Psychiatric electronic medical record generation device based on AI voice recognition

Abstract

The application relates to the technical field of medical informatics, and particularly relates to an AI voice recognition-based psychiatric electronic medical record generation device. A voice processing module collects doctor-patient conversation audio signals, generates a time-stamped text sequence through sound source separation and voice recognition. A semantic analysis module performs context modeling, entity recognition and relationship extraction on the text sequence, and outputs clinical entities and semantic relationships. A knowledge graph module performs graph reasoning based on psychiatric clinical knowledge data, and generates a diagnosis hypothesis. A template adaptation module selects a medical record template and fills in fields according to the analysis result and the diagnosis hypothesis, and generates a preliminary medical record document. A quality control module performs multistage checking on the document, and feeds back the checking result to each module to realize parameter optimization. Through data flow conversion and closed-loop feedback mechanism among the modules, the device realizes automatic conversion of unstructured doctor-patient conversation into standardized medical records, and effectively improves the accuracy and integrity of psychiatric medical record recording.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, and in particular to an AI-based voice recognition-based device for generating electronic medical records for psychiatry. Background Technology

[0002] Existing manual medical record recording technologies have the following technical pain points: In psychiatric clinical consultations, doctor-patient dialogues are often highly complex and unstructured. For example, patients may use vague descriptions, emotional expressions, or incoherent sentences, requiring doctors to spend extra time interpreting and organizing the information during manual recording. This reduces recording efficiency and increases the risk of missing key clinical information. For instance, when assessing patients with anxiety disorders, suicide risk cues or details of emotional fluctuations in the dialogue may be simplified or ignored due to recording delays, affecting the accuracy and completeness of the medical record. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an AI-based voice recognition-based electronic medical record generation device for psychiatry. This invention solves the technical problem that the complexity and unstructured nature of psychiatric doctor-patient dialogues lead to low efficiency and easy omission of key clinical information in existing manual medical record recording.

[0004] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows:

[0005] The present invention provides an AI-based speech recognition-based electronic medical record generation device for psychiatry, comprising: The speech processing module collects audio signals of doctor-patient dialogue, performs sound source separation and speech recognition on the audio signals, and transmits the generated timestamped text sequence to the semantic analysis module. The semantic analysis module receives the text sequence transmitted by the speech processing module, performs context modeling, entity recognition and relation extraction on the text sequence, and outputs the analysis results, including clinical entities and semantic relations, to the knowledge graph module and the template adaptation module. The knowledge graph module receives the analysis results output by the semantic analysis module, performs graph reasoning based on the stored psychiatric clinical knowledge data, generates diagnostic hypotheses, and transmits them to the template adaptation module. The template adaptation module receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module, selects a medical record template and fills in the fields, and sends the generated preliminary medical record document to the quality control module. The quality control module receives the preliminary medical record documents sent by the template adaptation module, performs multi-level verification on the documents, and feeds back the verification results, including error types and quality scores, to the speech processing module, semantic analysis module, knowledge graph module, and template adaptation module. The speech processing module adjusts the sound source separation parameters and speech recognition model parameters based on the verification results fed back by the quality control module. The semantic analysis module updates the professional dictionary and entity recognition model based on the verification results. The knowledge graph module corrects the node weights and relationships based on the verification results. The template adaptation module optimizes the template selection strategy and field filling rules based on the verification results. Through data flow between modules, the conversion from unstructured dialogue to standardized medical records is completed.

[0006] Furthermore, in the AI-based speech recognition psychiatric electronic medical record generation device of the present invention, the speech processing module includes: The sound source separation unit uses a directional microphone array to collect multi-channel mixed audio signals, processes the mixed audio signals using a generalized sidelobe canceller algorithm to enhance the target sound source, and uses an independent component analysis algorithm to separate the doctor's audio signal and the patient's audio signal. The speaker recognition unit receives the doctor's audio signal and the patient's audio signal output by the sound source separation unit, and performs speaker clustering on the separated audio signals using a Gaussian mixture model. It labels the doctor's audio signal with doctor's identity attributes and the patient's audio signal with patient's identity attributes. The speech recognition unit receives audio signals with identity attribute labels output by the speaker recognition unit. It adopts an end-to-end speech recognition model with encoder and decoder architecture. The encoder converts acoustic features into hidden representations, and the decoder generates corresponding text sequences based on the attention mechanism. It outputs doctor dialogue text and patient dialogue text with timestamps and confidence scores. The quality monitoring unit receives the text sequence and confidence score output by the speech recognition unit, calculates the audio signal-to-noise ratio in real time, and sends a re-acquisition command to the sound source separation unit when the signal-to-noise ratio is lower than a preset threshold or the confidence score of the text sequence is lower than a preset score. The sound source separation unit then re-acquisitions the audio signal according to the re-acquisition command.

[0007] Furthermore, in the AI-based speech recognition psychiatric electronic medical record generation device of the present invention, the semantic analysis module includes: The text preprocessing unit receives the timestamped text sequence transmitted by the speech processing module, performs sentence boundary detection and term normalization operations on the text sequence, including word segmentation, part-of-speech tagging and dependency parsing, and outputs the preprocessed text data. The context modeling unit receives the preprocessed text data output by the text preprocessing unit, adopts a bidirectional encoder with a Transformer architecture, and calculates context-aware word vector representations through a self-attention mechanism to generate a word vector sequence that includes context information. The entity recognition unit receives the word vector sequence output by the context modeling unit, and adopts a hybrid architecture of bidirectional long short-term memory network combined with conditional random field to identify clinical entities such as symptoms, drugs, and examination items, output the entity recognition results and maintain the entity coreference resolution chain; The relation extraction unit receives the entity recognition results output by the entity recognition unit, uses dependency path features and entity relative position features, determines the semantic relationship between entities based on a neural network classifier, establishes symptom, degree and time association structure, and outputs the analysis results including clinical entities and semantic relationships to the knowledge graph module and template adaptation module.

[0008] Furthermore, in the AI-based speech recognition psychiatric electronic medical record generation device of the present invention, the knowledge graph module includes: The graph storage unit uses a graph database to store psychiatric clinical pathway data and constructs a symptom-diagnosis-treatment association network. The network nodes include medical concept codes and clinical weight parameters. The graph query unit receives the analysis results, including clinical entities and semantic relationships, output by the semantic analysis module, extracts entity sets from the analysis results, retrieves relevant subgraphs through graph pattern matching, and transmits the retrieved subgraphs to the graph reasoning unit. The graph reasoning unit receives the relevant subgraphs transmitted by the graph query unit, performs path reasoning using a random walk algorithm, calculates the probability of diagnostic hypotheses based on a path sorting algorithm, generates a candidate diagnosis list, and transmits it to the template adaptation module. The dynamic update unit receives the analysis results from the semantic analysis module and the field filling results from the template adaptation module. Based on the analysis results and field filling results, it dynamically adjusts the node weights and edge strengths, discovers potential association patterns through meta-path similarity calculation, and updates the association network in the graph storage unit.

[0009] Furthermore, in the AI-based speech recognition psychiatric electronic medical record generation device of the present invention, the template adaptation module includes: The template matching unit receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module, calculates the similarity score between the analysis results and the medical record template fields, selects the optimal template based on clinical priority rules, and transmits the selected template information to the field filling unit. The field filling unit receives template information transmitted by the template matching unit, performs data type conversion and format standardization operations, performs weighted fusion processing on conflicting content, uses the beam search algorithm for multi-sequence optimization, and generates field filling results. The integrity verification unit receives the field filling results generated by the field filling unit, checks the coverage of required fields, generates a missing field report, and sends the verified medical record documents to the quality control module. The reinforcement learning unit receives the verification results from the quality control module, uses the quality of medical record generation as the reward function, and dynamically adjusts the template selection strategy and field mapping rules through the near-end strategy optimization algorithm. The adjusted strategy and rules are then fed back to the template matching unit and the field filling unit.

[0010] Furthermore, in the AI-based speech recognition psychiatric electronic medical record generation device of the present invention, the quality control module includes: The rule verification unit receives the preliminary medical record document sent by the template adaptation module, performs syntax standardization checks, medical terminology accuracy verification, and numerical range logic constraint checks, and outputs the first-level verification results. The anomaly detection unit receives the first-level verification result output by the rule verification unit, uses the isolated forest algorithm to identify outliers, combines the adversarial generative network to identify anomaly patterns, marks anomalous data items, and outputs the second-level verification result. The clinical validation unit receives the second-level validation results output by the anomaly detection unit, compares the generated content with the medical guidelines, performs clinical rationality validation through the expert knowledge base, calculates the clinical rationality score, and outputs the third-level validation results. The feedback control unit receives the third-level verification results output by the clinical validation unit, assigns correction strategies according to the error type, and feeds back the verification results, including the error type and quality score, to the speech processing module, semantic analysis module, knowledge graph module, and template adaptation module. The feedback control unit records error tracking information and performance monitoring data.

[0011] Furthermore, the AI-based speech recognition psychiatric electronic medical record generation device of the present invention further includes: The speech processing module and the semantic analysis module establish a two-way feedback channel: The speech recognition unit transmits the timestamp and confidence score attached to the recognition result to the text preprocessing unit of the semantic analysis module; The entity recognition unit feeds back the entity recognition results to the speech recognition unit in real time, and the speech recognition unit adjusts the weights of the domain language model based on the entity recognition results. When the quality monitoring unit detects a low-confidence text segment, it sends a reprocessing request to the relation extraction unit of the semantic analysis module. The relation extraction unit returns context information to the quality monitoring unit, which then generates an audio re-acquisition decision instruction based on the context information.

[0012] Furthermore, the AI-based speech recognition psychiatric electronic medical record generation device of the present invention further includes: The semantic analysis module and the knowledge graph module establish an iterative optimization mechanism: The entity recognition unit sends an entity query request to the atlas query unit, and the atlas query unit retrieves relevant clinical pathway constraints from the atlas storage unit. The graph reasoning unit returns the set of diagnostic hypotheses to the entity recognition unit, which then corrects the entity recognition bias based on the set of diagnostic hypotheses. The relation extraction unit outputs the corrected entity relations to the dynamic update unit, which then updates the node association strength and edge weight parameters based on the corrected entity relations.

[0013] Furthermore, the AI-based speech recognition psychiatric electronic medical record generation device of the present invention further includes: The knowledge graph module and the template adaptation module establish a constraint propagation path: The graph reasoning unit transmits the clinical pathway constraints of symptoms, diagnosis, and treatment to the template matching unit, which uses the clinical pathway constraints as the priority for template selection. The field population unit feeds back the field population results to the dynamic update unit, which uses the field population results to verify the accuracy of the diagnostic assumptions. The dynamic update unit fills in the consistency score based on the fields and dynamically adjusts the node weights and association strengths in the graph storage unit.

[0014] Furthermore, the AI-based speech recognition psychiatric electronic medical record generation device of the present invention further includes: The template adaptation module and the quality control module form a closed-loop verification circuit: The integrity verification unit sends the verified medical record documents to the rule verification unit of the quality control module; The feedback control unit generates a verification report that includes error localization and quality scoring, and sends the verification report to the reinforcement learning unit; The reinforcement learning unit adjusts the template selection strategy and field mapping rules based on the verification report, and feeds back the adjusted strategy and rules to the template matching unit and field filling unit; The feedback control unit collects cross-institutional optimization data and updates the model parameters of the context modeling unit and entity recognition unit of the semantic analysis module through a federated learning strategy.

[0015] Beneficial effects of this invention; This invention effectively solves the problems of low efficiency and omission of key information in psychiatric clinical consultations caused by the complexity and unstructured nature of doctor-patient dialogues through the collaborative work and bidirectional data flow of a speech processing module, a semantic analysis module, a knowledge graph module, a template adaptation module, and a quality control module. The speech processing module accurately distinguishes doctor-patient audio and generates timestamped text sequences by combining sound source separation and speaker recognition technologies. The semantic analysis module transforms unstructured dialogues into structured clinical data through context modeling and entity relation extraction. The knowledge graph module generates diagnostic hypotheses and provides clinical pathway support based on symptom diagnosis and treatment association networks. The template adaptation module dynamically optimizes template selection and field filling strategies using reinforcement learning. The quality control module continuously improves the quality of generated medical records through multi-level verification and feedback mechanisms. The closed-loop optimization mechanism between the modules enables the device to adaptively learn and optimize the processing flow, ultimately achieving efficient and accurate conversion from unstructured dialogues to standardized medical records, significantly improving the completeness, accuracy, and clinical applicability of psychiatric medical records. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0017] Figure 1 This is a system architecture diagram of the AI ​​speech recognition-based psychiatric electronic medical record generation device of the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] The present invention provides an AI-based speech recognition-based electronic medical record generation device for psychiatry, comprising: The speech processing module collects audio signals of doctor-patient dialogue, performs sound source separation and speech recognition on the audio signals, and transmits the generated timestamped text sequence to the semantic analysis module. The semantic analysis module receives the text sequence transmitted by the speech processing module, performs context modeling, entity recognition and relation extraction on the text sequence, and outputs the analysis results, including clinical entities and semantic relations, to the knowledge graph module and the template adaptation module. The knowledge graph module receives the analysis results output by the semantic analysis module, performs graph reasoning based on the stored psychiatric clinical knowledge data, generates diagnostic hypotheses, and transmits them to the template adaptation module. The template adaptation module receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module, selects a medical record template and fills in the fields, and sends the generated preliminary medical record document to the quality control module. The quality control module receives the preliminary medical record documents sent by the template adaptation module, performs multi-level verification on the documents, and feeds back the verification results, including error types and quality scores, to the speech processing module, semantic analysis module, knowledge graph module, and template adaptation module. The speech processing module adjusts the sound source separation parameters and speech recognition model parameters based on the verification results fed back by the quality control module. The semantic analysis module updates the professional dictionary and entity recognition model based on the verification results. The knowledge graph module corrects the node weights and relationships based on the verification results. The template adaptation module optimizes the template selection strategy and field filling rules based on the verification results. Through data flow between modules, the conversion from unstructured dialogue to standardized medical records is completed.

[0020] The AI-based speech recognition-based electronic medical record generation device for psychiatry automatically generates standardized medical records from doctor-patient dialogue audio through the collaborative work of multiple modules. The speech processing module first acquires the audio signal of the doctor-patient dialogue, using a directional microphone array to capture multi-channel mixed audio. Then, it enhances the target sound source using a generalized sidelobe canceller algorithm and separates the doctor's and patient's audio streams using independent component analysis. The speaker recognition unit clusters the separated audio using a Gaussian mixture model, labeling each speech segment with its identity attributes. The speech recognition unit employs an end-to-end model with an encoder and decoder architecture, converting acoustic features into hidden representations and generating text sequences with timestamps and confidence scores based on an attention mechanism. The quality monitoring unit calculates the audio signal-to-noise ratio (SNR) in real time. When the SNR falls below a preset threshold or the text confidence score is insufficient, it triggers an audio re-acquisition process to ensure the reliability of the input data.

[0021] The semantic analysis module receives the text sequence transmitted by the speech processing module and performs text preprocessing operations, including sentence boundary detection, terminology normalization, word segmentation, part-of-speech tagging, and dependency parsing. The context modeling unit employs a bidirectional encoder based on a Transformer architecture, using a self-attention mechanism to compute context-aware word vector representations and generate word vector sequences that include dialogue history information. The entity recognition unit, based on a hybrid architecture combining bidirectional long short-term memory networks and conditional random fields, identifies clinical entities such as symptoms, medications, and examination items, and maintains entity coreference resolution chains to ensure contextual consistency. The relation extraction unit utilizes dependency path features and entity relative position features, based on a neural network classifier, to determine semantic relationships between entities, establishes a symptom severity temporal correlation structure, and outputs the analysis results to downstream modules.

[0022] The knowledge graph module constructs a symptom diagnosis and treatment association network based on stored psychiatric clinical knowledge data. It uses a graph database to store node and edge information, with nodes including medical concept codes and clinical weight parameters. The graph query unit extracts entity sets from semantic analysis results and retrieves relevant subgraphs through graph pattern matching. The graph reasoning unit uses a random walk algorithm for path reasoning and a path ranking algorithm to calculate diagnostic hypothesis probabilities, generating a candidate diagnosis list. The dynamic update unit dynamically adjusts node weights and edge strengths based on semantic analysis output and template adaptation results, discovers potential association patterns through meta-path similarity calculation, and optimizes the graph structure.

[0023] The template adaptation module receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module. The template matching unit calculates the similarity score between the analysis results and the fields of the medical record template, and selects the optimal template based on clinical priority rules. The field filling unit performs data type conversion and format standardization operations, performs weighted fusion processing on conflicting content, and uses a bundle search algorithm for multi-sequence optimization to generate field filling results. The integrity verification unit checks the coverage of required fields, generates a missing field report, and sends the verified medical record documents to the quality control module. The reinforcement learning unit uses the quality of medical record generation as a reward function and dynamically adjusts the template selection strategy and field mapping rules through a proximal strategy optimization algorithm.

[0024] The quality control module performs multi-level validation on preliminary medical record documents. The rule validation unit performs grammatical accuracy checks, medical terminology accuracy verification, and numerical range logical constraint checks. The anomaly detection unit uses the isolated forest algorithm to identify outliers and combines it with generative adversarial networks for anomaly pattern recognition, marking anomalous data items. The clinical validation unit compares the generated content with medical guidelines, performs clinical rationality validation using an expert knowledge base, and calculates a clinical rationality score. The feedback control unit assigns correction strategies based on error type and feeds back the validation results, including error type and quality score, to the upstream module.

[0025] The device employs a closed-loop optimization mechanism through data flow between modules. The speech processing module adjusts sound source separation parameters and speech recognition model parameters based on feedback; the semantic analysis module updates the specialized dictionary and entity recognition model; the knowledge graph module corrects node weights and relationships; and the template adaptation module optimizes template selection strategies and field filling rules. This iterative optimization process, based on error tracking and performance monitoring data from real-world applications, gradually improves the accuracy and efficiency of the device in processing unstructured dialogues, ultimately achieving the generation of standardized medical records. In psychiatric clinical consultation scenarios, the device can handle patients' vague descriptions or emotional expressions, reducing the risk of missing key information. For example, it can capture details such as suicide risk cues when assessing anxiety disorders, enhancing the completeness and clinical value of medical records.

[0026] In the speech processing module, the sound source separation unit uses a directional microphone array to acquire multi-channel mixed audio signals. The microphone array uses beamforming technology to directionally capture the speech signals of doctors and patients. A generalized sidelobe canceller algorithm processes the mixed audio signals, calculating beam weights through adaptive filters to enhance the speech energy in the direction of the target sound source while suppressing environmental noise and reverberation interference. An independent component analysis algorithm is then applied to the enhanced signal, using the non-Gaussianity assumption of the signal to extract statistically independent doctor and patient audio signals by iteratively optimizing the separation matrix. The speaker recognition unit receives the separated audio signals and uses a Gaussian mixture model to cluster the voiceprint features. The feature extraction process includes calculating Mel-frequency cepstral coefficients to label the doctor's audio signal with doctor identity attributes and the patient's audio signal with patient identity attributes. The speech recognition unit adopts an end-to-end model with an encoder and decoder architecture. The encoder converts the acoustic feature sequence into a hidden representation, and the decoder generates the corresponding text sequence based on an attention mechanism, outputting doctor and patient dialogue texts with timestamps and confidence scores. The quality monitoring unit calculates the audio signal-to-noise ratio (SNR) in real time and monitors the confidence score of the text sequence. When the SNR is lower than the preset threshold or the confidence score is insufficient, it sends a re-acquisition command to the sound source separation unit. The sound source separation unit re-initializes the audio acquisition process according to the command to improve the reliability of the input data.

[0027] The semantic analysis module's text preprocessing unit receives timestamped text sequences from the speech processing module, performs sentence boundary detection, identifies sentence breaks in the dialogue, and performs terminology normalization, including word segmentation, part-of-speech tagging, and dependency parsing, outputting structured preprocessed text data. The context modeling unit employs a Transformer-based bidirectional encoder, using a self-attention mechanism to compute word vector representations, capture long-distance dependencies, and generate word vector sequences including contextual information. The entity recognition unit, based on a hybrid architecture combining bidirectional long short-term memory networks and conditional random fields, performs sequence labeling on the word vector sequences, identifies clinical entities such as symptoms, drugs, and examination items, and maintains entity coreference resolution chains to track multiple mentions of the same entity. The relation extraction unit utilizes dependency path features and entity relative position features, based on a neural network classifier, to determine semantic relationships between entities, such as the modification relationship between symptoms and severity or the association structure between symptoms and time, outputting the analysis results, including clinical entities and semantic relationships, to downstream modules.

[0028] The knowledge graph module's graph storage unit uses a graph database to store psychiatric clinical pathway data, constructing a symptom-diagnosis-treatment association network. Network nodes include medical concept encodings and clinical weight parameters, with edges representing the strength of associations between entities. The graph query unit extracts entity sets from the analysis results output by the semantic analysis module and retrieves relevant subgraphs through graph pattern matching, such as finding possible diagnostic paths based on symptom nodes. The graph reasoning unit receives the retrieved subgraphs, performs path reasoning using a random walk algorithm, calculates diagnostic hypothesis probabilities based on a path ranking algorithm, and generates a candidate diagnosis list. The dynamic update unit dynamically adjusts node weights and edge strengths based on the analysis results from the semantic analysis module and the field filling results from the template adaptation module. It discovers potential association patterns through meta-path similarity calculations, such as updating the weights of high-frequency query paths, and optimizing the association network in the graph storage unit.

[0029] The template matching unit of the template adaptation module receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module. It calculates the similarity score between the analysis results and the fields of the medical record template, and selects the optimal template based on clinical priority rules, such as prioritizing template types related to the diagnostic hypotheses. The field filling unit performs data type conversion and format standardization operations, performs weighted fusion processing on conflicting content, and uses a bundle search algorithm for multi-sequence optimization to generate field filling results. The integrity verification unit checks the coverage of required fields, generates a missing field report, and sends the verified medical record documents to the quality control module. The reinforcement learning unit uses the quality of medical record generation as a reward function and dynamically adjusts the template selection strategy and field mapping rules through a proximal policy optimization algorithm, such as optimizing field weight allocation based on historical feedback.

[0030] The quality control module's rule verification unit receives preliminary medical record documents from the template adaptation module, performs grammatical compliance checks, medical terminology accuracy verification, and numerical range logic constraint checks, outputting the first-level verification result. The anomaly detection unit uses the isolated forest algorithm to identify outliers, combines it with a generative adversarial network for anomaly pattern recognition, marks anomalous data items, and outputs the second-level verification result. The clinical validation unit compares the generated content with medical guidelines, performs clinical rationality verification using an expert knowledge base, calculates a clinical rationality score, and outputs the third-level verification result. The feedback control unit assigns correction strategies based on error type, feeds back the verification results, including error type and quality score, to the upstream module, and records error tracking information and performance monitoring data.

[0031] The speech processing module and the semantic analysis module establish a two-way feedback channel. The speech recognition unit transmits the timestamp and confidence score of the recognition result to the text preprocessing unit of the semantic analysis module for adjusting text alignment. The entity recognition unit feeds back the entity recognition result to the speech recognition unit in real time. The speech recognition unit adjusts the weights of the domain language model based on the entity recognition result, for example, increasing the weight of psychiatric terms. When the quality monitoring unit detects a low-confidence text segment, it sends a reprocessing request to the relation extraction unit of the semantic analysis module. The relation extraction unit returns context information to the quality monitoring unit, which generates an audio re-acquisition decision instruction based on the context information to improve recognition accuracy.

[0032] The semantic analysis module and the knowledge graph module establish an iterative optimization mechanism. The entity recognition unit sends entity query requests to the graph query unit, which retrieves relevant clinical pathway constraints from the graph storage unit to correct entity boundaries. The graph reasoning unit returns a set of diagnostic hypotheses to the entity recognition unit, which corrects entity recognition biases based on the set of diagnostic hypotheses, such as adjusting the confidence level of symptom entities. The relation extraction unit outputs the corrected entity relations to the dynamic update unit, which updates the node association strength and edge weight parameters based on the corrected entity relations to enhance the accuracy of the graph.

[0033] The knowledge graph module and template matching module establish a constraint transmission path. The knowledge graph reasoning unit transmits clinical pathway constraints related to symptoms, diagnoses, and treatment to the template matching unit. The template matching unit prioritizes template selection based on these clinical pathway constraints, for example, prioritizing templates that match the diagnostic pathway. The field filling unit feeds back the field filling results to the dynamic update unit, which uses these results to verify the accuracy of the diagnostic hypothesis, for example, by comparing the filled content with the reasoning results. Based on the field filling consistency score, the dynamic update unit dynamically adjusts the node weights and association strengths in the knowledge graph storage unit, optimizing the reasoning logic.

[0034] The template adaptation module and the quality control module form a closed-loop verification circuit. The integrity verification unit sends verified medical records to the rule verification unit of the quality control module, initiating a multi-level verification process. The feedback control unit generates a verification report including error localization and quality scoring, and sends the report to the reinforcement learning unit. The reinforcement learning unit adjusts the template selection strategy and field mapping rules based on the verification report, and feeds the adjusted strategy and rules back to the template matching unit and field filling unit. The feedback control unit collects cross-institutional optimization data and updates the model parameters of the context modeling unit and entity recognition unit of the semantic analysis module through a federated learning strategy, improving the device's generalization ability.

[0035] When the AI-based speech recognition-based electronic medical record generation device for psychiatric clinics is deployed in psychiatric clinical consultation scenarios, the speech processing module collects the raw audio signals of doctor-patient dialogues through a directional microphone array. In the actual treatment environment, the microphone array uses beamforming technology to directionally capture the speech signals from the doctor and patient, effectively isolating interference from noisy ambient sounds in the waiting area. A generalized sidelobe canceller algorithm preprocesses the mixed audio signal, calculates the optimal beam weight through an adaptive filter, and enhances the speech energy in the direction of the target sound source. An independent component analysis algorithm then performs blind source separation on the enhanced signal, utilizing the non-Gaussian characteristics of the signal to iteratively optimize the separation matrix, ultimately outputting independent audio streams for the doctor and patient.

[0036] The speaker recognition unit extracts features from the separated audio signals, calculates Mel-frequency cepstral coefficients as voiceprint features, and performs speaker clustering using a Gaussian mixture model. For example, during a consultation with a patient with depression, the device can accurately distinguish the doctor's tone of voice from the patient's low-pitched responses, labeling each speech segment with its corresponding identity attribute. The speech recognition unit employs an end-to-end model with an encoder and decoder architecture. The encoder converts the acoustic feature sequence into a hidden representation, and the decoder generates a timestamped text sequence based on an attention mechanism. When a patient's narration is disjointed, the quality monitoring unit monitors the audio signal-to-noise ratio and recognition confidence in real time. When the indicators fall below a set threshold, a re-acquisition mechanism is automatically triggered, and the sound source separation unit re-initializes the acquisition process to improve data quality.

[0037] After receiving the text sequence, the semantic analysis module performs sentence boundary detection to identify natural sentence breaks in the dialogue. Terminology normalization converts colloquial expressions into standard medical terms, for example, normalizing "can't sleep" to "insomnia symptoms." The context modeling unit calculates word vector representations using a bidirectional encoder based on a Transformer architecture, capturing long-distance semantic dependencies. The entity recognition unit employs a hybrid architecture combining a bidirectional long short-term memory network and a conditional random field. When recognizing clinical entities such as "depressed mood" and "loss of appetite," it can accurately associate multiple mentions of the same symptom description by maintaining entity coreference unlinking chains. The relation extraction unit uses dependency path analysis to establish semantic relationships between entities, building a relational structure between symptoms and factors such as degree and time.

[0038] The knowledge graph module's graph storage unit constructs a symptom-diagnosis-treatment association network, including clinical pathway data for psychiatric disorders such as depression and anxiety. The graph query unit retrieves relevant subgraphs based on the entity set output by semantic analysis through graph pattern matching. For example, when the input is the entity combination "insomnia + depressed mood + loss of interest," the graph reasoning unit uses a random walk algorithm for path reasoning and calculates the probability weight of the depression diagnosis hypothesis based on a path ranking algorithm. The dynamic update unit dynamically adjusts the node association strength based on the field filling results from the template adaptation module during actual consultations. If a strong association is found between a symptom combination and the diagnosis result, the weight coefficient of that path is increased accordingly.

[0039] The template matching unit of the template adaptation module calculates the similarity score between the template and the medical record template fields, combining semantic analysis results and diagnostic hypotheses. The field filling unit uses a bundle search algorithm for multi-sequence optimization, performing weighted fusion processing on potentially conflicting clinical descriptions. The integrity verification unit checks the coverage of required fields and generates supplementary collection suggestions when important clinical information is missing. The reinforcement learning unit uses the quality of medical record generation as a reward function and continuously adjusts the field mapping rules through a proximal strategy optimization algorithm. For example, when processing bipolar disorder cases, the device prioritizes matching specialized templates related to mood fluctuations to ensure accurate recording of the characteristics of alternating manic and depressive episodes.

[0040] The quality control module implements a multi-level verification mechanism. The rule verification unit checks the standardization of medical terminology usage, while the anomaly detection unit uses the isolated forest algorithm to identify abnormal data that does not conform to clinical practice. The clinical validation unit compares the generated content with the psychiatric treatment guidelines and calculates a clinical rationality score. The feedback control unit assigns correction strategies based on the error type; if insufficient diagnostic evidence is found, it suggests supplementing the collection of specific symptom information. The device, through a federated learning strategy, optimizes model parameters using cross-institutional data while protecting patient privacy, continuously improving its ability to handle complex clinical manifestations.

[0041] During device operation, a complete two-way feedback mechanism is established between the modules. The speech processing module adjusts the weights of the domain language model based on the semantic analysis results, while the knowledge graph module optimizes the reasoning path based on the field filling results, forming a continuously self-improving closed-loop device. This architecture is particularly suitable for unstructured dialogue scenarios commonly encountered in psychiatric consultations, effectively capturing clinical value information from patients' emotional expressions and significantly improving the completeness and accuracy of medical records.

[0042] Embodiment 1 of the present invention: In clinical consultations with patients with depression, the device collects audio signals of doctor-patient conversations via a directional microphone array. The sound source separation unit uses a generalized sidelobe canceller algorithm to process the mixed audio, effectively separating the doctor's questions from the patient's responses. The speaker recognition unit uses a Gaussian mixture model to cluster voiceprint features in the separated audio, accurately labeling the identity attributes of both parties. The speech recognition unit converts the audio into a timestamped text sequence using an end-to-end model. When a patient exhibits slow speech or repetitive expressions, the quality monitoring unit monitors the confidence index in real time, triggering an audio re-acquisition mechanism to ensure information integrity.

[0043] The semantic analysis module performs terminology normalization on the text sequence, converting patient-described symptoms such as depression and insomnia into standard medical terms. The entity recognition unit identifies clinical entities using a bidirectional long short-term memory network, while the relation extraction unit establishes association structures such as symptom duration and severity. The knowledge graph module performs path reasoning based on symptom combinations, generating diagnostic hypotheses and differential diagnosis suggestions for depression. The template adaptation module prioritizes matching medical record templates related to mood disorders, and the field filling unit uses a bundle search algorithm to optimize the logical order of symptom descriptions.

[0044] The quality control module identifies potential issues in medical records through a multi-level verification mechanism. The rule verification unit detects logical inconsistencies between the patient's chief complaint and present illness history over time; the anomaly detection unit identifies abnormal drug dosage data; and the clinical validation unit assesses the adequacy of the diagnostic evidence based on treatment guidelines. The feedback control unit relays the verification results to each module, enabling the semantic analysis module to update its professional dictionary and the knowledge graph module to adjust diagnostic path weights, thus forming a closed-loop optimization. The device ultimately generates a complete document conforming to psychiatric medical record standards, including core elements such as symptomatological analysis, diagnostic evidence, and treatment recommendations.

[0045] Embodiment 2 of the present invention: In acute assessment scenarios for patients with bipolar disorder, the device faces the challenge of significant mood swings and disjointed speech. The speech processing module enhances the target speech signal using adaptive beamforming technology, while independent component analysis (ICA) effectively separates the rapid speech during manic episodes from the subdued responses of depressive phases. The speech recognition unit adjusts the acoustic model parameters for emotional expressions, and the quality monitoring unit dynamically adjusts the confidence threshold based on audio spectral characteristics to ensure recognition accuracy under specific emotional states.

[0046] The semantic analysis module employs an attention mechanism to capture key information points in the patient's narrative, while the entity recognition unit uses a context-aware model to distinguish between grandiose ideas and reality testing abilities during manic episodes. The knowledge graph module performs dynamic reasoning based on symptom temporal features to identify cyclical patterns in mood swings. The template adaptation module calls upon templates specifically designed for bipolar disorder, and the field filling unit performs weighted fusion processing on contradictory symptom descriptions. The reinforcement learning unit optimizes the template selection strategy based on feedback from historical medical records, improving adaptability to complex clinical manifestations.

[0047] The quality control module detected that the descriptions of emotion transitions in the medical record documents were not clear enough, and the feedback control unit initiated a multi-round correction process. The semantic analysis module increased the weight of emotion vocabulary recognition, the knowledge graph module supplemented the typical feature library of emotion transitions, and the template adaptation module adjusted the temporal granularity of symptom recordings. By integrating multi-center clinical data through a federated learning strategy, the device continuously optimized its ability to identify specific manifestations of bipolar disorder, ultimately generating detailed medical records with clinical reference value, accurately recording mixed episode characteristics and longitudinal disease progression.

[0048] In the sound source separation unit of the speech processing module, the generalized sidelobe canceller algorithm calculates beam weights through an adaptive filter to enhance the speech energy in the direction of the target sound source. The core formula of this algorithm is as follows: ; in, Indicates at a point in time The filter weight vector, It is the input signal vector. It is an error signal. It is the step size parameter. It is a small constant used to avoid division by zero. The formula minimizes the error between the filter output and the desired signal by iteratively updating the weights.

[0049] Independent component analysis (ICA) algorithms are used to separate mixed audio signals, with the goal of optimizing the separation matrix. This makes the output signal The statistical independence is maximized. A commonly used optimization criterion is to minimize mutual information, with the formula: ; in, It is the output component entropy, It is the determinant of the separation matrix. The algorithm iteratively updates it using gradient descent. Until it converges.

[0050] In the speaker recognition unit, a Gaussian mixture model is used to cluster voiceprint features. The model probability density function is: ; in, It is an eigenvector (such as Mel frequency cepstral coefficients). It is the number of hybrid components. It is the first The weight of each component, Indicates a Gaussian distribution. and These are the mean and covariance matrices, respectively. The model parameters are estimated using the expectation-maximization algorithm.

[0051] In the context modeling unit of the semantic analysis module, the self-attention mechanism of the Transformer architecture is used to compute word vector representations. The formula for calculating attention weights is: ; in, , and These are query, key, and value matrices, respectively. This is the dimension of the key vector. The softmax function ensures weight normalization, thereby capturing long-distance dependencies.

[0052] The entity recognition unit employs a hybrid architecture combining a bidirectional long short-term memory network and a conditional random field (CRF). The CRF, given an input sequence... At that time, output the label sequence The probability is: ; in, It is a normalization factor. It is a characteristic function. These are the weight parameters. The model is trained by maximizing the log-likelihood.

[0053] In the graph reasoning unit of the knowledge graph module, a random walk algorithm is used for path reasoning. Node visit probabilities are determined by the transition matrix. Calculation, where Indicates from node To the node The transition probability. After... After the walk, the node probability distribution is as follows: ; in, This is the initial probability distribution. The path sorting algorithm then calculates the score of the diagnostic hypothesis based on the probability.

[0054] In the field filling unit of the template adaptation module, the beam search algorithm is used for multi-sequence optimization. The algorithm retains [data / values] at each step. There are candidate sequences, among which This is the beam width. The scoring function is typically a sum of log probabilities: ; in, It's time to step sequence, This is the input data.

[0055] The reinforcement learning unit uses a proximate policy optimization algorithm to dynamically adjust the template selection strategy. The objective function is: ; in, These are strategy parameters. It's a probability ratio. It is the dominant function. These are the pruning parameters. The algorithm uses a gradient ascent optimization strategy.

[0056] In the anomaly detection unit of the quality control module, the Isolation Forest algorithm is used to identify outliers. Anomaly scores are calculated based on path length. ; in, It is the path length from the root node to the leaf node. It is the expected value. The average path length is used for normalization.

[0057] The speech processing module's source separation unit employs a generalized sidelobe canceller algorithm to process mixed audio signals. The algorithm's data processing path is as follows: first, time-frequency transformation is performed on the multi-channel signals acquired by the microphone array, converting the time-domain signal into a frequency-domain sub-band signal; then, the autocorrelation matrix of the input signal is calculated at each frequency point, and the optimal beam weight vector is solved using the minimum mean square error criterion; finally, the weight vector is applied to the input signal to synthesize the enhanced target source signal. The independent component analysis algorithm's data processing path includes: centering and whitening preprocessing of the enhanced signal; estimating the separation matrix using a fixed-point iterative algorithm; approximating the statistical characteristics of the source signal using a nonlinear function; and finally, separating the doctor's and patient's audio signals through matrix multiplication.

[0058] The Gaussian mixture model computation path of the speaker recognition unit is as follows: First, Mel-frequency cepstral coefficient feature vectors are extracted from the separated audio. The model parameters, including the mean, covariance, and mixture weights of each Gaussian component, are estimated using the expectation-maximization algorithm. In the recognition stage, the likelihood probability of the input feature vector relative to each speaker model is calculated, and the identity corresponding to the model with the highest probability is selected as the recognition result. The data processing path of the encoder and decoder architecture of the speech recognition unit includes: The encoder converts the acoustic feature sequence into a hidden state sequence through a multi-layer long short-term memory network. The decoder calculates the weight distribution of the encoder's hidden states based on an attention mechanism, obtains the context vector through weighted summation, and finally predicts the probability distribution of the output word by combining the current state.

[0059] The context modeling unit of the semantic analysis module adopts a Transformer architecture. Its self-attention mechanism's computational path is as follows: the input word vectors are linearly transformed into query, key, and value matrices, respectively; the dot product of the query and key is calculated and its numerical range is adjusted using a scaling factor; the softmax function is applied to obtain the attention weight distribution; and finally, the weights are multiplied by the value matrix to obtain the weighted output. The conditional random field model computational path of the entity recognition unit includes: extracting contextual features of the input sequence through a bidirectional long short-term memory network; calculating the weighted sum of the state feature function and the transition feature function; calculating the normalization factor using forward and backward algorithms; and decoding the optimal label sequence using the Viterbi algorithm.

[0060] The graph reasoning unit of the knowledge graph module employs a random walk algorithm. Its computational path is as follows: Initialize the distribution of starting nodes based on the entity set output from semantic analysis; simulate multi-step walks using the transition probability matrix; statistically analyze the frequency distribution of each node's visits; and calculate the confidence score of the diagnostic hypothesis based on a path ranking algorithm. The meta-path similarity calculation path of the dynamic update unit includes: extracting different types of meta-path instances connecting two nodes; calculating the topological feature value of each meta-path; fusing the similarity scores of different meta-paths through linear weighting; and adjusting the association strength between nodes based on the score results.

[0061] The field filling unit of the template adaptation module uses a beam search algorithm for multi-sequence optimization. Its computational path is as follows: maintain a fixed-size candidate sequence set; at each time step, expand all candidate sequences with possible next words; calculate the cumulative score of each expanded sequence based on the language model score and template constraints; retain the k highest-scoring sequences for further expansion until all sequences generate an end marker. The near-end policy optimization algorithm of the reinforcement learning unit includes the following computational path: collect multi-round interaction data using the current policy model; calculate the probability ratio of the new policy to the old policy; control the policy update magnitude by pruning the probability ratio; and adjust the policy gradient update direction based on the advantage function estimation.

[0062] The anomaly detection unit of the quality control module employs the isolated forest algorithm. Its computational path involves: randomly selecting feature dimensions to construct multiple isolated trees; calculating the path length from the root node to the leaf node for each data point; estimating the anomaly score by averaging the path lengths from multiple samples; and determining whether a data point is an outlier based on a preset threshold. The federated learning strategy's computational path includes: each participant training and updating their model on their local dataset; encrypting the model update and sending it to the aggregation server; the server calculating the global model update using a secure aggregation algorithm; and distributing the aggregation result to each participant to update their local model.

Claims

1. A psychiatric electronic medical record generation device based on AI speech recognition, characterized in that, include: The speech processing module collects audio signals of doctor-patient dialogue, performs sound source separation and speech recognition on the audio signals, and transmits the generated timestamped text sequence to the semantic analysis module. The semantic analysis module receives the text sequence transmitted by the speech processing module, performs context modeling, entity recognition and relation extraction on the text sequence, and outputs the analysis results, including clinical entities and semantic relations, to the knowledge graph module and the template adaptation module. The knowledge graph module receives the analysis results output by the semantic analysis module, performs graph reasoning based on the stored psychiatric clinical knowledge data, generates diagnostic hypotheses, and transmits them to the template adaptation module. The template adaptation module receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module, selects a medical record template and fills in the fields, and sends the generated preliminary medical record document to the quality control module. The quality control module receives the preliminary medical record documents sent by the template adaptation module, performs multi-level verification on the documents, and feeds back the verification results, including error types and quality scores, to the speech processing module, semantic analysis module, knowledge graph module, and template adaptation module. The speech processing module adjusts the sound source separation parameters and speech recognition model parameters based on the verification results fed back by the quality control module. The semantic analysis module updates the professional dictionary and entity recognition model based on the verification results. The knowledge graph module corrects the node weights and relationships based on the verification results. The template adaptation module optimizes the template selection strategy and field filling rules based on the verification results. Through data flow between modules, the conversion from unstructured dialogue to standardized medical records is completed.

2. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 1, characterized in that, The voice processing module includes: The sound source separation unit uses a directional microphone array to collect multi-channel mixed audio signals, processes the mixed audio signals using a generalized sidelobe canceller algorithm to enhance the target sound source, and uses an independent component analysis algorithm to separate the doctor's audio signal and the patient's audio signal. The speaker recognition unit receives the doctor's audio signal and the patient's audio signal output by the sound source separation unit, and performs speaker clustering on the separated audio signals using a Gaussian mixture model. It labels the doctor's audio signal with doctor's identity attributes and the patient's audio signal with patient's identity attributes. The speech recognition unit receives audio signals with identity attribute labels output by the speaker recognition unit. It adopts an end-to-end speech recognition model with encoder and decoder architecture. The encoder converts acoustic features into hidden representations, and the decoder generates corresponding text sequences based on the attention mechanism. It outputs doctor dialogue text and patient dialogue text with timestamps and confidence scores. The quality monitoring unit receives the text sequence and confidence score output by the speech recognition unit, calculates the audio signal-to-noise ratio in real time, and sends a re-acquisition command to the sound source separation unit when the signal-to-noise ratio is lower than a preset threshold or the confidence score of the text sequence is lower than a preset score. The sound source separation unit then re-acquisitions the audio signal according to the re-acquisition command.

3. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 2, characterized in that, The semantic analysis module includes: The text preprocessing unit receives the timestamped text sequence transmitted by the speech processing module, performs sentence boundary detection and term normalization operations on the text sequence, including word segmentation, part-of-speech tagging and dependency parsing, and outputs the preprocessed text data. The context modeling unit receives the preprocessed text data output by the text preprocessing unit, adopts a bidirectional encoder with a Transformer architecture, and calculates context-aware word vector representations through a self-attention mechanism to generate a word vector sequence that includes context information. The entity recognition unit receives the word vector sequence output by the context modeling unit, and adopts a hybrid architecture of bidirectional long short-term memory network combined with conditional random field to identify clinical entities such as symptoms, drugs, and examination items, output the entity recognition results and maintain the entity coreference resolution chain; The relation extraction unit receives the entity recognition results output by the entity recognition unit, uses dependency path features and entity relative position features, determines the semantic relationship between entities based on a neural network classifier, establishes symptom, degree and time association structure, and outputs the analysis results including clinical entities and semantic relationships to the knowledge graph module and template adaptation module.

4. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 3, characterized in that, The knowledge graph module includes: The graph storage unit uses a graph database to store psychiatric clinical pathway data and constructs a symptom-diagnosis-treatment association network. The network nodes include medical concept codes and clinical weight parameters. The graph query unit receives the analysis results, including clinical entities and semantic relationships, output by the semantic analysis module, extracts entity sets from the analysis results, retrieves relevant subgraphs through graph pattern matching, and transmits the retrieved subgraphs to the graph reasoning unit. The graph reasoning unit receives the relevant subgraphs transmitted by the graph query unit, performs path reasoning using a random walk algorithm, calculates the probability of diagnostic hypotheses based on a path sorting algorithm, generates a candidate diagnosis list, and transmits it to the template adaptation module. The dynamic update unit receives the analysis results from the semantic analysis module and the field filling results from the template adaptation module. Based on the analysis results and field filling results, it dynamically adjusts the node weights and edge strengths, discovers potential association patterns through meta-path similarity calculation, and updates the association network in the graph storage unit.

5. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 4, characterized in that, The template adaptation module includes: The template matching unit receives the analysis results from the semantic analysis module and the diagnostic hypotheses from the knowledge graph module, calculates the similarity score between the analysis results and the medical record template fields, selects the optimal template based on clinical priority rules, and transmits the selected template information to the field filling unit. The field filling unit receives template information transmitted by the template matching unit, performs data type conversion and format standardization operations, performs weighted fusion processing on conflicting content, uses the beam search algorithm for multi-sequence optimization, and generates field filling results. The integrity verification unit receives the field filling results generated by the field filling unit, checks the coverage of required fields, generates a missing field report, and sends the verified medical record documents to the quality control module. The reinforcement learning unit receives the verification results from the quality control module, uses the quality of medical record generation as the reward function, and dynamically adjusts the template selection strategy and field mapping rules through the near-end strategy optimization algorithm. The adjusted strategy and rules are then fed back to the template matching unit and the field filling unit.

6. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 5, characterized in that, The quality control module includes: The rule verification unit receives the preliminary medical record document sent by the template adaptation module, performs syntax standardization checks, medical terminology accuracy verification, and numerical range logic constraint checks, and outputs the first-level verification results. The anomaly detection unit receives the first-level verification result output by the rule verification unit, uses the isolated forest algorithm to identify outliers, combines the adversarial generative network to identify anomaly patterns, marks anomalous data items, and outputs the second-level verification result. The clinical validation unit receives the second-level validation results output by the anomaly detection unit, compares the generated content with the medical guidelines, performs clinical rationality validation through the expert knowledge base, calculates the clinical rationality score, and outputs the third-level validation results. The feedback control unit receives the third-level verification results output by the clinical validation unit, assigns correction strategies according to the error type, and feeds back the verification results, including the error type and quality score, to the speech processing module, semantic analysis module, knowledge graph module, and template adaptation module. The feedback control unit records error tracking information and performance monitoring data.

7. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 6, characterized in that, Also includes: The speech processing module and the semantic analysis module establish a two-way feedback channel: The speech recognition unit transmits the timestamp and confidence score attached to the recognition result to the text preprocessing unit of the semantic analysis module; The entity recognition unit feeds back the entity recognition results to the speech recognition unit in real time, and the speech recognition unit adjusts the weights of the domain language model based on the entity recognition results. When the quality monitoring unit detects a low-confidence text segment, it sends a reprocessing request to the relation extraction unit of the semantic analysis module. The relation extraction unit returns context information to the quality monitoring unit, which then generates an audio re-acquisition decision instruction based on the context information.

8. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 7, characterized in that, Also includes: The semantic analysis module and the knowledge graph module establish an iterative optimization mechanism: The entity recognition unit sends an entity query request to the atlas query unit, and the atlas query unit retrieves relevant clinical pathway constraints from the atlas storage unit. The graph reasoning unit returns the set of diagnostic hypotheses to the entity recognition unit, which then corrects the entity recognition bias based on the set of diagnostic hypotheses. The relation extraction unit outputs the corrected entity relations to the dynamic update unit, which then updates the node association strength and edge weight parameters based on the corrected entity relations.

9. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 8, characterized in that, Also includes: The knowledge graph module and the template adaptation module establish a constraint propagation path: The graph reasoning unit transmits the clinical pathway constraints of symptoms, diagnosis, and treatment to the template matching unit, which uses the clinical pathway constraints as the priority for template selection. The field population unit feeds back the field population results to the dynamic update unit, which uses the field population results to verify the accuracy of the diagnostic assumptions. The dynamic update unit fills in the consistency score based on the fields and dynamically adjusts the node weights and association strengths in the graph storage unit.

10. The psychiatric electronic medical record generation device based on AI speech recognition as described in claim 9, characterized in that, Also includes: The template adaptation module and the quality control module form a closed-loop verification circuit: The integrity verification unit sends the verified medical record documents to the rule verification unit of the quality control module; The feedback control unit generates a verification report that includes error localization and quality scoring, and sends the verification report to the reinforcement learning unit; The reinforcement learning unit adjusts the template selection strategy and field mapping rules based on the verification report, and feeds back the adjusted strategy and rules to the template matching unit and field filling unit; The feedback control unit collects cross-institutional optimization data and updates the model parameters of the context modeling unit and entity recognition unit of the semantic analysis module through a federated learning strategy.

Images