can be directly quantified in the time domain without the need for complex frequency-domain Fourier transforms.

[0035] S403: Construct a joint feature vector based on the ratio of the comprehensive score of the disease condition and the crossover rate of the waveform. Call the preset kernel function to calculate the dot product of the joint feature vector and the sample set. Combine the dot product and the Lagrange multiplier to obtain the hyperplane decision distance. Project the joint feature vector to the risk quadrant to establish the patient status risk matrix. The comprehensive score ratio of the disease condition and the waveform crossover rate are concatenated to form a joint feature vector with two dimensions. For example, the previously calculated 8.0 and 1.5 are concatenated to form a two-dimensional array. The built-in support vector machine classification model's preset parameters are called, and the kernel function algorithm program in the model's memory is read. This step uses the radial basis function kernel to perform spatial mapping. A support vector sample set pre-trained on a large number of confirmed cases in the cloud is obtained. The currently generated joint feature vector is input into the radial basis function kernel and a dot product operation is performed with the support vectors in the sample set. The specific internal calculation logic of the kernel function is as follows: the joint feature vector is subtracted from the multiple dimensions of the support vectors, and the sum of the squares is obtained to get the squared Euclidean distance. The squared Euclidean distance is divided by the model's preset variance parameter, and a negative sign is added. Then, a natural exponential operation is performed to obtain the mapped value. Finally, the multiple mapped values ​​are added together to complete the dot product operation. The pre-trained Lagrange multiplier coefficient array of the model is extracted. The mapped value output from the kernel function dot product calculation in the previous step is multiplied by the corresponding Lagrange multiplier coefficient. The product results are then summed, and finally, the preset bias constant parameter of the classification model is added to obtain a specific floating-point value, which is defined as the hyperplane decision distance. Based on the sign and absolute value of the hyperplane decision distance, the joint feature vector is projected into the risk quadrant of a 2D Cartesian coordinate system. Distance values ​​greater than 0 are projected into the high-risk quadrant, and those less than 0 are projected into the low-risk quadrant. Based on this, a matrix structure is generated to establish a patient state risk matrix containing the patient's real-time risk tendency indicator and specific risk distance value. For example, the joint feature vector is an array containing 8.0 and 1.5. After calculating the Euclidean distance with the sample set and performing a natural exponential transformation, it is multiplied by the Lagrange multipliers 1.2 and 2.5 and summed. Finally, a bias constant of -0.8 is added to calculate the hyperplane decision distance as 3.2. Since the value of 3.2 is greater than 0 and the absolute value is large, its coordinates are projected to the first quadrant, i.e., the extremely high risk quadrant, and finally a patient status risk matrix is ​​generated that identifies high-risk status and is accompanied by a distance value of 3.2.

[0036] Table 2: Example Table of Joint Feature Decision Matrix

[0037] As shown in Table 2, the entire process of obtaining the hyperplane decision distance by substituting the joint feature vectors into the Lagrange multipliers and the bias constants and performing a summation logical operation is illustrated.

[0038] Please see Figure 6 The specific steps of S5 are as follows: S501: Call the patient status risk matrix to extract diagonal values ​​to construct a feature dimensionality reduction sequence, obtain the health baseline distribution function of the preset sample, calculate the absolute deviation between the feature dimensionality reduction sequence and the mean of the health baseline distribution function, extract the deviation out-of-bounds sequence to locate high-risk vital signs parameters, and establish a status assessment vector. The patient status risk matrix is ​​extracted, and the diagonal values ​​at the main diagonal position are arranged sequentially to construct a feature dimensionality reduction sequence. The health baseline distribution function of a preset sample is retrieved from the database, and the mean of the health baseline distribution function under the same time window is calculated and output. The corresponding mean is subtracted from the values ​​in the feature dimensionality reduction sequence, and the absolute value of the difference is obtained by removing the negative sign. A preset threshold of 1.5 is set, and absolute deviations greater than 1.5 are selected and combined with their corresponding original physiological feature names to extract the deviation out-of-bounds sequence. Abnormal parameters are reverse-positioned and labeled as high-risk vital signs, and packaged with the severely deviated values ​​to establish a status assessment vector. For example, if the feature dimensionality reduction sequence value is 9.5, the baseline mean is 5.0, and the absolute value of the subtraction is 4.5, since 4.5 is greater than 1.5, it is extracted to form a deviation out-of-bounds sequence. The tachycardia label is located and packaged with the value 4.5 to establish a status assessment vector.

[0039] S502: Based on the state assessment vector, retrieve the emergency triage standard mapping table, extract the corresponding upper and lower limit boundary conditions in the emergency triage standard mapping table, detect the hierarchical position of high-risk vital sign parameters falling into the interval of the upper and lower limit boundary conditions, allocate discrete integer variables according to the hierarchical position, and obtain priority codes; The generated state assessment vector, containing the abnormality type and severity values, is extracted. Using this vector as a keyword query, a retrieval request is sent to the triage database to retrieve the fixed emergency triage standard mapping table. The multi-level assessment rules configured within the emergency triage standard mapping table are read, and the upper and lower boundary condition value ranges corresponding to the located high-risk vital sign parameters are extracted. For example, if the high-risk vital sign parameter is the overall absolute deviation of the condition, the lower boundary condition extracted from the mapping table is a deviation greater than or equal to 8.0 for Level 1 critical, between 5.0 and 8.0 for Level 2 severe, and between 2.0 and 5.0 for Level 3 acute. The severe deviation value carried in the state assessment vector is input into the detection engine, which checks from the highest level downwards whether the value falls within the extracted upper and lower boundary condition ranges. When the detection engine determines that the current deviation value is simultaneously greater than or equal to the lower boundary condition and strictly less than the upper boundary condition, it locks the level position of the value within that range. Based on the successfully matched hierarchical position, a pre-defined dictionary of discrete integer variables is used for mapping and allocation. Discrete integer 1 is assigned to level 1, discrete integer 2 to level 2, discrete integer 3 to level 3, and discrete integer 4 to level 4. The final assigned discrete integer variable is directly output as the result. This discrete integer represents the priority logic for the current patient's emergency treatment, successfully obtaining a general priority code. For example, if the input state evaluation vector carries an absolute deviation value of 4.5, the detection engine finds that 4.5 does not meet the level 1 condition of being greater than 8.0, nor the level 2 condition of being between 5.0 and 8.0. Further comparison reveals that 4.5 is greater than the lower boundary condition 2.0 and less than the upper boundary condition 5.0, precisely falling into the level position of the level 3 interval. Subsequently, according to the hierarchical mapping rules, discrete integer variable 3 is directly assigned to it, successfully obtaining a priority code of value 3 for subsequent process scheduling.

[0040] S503: Reads the timestamp of the clock generator's operating cycle, concatenates the priority code and the timestamp bit by bit to construct a basic code segment, performs modulo-2 division on the basic code segment to extract the remainder parameter as an additional check segment, combines the basic code segment and the additional check segment to generate a dynamic triage level identifier. The hardware clock generator is invoked to extract a pure digital timestamp sequence accurate to milliseconds within the current operating cycle. Priority encoding values ​​are shifted to the high-order bits, and the timestamp sequence is placed in the low-order bits. A binary bit-by-bit concatenation instruction is executed to physically concatenate the sequences, constructing the basic code segment. A preset 16-bit fixed generator polynomial divisor constant is used. Sixteen zero placeholders are appended to the end of the basic code segment as the dividend. A bitwise XOR operation is used to perform modulo-2 division instead of subtraction, and the remaining remainder parameter is extracted as an additional check segment. Finally, a string concatenation operation is performed, using the basic code segment containing triage level and time information as the data body, seamlessly concatenating the aforementioned additional check segment at the end. The above information segments are packaged and compiled, ultimately generating a dynamic triage level identifier code that includes triage level instructions, a time-based traceability label, and a function to prevent garbled characters during transmission. This code is then pushed to the nurse station terminal screen for patient reception guidance. For example, if the priority code is an integer 3 and the timestamp is 1634567890, the basic code segment constructed by bit concatenation is 31634567890. After being converted to binary, zeros are padded to the end, and a bitwise XOR operation of modulo 2 division is performed. After multiple rounds of shift calculations, the remainder parameter value is obtained, for example, 1011, as an additional check segment. Finally, 31634567890 is concatenated with the check segment and formatted to generate the final dynamic triage level identifier, ensuring the accuracy and robustness of the underlying data scheduling when facing high-concurrency triage tasks. The advantage of this operation logic is that by introducing timestamp concatenation and hardware-level modulo 2 division, the risk of packet loss or scrambling of rating instructions caused by network congestion in the emergency room is completely eliminated, significantly improving the fault tolerance of triage execution commands.

[0041] Please see Figure 7 An emergency triage system based on timeline interpretation and clinical decision support includes: The physiological time series acquisition module collects the patient's voice complaint signal and inputs it into a pre-trained acoustic-language joint model to extract the semantic vocabulary and timestamps of the complaint symptoms, obtains the semantic feature sequence of the complaint, obtains the comprehensive disease score and multidimensional physiological objective indicators, aligns them with the timestamps of the semantic feature sequence of the complaint at the millisecond level, and generates a multimodal physiological time series dataset. The dynamic feature warping module inputs the multimodal physiological time series dataset into the dynamic time warping algorithm for time axis calibration, calculates the rate of change of the comprehensive disease score, extracts the extreme value difference of abnormal fluctuations of vital signs in multidimensional physiological objective indicators, and generates the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. The time-domain cross-correlation analysis module calculates the Pearson correlation coefficient based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. When the correlation coefficient exceeds the preset threshold and the peak time difference is less than the set duration, it is marked as a high-risk coupling event. The risk space mapping module processes high-risk coupled events in segments with fixed durations, calculates the ratio of the time integral to the time length of the comprehensive disease score, statistically analyzes the crossover rate of the waveforms of multidimensional physiological objective indicators, and uses the support vector machine algorithm to perform feature space mapping on the ratio of the comprehensive disease score to generate a patient state risk matrix. The emergency triage decision module calls upon the patient status risk matrix grading results to assess the patient's current health status, converts it into the corresponding priority code according to the emergency four-level triage standard, and generates a dynamic triage level identifier.

[0042] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the technical solution.

Claims

1. An emergency triage method based on time-axis audio-visual interpretation and clinical decision support, characterized in that, Includes the following steps: S1: Collect the patient's voice complaint signal and input it into the pre-trained acoustic-language joint model to extract the semantic vocabulary and timestamps of the complaint symptoms, obtain the semantic feature sequence of the complaint, obtain the comprehensive disease score and multidimensional physiological objective indicators, align them with the timestamps of the semantic feature sequence of the complaint at the millisecond level, and generate a multimodal physiological time series dataset; S2: Input the multimodal physiological time series dataset into the dynamic time warping algorithm for time axis calibration, calculate the rate of change of the comprehensive disease score, extract the extreme difference of abnormal fluctuations of vital signs in multidimensional physiological objective indicators, and generate the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. S3: Perform time-domain cross-correlation analysis based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological signs to calculate the Pearson correlation coefficient. When the correlation coefficient exceeds the preset threshold and the peak time difference is less than the set duration, it is marked as a high-risk coupling event. S4: The high-risk coupling events are processed in segments with fixed durations. The ratio of the time integral to the time length of the comprehensive disease score is calculated. The crossover rate of the waveforms of multidimensional physiological objective indicators is statistically analyzed. The ratio of the comprehensive disease score is mapped to the feature space using the support vector machine algorithm to generate a patient state risk matrix.

2. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 1, characterized in that, The multimodal physiological time-series dataset includes speech fundamental frequency features, MEWS modified early warning score, and vital sign cycle diagram. The comprehensive disease score gradient vector and physiological sign fluctuation amplitude vector include the first derivative of the score, the score range, the duration of abnormal heart rhythm, and the hemodynamic index drift. The high-risk coupling events include paroxysmal tachycardia segments, sympathetic nerve excitation periods, and hemodynamic abnormalities. The patient status risk matrix includes complication rate, disease deterioration probability, and clinical prognostic index.

3. The emergency triage method based on time-axis interpretation and clinical decision support according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect the patient's voice complaint signal, perform inner product operation according to the language transformation matrix, extract the semantic words and timestamps of the complaint symptoms, construct a discrete vector array by mapping the semantic words of the complaint symptoms to the space, and concatenate the discrete vector arrays according to the timestamps to generate the semantic feature sequence of the complaint. S102: Obtain the comprehensive disease score and multidimensional physiological signals through physiological monitoring equipment, extract the periodic sequence of physiological signals, call the semantic feature sequence of the chief complaint to calculate the dimension and truncate the periodic sequence to construct the truncated electrocardiogram sequence, aggregate the comprehensive disease score and truncated sign sequence, and generate a set of basic physiological indicators. S103: Construct a clock baseline based on the timestamp of the chief complaint semantic feature sequence, extract collection nodes for the set of basic physiological indicators, compare the deviation between the collection nodes and the timestamp, establish a key-value pair mapping relationship when the deviation is less than the tolerance baseline value and write it into the multimodal time series feature matrix to generate a multimodal physiological time series dataset.

4. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Extract multidimensional timestamps from the multimodal physiological time series dataset, calculate spatial Euclidean distance to construct a two-dimensional distance matrix, traverse the two-dimensional distance matrix to calculate the cumulative path cost of adjacent nodes, select the sequence with the minimum cumulative path cost as the time series normalization path and resample the values ​​to generate a calibrated physiological feature sequence. S202: Call the calibration physiological feature sequence to extract the comprehensive disease score data, read the score values ​​of adjacent nodes in the comprehensive disease score data to calculate the absolute difference, calculate the first derivative based on the absolute difference and the corresponding time interval, arrange the first derivative in ascending order according to the timestamp, and obtain the gradient vector of the comprehensive disease score. S203: Based on the calibrated physiological feature sequence, retrieve the physiological sign dimension features, search for local maxima within the physiological sign dimension features to determine the peak nodes, quantify the absolute difference of amplitude between adjacent nodes to construct an amplitude feature array, perform dimension alignment processing according to the time axis coordinates, and obtain the physiological sign fluctuation amplitude vector.

5. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations, extract the time-domain segmented sequence, quantify the dispersion and co-variation trend of the node distribution within the time-domain sequence, and perform time-domain cross-correlation analysis based on the dispersion and co-variation trend to obtain the Pearson correlation coefficient. S302: Call the gradient vector of the comprehensive disease score to traverse the time window values, compare the values ​​of adjacent nodes to filter local maxima points and extract the gradient peak timestamp parameter, extract the amplitude peak timestamp parameter for the amplitude vector of physiological sign fluctuations, construct a pairing combination of gradient peak timestamp parameter and amplitude peak timestamp parameter, calculate the absolute value of the difference in the horizontal axis coordinate, and generate the peak time difference; S303: Based on the Pearson correlation coefficient and peak time difference triggering dual judgment logic, the Pearson correlation coefficient is compared with the preset correlation threshold, the peak time difference is compared with the set duration, the segmented feature node allocation status identifier that satisfies the dual judgment logic is extracted, and a high-risk coupling event is established.

6. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 5, characterized in that, The preset association threshold is set based on the following: retrieve the normal disease score benchmark vector and the normal sign amplitude benchmark vector from the historical sample database, perform Pearson association calculation on the normal disease score benchmark vector and the normal sign amplitude benchmark vector to generate a benchmark association coefficient sequence, extract the upper limit edge value of the benchmark association coefficient sequence in the pre-configured confidence interval, and configure the upper limit edge value as the preset association threshold. The set duration is based on the following: retrieve the set of peak timestamp parameters of stress symptoms and the set of peak timestamp parameters of stress signs from the historical sample database, calculate the absolute value of the difference between the elements in the set of peak timestamp parameters of stress symptoms and the corresponding elements in the set of peak timestamp parameters of stress signs, generate a baseline time difference sequence, extract the maximum distribution limit parameter of the baseline time difference sequence, and configure the maximum distribution limit parameter as the set duration. The condition that the dual-judgment logic is satisfied is that the Pearson correlation coefficient is greater than the preset correlation threshold and the peak time difference is less than the set duration.

7. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: The high-risk coupling event is divided into fixed-length analysis windows at equal intervals along the time axis. The time series parameters of the comprehensive disease score inside the fixed-length analysis window are retrieved. The time series parameters of the comprehensive disease score are summed to obtain the cumulative area. The average value of the time span of the fixed-length analysis window is then processed to obtain the ratio of the comprehensive disease score. S402: Call the high-risk coupling event to read the physiological waveform parameters, configure the upper and lower limits of the safety threshold baseline to traverse the physiological waveform parameters, detect out-of-bounds flip nodes, count the total number of out-of-bounds flip nodes inside the fixed-length analysis window, and calculate the ratio with the time span value to obtain the waveform out-of-bounds crossover rate. S403: Construct a joint feature vector based on the comprehensive score ratio of the condition and the waveform crossover rate, call the preset kernel function to calculate the dot product of the joint feature vector and the sample set, aggregate the dot product and the Lagrange multiplier to obtain the hyperplane decision distance, project the joint feature vector to the risk quadrant, and establish the patient status risk matrix.

8. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 1, characterized in that, The method also includes step S5: S5: Call the patient status risk matrix classification results to assess the patient's current health status, convert it into the corresponding priority code according to the emergency four-level triage standard, and generate a dynamic triage level identifier; The dynamic triage level identifier includes emergency response time limit, triage level code, and priority order of treatment.

9. The emergency triage method based on time-axis audio-visual translation and clinical decision support according to claim 8, characterized in that, The specific steps of S5 are as follows: S501: Call the patient status risk matrix to extract diagonal values ​​to construct a feature dimensionality reduction sequence, obtain the health baseline distribution function of the preset sample, calculate the absolute deviation between the feature dimensionality reduction sequence and the mean of the health baseline distribution function, extract the deviation out-of-bounds sequence to locate high-risk vital signs parameters, and establish a status assessment vector. S502: Based on the state assessment vector, retrieve the emergency triage standard mapping table, extract the corresponding upper and lower limit boundary conditions in the emergency triage standard mapping table, detect the hierarchical position of the high-risk vital sign parameters falling into the interval of the upper and lower limit boundary conditions, allocate discrete integer variables according to the hierarchical position, and obtain priority codes; S503: Read the timestamp of the clock generator's operating cycle, concatenate the priority code and the timestamp bit by bit to construct a basic code segment, perform modulo-2 division on the basic code segment to extract the remainder parameter as an additional check segment, combine the basic code segment and the additional check segment to generate a dynamic triage level identifier.

10. An emergency triage system based on time-axis audio-visual translation and clinical decision support, characterized in that, The system is used to implement the emergency triage method based on time-axis interpretation and clinical decision support as described in any one of claims 1-9, and the system includes: The physiological time series acquisition module collects the patient's voice complaint signal and inputs it into a pre-trained acoustic-language joint model to extract the semantic vocabulary and timestamps of the complaint symptoms, obtains the semantic feature sequence of the complaint, obtains the comprehensive disease score and multidimensional physiological objective indicators, aligns them with the timestamps of the semantic feature sequence of the complaint at the millisecond level, and generates a multimodal physiological time series dataset. The dynamic feature warping module inputs the multimodal physiological time series dataset into the dynamic time warping algorithm for time axis calibration, calculates the rate of change of the comprehensive disease score, extracts the extreme difference of abnormal fluctuations of vital signs in multidimensional physiological objective indicators, and generates the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. The time-domain cross-correlation analysis module calculates the Pearson correlation coefficient based on the gradient vector of the comprehensive disease score and the amplitude vector of physiological sign fluctuations. When the correlation coefficient exceeds the preset correlation threshold and the peak time difference is less than the set duration, it is marked as a high-risk coupling event. The risk space mapping module processes the high-risk coupled events in segments with fixed durations, calculates the ratio of the time integral to the time length of the comprehensive disease score, statistically analyzes the crossover rate of the waveforms of multidimensional physiological objective indicators, and performs feature space mapping on the ratio of the comprehensive disease score using a support vector machine algorithm to generate a patient state risk matrix. The emergency triage decision module calls the patient status risk matrix grading results to assess the patient's current health status, converts it into the corresponding priority code according to the emergency four-level triage standard, and generates a dynamic triage level identifier.

Images