A pediatric emergency pre-examination triage method based on crying sound emotion recognition

Abstract

The application discloses a pediatric emergency pre-examination triage method based on crying sound emotion recognition and belongs to the technical field of emergency triage, and specifically comprises the following steps: constructing a sick child crying sound and health information dataset; extracting the frequency, intensity and rhythm characteristics in the audio to form standardized audio feature data; dividing the feature categories according to the age of the sick children, associating the standardized audio features of the same age group with corresponding emotion labels, and establishing an age-based emotion feature library; training an emotion recognition model based on the age-based emotion feature library, determining the corresponding relationship between different emotion types and the emergency degree of pediatric emergency conditions; collecting the crying sound of a newly admitted sick child in real time, inputting the trained emotion recognition model, and outputting the emotion type and emergency degree corresponding to the crying sound of the sick child; calling the basic health information of the newly admitted sick child, combining the emergency degree obtained by the emotion recognition model, generating a targeted pre-examination triage scheme, and synchronizing the pre-examination triage scheme to the pediatric emergency triage management system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of emergency triage, in particular to a pediatric emergency pre-examination triage method based on cry emotion recognition. BACKGROUND

[0002] Pediatric emergency is a high-demand scenario in medical services, with rapid changes in the condition of sick children and anxious emotions of their families. Efficient and accurate pre-examination triage is a key link to ensure the order of diagnosis and treatment and the safety of sick children. The existing pediatric emergency pre-examination triage method mainly relies on the clinical experience of medical staff, combined with the description of the symptoms of sick children by their families and basic physiological signs such as body temperature and heart rate for judgment. Some medical institutions may use standardized triage scales to assist in evaluation, but the core still relies on subjective judgment and single-dimensional objective signs as the main basis, and has not fully combined the unique non-verbal expression information of sick children for triage decision-making.

[0003] The core problem of the existing pediatric emergency pre-examination triage technology is that it fails to effectively integrate the key information of the cries of sick children, which directly reflects the emotional and disease state, and lacks targeted consideration of the physiological characteristics of children of different ages, resulting in limitations such as strong subjectivity and incomplete judgment dimensions in triage decision-making. The description of symptoms by families is easily influenced by emotions, and a unified evaluation standard is difficult to adapt to the differences in cries and disease manifestations of children of different ages, ultimately resulting in insufficient accuracy in judging the emergency level of the disease, which may either delay the disposal of emergency cases or cause non-emergency cases to occupy high-quality medical resources, affecting the overall diagnosis and treatment efficiency and service quality of pediatric emergency. SUMMARY

[0004] The purpose of the present application is to provide a pediatric emergency pre-examination triage method based on cry emotion recognition, which solves the problems in the background art:

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A pediatric emergency pre-examination triage method based on cry emotion recognition, comprising the following steps:

[0007] S1, collecting the cry audio data and basic health information of the pediatric emergency sick children to be triaged, and associating the two types of information, to construct a sick child cry and health information dataset;

[0008] S2, performing noise removal processing on the cry audio in the dataset, extracting the frequency, intensity and rhythm characteristics in the audio, and forming standardized audio feature data;

[0009] S3, dividing the features into categories according to the age of the sick children, associating the standardized audio features of the same age group with the corresponding emotion labels, and establishing an age-specific emotion feature library;

[0010] S4, training the emotion recognition model based on the age-based emotion feature library, to determine the corresponding relationship between different emotion types and the emergency degree of the pediatric emergency condition;

[0011] S5, real-time collection of the crying audio of the newly admitted child, inputting the trained emotion recognition model, and outputting the corresponding emotion type and emergency degree of the child's crying;

[0012] S6, retrieving the basic health information of the newly admitted child, combining the emergency degree obtained by the emotion recognition model, generating a targeted pre-examination triage scheme, and synchronizing the pre-examination triage scheme to the pediatric emergency triage management system.

[0013] As a further scheme of the application: in the step S2, the process of extracting the frequency, intensity and rhythm characteristics in the audio to form the standardized audio feature data is:

[0014] The noise-removed crying audio is cut into continuous audio segments according to a fixed time length, ensuring that each segment contains a complete crying rhythm, and defining a uniform analysis unit for feature extraction;

[0015] Perform feature extraction on each audio segment, obtain frequency characteristics through segmented spectrum analysis, calculate the ratio of peak value to average amplitude in the segment to obtain intensity characteristics, and record the time interval sequence of adjacent peaks as rhythm characteristics;

[0016] Integrate the extracted frequency, intensity and rhythm characteristics into a feature vector according to a predetermined order, adjust the value range of each index in the feature vector through a unified numerical conversion rule, and form standardized audio feature data.

[0017] As a further scheme of the application: in the step S3, the process of dividing the feature categories according to the age of the child, associating the standardized audio features of the same age group with the corresponding emotion labels, and establishing the age-based emotion feature library is:

[0018] Pretreat the standardized audio features, use unsupervised clustering algorithm, preliminarily group according to the similarity between feature vectors, and form several basic feature clusters;

[0019] For each basic feature cluster, analyze the age distribution of the children in it, determine the main age range corresponding to each cluster, and take this range as the age identification information of the cluster;

[0020] Associate the basic feature clusters with age identification information with the corresponding emotion labels, classify and integrate the associated feature clusters according to the age identification, and construct the age-based emotion feature library.

[0021] As a further scheme of the present application: the specific manner of the preprocessing of the standardized audio features, adopting an unsupervised clustering algorithm, and preliminarily grouping according to the similarity between the feature vectors to form a plurality of basic feature clusters is as follows:

[0022] For each feature vector in the standardized audio features, the Euclidean distance between the feature vector and all other feature vectors in the data set is calculated to generate a distance matrix as the basic data for feature enhancement;

[0023] According to the distance matrix, the elements of each feature vector are weighted and adjusted, and the smaller the distance, the greater the weight corresponding to the feature vector, so as to highlight the common information of similar features and obtain an enhanced feature vector;

[0024] A density-based clustering algorithm is adopted, the enhanced feature vector is taken as input, the density threshold and neighborhood radius parameters are set, the feature vectors with connected density in the feature space are classified into a class, and a basic feature cluster is formed.

[0025] As a further scheme of the present application: in the step S4, the process of training the emotion recognition model based on the age-based emotion feature library and determining the corresponding relationship between different emotion types and the emergency degree of the pediatric emergency condition is as follows:

[0026] The age-based emotion feature library is split into independent training sets and test sets according to age groups to ensure that the distribution of emotion labels in each set is consistent with the original library;

[0027] A multi-branch neural network is constructed, the basic layer shares the extraction of general audio features, each age group corresponds to an independent branch layer, and the emotion type of the age group is output;

[0028] The corresponding branch is trained using the training set of each age group, the model recognition accuracy is verified through the test set, and the network parameters are adjusted until the model performance is stable;

[0029] The actual priority data of the emergency treatment of the children corresponding to each emotion type is collected, and a fixed corresponding relationship between the emotion type and the emergency degree is established by using a statistical analysis method.

[0030] As a further scheme of the present application: the specific manner of constructing the multi-branch neural network, the basic layer sharing the extraction of general audio features, each age group corresponding to an independent branch layer, and the emotion type of the age group being output is as follows:

[0031] The standardized audio features are taken as input, a convolution layer is constructed as the basic layer, and general audio features are extracted through convolution operation;

[0032] An independent branch is constructed for each age group, the general features output by the basic layer are input into the fully connected layer of the branch together with the numerical code of the age group, and feature adaptation is performed;

[0033] A multi-classification output layer is arranged at the end of each branch, and a multi-classification activation function is used to output the probability of each emotion type in the age range, so as to realize emotion type output.

[0034] As a further scheme of the present application, in the step S6, the process of generating a targeted pre-examination and triage scheme is as follows:

[0035] The unique identifier of the newly admitted child is taken as the retrieval basis to retrieve the corresponding basic health information from the child crying and health information dataset, and accurate information matching is completed.

[0036] The key items in the basic health information are extracted and input into a feature fusion module together with the emergency degree output by the emotion recognition model for correlation analysis to obtain a comprehensive evaluation result.

[0037] According to the comprehensive evaluation result, the preset pediatric emergency triage rule library is matched to determine the treatment order, corresponding diagnosis room and preliminary nursing suggestion, and a targeted pre-examination and triage scheme is formed.

[0038] As a further scheme of the present application, the specific way of extracting the key items in the basic health information and inputting them into the feature fusion module together with the emergency degree output by the emotion recognition model for correlation analysis to obtain a comprehensive evaluation result is as follows:

[0039] The past medical history, current body temperature and allergy history in the basic health information are extracted as key items, and each key item is converted into a standardized numerical code.

[0040] The coded key items and the emergency degree output by the emotion recognition model are assigned a preset weight, and are integrated in proportion to the weight to form a fusion feature vector input module.

[0041] The feature fusion module performs splicing and logical judgment on the fusion feature vector to output a comprehensive evaluation result containing health risks and emergency levels.

[0042] The present application has the following advantages:

[0043] The application significantly improves the accuracy and objectivity of the triage decision by deeply combining the infant crying sound emotion recognition with the pediatric emergency pre-examination triage. By collecting infant crying sound audio data, extracting frequency, intensity and rhythm characteristics and standardizing processing, and constructing an age-based emotion feature library according to age, the physiological development differences of infants of different ages are targetedly adapted, and the problem of insufficient adaptability caused by uniform evaluation standard is avoided. The construction of the multi-branch neural network enables the model to accurately identify the corresponding emotion type of the crying sound of infants of different ages, and combined with the prior medical history, body temperature and other basic health information for multi-dimensional fusion analysis, breaking the limitations of traditional triage relying on artificial subjective judgment and single physiological indicators, effectively reducing the triage errors caused by family description deviation or symptom misjudgment, providing a more comprehensive and reliable basis for the judgment of the emergency degree, and ensuring that emergency cases can be identified and disposed of in time.

[0044] The application significantly shortens the time cost of pre-examination triage by collecting real-time new infant crying sound audio, the trained model quickly outputs the emotion type and emergency degree, and then automatically generates a targeted pre-examination triage scheme combined with the basic health information and synchronizes it to the management system. The whole realizes the intelligentization and standardization of triage decision, reduces the dependence on individual experience of medical staff, helps to improve the overall diagnosis and treatment order and service quality of pediatric emergency, and provides more efficient and accurate medical protection for infants. BRIEF DESCRIPTION OF DRAWINGS

[0045] The application will be further described below in conjunction with the accompanying drawings.

[0046] Figure 1 It is a flowchart of a pediatric emergency pre-examination triage method based on crying sound emotion recognition. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0048] Please refer to Figure 1 The application is a pediatric emergency pre-examination triage method based on crying sound emotion recognition, which comprises the following steps:

[0049] S1, collecting the crying sound audio data and basic health information of the pediatric emergency patients to be triaged, associating the two types of information, and constructing a crying sound and health information data set;

[0050] S2, noise removal processing is performed on the crying audio in the data set, frequency, intensity and rhythm characteristics in the audio are extracted, and standardized audio characteristic data is formed;

[0051] S3, the characteristic categories are divided according to the age of the children, the standardized audio characteristics of the same age group are associated with the corresponding emotional labels, and an age-based emotional feature library is established;

[0052] S4, training an emotional recognition model based on the age-based emotional feature library, determining the corresponding relationship between different emotional types and the emergency degree of pediatric emergency conditions;

[0053] S5, real-time collection of the crying audio of the new children in the hospital, inputting the trained emotional recognition model, and outputting the corresponding emotional type and emergency degree of the crying audio of the children;

[0054] S6, calling the basic health information of the new children in the hospital, combining the emergency degree obtained by the emotional recognition model, generating a targeted pre-examination triage scheme, and synchronizing the pre-examination triage scheme to the pediatric emergency triage management system.

[0055] In a preferred embodiment of the present application, in step S1, the crying audio data and basic health information of the children in the pediatric emergency triage are collected, and the two types of information are associated, and the specific content of the children's crying and health information data set is:

[0056] Deploy the device suitable for collecting children's crying in the pediatric emergency triage area, when the children to be triaged cry, the medical staff starts the device to record the crying audio in time, avoids environmental interference during recording, and ensures that the audio is clear and identifiable. The basic health information is obtained by the medical staff through communication with the parents, and is synchronized and checked with the existing records in the hospital emergency pre-examination registration system. The core collects the age, gender, current temperature, medical history, allergy and current main symptoms of the children, and accurately enters the system after the collection is completed. The unique medical identification of the children is used as the association basis, and the recorded crying audio and the basic health information entered into the system are matched one by one, so that each audio data can be accurately associated with the complete health information of the corresponding children. After the matching is correct, all the associated data are sorted and summarized according to the unified rules, and finally the crying and health information data set of the children is constructed.

[0057] In another preferred embodiment of the present application, in step S2, the process of extracting the frequency, intensity and rhythm characteristics in the audio to form standardized audio characteristic data is:

[0058] In the audio feature extraction stage, first, the crying audio after noise removal is segmented. The complete audio is divided into continuous segments according to a fixed time length, and the setting of the time length is based on the principle of ensuring that each segment contains a complete crying syllable. For example, if the crying syllable of the child usually lasts about 1.5 seconds, the segmentation time can be set to 2 seconds, so that the syllable can be completely covered, and the feature redundancy caused by too long segments can be avoided. In this way, a unified and standardized analysis unit is defined for subsequent feature extraction, so that each segment is comparable.

[0059] Then, specific feature extraction operations are performed on each audio segment. For frequency features, a segmented spectrum analysis method is used to convert the audio segment into a spectrum graph through specific signal processing means. From the spectrum graph, the distribution of different frequency components can be observed directly, and then the key information reflecting the frequency characteristics of the crying sound can be extracted. Taking intensity feature extraction as an example, first find the maximum amplitude of the sound wave in the segment, then calculate the average amplitude of the entire segment, and the ratio of the two is the intensity feature value, which can reflect the strength change of the crying sound. In the rhythm feature extraction, the position of the sound wave peak is detected, and the time interval between the adjacent two peaks is recorded. The sequence formed by these intervals constitutes the rhythm feature, for example, the interval between the peaks of the fast and continuous crying sound is short, and the rhythm feature value will show the corresponding regularity.

[0060] Finally, the features are integrated and standardized. The extracted frequency, intensity, and rhythm features are combined into a multi-dimensional feature vector in a predetermined order, such as frequency features first, then intensity features, and finally rhythm features. Since the value range of different features may differ greatly, such as the frequency value may be in the range of hundreds to thousands of hertz, and the intensity feature ratio may be in the range of several times to dozens of times, which will affect the training effect of the subsequent model. Therefore, a unified numerical conversion rule is needed to adjust the values of all feature indicators to a same reasonable range, such as between 0 and 1, so as to form standardized audio feature data.

[0061] In another preferred embodiment of the present application, in the step S3, the feature categories are divided according to the age of the child, and the standardized audio features of the same age group are associated with the corresponding emotion labels to establish a divided-age emotion feature library.

[0062] In the initial stage of constructing the divided-age emotion feature library, the preprocessed audio feature data is first preprocessed to ensure that the data quality meets the clustering requirements. These standardized features have adjusted the value range through a unified rule, eliminating the dimensional differences. Next, an unsupervised clustering algorithm is used to preliminarily group the data according to the similarity between feature vectors, forming several basic feature clusters, and the specific process is as follows:

[0063] First, the Euclidean distance between each feature vector and all other feature vectors in the dataset is calculated. The Euclidean distance is a common way to measure the similarity of two multi-dimensional feature vectors, with a smaller distance indicating that the two feature vectors correspond to cry audio with similar characteristics in terms of frequency, intensity, rhythm, etc. For example, if feature vector A corresponds to a cry audio emitted by a baby due to abdominal pain, and feature vector B corresponds to a cry audio emitted by another baby due to a fever, the two differ in terms of frequency peak and rhythm interval, and the calculated Euclidean distance will be larger. If feature vector C also corresponds to a cry audio emitted by a baby due to abdominal pain, its Euclidean distance with vector A will be smaller. By calculating the distance between all pairs of feature vectors, a complete distance matrix is generated, which clearly presents the similarity relationship between any two feature vectors and serves as the basis data for subsequent feature enhancement.

[0064] Subsequently, the elements of each feature vector are adjusted by weighting according to the distance matrix. The adjustment rule is that the smaller the distance, the greater the weight of the corresponding feature vector. The core purpose of this design is to highlight the common information of similar features. For example, in the distance matrix, the k feature vectors closest to feature vector D often represent audio with highly similar emotional expressions to D, such as both being intense and painful. By giving these similar vectors a higher weight, the adjustment of the feature elements of D can incorporate more of these common features, resulting in an enhanced feature vector. This enhancement process can effectively reduce the interference of individual differences on feature clustering, allowing cry features of the same emotion to be more concentrated.

[0065] Finally, a density-based clustering algorithm is used, with the enhanced feature vectors as input data. Two key parameters need to be set during clustering: the density threshold and the neighborhood radius. The neighborhood radius is used to define the neighbor range of a feature vector, i.e., other vectors within the neighborhood radius of the vector are considered neighbors. The density threshold is the standard for determining whether a region forms a cluster. If the number of neighbors of a feature vector reaches or exceeds the density threshold, the region is identified as a core region, and all density-connected core regions and surrounding neighbor vectors are classified into a class. For example, if the neighborhood radius is set to r and the density threshold is set to m, when the number of neighbors of feature vector E reaches m, E becomes a core vector, and its neighbor vectors, including the neighbors of other core vectors, will gradually aggregate to form a basic feature cluster. Through this process, all feature vectors are divided into several non-overlapping basic feature clusters, and the feature vectors within each cluster have a high degree of similarity, corresponding to cry audio with a certain type of emotional expression.

[0066] After the division of the basic feature clusters, age distribution analysis needs to be performed on each cluster to determine its corresponding main age range, and this range is taken as the age identification information of the cluster. Specifically, first, the age information of all children in each basic feature cluster is extracted, which comes from the previously constructed children's crying and health information dataset and has been accurately associated with the audio feature data.

[0067] Next, statistical analysis is performed on the age data within each cluster, such as calculating the proportion of children in different age groups (e.g. 0-1 years old, 1-3 years old, 3-6 years old, etc.) within the cluster. For example, in a certain basic feature cluster, there are a total of 100 children, of which 65 are 0-1 years old, accounting for 65%; 25 are 1-3 years old, accounting for 25%; and 10 are 3-6 years old, accounting for 10%. Through this statistics, it can be clearly seen that the age distribution of children in this cluster is mainly 0-1 years old, so 0-1 years old is determined as the main age range of this cluster.

[0068] If the age distribution of a certain cluster is relatively dispersed, such as 0-1 years old accounting for 40%, 1-3 years old accounting for 35%, and 3-6 years old accounting for 25%, further analysis of the feature commonality of the cluster is needed, combined with clinical experience to determine which age group it is more inclined to express emotions. For example, if the feature vector of this cluster shows the typical high-frequency characteristics of infant crying in frequency features, even if the proportion of 1-3 year-old children is close, it can still be determined that the main age range is 0-1 years old. Finally, each basic feature cluster is given a unique age identification information, establishing a clear correspondence between the cluster and the age range of the children.

[0069] After adding age identification to the basic feature clusters, the next step is to associate these clusters with corresponding emotional labels. Emotional labels are used to describe the types of emotions expressed by the crying audio, such as pain, hunger, fear, irritability, etc. The determination of these labels is based on clinical observation and professional judgment of medical staff. For example, by analyzing the medical record information and treatment scene of a certain basic feature cluster, if most children in this cluster show obvious symptoms of pain (such as severe abdominal pain, high fever convulsions, etc.) at that time, and the intensity and rhythm characteristics of their crying audio are highly consistent with the typical manifestations of pain emotions, then the pain emotion label is associated with this cluster.

[0070] After the association of the feature clusters and the emotional labels is completed, all the associated feature clusters are classified and integrated according to the age identifiers. For example, all the feature clusters with an age identifier of 0-1 year old are classified into one category and integrated into a 0-1 year old emotional feature set; all the feature clusters with an age identifier of 1-3 years old are classified into another category and integrated into a 1-3 year old emotional feature set, and so on. In the integration process, the feature clusters of each age group also need to be combed again to ensure that the clusters of the same emotional type have consistency in features and the clusters of different emotional types have obvious distinguishing degrees in features. Finally, these classified and integrated feature sets according to the ages jointly constitute the age-specific emotional feature library.

[0071] In another preferred embodiment of the present application, in the step S4, the process of training an emotional recognition model based on the age-specific emotional feature library and determining the corresponding relationship between different emotional types and the emergency degree of the pediatric emergency condition is as follows:

[0072] Before the training of the emotional recognition model is carried out, the data of the constructed age-specific emotional feature library is first split, and the age-specific emotional feature library is divided into independent training sets and test sets according to the age groups. The age-specific emotional feature library is classified according to the age groups, and each age group contains feature data corresponding to multiple emotional types. The splitting process needs to strictly follow the principle of independent splitting of data within the same age group to avoid the mixing of data of different age groups into the same set.

[0073] When splitting, the distribution of emotional labels in the training set and the test set needs to be ensured to be consistent with the original age-specific feature library. For example, in the original feature library of a certain age group, the proportion of pain emotional features is the highest, and the proportions of hunger, irritability and fear emotional features are distributed in turn. When splitting, the proportions of various emotional features in the training set and the test set need to be matched with the original library through random sampling. The purpose of this is to avoid the model training from being biased towards a certain type of emotion due to unbalanced data distribution, and to ensure that the model can uniformly learn the features of all emotional types of the age group. After splitting, each age group will have exclusive training sets and test sets.

[0074] In the model construction stage, a multi-branch neural network structure is adopted to extract common features through a basic layer, and then adapt to the feature differences of different age groups through independent branch layers, so as to finally realize the accurate output of emotional types. The specific process is as follows:

[0075] First, standardized audio features are taken as input, and a convolutional layer is constructed as the basic layer of the neural network. The core role of the convolutional layer is to extract general audio features, which are the basic attributes common to crying audio of different age groups, such as the basic spectral pattern of crying sound and the common rules of sound wave vibration. Through convolutional operation, the basic layer will filter and refine the input standardized features layer by layer, filter out irrelevant minor differences, and retain core common information that reflects emotional expression, laying the foundation for subsequent feature adaptation for different age groups. This step is like first extracting the essential features of "crying" from all crying sounds, and then optimizing them individually for different age groups.

[0076] Next, an independent branch layer is constructed for each age group to ensure that the model can accurately adapt to the crying feature differences of different age groups. First, each age group is converted into a corresponding numerical code, which is only used to identify age information and does not participate in feature operation, but only serves as a basis for branch layer identification of the corresponding age group. Then, the general audio features output by the basic layer are input into the branch's fully connected layer along with the numerical code of the age group. The fully connected layer will perform deep fusion and feature adaptation on these two types of information. For example, the crying frequency of 0-1 year old infants is generally high, and the fully connected layer will strengthen the weight of high-frequency related features based on the general features and weaken features that do not conform to the characteristics of this age group. The crying frequency of 3-6 year old children is relatively low, and the rhythm is more regular, so the fully connected layer will adjust the feature adaptation direction accordingly to make the features more suitable for the crying characteristics of this age group. In this way, each branch layer can form a feature processing logic that adapts to the corresponding age group.

[0077] Finally, a multi-class output layer is set at the end of each branch, and a multi-class activation function is used to complete the emotion type output. The multi-class output layer will receive the feature data adapted by the fully connected layer, calculate the matching degree of each emotion type through the activation function, and finally output the corresponding probability of all preset emotion types for this age group. For example, the output layer of the 0-1 year old branch will output the probability values of pain, hunger, irritability, and fear, and the probability values represent the matching degree of the current input audio with the emotion type. The model will determine the most likely emotion type corresponding to the current crying sound according to the distribution of probability values, thereby realizing the emotion recognition function for a specific age group. The entire neural network structure not only ensures the efficient reuse of general features, but also realizes the individual adaptation of age groups through independent branches, balancing the training efficiency and recognition accuracy of the model.

[0078] The model training stage adopts the method of "age training and age verification" to ensure that each age branch can achieve stable recognition performance. For each age group, input its exclusive training set into the corresponding branch layer, and let the branch layer gradually learn the feature rules of different emotion types in this age group through continuous iteration. During the training process, the model will continuously adjust the internal parameters to optimize the correspondence between features and emotion types, such as making the high-frequency and high-intensity features of the "pain" emotion type form a more stable association with the "pain" label.

[0079] After a certain stage of training, the test set of this age group is introduced to verify the model performance. Input the feature data in the test set into the trained model to obtain the emotion type recognition result, and then compare it with the real emotion label in the test set to judge the accuracy of the model recognition. If the verification finds that the recognition result of a certain emotion type deviates greatly from the real label, such as frequently identifying "fear" type crying as "irritation", it needs to adjust the network parameters, such as optimizing the feature adaptation logic of the fully connected layer or adjusting the feature extraction granularity of the convolution layer. Through the cycle of training-verification-adjustment, repeatedly optimize the network parameters until the recognition result of the model on the test set tends to be stable and no longer fluctuates greatly. At this time, it is considered that the branch training of this age group is completed. After all age branches complete the training, the whole multi-branch emotion recognition model can be put into subsequent use.

[0080] After the model training is completed, a fixed correspondence between different emotion types and the emergency degree of pediatric emergency conditions needs to be established to provide a basis for pre-examination triage. First, collect the actual handling priority data of children corresponding to each emotion type. These data come from the clinical diagnosis and treatment records of pediatric emergency, including the final diagnosis of children corresponding to different emotion types, the arrangement of treatment order, the allocation of medical resources, and other information. For example, collect the emergency handling records of all children identified by the model as "pain" emotion, and collect the handling records of other emotion types such as "hunger", "irritation", "fear", etc. to ensure that the data covers all preset emotion types.

[0081] Then, statistical analysis methods are used to organize and analyze the collected data. For each emotion type, summarize the actual handling priority of all children corresponding to it, and analyze the distribution rule of the emergency degree of children under this emotion type. For example, statistics show that among the children corresponding to the "pain" emotion, most have high fever convulsions, severe abdominal pain, trauma, and other urgent conditions that need to be treated immediately, and their handling priority is generally high. Children corresponding to the "hunger" emotion mostly have unmet physiological needs and no urgent conditions, and their handling priority is low. Children corresponding to the "fear" emotion have mild trauma or sudden discomfort, and their handling priority is between the two. Through this statistical analysis, the correspondence between each emotion type and the emergency degree of the condition is clear, and a fixed correspondence is finally established.

[0082] In a preferred embodiment of the present application, in step S6, the process of combining the basic health information of the newly admitted child with the emergency level obtained from the emotion recognition model to generate a targeted pre-examination triage plan is as follows:

[0083] When the newly admitted child completes the crying sound audio collection and emotion recognition, the process of retrieval and matching of basic health information is first started. The newly admitted child will obtain a unique identifier when registering in the emergency department, which can be a registration number, an electronic registration card number, etc. This identifier runs through the entire emergency process, ensuring the uniqueness and traceability of the child's information. Taking this unique identifier as the core retrieval basis, the system will automatically associate to the pre-constructed child crying sound and health information dataset and initiate a precise retrieval request. During the retrieval process, the system will traverse all associated records in the dataset according to the unique identifier to locate the health information item corresponding to the child. To avoid matching errors caused by identifier input errors or data duplication, the system will automatically check the key information after retrieval, such as the child's age, gender, and other basic information recorded in the dataset, and the information entered by the newly admitted child during registration. Only when the two pieces of information are completely consistent, the precise matching is confirmed, and the corresponding basic health information is completely retrieved to the current triage process; if the information is inconsistent, the system will prompt the medical staff to manually check and confirm to ensure that the information used for subsequent triage is accurate and reliable.

[0084] After the basic health information is retrieved, the key items need to be extracted and combined with the emergency level output by the emotion recognition model to construct a fusion feature vector. The specific process is as follows:

[0085] First, the key items are extracted. From the retrieved basic health information, three items that have a core impact on triage decision are selected, which are past medical history, current body temperature, and allergy history. When extracting the past medical history, the information related to emergency conditions is focused on, such as whether there are repeated respiratory infections, gastrointestinal discomfort, and other past emergency records, to ensure that the extracted information can reflect the child's basic health status and potential risks. The current body temperature information is directly taken from the real-time measurement results after the child is admitted to the hospital, without additional processing, only to ensure the authenticity and effectiveness of the data. The extraction of allergy history focuses on clear drug allergies, food allergies, or contact allergies. If the child has no relevant allergy records, it should also be extracted as clear information.

[0086] Next, the extracted key items are standardized numerical encoding. The core principle of encoding is to convert non-numeric information into a unified format of numerical value, which is convenient for subsequent weight allocation and feature integration. For the past medical history, according to its influence on the emergency condition, it is encoded, such as a serious chronic disease history is encoded as a numerical value, an ordinary past medical history is encoded as another numerical value, and no past medical history is encoded as a third numerical value. The current body temperature is encoded according to the clinical body temperature division standard, such as a normal body temperature range corresponds to a numerical value, mild fever corresponds to a numerical value, and moderate and above fever corresponds to a higher numerical value. The encoding logic of the allergy history is more concise, and the definite allergy record is encoded as a numerical value, and the non-allergy record is encoded as another numerical value. In this way, all key items are converted into standardized numerical values that can participate in subsequent calculations.

[0087] Then the encoded key items and the emergency degree output by the emotion recognition model are allocated with preset weights. The basis of weight allocation comes from the clinical diagnosis and treatment experience of pediatric emergency, which gives priority to indicators that have greater impact on the emergency degree. For example, the current body temperature as a real-time physiological indicator directly reflects the current physical condition of the child, and the weight allocation is relatively high; the emergency degree output by the emotion recognition model is the core judgment basis based on the crying sound emotion features, and also allocates a high weight; the past medical history as the basic health background, the weight is slightly lower than the first two; the allergy history mainly affects the safety of the subsequent treatment and nursing plan, and the weight is reasonably set according to the clinical importance. After the allocation is completed, all indicators are integrated according to the weight proportion, such as multiplying the encoding numerical value of each indicator with the corresponding weight, and then combining all the results in order to form a fusion feature vector containing multi-dimensional information, and finally inputting the vector into the feature fusion module.

[0088] After receiving the input fusion feature vector, the feature fusion module first performs a splicing operation. The splicing process is to combine the past medical history code, current body temperature code, allergy history code, and emergency level value contained in the fusion feature vector into a continuous comprehensive feature sequence in a predetermined order. The purpose of this step is to break the independence between indicators, so that the module can consider all relevant factors at the same time, and avoid decision bias caused by single indicator analysis. After splicing is completed, the module starts the logical judgment process, which is based on the clinical diagnosis and treatment specifications and triage experience of pediatric emergency department. For example, if the fusion feature vector shows that the current body temperature code corresponds to moderate and above fever, the emotional recognition emergency level value is high, and the past medical history code shows that the child has a history of epilepsy, the module will combine the medical knowledge that high fever can induce epilepsy to judge that the child has a high health risk; if the allergy history code shows that the child is allergic to certain common antipyretics, the module will take this factor into account and mark the risk points that need to be avoided in subsequent treatment. In the logical judgment process, the module will check the value combination of each indicator one by one, such as the combination of high emergency level and normal body temperature, no special past medical history, and the combination of high emergency level and high fever, chronic disease history, which will be judged as different health risk levels. Finally, the module outputs two core results, one is the health risk assessment, which clearly shows the possible disease risk and potential hazards of the child; the other is the emergency level judgment, which determines the priority of the child in emergency treatment, and the two together constitute the comprehensive evaluation result, which provides a direct basis for the generation of subsequent triage plan.

[0089] After getting the comprehensive evaluation result, the system will automatically match the pediatric emergency preset triage rule library to gradually generate a targeted pre-examination triage plan. The triage rule library is based on the diagnosis and treatment process of pediatric emergency department, division of department functions and clinical nursing specifications, and contains standardized processing rules corresponding to different health risks and different emergency levels, covering the core content of treatment order, clinic allocation, preliminary nursing suggestion, etc.

[0090] When determining the treatment order, the system will match the corresponding priority standard in the rule library according to the emergency level in the comprehensive evaluation result. For example, children with the highest emergency level will be included in the "priority treatment" sequence and directly arranged at the front of the available diagnosis and treatment resources; children with medium emergency level will be sorted according to the registration order and disease severity; children with lower emergency level will wait for treatment in the regular order. At the same time, the rule library will consider the type of health risk of the child, such as children with shortness of breath risk, even if the emergency level is the same as other children, the priority of treatment will be appropriately improved to ensure timely intervention.

[0091] In the clinic allocation link, the system matches the health risk direction in the comprehensive evaluation result with the diagnosis and treatment range of each clinic. For example, if the comprehensive evaluation result shows that the child patient may have a respiratory infection risk and is accompanied by high fever symptoms, the child patient is preferentially allocated to the pediatric emergency internal medicine clinic; if the evaluation result shows that the child patient may have a trauma or surgery-related condition, the child patient is matched to the surgery clinic; for a child patient with a high health risk that requires multidisciplinary collaborative diagnosis and treatment, the child patient is directly allocated to the emergency critical observation room, and the medical staff of the related departments are simultaneously notified to prepare.

[0092] The generation of the preliminary nursing suggestion focuses on relieving the discomfort of the child patient at the moment and avoiding potential risks. The system extracts corresponding rules according to the key health information and the comprehensive evaluation result. For example, for a child patient with high fever, it is suggested to perform physical cooling and keep the environment ventilated; for a child patient with a history of allergy, it is clearly suggested to avoid contact with the relevant allergens during the nursing process; for a child patient with a history of heart disease, it is suggested to monitor the heart rate and respiratory status and avoid crying to increase the heart burden.

[0093] The determined order of treatment, the corresponding clinic and the preliminary nursing suggestion are integrated and summarized to form a complete and targeted pre-examination triage plan. After the plan is generated, it is simultaneously displayed on the work terminal of the medical staff and pushed to the pediatric emergency triage management system to ensure that the triage information is shared in real time in the entire diagnosis and treatment process, provide clear work guidance for the medical staff, and also let the parents know the treatment arrangement and matters needing attention of the child patient.

[0094] The above describes one embodiment of the present application in detail, but the content described is only the preferred embodiment of the present application and cannot be considered to limit the implementation range of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A pediatric emergency pre-examination triage method based on crying emotional recognition, characterized in that, The method comprises the following steps: S1, collecting pediatric emergency waiting triage children's crying audio data and basic health information, corresponding association of the two types of information, and constructing children's crying and health information dataset; S2, noise removal processing is performed on the crying audio in the dataset, the frequency, intensity and rhythm characteristics in the audio are extracted, and standardized audio feature data is formed; S3, the feature categories are divided according to the age of the children, the standardized audio features of the same age group are associated with the corresponding emotional labels, and an age-based emotional feature library is established; S4, training an emotional recognition model based on the age-based emotional feature library to determine the corresponding relationship between different emotional types and the emergency degree of pediatric emergency; The specific process is as follows: The age-based emotional feature library is split into independent training set and test set according to age group, and the distribution of emotional labels in each set is consistent with the original library; A multi-branch neural network is constructed, the basic layer shares the extraction of general audio features, each age group has an independent branch layer, and the emotional type of the age group is output; The corresponding branch is trained using the training set of each age group, the model recognition accuracy is verified through the test set, and the network parameters are adjusted until the model performance is stable; The actual priority data of the children in emergency treatment corresponding to each emotional type is collected, and a fixed corresponding relationship between emotional type and emergency degree is established by statistical analysis method; The specific way of constructing a multi-branch neural network, the basic layer shares the extraction of general audio features, and each age group has an independent branch layer to output the emotional type of the age group is as follows: Take the standardized audio features as input, construct a convolution layer as the basic layer, and extract general audio features through convolution operation; For each age group, an independent branch is constructed, the general features output by the basic layer are input into the fully connected layer of the branch together with the numerical coding of the age group, and feature adaptation is performed; A multi-classification output layer is set at the end of each branch, a multi-classification activation function is used to output the probability of each emotional type in the age group, and the emotional type output is realized; S5, real-time collection of new children's crying audio, input into the trained emotional recognition model, output of the emotional type and the emergency degree corresponding to the children's crying; S6, calling the basic health information of the new children, combining with the emergency degree obtained by the emotional recognition model, generating a targeted pre-examination and triage scheme, and synchronizing the pre-examination and triage scheme to the pediatric emergency triage management system.

2. The pediatric emergency pre-check triage method based on cry emotion recognition according to claim 1, characterized in that, In step S2, the process of extracting the frequency, intensity and rhythm characteristics in the audio to form standardized audio feature data is as follows: The noise-removed crying audio is cut into continuous audio segments according to a fixed time length, ensuring that each segment contains complete crying rhythm, and defining a unified analysis unit for feature extraction; Perform feature extraction on each audio segment, obtain frequency characteristics through segmented spectrum analysis, calculate the ratio of sound wave peak value to average amplitude in the segment to obtain intensity characteristics, and record the time interval sequence of adjacent peaks as rhythm characteristics; The extracted frequency, intensity and rhythm characteristics are integrated into a feature vector in a predetermined order, the value range of each index in the feature vector is adjusted through a unified numerical conversion rule, and standardized audio feature data is formed.

3. The pediatric emergency pre-check triage method based on cry emotion recognition according to claim 1, characterized in that, In the step S3, the process of dividing the feature categories according to the age of the sick children, associating the standardized audio features of the same age group with the corresponding emotion labels, and establishing the age-specific emotion feature library is as follows: The standardized audio features are preprocessed, an unsupervised clustering algorithm is adopted, and the similarity between the feature vectors is used for preliminary grouping to form a plurality of basic feature clusters; For each basic feature cluster, the age distribution of the sick children in the cluster is analyzed, and the main age range corresponding to each cluster is determined as the age identification information of the cluster; The basic feature clusters with age identification information are associated with the corresponding emotion labels, and the associated feature clusters are classified and integrated according to the age identification to form the age-specific emotion feature library.

4. The pediatric emergency pre-check triage method based on cry emotion recognition according to claim 3, characterized in that, The specific way of preprocessing the standardized audio features, adopting an unsupervised clustering algorithm, and preliminarily grouping according to the similarity between the feature vectors to form a plurality of basic feature clusters is as follows: For each feature vector in the standardized audio features, the Euclidean distance between it and all other feature vectors in the data set is calculated to generate a distance matrix as the basic data for feature enhancement; According to the distance matrix, the elements of each feature vector are weighted and adjusted, and the feature vectors with smaller distances have larger weights, highlighting the common information of similar features, to obtain enhanced feature vectors; A density-based clustering algorithm is adopted, the enhanced feature vectors are taken as input, the density threshold and neighborhood radius parameters are set, and the feature vectors with connected density in the feature space are classified into a class to form a basic feature cluster.

5. The pediatric emergency pre-check triage method based on cry emotion recognition according to claim 1, characterized in that, In the step S6, the process of retrieving the basic health information of the newly admitted sick children and combining it with the emergency level obtained by the emotion recognition model to generate a targeted pre-examination triage plan is as follows: The unique identifier of the newly admitted sick children is used as the retrieval basis to retrieve the corresponding basic health information from the sick children's cry and health information data set, and accurate information matching is completed; The key items in the basic health information are extracted and input into the feature fusion module together with the emergency level output by the emotion recognition model for correlation analysis to obtain a comprehensive evaluation result; According to the comprehensive evaluation result, the pediatric emergency preset triage rule library is matched to determine the treatment order, corresponding examination room, and preliminary nursing suggestion to form a targeted pre-examination triage plan.

6. The pediatric emergency pre-check triage method based on cry emotion recognition according to claim 5, characterized in that, The specific way of extracting the key items in the basic health information and inputting them into the feature fusion module together with the emergency level output by the emotion recognition model for correlation analysis to obtain a comprehensive evaluation result is as follows: The past medical history, current body temperature, and allergy history in the basic health information are extracted as key items, and each key item is converted into a standardized numerical code; The coded key items and the emergency level output by the emotion recognition model are assigned with preset weights, and the integrated feature vector is input into the module according to the weight proportion; The feature fusion module performs splicing and logical judgment on the integrated feature vector to output a comprehensive evaluation result containing health risks and emergency levels.

Images