Child chief complaint symptom automatic extraction method and system based on natural language processing algorithm

Abstract

The invention provides a natural language processing algorithm-based child chief complaint symptom automatic extraction method and system, and relates to the technical field of data intelligent analysis, and the method comprises the steps: receiving child chief complaint text data containing parents' commission mixed information, and collecting the multi-modal behavior data generated in the commission process of parents; identifying and extracting A-class semantic anchor points and B-class semantic anchor points according to the chief complaint text data and the multi-modal behavior data of the children; the A-class semantic anchor points are derived from non-standardized expression of children and are intensively distributed in oral mucosa, throat and abdomen symptom areas; the class-B semantic anchor points are derived from information given by parents, and are concentrated in three dimensions of an eye staring mode, a voice rhythm characteristic and a limb movement amplitude; and constructing a compact semantic space corresponding to the A-class anchor point set through the A-class semantic anchor points. The method is suitable for accurate extraction and standardization processing of symptom information in a child medical diagnosis and treatment scene.

Description

Technical Field

[0001] This invention relates to the field of data intelligent analysis technology, and in particular to a method and system for automatically extracting children's subjective symptoms based on natural language processing algorithms. Background Technology

[0002] In pediatric medical settings, accurate extraction of chief complaints is a crucial prerequisite for disease diagnosis. Because children's language development is not yet mature, they cannot clearly and systematically describe their discomfort; therefore, clinical practice often relies on parents to describe symptoms. Existing symptom extraction methods based on natural language processing are primarily designed for standardized adult language and are ill-suited to the unique characteristics of children's complaints.

[0003] Existing technologies have the following shortcomings: For example, the methods rely solely on text data for semantic analysis, failing to fully integrate children's non-standardized expressive features with multimodal behavioral information from parents' descriptions, resulting in insufficient accuracy in symptom extraction. Children's symptom descriptions are often colloquial and fragmented, concentrated in core areas such as the oral mucosa, pharynx, and abdomen. Existing corpora lack targeted adaptation, making it difficult to accurately identify such non-standardized semantics. Furthermore, behavioral data such as eye gaze patterns, vocal prosody features, and body movement amplitude during parental descriptions contain supplementary semantic information related to symptoms; existing technologies do not incorporate this into the semantic analysis framework, making it impossible to correct semantic biases through multi-dimensional information.

[0004] The aforementioned problems directly lead to a large mapping error between children's non-standard expressions and standard medical terminology, making it difficult to effectively distinguish between core symptoms and accompanying phenomena, thereby affecting the accuracy of subsequent diagnosis and treatment decisions. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an automatic method for extracting children's chief complaints based on natural language processing algorithms, which realizes accurate extraction, standardized conversion and core symptom screening of children's chief complaints, and effectively improves the accuracy and practicality of children's symptom extraction.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for automatically extracting children's subjective symptoms based on natural language processing algorithms, the method comprising: It receives text data of children's complaints containing mixed information from parents' statements, and simultaneously collects multimodal behavioral data generated by parents during the statement process; Based on children's chief complaint text data and multimodal behavioral data, we identified and extracted type A and type B semantic anchors. Type A semantic anchors are derived from children's non-standardized expressions and are concentrated in the oral mucosa, pharynx and abdominal symptom areas. Type B semantic anchors are derived from information provided by parents and are concentrated in three dimensions: eye gaze patterns, speech rhythm features and body movement amplitude. A compact semantic space corresponding to the set of A-type semantic anchors is constructed using A-type semantic anchors; a wide-area semantic space corresponding to the set of B-type semantic anchors is constructed using B-type semantic anchors; and a calibration semantic anchor set is generated by calculating the intersection of the compact semantic space and the wide-area semantic space. The child's chief complaint text is semantically calibrated based on the set of calibrated semantic anchors to generate a semantically optimized chief complaint text; the semantically optimized chief complaint text is then preprocessed to generate an initial set of symptom entities. Based on a pre-defined corpus of children's language, dynamic semantic mapping is performed on the initial set of symptom entities to convert children's non-standardized expressions into standard medical terms, forming a set of standard symptom entities. Information filtering is performed on the standard symptom entity set to distinguish between core symptoms and accompanying phenomena, resulting in the core symptom entity set.

[0007] Furthermore, based on the children's subjective complaint text data and multimodal behavioral data, type A and type B semantic anchors were identified and extracted. Type A semantic anchors originated from children's non-standardized expressions and were concentrated in the oral mucosa, pharynx, and abdominal symptom areas. Type B semantic anchors originated from information provided by parents and were concentrated in three dimensions: eye gaze patterns, vocal prosody features, and body movement amplitude, including: The text data of children's complaints is preprocessed by word segmentation and entity recognition to extract symptom description fragments from the text of children's complaints; Based on the symptom description fragments in children's chief complaint text, we identify children's habitual non-standardized expressions and generate a candidate set of Class A semantic anchors by adopting dynamic matching rules based on children's developmental stages. Class A semantic anchors are located in three core symptom areas: oral mucosa, pharynx, and abdomen. Based on multimodal behavioral data, behavioral features in three dimensions—eye gaze pattern, vocal prosody features, and body movement amplitude—were extracted from the process of parents recounting stories. Based on three dimensions of behavioral characteristics, combined with the text content narrated by parents, a candidate set of semantic anchor points of type B is generated. The candidate sets of semantic anchors of type A and type B are jointly verified to form the final set of semantic anchors of type A and the final set of semantic anchors of type B.

[0008] Furthermore, a compact semantic space corresponding to the set of A-type semantic anchors is constructed using A-type semantic anchors; a wide-area semantic space corresponding to the set of B-type semantic anchors is constructed based on B-type semantic anchors. By calculating the intersection of the compact semantic space and the wide-area semantic space, a calibration semantic anchor set is generated, including: A compact semantic space is constructed based on the final set of A-type semantic anchor points. The compact semantic space is mapped to a first multidimensional sphere, where the center of the sphere is determined by the semantic anchor point clustering center of three symptom regions: oral mucosa, pharynx, and abdomen. The radius of the sphere is dynamically adjusted by the semantic distribution density of the child's developmental stage. A wide-area semantic space is constructed based on the final set of B-type semantic anchor points. The wide-area semantic space is mapped to a second multi-dimensional sphere, where the center of the sphere is determined by the weighted center of semantic anchor points in three dimensions: eye gaze pattern, speech prosody features, and body movement amplitude. The radius of the sphere is dynamically adjusted by the degree of semantic dispersion of the information narrated by the parents. Calculate the spatial overlap region between the first and second multidimensional spheres. By performing geometric calculations of the sphere intersection, determine the intersection space between the first and second multidimensional spheres. The boundary of the intersection space is determined by the relative position and radius of the first and second multidimensional spheres. Semantic anchors located in the intersection space are extracted to generate a set of calibrated semantic anchors, which simultaneously satisfy the dual semantic constraints of the compact semantic space and the wide-area semantic space.

[0009] Furthermore, the spatial overlap region between the first and second multidimensional spheres is calculated. By performing geometric calculations of the sphere intersection, the intersection space of the first and second multidimensional spheres is determined. The boundary of the intersection space is jointly determined by the relative positions and radii of the first and second multidimensional spheres, including: Calculate the spatial vector distance between the center of the first multidimensional sphere and the center of the second multidimensional sphere to obtain the relative positional relationship between the centers of the first and second multidimensional spheres; Based on the relative positional relationship between the center of the first multidimensional sphere and the center of the second multidimensional sphere, the spatial vector distance is compared with the sum of the radii of the first and second multidimensional spheres. When the spatial vector distance is less than the sum of the radii of the first and second multidimensional spheres, it is determined that there is a spatial overlap between the first and second multidimensional spheres, and a determination result is generated simultaneously. Based on the judgment results, the intersection region of the first and second multidimensional spheres is calculated by performing cross-volume operations based on the geometric features of the spheres. The cross-volume of the first and second multidimensional spheres in the semantic space is calculated by establishing the geometric parameter equations of the sphere intersection. Based on the boundary features of the cross-volume, a mathematical description of the intersection space is generated, providing geometric constraints for the selection of semantic anchor points.

[0010] Furthermore, the child's chief complaint text is semantically calibrated based on the set of calibrated semantic anchors to generate a semantically optimized chief complaint text; the semantically optimized chief complaint text is then preprocessed to generate an initial set of symptom entities, including: Based on the set of calibrated semantic anchors, the received original child complaint text is subjected to semantic reconstruction processing, including semantic fragment replacement and semantic weight adjustment, and the semantic reconstruction result is obtained. Based on the semantic reconstruction results, a semantically optimized chief complaint text is generated; the semantically optimized chief complaint text retains the core symptom descriptions corresponding to the calibrated semantic anchors, while removing semantic fragments that do not meet the geometric constraints. The semantically optimized chief complaint text is preprocessed by word segmentation and entity recognition to extract symptom description fragments from the semantically optimized chief complaint text. Based on semantically optimized symptom description fragments in the chief complaint text, and combined with the confidence weights of calibrated semantic anchors, an initial set of symptom entities is generated.

[0011] Furthermore, based on a pre-defined corpus specifically for children, dynamic semantic mapping is performed on the initial set of symptom entities to convert non-standardized expressions in children into standard medical terms, forming a standard set of symptom entities, including: Based on a set of children-specific corpus, a standard medical terminology semantic space is constructed. Each symptom entity in the initial symptom entity set is mapped to the semantic space to form a corresponding semantic coordinate point, and the mapping results are recorded synchronously. Calculate the projection position of each semantic coordinate point in the standard medical terminology semantic space, that is, determine the projection direction by calculating the angle between the semantic coordinate point and the basis vector of the medical terminology space; Based on the projection location, a nearest neighbor search is performed in the semantic space of standard medical terms to match the nearest standard medical term for each semantic coordinate point, and matching results are generated. Based on the matching results, the accuracy of the mapping results is verified by semantic similarity calculation. The semantic similarity is calculated by using the vector inner product method to generate a standard symptom entity set.

[0012] Furthermore, information filtering is performed on the standard symptom entity set to distinguish between core symptoms and accompanying phenomena, resulting in the core symptom entity set, which includes: Symptom weights are assigned to the standard symptom entity set, with core symptoms assigned higher weights and accompanying phenomena assigned lower weights, to obtain the symptom weight assignment results. Based on the symptom weight allocation results, the standard symptom entity set is filtered by setting a weight threshold to remove accompanying phenomena with weights below the threshold, and the filtering results are generated at the same time. Based on the screening results, target symptom entities with weights greater than or equal to the threshold are extracted to form a core symptom entity set.

[0013] Secondly, an automatic symptom extraction system for children based on natural language processing algorithms, wherein the system performs the following methods: The receiving module is used to receive text data of children's complaints containing mixed information of parents' narration, and at the same time collect multimodal behavioral data generated by parents during the narration process; The extraction module is used to identify and extract Class A and Class B semantic anchors based on children's chief complaint text data and multimodal behavioral data. Class A semantic anchors are derived from children's non-standardized expressions and are concentrated in the oral mucosa, pharynx and abdominal symptom areas. Class B semantic anchors are derived from information provided by parents and are concentrated in three dimensions: eye gaze patterns, speech rhythm features and body movement amplitude. The construction module is used to construct a compact semantic space corresponding to the set of A-type semantic anchors; construct a wide-area semantic space corresponding to the set of B-type semantic anchors based on the set of B-type semantic anchors; and generate a set of calibration semantic anchors by calculating the intersection of the compact semantic space and the wide-area semantic space. The processing module is used to perform semantic calibration processing on the child's chief complaint text based on the set of calibrated semantic anchors to generate semantically optimized chief complaint text; and to preprocess the semantically optimized chief complaint text to generate an initial set of symptom entities. The conversion module is used to perform dynamic semantic mapping on the initial symptom entity set based on a set corpus specifically for children, so as to convert children's non-standardized expressions into standard medical terms and form a standard symptom entity set. The execution module is used to filter information from the standard symptom entity set to distinguish between core symptoms and accompanying phenomena, thereby obtaining the core symptom entity set.

[0014] Thirdly, a computing device including a memory and a processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in the first aspect.

[0015] Fourthly, a computer-readable storage medium for storing a computer program for performing the method as described in the first aspect.

[0016] The above-described solution of the present invention has at least the following beneficial effects: By employing a technical solution that integrates children's subjective complaint text data with multimodal behavioral data from parents' accounts, this approach extracts type A semantic anchors corresponding to children's non-standardized expressions and type B semantic anchors corresponding to parents' behavioral characteristics. It constructs compact and wide-area semantic spaces, calculates their intersection to generate calibrated semantic anchors, and combines this with a children-specific corpus to achieve dynamic semantic mapping from non-standardized expressions to standard medical terms. Furthermore, it distinguishes between core symptoms and accompanying phenomena through weighted filtering. This overcomes the technical problems of existing technologies that rely solely on single text data, fail to adapt to children's non-standardized expression characteristics and parents' multimodal behavioral information, resulting in insufficient symptom extraction accuracy, large terminology mapping errors, and difficulty in distinguishing between core symptoms and accompanying phenomena. Consequently, it achieves accurate extraction, standardized conversion, and efficient screening of children's subjective symptoms, effectively improving the accuracy and practicality of symptom extraction and providing reliable data support for medical diagnostic decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for automatically extracting children's subjective symptoms based on natural language processing algorithms. Figure 2 A schematic diagram of an automatic symptom extraction system for children based on natural language processing algorithms; Figure 3 This is a schematic diagram of a computing device. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] Embodiments of the present invention propose an automatic extraction method for children's subjective symptoms based on natural language processing algorithms, the method comprising: Step 1: Receive text data of the child's account containing mixed information from the parent's narration, and simultaneously collect multimodal behavioral data generated by the parent during the narration process; Step 2: Based on the child's chief complaint text data and multimodal behavioral data, identify and extract type A semantic anchors and type B semantic anchors. Type A semantic anchors are derived from the child's non-standardized expressions and are concentrated in the oral mucosa, pharynx and abdominal symptom areas. Type B semantic anchors are derived from the information conveyed by the parents and are concentrated in three dimensions: eye gaze patterns, speech rhythm features and body movement amplitude. Step 3: Construct a compact semantic space corresponding to the set of A-type semantic anchors; construct a wide-area semantic space corresponding to the set of B-type semantic anchors based on the set of B-type semantic anchors; and generate a set of calibration semantic anchors by calculating the intersection of the compact semantic space and the wide-area semantic space. Step 4: Perform semantic calibration on the child's chief complaint text based on the set of calibrated semantic anchors to generate a semantically optimized chief complaint text; preprocess the semantically optimized chief complaint text to generate an initial set of symptom entities; Step 5: Based on the established corpus specifically for children, perform dynamic semantic mapping on the initial symptom entity set to convert children's non-standardized expressions into standard medical terms, forming a standard symptom entity set; Step 6: Filter the standard symptom entity set to distinguish between core symptoms and accompanying phenomena, thus obtaining the core symptom entity set.

[0020] In this embodiment of the invention, the textual data of children's complaints and the multimodal behavioral information during the parents' retelling process are integrated simultaneously. This fully explores the symptom-related semantics contained in children's non-standardized expressions and parents' behavioral characteristics, overcoming the limitations of single text analysis. By accurately capturing A-type semantic anchors focusing on core symptom areas and B-type semantic anchors reflecting the characteristics of parents' retelling, semantic calibration is completed by combining the intersection operation of the two semantic spaces, effectively avoiding semantic recognition bias caused by children's colloquial and fragmented expressions. Relying on a set of children-specific corpora, the accurate conversion of non-standardized expressions into standard medical terms is achieved. At the same time, by screening core symptoms and accompanying phenomena, a highly targeted set of core symptoms is output. This method effectively improves the accuracy and standardization of children's complaints, provides high-quality data support for clinical diagnosis and treatment decisions, helps to shorten the diagnosis cycle, reduce the probability of misdiagnosis, adapts to the actual needs of pediatric medical scenarios, and has high clinical application value.

[0021] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Receive the child's chief complaint text data input by the parent through the medical self-service terminal device. The child's chief complaint text data includes symptom descriptions and non-core information, specifically: The input interface of the medical self-service terminal device adopts a segmented design. The interface displays guiding questions in batches through pop-up windows, covering core symptom directions such as oral pain, throat sensation, and abdominal distension; Parents can choose to manually input via touchscreen or voice input. When inputting via voice, the real-time speech-to-text function is enabled to support the recognition of Mandarin with regional accents. Through multiple rounds of voice feature comparison, the recognition accuracy is ensured to be no less than 95%; The text input area is set with a character length limit of 10 to 500 characters, and the input content is monitored in real time, automatically filtering special symbols and illegal characters. At the same time, it is verified whether the text contains at least one symptom-related description. If not, a secondary guiding prompt will pop up until the received text data is complete and meets the symptom extraction requirements.

[0022] Step 1.2: Simultaneously initiate multimodal behavioral data collection. This involves collecting behavioral data in real time during the parent's narration, including eye gaze trajectory, speech waveform signals, and body movement videos. Specifically, this includes: after the parent triggers the narration input operation, the multimodal data collection operation is initiated simultaneously; the eye gaze trajectory is captured by the terminal's built-in infrared camera, which is pre-adjusted to a shooting angle covering the screen and surrounding area within a certain and appropriate range, and the coordinate changes, switching frequency, and single dwell time of the parent's gaze focus are recorded in real time; speech signals are collected through a high-sensitivity microphone, simultaneously capturing prosodic features such as speech speed, pitch, and pause intervals; and body movements are captured by a high-definition camera, focusing on capturing hand pointing movements, body posture adjustments, and limb swing amplitudes during symptom description. All collected raw data is transmitted to the terminal's local storage in real-time stream format to avoid data loss due to transmission delays.

[0023] Step 1.3 involves performing time-series alignment processing on the multimodal behavioral data. Timestamp synchronization is used to ensure temporal consistency across different modalities, resulting in processed multimodal behavioral data. Specifically, this includes: using the time of the medical self-service terminal as a unified benchmark, generating millisecond-accurate timestamps for each frame of the three data types: eye gaze trajectory, voice waveform signal, and body movement video; setting the trigger time of the parent's first character input as the zero point; comparing the timestamp sequences of the three data types and setting a ±10 millisecond time deviation threshold; adjusting the timeline of voice data exceeding the threshold using frame interpolation or deletion; calibrating body movement videos by adjusting frame intervals and playback order; and for eye gaze trajectory data, first locating valid coordinate data before and after the missing time point, assigning weights based on the time interval between the missing time point and the valid data points before and after it (shorter intervals, higher weights), and then calculating the coordinate information of the missing time point through weighted averages. Through targeted adjustments to the three data types, the starting point and duration of all multimodal data are ensured to be perfectly matched, achieving a high degree of consistency in the time dimension.

[0024] Step 1.4 involves associating and storing the text data with the processed multimodal behavioral data to form a complete dataset to be processed. This includes: storing data using a dual backup method of encrypted cloud storage and local caching; using the child's medical card number as a unique index to associate and bind the child's chief complaint text data with the time-aligned multimodal data; storing the text data in a structured manner, dividing it into fields such as symptom description, non-core information, and input method; storing eye gaze trajectory data as coordinate sequence files; saving speech waveform signals in WAV format; storing body movement videos in H.264 compression format, and retaining a backup of the original uncompressed version; synchronously recording metadata such as the child's age, parent's narration method, acquisition device number, and ambient temperature during storage, ultimately forming a complete dataset to be processed containing text information, multimodal behavioral information, and metadata; and supporting fast index retrieval and multi-dimensional data linkage to meet the efficient data retrieval needs of subsequent processes.

[0025] In this embodiment of the invention, the text data of the child's chief complaint input by the parent is conveniently received through a medical self-service terminal. Simultaneously, multimodal behavioral data such as eye gaze trajectory, voice waveform signal, and body movement video are collected during the parent's description process. Time-series alignment processing ensures the temporal consistency of different modal data. The text data and the processed multimodal behavioral data are then associated and stored to form a complete dataset to be processed. This not only enriches the collection dimensions of symptom-related information but also avoids association analysis errors caused by time deviations. It provides high-quality and highly consistent basic data support for the accurate extraction of A-type and B-type semantic anchors, the construction of dual semantic spaces, and semantic calibration processes. This effectively enhances the rigor and reliability of the overall extraction method and fully adapts to the actual data collection needs of pediatric medical scenarios.

[0026] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1 involves preprocessing the child's complaint text data. This is achieved through word segmentation and entity recognition to extract symptom description fragments from the child's complaint text. Specifically, this includes: performing a hierarchical preprocessing procedure on the received child's complaint text data; firstly, semantic filtering to remove meaningless interjections, repetitive expressions, and redundant content unrelated to the symptoms; secondly, adapting to the characteristics of children's colloquial expressions during word segmentation, prioritizing semantic integrity for special expressions such as reduplicated words and elliptical sentences, avoiding fragmented expressions related to core symptoms such as stomach pain and dry throat; and thirdly, combining contextual semantic association analysis in the entity recognition stage, constructing a symptom-related vocabulary dictionary to filter core expressions such as pain, itching, swelling, and abnormal secretions, while excluding non-symptom information such as descriptions of daily activities. Finally, symptom description fragments focusing on the child's physical discomfort are extracted, ensuring that the fragments are semantically complete and highly relevant to the symptom extraction target.

[0027] Step 2.2 identifies habitual non-standardized expressions in children based on symptom description fragments in their chief complaint text. A candidate set of Class A semantic anchors is generated using dynamic matching rules based on children's developmental stages. Class A semantic anchors are located in three core symptom regions: oral mucosa, pharynx, and abdomen. Specifically, this involves dividing the child's development into multiple developmental stages according to age gradients. For each stage, a dedicated non-standardized expression reference library is constructed based on clinically collected symptom description data. This library covers frequently used colloquial and fragmented symptom expressions for that stage. The extracted symptom description fragments are progressively compared with the reference libraries for each stage. The dynamic matching rules adjust the core matching dimensions according to the developmental stage: for younger children, emphasis is placed on morphological similarity, matching reduplicated words like "toothache" and "bloated stomach"; for school-aged children, emphasis is placed on semantic consistency, matching simple phrases like "something in the throat" and "cracked mouth." Combining the semantic orientation of the description fragments, successfully matched expressions are precisely located in the three core symptom regions: oral mucosa, pharynx, and abdomen. These are then categorized and summarized by region to form the candidate set of Class A semantic anchors.

[0028] Step 2.3: Based on multimodal behavioral data, extract behavioral features from the parents' description process in three dimensions: eye gaze patterns, speech prosody features, and body movement amplitude. Specifically, this includes: extracting dimensional features from the time-aligned multimodal behavioral data; for eye gaze patterns, first dividing the screen symptom prompt area, including corresponding sections for the mouth, throat, and abdomen, analyzing the focus distribution, single dwell time, and area switching frequency of each section during the parents' description, and extracting the features of target areas where the gaze duration exceeds a set proportion; for speech prosody features, by analyzing the frequency changes of the speech signal, the amplitude of tone fluctuations, and the pause intervals in sentences, capturing the emphasis features of slowed speech and increased tone when describing symptoms, and distinguishing the differences between everyday expressions and the speech of symptom emphasis descriptions; for body movement amplitude, by analyzing video frames to screen targeted actions related to symptoms, excluding unconscious body movements, extracting action features such as hands pointing to the corresponding area of ​​the body and body posture tilting towards the symptom area, and recording the duration and amplitude parameters of the actions.

[0029] Step 2.4: Based on the three dimensions of behavioral features and combined with the text content narrated by the parents, a candidate set of Class B semantic anchors is generated. Specifically, this includes: establishing a temporal correlation mapping table between behavioral features and text content, and mapping the extracted three types of behavioral features to the text fragments narrated by the parents one by one according to the timestamp; when a specific behavioral feature appears in a certain time period, the text description content in that time period is located simultaneously: if the parent is looking at the oral symptom area and the tone of voice rises, the oral-related description in the corresponding text is marked as a candidate anchor; if the parent points to the abdomen and the amplitude of the movement increases, a candidate anchor is generated by combining the description of abdominal discomfort in the text; the effectiveness of the anchors is strengthened through two-way verification: the behavioral features must be able to support the symptom orientation of the text description, and the text description must be able to explain the logic of the occurrence of the behavioral features. If the two match, they are included in the candidate set, and finally, a candidate set of Class B semantic anchors is formed.

[0030] Step 2.5 involves joint verification of the candidate sets of semantic anchors in categories A and B to form the final sets of semantic anchors in categories A and B. Specifically, this includes: performing a three-level joint verification process on the candidate sets of semantic anchors in categories A and B; the first level performs regional consistency verification, comparing the core symptom regions corresponding to anchors in the two sets. If a category A anchor points to the throat, it needs to be confirmed that there are throat-related behavioral features supporting anchors in category B. Candidate anchors with inconsistent regions are directly eliminated; the second level performs semantic rationality verification, combining the physiological logic of children's symptoms to eliminate contradictory anchors, such as a category A anchor pointing to oral pain while a category B anchor points to abdominal movements; the third level implements confidence assessment, assigning confidence scores based on the accuracy of anchor matching and the amount of supporting evidence, i.e., the amount of textual or behavioral features. Anchors with scores meeting the set standards are retained, ultimately forming the final sets of semantic anchors in categories A and B that combine accuracy and relevance.

[0031] In this embodiment of the invention, by preprocessing the child's complaint text data through word segmentation and entity recognition, symptom description fragments are accurately extracted and redundant information is filtered out. Non-standardized expressions are identified based on dynamic matching rules adapted to the child's developmental stage, generating a candidate set of Class A semantic anchor points focusing on the core symptom region. Simultaneously, three dimensions of behavioral features are extracted from the parents' multimodal behavioral data, and combined with the content of the substitute text to construct a candidate set of Class B semantic anchor points. Then, invalid candidates are eliminated through joint verification, and the anchor point relevance is strengthened. This solves the problems of difficulty in identifying non-standardized expressions in children and ambiguity in locating the core symptom region, while also compensating for the limitation of single text analysis lacking behavioral semantic supplementation. The resulting two sets of semantic anchor points are both targeted and comprehensive, providing accurate and reliable foundational support for the construction of a dual semantic space and semantic calibration, effectively improving the semantic recognition accuracy of the overall symptom extraction process.

[0032] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Construct a compact semantic space based on the final set of Class A semantic anchors. Map the compact semantic space to a first multidimensional sphere, where the center of the sphere is determined by the clustering centers of semantic anchors in the three symptom regions of oral mucosa, pharynx, and abdomen. The radius of the sphere is dynamically adjusted by the semantic distribution density during the child's developmental stage. Specifically, this includes: first, accurately classifying the final set of Class A semantic anchors according to the three core symptom regions of oral mucosa, pharynx, and abdomen; second, performing semantic feature similarity clustering analysis on the semantic anchors in each region. Based on the semantic vectors of the anchors, calculate the semantic similarity between any two anchors, aggregating anchors with close semantic distances into a cluster, while removing isolated abnormal anchors whose semantic distance exceeds the reasonable range of the average distance within the region, ensuring that the anchors in each region focus on the same... The semantic features of each symptom region are used as the semantic anchor point cluster center. Weights are then assigned based on the frequency of occurrence and diagnostic relevance of each symptom region in clinical practice, with regions exhibiting higher frequency and stronger diagnostic relevance receiving higher weights. A weighted average is used to calculate the center of the first multidimensional sphere, ensuring the center better aligns with core clinical needs. The semantic distribution density of Class A semantic anchor points for each child's developmental stage is calculated, and the semantic distance dispersion from all anchor points within the region to the cluster center is statistically analyzed. Smaller dispersion indicates a more concentrated semantic distribution, requiring a smaller sphere radius to focus on core semantics; larger dispersion indicates a more dispersed semantic distribution, requiring a larger sphere radius to cover the entire semantic range. Ultimately, this achieves a precise mapping from a compact semantic space to the first multidimensional sphere.

[0033] Step 3.2: Construct a wide-area semantic space based on the final set of B-type semantic anchors. Map this wide-area semantic space to a second multi-dimensional sphere, where the center of the sphere is determined by the weighted center of the semantic anchors in three dimensions: eye gaze pattern, speech prosody features, and body movement amplitude. The radius of the sphere is dynamically adjusted based on the semantic dispersion of the parent's narration information. Specifically, this includes: classifying and organizing the final set of B-type semantic anchors according to three dimensions: eye gaze pattern, speech prosody features, and body movement amplitude, ensuring that each anchor retains complete semantic features and timestamp association information; extracting semantic features from the semantic anchors in each dimension, summarizing them to form a set of semantic features for each dimension, and calculating the mean of this set as the semantic feature center value for that dimension; assigning weights according to the support strength of each dimension for symptom semantics. Eye gaze pattern can directly correspond to the symptom prompt area on the screen, quickly locking the symptom association direction, therefore... The second multidimensional sphere has the highest weight; phonological prosody features, reflecting the importance of symptoms through changes in intonation and speech rate, have the next highest weight; and body movement amplitude, as supplementary information, has the lowest weight. A weighted summation is used to obtain the center of the second multidimensional sphere, highlighting the semantic contribution of key behavioral features. The semantic dispersion of parental descriptions is analyzed. At the text level, the frequency of repetition of symptom descriptions and the coherence of sentence logic are statistically analyzed. At the behavioral feature level, the synchronization and consistency of phonological prosody and body movements when describing the same symptom are judged. High repetition frequency, logical coherence, and strong multimodal behavioral synchronicity indicate low semantic dispersion; conversely, low repetition frequency indicates high semantic dispersion. A smaller sphere radius is set when semantic dispersion is low, focusing on core semantic anchors; a larger sphere radius is used when semantic dispersion is high, covering more potentially relevant semantic anchors to avoid missing effective information and achieving a reasonable mapping from the broad semantic space to the second multidimensional sphere.

[0034] Step 3.3: Calculate the spatial overlap region between the first multidimensional sphere and the second multidimensional sphere. By performing geometric calculations of the sphere intersection, the intersection space of the first multidimensional sphere and the second multidimensional sphere is determined. The boundary of the intersection space is determined by the relative position and radius of the first multidimensional sphere and the second multidimensional sphere.

[0035] Step 3.4: Extract semantic anchor points located in the intersection space to generate a set of calibrated semantic anchor points. These calibrated semantic anchor points simultaneously satisfy both compact and wide-area semantic constraints. Specifically, this includes: first, based on the semantic coordinates and radius parameters of the centers of the first and second multidimensional spheres, defining the overlapping semantic regions of the two spheres in the semantic space, forming an intersection space that simultaneously carries two types of semantic constraints; then, extracting each anchor point from the A-type and B-type semantic anchor point sets one by one, converting the semantic features of the anchor point into a semantic vector, and calculating the similarity between this vector and the center vector of the first multidimensional sphere to obtain the semantic distance from the anchor point to the center of the first sphere; the higher the similarity, the closer the distance. The semantic distance from the anchor point to the center of the second multidimensional sphere is calculated using the same method. It is then determined whether the two semantic distances are less than the radii of the corresponding spheres. Only when the distance from the anchor point to the center of the first sphere is less than the radius of the first sphere, and the distance to the center of the second sphere is less than the radius of the second sphere, is the anchor point considered to be located in the intersection space. Such anchor points not only strictly match the semantic orientation of the core symptom areas of children, but also receive semantic support from the multimodal behavioral characteristics of parents, thus possessing dual semantic validity. All anchor points that meet the conditions are classified and organized according to the three core symptom areas of oral mucosa, pharynx, and abdomen, ensuring that the calibration anchor points in each area are concentrated and semantically consistent, ultimately generating a set of calibration semantic anchor points that combines accuracy and relevance.

[0036] In this embodiment of the invention, a compact semantic space is constructed based on a set of A-type semantic anchors and mapped to a first multidimensional sphere. The center of the sphere is determined by the clustering center of anchors in the core symptom region, and the radius of the sphere is dynamically adjusted according to the semantic distribution density of the child's developmental stage. Simultaneously, a wide-area semantic space is constructed based on a set of B-type semantic anchors and mapped to a second multidimensional sphere. The center of the sphere is determined by the weighted center of anchors in three types of behavioral features, and the radius of the sphere is dynamically adjusted according to the degree of semantic dispersion of the parents' statements. Then, the intersection space is determined by calculating the overlapping area of ​​the two sphere spaces through a sphere intersection geometry algorithm. Semantic anchors that satisfy dual semantic constraints are extracted to generate a set of calibrated semantic anchors. This effectively integrates the core semantics of children's non-standardized expressions with the supplementary semantics of parents' multimodal behaviors, solving the problem that a single semantic space cannot simultaneously take into account both relevance and comprehensiveness, and that semantic deviations cannot be accurately calibrated. The calibrated semantic anchors focus on the core symptom region of the child and are supported by the semantics of the parents' behavioral information. This provides a high-precision and high-reliability anchor foundation for the semantic calibration of children's complaints and the extraction of symptom entities, further improving the accuracy and rigor of the overall symptom extraction process.

[0037] In a preferred embodiment of the present invention, step 3 above may include: Step 3.31: Calculate the spatial vector distance between the center of the first multidimensional sphere and the center of the second multidimensional sphere to obtain their relative positional relationship. Specifically, this includes: first, clarifying the n-dimensional semantic space in which the first and second multidimensional spheres reside, where n corresponds to the total dimension of semantic features, such as the symptom region semantic dimension of anchor points A and the behavioral feature semantic dimension of anchor points B; let the center of the first multidimensional sphere be O1, and its semantic coordinate vector be O1=(x... 11 x 12 , ..., x 1n Each component x 1i The quantized value corresponding to a certain semantic feature, such as x 11 Corresponding semantics of oral mucosal symptoms, x 12 Corresponding semantics of throat symptoms; the center of the second multidimensional sphere is O2, and its semantic coordinate vector is O2=(x 21 x 22 , ..., x 2n Each component x 2i The semantic quantization value corresponding to a certain behavioral feature, such as x 21 Corresponding eye gaze pattern semantics, x 22 Corresponding to the semantic features of speech prosody; when calculating the spatial vector distance d between the centers of two spheres, the Euclidean distance formula is used, and the coordinate difference (x) is calculated dimension by dimension. 1i -x 2i Then, square each difference and sum them up to get... Finally, take the square root of the sum, i.e., d = This calculation not only yields the distance value, but also clarifies the offset direction of O1 relative to O2 in each semantic dimension based on the positive and negative distribution of the coordinate differences in each dimension, thus fully obtaining the relative positional relationship between the two sphere centers.

[0038] Step 3.32: According to the relative position relationship between the center of the first multi-dimensional sphere and the center of the second multi-dimensional sphere, compare the magnitude relationship between the spatial vector distance and the sum of the radii of the first multi-dimensional sphere and the second multi-dimensional sphere. When the spatial vector distance is less than the sum of the radii of the first multi-dimensional sphere and the second multi-dimensional sphere, it is determined that there is a spatial overlapping region between the first multi-dimensional sphere and the second multi-dimensional sphere, and a determination result is generated synchronously, which specifically includes: first extract the dynamic radius r1 of the first multi-dimensional sphere (determined by the semantic distribution density of the child's development stage) and the dynamic radius r2 of the second multi-dimensional sphere (determined by the semantic dispersion degree of the parent's reported information), and calculate the numerical sum S = r1 + r2 of the two; compare the obtained spatial vector distance d with S numerically, and follow the geometric determination logic during the comparison: if d < S, it means that there is a partial overlapping region between the two multi-dimensional spheres in the semantic space, and a determination result including the specific value of d, the specific value of S, the overlapping determination conclusion, and the relative position offset information of the two sphere centers is generated synchronously; if d = S, it means that the two spheres are tangent in the semantic space and there is only one common point, and it is determined that there is no effective overlapping region, and a tangent determination result is generated; if d > S, it means that the two spheres have no intersection in the semantic space, and a non-overlapping determination result is generated. The determination result needs to clearly mark the core calculation parameters, providing a direct basis for the calculation of the intersection region later.

[0039] Step 3.33: Based on the determination result, calculate the intersection region of the first multi-dimensional sphere and the second multi-dimensional sphere by performing a cross-volume operation based on the geometric characteristics of the sphere, and calculate the cross-volume of the first multi-dimensional sphere and the second multi-dimensional sphere in the semantic space by establishing a geometric parameter equation for the intersection of the spheres; based on the boundary characteristics of the cross-volume, generate a mathematical description of the intersection space, providing geometric constraint conditions for the semantic anchor point screening, which specifically includes: based on the overlapping confirmation information in the determination result, calculate the intersection region using the cross-volume algorithm of the geometric characteristics of the sphere; first establish the geometric parameter equations of the two multi-dimensional spheres. Let the coordinate vector of any point P in the semantic space be P=(x1, x2,..., x n ), then the equation of the first multi-dimensional sphere is: = (i from 1 to n), this equation represents the set of all points whose distance to O1 is equal to r1; the equation of the second multi-dimensional sphere is = (i from 1 to n), representing the set of all points whose distance to O2 is equal to r2; to determine the intersection boundary of the two spheres, solve the two equations simultaneously. First expand the two equations respectively: -2 x 1i + = , -2 x 2i + = Subtract the two equations to eliminate the difference. The term yields -2. x 1i + +2 x 2i - = - After simplification, the equations of the intersecting planes are obtained: 2 (x) 2i -x 1i )= - - + This plane is the boundary plane of the intersection region of the two spheres, and all intersection points lie on this plane. Based on the equations of the two spheres and the intersection plane, the mathematical description of the intersection space can be clearly defined: any point P in the space must simultaneously satisfy... ≤ , ≤ And satisfy 2 (x 2i -x 1i )= - - + This description transforms the intersection space into explicit geometric constraints. When selecting semantic anchors, it is only necessary to verify whether the semantic coordinates of the anchors simultaneously satisfy the above three expressions to accurately determine whether they are located within the intersection space, thus providing a rigorous quantitative basis for anchor selection.

[0040] In this embodiment of the invention, the spatial vector distance between the centers of the first and second multidimensional spheres is first calculated to clarify the relative positional relationship between the two centers. Then, the existence of a spatial overlap region is determined by comparing the spatial vector distance with the sum of the radii of the two spheres, and a determination result is generated. Finally, based on the determination result, a cross-volume algorithm based on the geometric features of the spheres is used to establish the geometric parameter equations of the intersection of the spheres to calculate the cross-volume. Combined with the boundary features of the cross-volume, a mathematical description of the intersection space is generated. This overcomes the problem of insufficient accuracy in semantic anchor point selection due to the fuzzy definition of the overlapping region of the dual semantic space and the lack of clear geometric constraints. It provides a rigorous and quantifiable geometric basis for the accurate extraction of semantic anchor points in the intersection space, ensuring that the selected calibration semantic anchor points simultaneously meet the dual constraints of the compact semantic space and the wide-area semantic space. This further enhances the scientificity and reliability of the semantic calibration process and provides solid spatial analysis support for the accurate extraction of children's chief complaints.

[0041] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Based on the calibrated semantic anchor set, perform semantic reconstruction processing on the received original child complaint text, including semantic fragment replacement and semantic weight adjustment, and obtain the semantic reconstruction results. Specifically, this includes: first, extracting the core semantic features of each anchor in the calibrated semantic anchor set; then, constructing a multi-dimensional semantic matching template based on clinical symptom description standards. The template covers key semantic dimensions such as symptom location, discomfort type, and manifestation characteristics; and performing semantic association analysis on the template and the original child complaint text sentence by sentence and word by word. By calculating the fit between text fragments and anchor semantic features, non-standardized symptoms such as abdominal pain and dry throat are located. The text describes a process for reconstructing a child's complaints. It involves replacing colloquial and fragmented phrases with logically complete, standardized expressions, such as replacing "stomach ache" with "abdominal pain" and "dry throat" with "throat dryness and discomfort." Simultaneously, it extracts the confidence weight of each calibration anchor point and assigns semantic priorities based on these weights. Anchor points with higher weights occupy a central position in the text, are prioritized for complete expression, and are presented earlier. Lower-weighted phrases are presented as supplementary information. This collaborative process of semantic replacement and priority adjustment results in a semantically coherent and clearly defined semantic reconstruction.

[0042] Step 4.2: Based on the semantic reconstruction results, generate a semantically optimized chief complaint text. This text retains the core symptom descriptions corresponding to the calibrated semantic anchors while removing semantic fragments that do not meet geometric constraints. Specifically, this includes: based on the semantic reconstruction results, firstly selecting core symptom descriptions directly associated with the calibrated semantic anchors, clarifying the symptom region and semantic orientation corresponding to each description; referring to the geometric constraints determined in Step 3, converting each semantic fragment into corresponding semantic coordinates, and verifying whether each coordinate simultaneously satisfies the conditions that the distance to the center of the first multidimensional sphere is less than the first radius and the distance to the center of the second multidimensional sphere is less than the second radius; retaining the core symptom descriptions that pass the verification, and removing three types of content that do not meet the constraints: firstly, descriptions of daily activities unrelated to the core symptoms; secondly, semantically ambiguous statements without anchor support; and thirdly, repetitive and redundant similar descriptions; logically organizing the retained core symptom descriptions, sorting them according to the chronological order of symptom occurrence or the priority of confidence weight, ensuring clear text expression, and finally generating a semantically optimized chief complaint text that focuses on the core symptoms and is semantically accurate.

[0043] Step 4.3 involves preprocessing the semantically optimized complaint text. This is achieved through word segmentation and entity recognition to extract symptom description fragments. Specifically, this includes: performing a hierarchical preprocessing procedure on the semantically optimized complaint text; firstly, semantic filtering to remove meaningless interjections like "ah" and "oh," repetitive expressions, and redundant formatting to ensure concise and organized text; secondly, word segmentation using a semantic unit-based splitting method, with complete symptom descriptions as the smallest unit, referencing the semantic boundaries of calibrated semantic anchors to avoid fragmenting core symptom phrases such as dry throat and paroxysmal abdominal pain; and thirdly, entity recognition combined with contextual semantic association analysis, calling a symptom-specific vocabulary dictionary, and associating the core semantic features of calibrated semantic anchors. Through keyword matching and semantic fit judgment, it prioritizes core symptom-related expressions such as pain, itching, swelling, and abnormal secretions, excluding non-symptom information such as diet and sleep, accurately extracting symptom description fragments focusing on the child's core discomfort, ensuring the fragments are semantically complete and highly consistent with the symptom extraction target.

[0044] Step 4.4: Based on semantic optimization of the symptom description fragments in the chief complaint text, and combined with the confidence weights of calibrated semantic anchors, an initial symptom entity set is generated. Specifically, this includes: performing secondary semantic matching between the extracted symptom description fragments and the calibrated semantic anchor set; accurately assigning corresponding anchor confidence weights to each symptom description fragment by comparing core semantic features with timestamp association information; sorting the symptom description fragments from high to low confidence weights, setting a weight threshold, removing secondary information below the threshold, and prioritizing the retention of core symptom fragments with higher weights; calculating the semantic similarity of the sorted fragments; if the semantic overlap of two fragments reaches the set standard, they are determined to describe the same symptom, and the version with the most complete description and highest weight is retained, completing the merging and deduplication; according to the semantic orientation of the fragments, they are classified into three core symptom regions: oral mucosa, pharynx, and abdomen, and sorted and organized within each region according to weight, finally generating an initial symptom entity set containing the location of symptom occurrence, specific manifestation, semantic priority, and confidence weight, ensuring that the symptom information in the set is accurate and without redundancy.

[0045] In this embodiment of the invention, the original child's chief complaint text is reconstructed by replacing semantic fragments and adjusting semantic weights through a set of calibrated semantic anchors. This transforms the child's non-standardized expression into standardized semantics that better fits symptom extraction. Based on the semantic reconstruction results, a semantically optimized chief complaint text is generated, retaining the core symptom description and removing invalid fragments that do not conform to geometric constraints. Subsequently, symptom description fragments in the optimized text are accurately extracted through word segmentation and entity recognition preprocessing. Finally, an initial symptom entity set is generated by combining the confidence weights of the calibrated semantic anchors. This effectively corrects the semantic bias caused by the child's colloquial and fragmented expression, filters out redundant and invalid information, strengthens the semantic orientation of the core symptoms, and makes the generated initial symptom entity set both accurate and relevant. This provides high-quality basic data for the conversion of non-standardized expressions into standard medical terms, further improving the standardization and reliability of child's chief complaint symptom extraction.

[0046] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Based on the established pediatric-specific corpus, construct a standard medical terminology semantic space. Map each symptom entity in the initial symptom entity set to the semantic space, forming corresponding semantic coordinate points, and simultaneously record the mapping results. Specifically, this includes: first, performing a comprehensive semantic analysis on the established pediatric-specific corpus, which contains standard medical terms for common childhood symptoms and corresponding colloquial, fragmented, and non-standardized expressions; extracting the core semantic features of each standard medical term in the corpus, clarifying five fixed semantic dimensions: symptom location, nature, severity, frequency of onset, and duration; and setting quantification rules for each dimension, assigning specific values ​​to symptom locations based on core areas such as the oral mucosa, pharynx, and abdomen, and classifying symptom nature based on pain, itching, etc. The symptoms, such as itching and swelling, are quantified, and their severity is graded as mild, moderate, and severe. The frequency and duration of attacks are converted into numerical values ​​according to commonly used clinical standards. The quantification results of the five dimensions are integrated into a five-dimensional semantic vector. The semantic vectors of all standard medical terms are used to construct a multi-dimensional standard medical terminology semantic space. For each symptom entity in the initial symptom entity set, semantic features are extracted according to the same dimension and quantified to obtain a semantic vector consistent with the dimensions of the standard terms. Based on the specific values ​​of this vector in the five dimensions, the corresponding unique position is located in the standard medical terminology semantic space to form a semantic coordinate point. The original description, quantization vector, semantic coordinate point, and values ​​of each dimension of each symptom entity are recorded simultaneously to ensure that the mapping process is traceable.

[0047] Step 5.2: Calculate the projection position of each semantic coordinate point in the standard medical terminology semantic space. This involves determining the projection direction by calculating the angle between the semantic coordinate point and the basis vectors of the medical terminology space. Specifically, this includes: clarifying the basis vector composition of the standard medical terminology semantic space, where each of the five core semantic dimensions corresponds to a unit basis vector. Each basis vector has a value of 1 for only the corresponding dimension and 0 for the other dimensions, pointing to the semantic directions of symptom location, nature, degree, frequency of onset, and duration, respectively; converting each generated semantic coordinate point into a five-dimensional semantic vector V; and calculating the cosine of the angle between vector V and each basis vector Bi. The calculation process is as follows: divide the dot product of vector V and Bi by the product of the magnitudes of vector V (|V|) and Bi (|Bi|), i.e., cosθi = (V·Bi) / (|V|×|Bi|), where cosθi represents the cosine of the angle θi; θi specifically refers to the five-dimensional semantic space. The angle between the semantic vector V and the i-th unit basis vector Bi; i is the index used to identify the core semantic dimension / corresponding unit basis vector, corresponding to the five core semantic dimensions mentioned above, namely symptom location, nature, degree, frequency of onset, and duration, with values ​​of 1, 2, 3, 4, and 5; since Bi is a unit vector with a magnitude of 1, it is simplified to cosθi = V·Bi; the direction fit between vector V and each basis vector is determined based on the magnitude of the five cosine values, with a higher fit as the cosine value is closer to 1; the direction corresponding to the basis vector with the largest cosine value is selected as the primary projection direction, and the remaining basis vectors are used as auxiliary projection directions; the projection component of vector V on each basis vector is calculated, i.e., projection value = |V| × cosθi; the projection components are weighted and summed according to the basis vector weights, with the weights being proportional to the cosine values, and finally the specific projection position of the semantic coordinate point in the standard medical terminology semantic space is determined.

[0048] Step 5.3: Based on the projection position, perform a nearest neighbor search in the standard medical terminology semantic space to match the nearest standard medical term for each semantic coordinate point and generate matching results. Specifically, this includes: using the determined projection position as the center, setting the search neighborhood radius according to the term density in the standard medical terminology semantic space; reducing the radius when terms are densely distributed and expanding it when they are sparsely distributed; traversing all semantic coordinate points corresponding to standard medical terms in the semantic space; for each standard term coordinate point Sj, calculating its Euclidean distance to the current projection position P; the calculation process involves calculating the coordinate difference between P and Sj dimension by dimension, squaring each difference, summing the results, and taking the square root of the sum to obtain the distance dj = Here, p1 to p5 are the five-dimensional coordinates of P, and sj1 to sj5 are the five-dimensional coordinates of Sj. The five-dimensional coordinates of p1 to p5 and sj1 to sj5 completely correspond to the five core semantic dimensions of symptom location, nature, severity, frequency of onset, and duration mentioned earlier. This is the core premise for dimensional uniformity and comparability in the entire semantic space calculation. Record the distance values ​​between all standard terms and P, and filter out the standard medical terms with the smallest dj. If there are multiple terms with the same smallest dj distance, prioritize the term with higher clinical application frequency as the matching object. Generate matching results containing the original symptom entity, matching standard term, distance value, and term application frequency to ensure that each symptom entity corresponds to a unique matching term.

[0049] Step 5.4: Based on the matching results, verify the accuracy of the mapping results through semantic similarity calculation. Calculate semantic similarity using the vector inner product method to generate a standard symptom entity set. Specifically, this includes: extracting the five-dimensional semantic vector Vt of each symptom entity and the five-dimensional semantic vector Vs of the corresponding matching standard term from the generated matching results; calculating the semantic similarity between the two using the vector inner product method. The calculation process is as follows: multiply the values ​​of the corresponding dimensions of Vt and Vs respectively to obtain the product of the five dimensions, and then sum the five product results to obtain the inner product value Sim = vt1×vs1 + vt2×vs2 + vt3×vs3 + vt4×vs4 + vt5×vs5; The larger the Sim value, the higher the semantic fit between the two vectors. Based on historical matching data from a children's corpus, a semantic similarity threshold is set, which is the average inner product of historical valid matching cases. If the current Sim value is higher than the threshold, the mapping result is verified to be accurate, and the matching standard term is retained. If the Sim value is lower than the threshold, the semantic quantization process and projection position calculation results of the symptom entity are re-examined to eliminate quantization errors or projection biases, and the matching process is re-executed. All verified standard medical terms are categorized and organized according to symptom location, duplicate matching items are removed, and the term with the highest Sim value is retained for multiple valid matching results of the same symptom entity. Finally, a standardized, accurate, and effective set of standard symptom entities is generated.

[0050] In this embodiment of the invention, a standard medical terminology semantic space is constructed to specifically adapt to the characteristics of children's symptom expressions. Each symptom entity in the initial symptom entity set is mapped to a semantic coordinate point, and the mapping results are recorded. This solves the problem of insufficient mapping adaptability caused by the lack of children-specificity in existing corpora. Then, a spatial projection algorithm is used to determine the projection direction by calculating the angle between the semantic coordinate point and the basis vector of the medical terminology space, clarifying the relative position of the symptom entity in the standard semantic space and providing a directional basis for accurate matching. Subsequently, a nearest neighbor search is performed based on the projection position to match the nearest standard medical term for each semantic coordinate point, effectively narrowing the mapping gap between children's non-standardized expressions and standard medical terms. Finally, the semantic similarity is calculated using the vector inner product method to verify the mapping accuracy, ensuring the reliability of the matching results. Ultimately, a standard symptom entity set is generated. This not only overcomes the defect of large single-text mapping error in the prior art, but also realizes the accurate conversion of children's subjective symptoms from non-standardized expressions to standard medical terms, improving the standardization and accuracy of symptom extraction.

[0051] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1 assigns symptom weights to the standard symptom entity set, assigning higher weights to core symptoms and lower weights to accompanying phenomena to obtain the symptom weight allocation results. Specifically, this involves: first, constructing a multi-dimensional symptom weight allocation rule system based on pediatric clinical data, combining disease diagnosis logic and symptom association principles to clarify the judgment logic for core symptoms and accompanying phenomena; the determination of core symptoms must meet three conditions: first, a direct causal relationship with the disease; second, specific manifestations at the lesion site; and third, inclusion in mandatory clinical diagnostic indicators—all three are indispensable; the judgment criteria for accompanying phenomena are to assist in describing the condition, lack specificity, and only have an indirect relationship with core symptoms, and cannot be used as a sole diagnostic basis; the weight allocation principle is based on the confidence weight of calibrated semantic anchors, integrating the strong clinical association of symptoms. A comprehensive assessment is conducted. The clinical relevance strength is assessed by statistically analyzing the frequency of symptom occurrence and its diagnostic contribution in corresponding disease cases. The higher the frequency of occurrence and the more crucial its supporting role in the diagnostic conclusion, the higher its weighting percentage. During the comprehensive assessment, the confidence weight and clinical relevance strength are first integrated according to the diagnostic logic, and then the weighting range is determined through hierarchical assessment. For each symptom entity in the standard symptom entity set, its type attribute is first clarified according to the judgment criteria. Core symptoms are assigned a higher value range after comprehensive assessment, and accompanying phenomena are assigned a lower value range. During the assignment process, the weight difference of symptoms of the same type must be consistent with the clinical importance. For example, in the same core symptom, the weight of symptoms with high lesion site specificity is higher than that of symptoms with low specificity. Finally, a complete allocation result is formed that includes each symptom entity, type attribute, weight value, and assignment basis.

[0052] Step 6.2: Based on the symptom weight allocation results, the standard symptom entity set is screened by setting a weight threshold to remove accompanying phenomena with weights below the threshold, and the screening results are generated. Specifically, this includes: based on the weight allocation results, using statistical analysis methods to analyze the weight distribution characteristics of all symptom entities, clarifying the concentration range, dispersion range, and peak position of the weight values; the threshold setting principle is guided by clinical diagnostic needs, combined with the weight distribution pattern and disease diagnosis logic; firstly, referring to the lowest weight range of core symptoms in the diagnosis of common childhood diseases, the quartiles of the weight distribution are extracted, and the values ​​corresponding to the upper quartiles are used as the initial threshold; subsequently, clinical experts from pediatrics, otolaryngology, gastroenterology, and other relevant core symptom areas are invited to conduct a multi-dimensional rationality assessment of the initial threshold; the expert evaluation scope... Based on the diagnostic focus of different disease types, the initial threshold is analyzed to determine if there is a risk of missing key core symptoms or retaining ineffective accompanying phenomena. The threshold value is adjusted according to the assessment opinions. After the threshold is determined, the screening process is initiated to perform weight verification on each symptom entity in the standard symptom entity set, comparing the symptom weight with the set threshold one by one. At the same time, symptom correlation verification is added to determine whether low-weight symptoms are strongly correlated with core symptoms and have auxiliary value in interpreting the condition. If they are only unrelated isolated accompanying phenomena, they are removed even if they are close to the threshold. Symptom entities with a weight greater than or equal to the threshold are directly retained. During the screening process, the judgment result, weight value, type attribute and correlation analysis conclusion of each symptom are recorded simultaneously, and finally a complete screening result including a retention list, a removal list, threshold explanation and verification basis is generated.

[0053] Step 6.3: Based on the screening results, extract target symptom entities with weights greater than or equal to the threshold to form a core symptom entity set. This includes: systematically reviewing the generated screening results, prioritizing the extraction of all symptom entities with weights greater than or equal to the threshold to form a core symptom candidate set; accurately classifying symptoms according to their location, placing them into three core regions: oral mucosa, pharynx, and abdomen; verifying the lesion site description and clinical diagnostic criteria for each symptom entity during classification to ensure accurate classification; and conducting secondary verification of candidate symptoms within each region, in accordance with pediatric disease diagnosis and treatment guidelines, checking whether the core attributes of each symptom conform to the guidelines. The clinical diagnostic logic prioritizes excluding accompanying symptoms that were included due to misjudgment of weights. After successful verification, symptoms are sorted from highest to lowest weight. In addition to weights, the sorting logic includes the chronological order of symptom appearance and the urgency of diagnosis, prioritizing core symptoms that appear first or those with urgent guiding significance for treatment intervention. Each symptom entity is clearly labeled with its standard medical terminology, weight value, clinically related diseases, time of appearance, and diagnostic priority. The labeling information is based on pediatric clinical case data and treatment guidelines. Finally, all labeled symptom entities are integrated to form a well-structured, focused set of core symptom entities that conforms to clinical application scenarios.

[0054] In this embodiment of the invention, by performing symptom weight allocation on the standard symptom entity set, the high priority of core symptoms and the low priority of accompanying phenomena are clearly defined, making the importance hierarchy of different types of symptoms clearly distinguishable. Then, based on the weight allocation results, a reasonable threshold is set to filter the standard symptom entity set, accurately eliminating interference information such as accompanying phenomena with weights below the threshold. Finally, based on the filtering results, target symptom entities with weights greater than or equal to the threshold are extracted to form the core symptom entity set. This effectively solves the problem in the prior art that it is difficult to effectively distinguish between the core symptoms and accompanying phenomena reported by children, significantly reduces the interference of irrelevant information on diagnosis and treatment decisions, makes the extracted symptom information more focused on the key points of disease diagnosis, improves the pertinence and accuracy of symptom extraction, provides logically clear and key-point-focused core symptom data support for clinical disease diagnosis, and further ensures the accuracy and efficiency of diagnosis and treatment decisions.

[0055] An automatic symptom extraction system for children based on natural language processing algorithms includes: The receiving module is used to receive text data of children's complaints containing mixed information of parents' narration, and at the same time collect multimodal behavioral data generated by parents during the narration process; The extraction module is used to identify and extract Class A and Class B semantic anchors based on children's chief complaint text data and multimodal behavioral data. Class A semantic anchors are derived from children's non-standardized expressions and are concentrated in the oral mucosa, pharynx and abdominal symptom areas. Class B semantic anchors are derived from information provided by parents and are concentrated in three dimensions: eye gaze patterns, speech rhythm features and body movement amplitude. The construction module is used to construct a compact semantic space corresponding to the set of A-type semantic anchors; construct a wide-area semantic space corresponding to the set of B-type semantic anchors based on the set of B-type semantic anchors; and generate a set of calibration semantic anchors by calculating the intersection of the compact semantic space and the wide-area semantic space. The processing module is used to perform semantic calibration processing on the child's chief complaint text based on the set of calibrated semantic anchors to generate semantically optimized chief complaint text; and to preprocess the semantically optimized chief complaint text to generate an initial set of symptom entities. The conversion module is used to perform dynamic semantic mapping on the initial symptom entity set based on a set corpus specifically for children, so as to convert children's non-standardized expressions into standard medical terms and form a standard symptom entity set. The execution module is used to filter information from the standard symptom entity set to distinguish between core symptoms and accompanying phenomena, thereby obtaining the core symptom entity set.

[0056] The automatic extraction system according to embodiments of the present invention can correspond to performing the method described in the embodiments of the present invention, and the above and other operations and / or functions of each module of the automatic extraction system are respectively for implementing Figure 1The corresponding process of the method in the illustrated embodiment will not be described in detail here for the sake of brevity.

[0057] This application also provides a computing device. This computing device can utilize a server.

[0058] like Figure 2 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 700 includes a bus 701, a processor 702, a communication interface 703, and a memory 704. The processor 702, the memory 704, and the communication interface 703 communicate with each other via the bus 701.

[0059] The 701 bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 2 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0060] The processor 702 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0061] Communication interface 703 is used for external communication. Memory 704 may include volatile memory, such as random access memory (RAM). Memory 704 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD). Executable code is stored in memory 704, and processor 702 executes this executable code to perform the aforementioned method for automatically extracting children's subjective symptoms based on natural language processing algorithms.

[0062] Specifically, in implementing the embodiments of the automatic extraction system for children's subjective symptoms based on natural language processing algorithms described above, and where each module or unit of the automatic extraction system for children's subjective symptoms based on natural language processing algorithms described above is implemented in software, the software or program code required to execute the functions of each module / unit in the automatic extraction system for children's subjective symptoms based on natural language processing algorithms described above can be partially or entirely stored in the memory 704. The processor 702 executes the program code corresponding to each unit stored in the memory 704 to execute the aforementioned automatic extraction method for children's subjective symptoms based on natural language processing algorithms.

[0063] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct a computing device to execute the above-described method for automatically extracting children's subjective symptoms based on natural language processing algorithms.

[0064] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0065] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0066] When the computer program product is executed by a computer, the computer performs any of the aforementioned methods of the automatic extraction method for children's subjective symptoms based on natural language processing algorithms. The computer program product can be a software installation package; when any of the aforementioned methods of the automatic extraction method for children's subjective symptoms based on natural language processing algorithms is required, the computer program product can be downloaded and executed on the computer.

[0067] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatic extraction of chief complaints symptoms in children based on natural language processing algorithms, characterized by, The method comprises: receiving child complaint text data containing parent paraphrase mixed information, while collecting multi-modal behavior data generated by parents during the paraphrasing process; According to the child complaint text data and the multi-modal behavior data, A-type semantic anchors and B-type semantic anchors are identified and extracted; A-type semantic anchors are derived from children's non-standardized expressions, and are concentrated in the oral mucosa, throat and abdominal symptom areas; B-type semantic anchors are derived from parent paraphrase information, and are concentrated in the three dimensions of eye gaze patterns, speech prosody characteristics and body movement amplitude; A compact semantic space corresponding to the A-type anchor set is constructed through the A-type semantic anchors; a wide-area semantic space corresponding to the B-type anchor set is constructed according to the B-type semantic anchors, and a calibration semantic anchor set is generated by calculating the intersection of the compact semantic space and the wide-area semantic space; According to the calibration semantic anchor set, the child complaint text is subjected to semantic calibration processing to generate a semantic optimized complaint text; the semantic optimized complaint text is preprocessed to generate an initial symptom entity set; Based on the set child exclusive corpus, the initial symptom entity set is subjected to dynamic semantic mapping processing to convert children's non-standardized expressions into standard medical terminology, forming a standard symptom entity set; The standard symptom entity set is subjected to information screening to distinguish core symptoms from accompanying phenomena, obtaining a core symptom entity set. 2.The method of claim 1, wherein the natural language processing algorithm is based on a rule-based algorithm. According to the child complaint text data and the multi-modal behavior data, A-type semantic anchors and B-type semantic anchors are identified and extracted; A-type semantic anchors are derived from children's non-standardized expressions, and are concentrated in the oral mucosa, throat and abdominal symptom areas; B-type semantic anchors are derived from parent paraphrase information, and are concentrated in the three dimensions of eye gaze patterns, speech prosody characteristics and body movement amplitude, including: The child complaint text data is preprocessed through word segmentation processing and entity recognition operations to extract symptom description fragments in the child complaint text; Based on the symptom description fragments in the child complaint text, children's habitual non-standardized expressions are identified, and an A-type semantic anchor candidate set is generated by using dynamic matching rules based on the child development stage; the A-type semantic anchors are located in the three core symptom areas of oral mucosa, throat and abdomen; Based on the multi-modal behavior data, behavior characteristics in the three dimensions of eye gaze patterns, speech prosody characteristics and body movement amplitude are extracted from the parent paraphrasing process; Based on the behavior characteristics in the three dimensions, combined with the text content of the parent paraphrase, a B-type semantic anchor candidate set is generated; The A-type semantic anchor candidate set and the B-type semantic anchor candidate set are jointly verified to form the final A-type semantic anchor set and the final B-type semantic anchor set. 3.The method of claim 2, wherein the natural language processing algorithm is based on a rule-based algorithm. A compact semantic space corresponding to the A-type anchor set is constructed through the A-type semantic anchors; a wide-area semantic space corresponding to the B-type anchor set is constructed according to the B-type semantic anchors, and a calibration semantic anchor set is generated by calculating the intersection of the compact semantic space and the wide-area semantic space, including: constructing a compact semantic space based on the final set of class A semantic anchors, mapping the compact semantic space into a first multi-dimensional sphere, wherein a sphere center of the first multi-dimensional sphere is determined by semantic anchor cluster centers of the three symptom areas of oral mucosa, throat, and abdomen, and a sphere radius of the first multi-dimensional sphere is dynamically adjusted by semantic distribution density of the child development stage; constructing a wide-area semantic space based on the final set of class B semantic anchors, mapping the wide-area semantic space into a second multi-dimensional sphere, wherein a sphere center of the second multi-dimensional sphere is determined by weighted centers of semantic anchors of the three dimensions of eye gaze patterns, speech prosody features, and body movement amplitudes, and a sphere radius of the second multi-dimensional sphere is dynamically adjusted by semantic dispersion degree of the parent proxy information; calculating a spatial overlap region of the first multi-dimensional sphere and the second multi-dimensional sphere to determine an intersection space of the first multi-dimensional sphere and the second multi-dimensional sphere, wherein a boundary of the intersection space is jointly determined by relative positions and radii of the first multi-dimensional sphere and the second multi-dimensional sphere; extracting semantic anchors located in the intersection space to generate a set of calibrated semantic anchors, wherein the calibrated semantic anchors simultaneously satisfy dual semantic constraints of the compact semantic space and the wide-area semantic space. 4.The method of claim 3, wherein the natural language processing algorithm is based on a rule-based algorithm. calculating a spatial overlap region of the first multi-dimensional sphere and the second multi-dimensional sphere to determine an intersection space of the first multi-dimensional sphere and the second multi-dimensional sphere, wherein a boundary of the intersection space is jointly determined by relative positions and radii of the first multi-dimensional sphere and the second multi-dimensional sphere, including: calculating a spatial vector distance between the sphere center of the first multi-dimensional sphere and the sphere center of the second multi-dimensional sphere to obtain a relative positional relationship between the sphere center of the first multi-dimensional sphere and the sphere center of the second multi-dimensional sphere; comparing the spatial vector distance with a sum of the radius of the first multi-dimensional sphere and the radius of the second multi-dimensional sphere according to the relative positional relationship between the sphere center of the first multi-dimensional sphere and the sphere center of the second multi-dimensional sphere, and when the spatial vector distance is smaller than the sum of the radius of the first multi-dimensional sphere and the radius of the second multi-dimensional sphere, determining that there is a spatial overlap region between the first multi-dimensional sphere and the second multi-dimensional sphere, and simultaneously generating a determination result; based on the determination result, calculating an intersection region of the first multi-dimensional sphere and the second multi-dimensional sphere, establishing a geometric parameter equation of sphere intersection to calculate an intersection volume of the first multi-dimensional sphere and the second multi-dimensional sphere in the semantic space, and based on the intersection volume, generating a mathematical description of the intersection space to provide geometric constraint conditions for semantic anchor screening. 5.The method of claim 4, wherein the natural language processing algorithm is based on a rule-based algorithm. performing semantic calibration processing on the child complaint text according to the set of calibrated semantic anchors to generate a semantic optimized complaint text; preprocessing the semantic optimized complaint text to generate an initial symptom entity set, including: based on the set of calibrated semantic anchors, performing semantic reconstruction processing including semantic fragment replacement and semantic weight adjustment on the received original child complaint text, and simultaneously obtaining a semantic reconstruction result; generating a semantic optimized complaint text according to the semantic reconstruction result; the semantic optimized complaint text retains the core symptom description corresponding to the calibrated semantic anchors, while removing semantic fragments that do not meet the geometric constraint conditions; preprocessing the semantic optimized complaint text to extract symptom description fragments in the semantic optimized complaint text through word segmentation processing and entity recognition operations; Based on the semantic optimization of the symptom description fragment in the chief complaint text, combined with the confidence weight of the calibrated semantic anchor, an initial symptom entity set is generated. 6.The method of claim 5, wherein the natural language processing algorithm is based on a rule-based algorithm. Based on the set of child-specific corpus, the initial symptom entity set is dynamically processed by semantic mapping to convert the child's non-standardized expression into standard medical terminology to form a standard symptom entity set, including: Based on the set of child-specific corpus, a standard medical terminology semantic space is constructed, and each symptom entity in the initial symptom entity set is mapped to the semantic space to form a corresponding semantic coordinate point, and the mapping results are recorded synchronously; Calculate the projection position of each semantic coordinate point in the standard medical terminology semantic space, that is, determine the projection direction by calculating the included angle between the semantic coordinate point and the medical terminology space base vector; Based on the projection position, the nearest neighbor search is performed in the standard medical terminology semantic space to match the nearest standard medical terminology for each semantic coordinate point, and the matching results are generated; Based on the matching results, the accuracy of the mapping results is verified by calculating the semantic similarity, the semantic similarity is calculated by using the vector inner product method, and the standard symptom entity set is generated. 7.The natural language processing algorithm based automatic extraction of children chief complaint symptoms method according to claim 6, characterized in that, The standard symptom entity set is screened to distinguish core symptoms from accompanying phenomena to obtain a core symptom entity set, including: The standard symptom entity set is subjected to symptom weight allocation, wherein the core symptoms are given a higher weight and the accompanying phenomena are given a lower weight to obtain a symptom weight allocation result; According to the symptom weight allocation result, the standard symptom entity set is screened by setting a weight threshold to remove the accompanying phenomena with a weight lower than the threshold, and a screening result is generated; Based on the screening result, the target symptom entity with a weight greater than or equal to the threshold is extracted to form the core symptom entity set.

8. An automatic extraction system of chief complaints symptoms in children based on natural language processing algorithms, characterized by, The system performs the method of any one of claims 1 to 7, comprising: A receiving module for receiving child complaint text data containing mixed information of parents' narration, and simultaneously collecting multi-modal behavior data generated by parents during the narration process; An extraction module for identifying and extracting A-type semantic anchors and B-type semantic anchors from the child complaint text data and multi-modal behavior data; A-type semantic anchors are derived from children's non-standardized expressions and are concentrated in the oral mucosa, throat and abdominal symptom areas; B-type semantic anchors are derived from parents' narration information and are concentrated in the eye gaze pattern, voice prosody characteristics and body movement amplitude; A construction module for constructing a compact semantic space corresponding to the A-type anchor set through the A-type semantic anchors; constructing a wide-area semantic space corresponding to the B-type anchor set according to the B-type semantic anchors, and generating a calibrated semantic anchor set by calculating the intersection of the compact semantic space and the wide-area semantic space; A processing module for performing semantic calibration processing on the child complaint text according to the calibrated semantic anchor set to generate a semantically optimized chief complaint text; and preprocessing the semantically optimized chief complaint text to generate an initial symptom entity set; A conversion module for dynamically processing the initial symptom entity set by semantic mapping based on the set of child-specific corpus to convert the child's non-standardized expression into standard medical terminology to form a standard symptom entity set; An execution module is configured to perform information screening on the standard symptom entity set to distinguish core symptoms from accompanying phenomena, and obtain a core symptom entity set.

9. A computing device, comprising: comprise a memory and a processor; wherein the memory stores one or more computer programs comprising instructions that, when executed by the processor, cause the computing device to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium is configured to store a computer program for performing the method of any one of claims 1 to 7.

Images