AI-based schizophrenia clinical thinking training simulation system

Abstract

The invention discloses an AI-based schizophrenia clinical thinking training simulation system, which relates to the technical field of artificial intelligence, and is characterized by comprising the steps of constructing a dynamic knowledge base, receiving and preprocessing multi-modal input, performing standardized mapping by extracting keywords, and then performing judgment of a rule layer and a semantic layer. A student question method and a standard answer method in the dynamic knowledge base are sent into a model for response generation, then four-dimensional quantitative scoring and clinical operation multi-dimensional scoring are carried out, error codes and positioning sentences thereof are output, and finally personalized training courses are recommended to students; and hot-updating the dynamic knowledge base continuous optimization model. According to the method, four dimensions of the ICD-10 diagnostic standard are converted into a quantitative scoring algorithm capable of being automatically executed, quantitative assessment of clinical operation specifications is integrated, a diagnosis and operation double-track evaluation system is constructed, the problems that a traditional method is lack of evaluation standards and high in subjectivity are solved, and the evaluation efficiency is improved. Therefore, the evaluation of the clinical thinking ability of the medical students becomes objective, comprehensive and evidence-based.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an AI-based clinical thinking training simulation system for schizophrenia. Background Technology

[0002] Schizophrenia is a complex and variable mental illness, and its clinical diagnosis and treatment place extremely high demands on physicians' professional skills and clinical reasoning abilities. However, there are currently many unresolved issues in the field of clinical reasoning training for schizophrenia.

[0003] On the one hand, traditional teaching models have significant limitations in training clinical thinking skills for schizophrenia. They mainly rely on classroom lectures, case discussions, and limited opportunities for clinical practice. While classroom lectures and case discussions can impart theoretical knowledge, they lack the interactivity and practicality of real clinical scenarios. Clinical practice opportunities are not only limited in number but also constrained by many factors such as the patient's condition, time arrangements, and medical resources. This makes it difficult for medical students to obtain sufficient practical training, and their clinical thinking is prone to being one-sided, making it difficult to meet the actual diagnostic and treatment needs.

[0004] On the other hand, the application of existing intelligent robot technology in the field of schizophrenia is not perfect. The current dialogue robot technology developed for schizophrenia is mostly focused on rehabilitation training and other fields. Its application in clinical thinking training for schizophrenia is still immature, with problems such as narrow coverage and insufficient simulation depth, making it difficult to effectively improve the clinical diagnostic ability of medical students.

[0005] Furthermore, the clinical research and diagnosis of schizophrenia are unique. Compared to other diseases, objective biological indicators are scarce, and diagnosis mainly relies on the patient's clinical manifestations and medical history. Its characteristic symptoms include positive symptoms (such as hallucinations and delusions) and negative symptoms (such as emotional blunting and diminished volition). These symptoms manifest in a complex and diverse manner and are easily affected by various factors such as individual patient differences and environmental factors. This requires doctors to have keen observation skills, rich clinical experience, and systematic clinical thinking ability to accurately identify and diagnose the disease.

[0006] Furthermore, my country's intelligent robot technology started relatively late, and its application and promotion in the medical field have lagged behind. This has also limited the application and development of intelligent robot technology in clinical thinking training for schizophrenia to some extent. At present, there is an urgent need for a training tool that can effectively improve medical students' clinical thinking ability for schizophrenia, so as to overcome the limitations of traditional training methods, meet the complex needs of clinical diagnosis of schizophrenia, and promote the application and development of intelligent robot technology in the field of clinical thinking training for schizophrenia in my country.

[0007] Therefore, the present invention aims to provide an AI-based clinical thinking training simulation system for schizophrenia to solve the above-mentioned problems. Summary of the Invention

[0008] The purpose of this invention is to provide an AI-based clinical thinking training simulation system for schizophrenia. This invention transforms the four dimensions of the ICD-10 diagnostic criteria into an automatically executable quantitative scoring algorithm, while incorporating quantitative assessments of clinical operation procedures. This constructs a dual-track evaluation system of diagnosis and operation, solving the problems of lack of evaluation standards and strong subjectivity in traditional methods. This makes the evaluation of medical students' clinical thinking ability more objective, comprehensive, and evidence-based.

[0009] The above-mentioned technical objective of the present invention is achieved through the following technical solution: an AI-based clinical thinking training simulation system for schizophrenia, comprising the following steps:

[0010] S1. Construct a dynamic knowledge base: Input the ICD-10 diagnostic criteria for schizophrenia, clinical operation guidelines, real desensitized medical records, and standardized patient scripts into a dynamic knowledge base with a graph-vector-inverted triple index structure;

[0011] S2. Multimodal input reception and preprocessing: Receives keyboard, voice, or touch input, performs medical synonym normalization, sensitive word filtering, and pinyin error correction, and generates standardized text;

[0012] S3. Keyword extraction and standardized mapping: Using a three-level strategy of AC automaton, edit distance and SBERT vector cosine similarity, symptom keywords are extracted from the standardized text and mapped to ICD-10 standardized symptom vectors Vkey;

[0013] S4. Two-layer matching judgment: First, perform rule layer matching, then perform semantic layer vector cosine matching, and trigger unknown problem clustering for missed problems;

[0014] S5. Deep Learning Response Generation: The student's question and the standard answer from the dynamic knowledge base are fed into the domain-trained and distilled Chinese-TinyBERT model to generate a simulated patient response and return it.

[0015] S6. Four-dimensional quantitative scoring: Based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions, the scores of the four items of symptoms, course of disease, severity, and exclusion are automatically calculated.

[0016] S7. Multi-dimensional scoring of clinical operation: quantitative scoring of the coverage of the seven domains of mental health examination, the frequency of humanistic care and empathy, the keyword recognition rate of risk assessment, and the proportion of open-ended questions.

[0017] S8. Defect localization and error coding: Map dimensions below the threshold and key information that has not been followed up to a preset error coding library, and output the error code and its localization sentence;

[0018] S9. Targeted training course recommendation: Based on the error code-course mapping matrix and historical improvement rate, recommend personalized training courses to students;

[0019] S10, Incremental Learning and Knowledge Base Update: After clustering unknown problems and undergoing double-blind manual review, the data is hot-updated to the dynamic knowledge base to achieve continuous model optimization.

[0020] The present invention is further configured such that: the keyword extraction and standardization mapping step includes: constructing a four-level tree of symptoms-keywords-synonyms-scores, using precise, fuzzy, and semantic three-level recall, and standardizing it through a medical synonym mapping table, outputting a 768-dimensional symptom vector Vkey, whose elements take scores of 0, 1, 2, and 5 to represent four intensities: not mentioned, possible, clear, and typical.

[0021] The present invention is further configured such that the two-layer matching judgment step includes: the rule layer directly checks whether the keyword "suicide" is contained and no follow-up question is asked; if so, the error code E305 is immediately marked; the semantic layer calculates the cosine of the mean of Vkey and the question library vector; if it is less than 0.78, it is marked as an unknown question and sent to HDBSCAN clustering.

[0022] The present invention is further configured such that: the deep learning response generation step includes: using Chinese-BERT-wwm-ext trained after continuing the MLM domain as the teacher model, distilling to obtain a 6-layer Chinese-TinyBERT student model, jointly training the matching task and the generation task, the loss function being a weighted sum of BCELoss and CrossEntropy, and using Redis caching and Triton batch inference in the inference stage to make the single sentence latency ≤120 m.

[0023] The present invention is further configured such that: the ICD-10 four-dimensional quantitative scoring step includes: symptom criteria are calculated by summing the elements of the positive / negative symptom vector and multiplying by 5 points; the course of the disease criteria are extracted by regular expression to extract the number of months in the answer, and ≥1 month is given 20 points; the severity is given 10 points each by keywords of unemployment and lack of insight; the exclusion criteria are given 10 points each by keywords of drug abuse and alcohol abuse, with a total score of 100 points for the four items.

[0024] The present invention is further configured such that: the multi-dimensional scoring of clinical operation includes: the completeness of mental examination is calculated as 3 points based on the number of hits of keywords in 7 major domains, with a maximum of 15 points; the humanistic care index is accumulated based on the number of hits in the preset empathy dictionary, with a maximum of 10 points; the risk assessment ability is given 0-15 points based on the depth of follow-up questions in the suicide / harm dictionary; and the communication skills are calculated by linear interpolation based on the proportion of open-ended questions, with a proportion >60% earning 10 points.

[0025] This invention also provides an AI-based clinical thinking training simulation system for schizophrenia, including a dynamic knowledge base module, a multimodal input preprocessing module, a keyword processing module, a two-layer matching judgment module, a deep learning response generation module, a quantitative scoring module, a defect analysis module, a targeted recommendation module, and an incremental learning module;

[0026] The dynamic knowledge base module is used to store the ICD-10 diagnostic criteria for schizophrenia, clinical operation guidelines, real desensitized medical records, and standardized patient scripts, and uses a graph-vector-inverted triple index structure for data organization and management.

[0027] The multimodal input preprocessing module is used to receive multimodal input from keyboard, voice or touch, and generate standardized text data through medical synonym normalization, sensitive word filtering and pinyin error correction, and transmit it to the keyword processing module.

[0028] The keyword processing module receives standardized text data transmitted by the multimodal input preprocessing module and uses a three-level strategy of AC automaton, edit distance and SBERT vector cosine similarity to extract symptom keywords from the standardized text and map them into ICD-10 standardized symptom vectors Vkey.

[0029] The dual-layer matching judgment module is communicatively connected to the keyword processing module and the dynamic knowledge base module, respectively. It is used to first perform rule-layer matching, then perform semantic-layer vector cosine matching, and trigger unknown question clustering for questions that are not matched.

[0030] The deep learning response generation module is communicatively connected to the two-layer matching judgment module, and is used to input the student's question and the standard answer in the dynamic knowledge base into the Chinese-TinyBERT model that has been trained and distilled by the domain, generate the patient's simulated answer and return it;

[0031] The quantitative scoring module is communicatively connected to the deep learning response generation module and includes a four-dimensional scoring submodule and a clinical operation scoring submodule. The four-dimensional scoring submodule is used to automatically calculate the scores for symptoms, course of disease, severity, and exclusion based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions. The clinical operation scoring submodule is used to quantitatively score the coverage of the seven domains of mental health examination, the frequency of empathy in humanistic care, the recognition rate of risk assessment keywords, and the proportion of open-ended questions.

[0032] The defect analysis module is communicatively connected to the quantitative scoring module and is used to map scoring dimensions below the threshold and key information that has not been followed up to a preset error coding library, and output error codes and their location sentences.

[0033] The targeted recommendation module is communicatively connected to the defect analysis module and is used to recommend personalized training courses to students based on the error code-course mapping matrix and historical improvement rate.

[0034] The incremental learning module is communicatively connected to the two-layer matching judgment module and the dynamic knowledge base module, respectively. It is used to hot update the clustering results of unknown problems to the dynamic knowledge base module after double-blind manual review, so as to realize continuous system optimization.

[0035] The present invention also provides an AI-based clinical thinking training simulation device for schizophrenia, comprising at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement an AI-based clinical thinking training simulation method for schizophrenia.

[0036] The present invention also provides a computer-readable storage medium storing computer instructions for execution by a computer to implement an AI-based clinical thinking training simulation method for schizophrenia.

[0037] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements an AI-based clinical thinking training simulation method for schizophrenia.

[0038] In summary, the present invention has the following beneficial effects:

[0039] 1. This invention achieves standardized and quantitative assessment of clinical reasoning training, greatly improving the objectivity and accuracy of the assessment. Specifically, it innovatively transforms the four dimensions of the ICD-10 diagnostic criteria (symptoms, course of disease, severity, and exclusion criteria) into an automatically executable quantitative scoring algorithm. Furthermore, it innovatively incorporates quantitative assessments of clinical operational procedures such as humanistic care and risk assessment, constructing a dual-track assessment system of "diagnosis" and "operation." This solves the problems of missing assessment standards and strong subjectivity in traditional methods, making the evaluation of medical students' clinical reasoning abilities objective, comprehensive, and evidence-based. Actual testing shows that this invention can increase the diagnostic accuracy rate to over 93% and the defect localization accuracy rate to 91%.

[0040] 2. This invention provides a precise and operable feedback mechanism, effectively improving training efficiency and teaching quality. Through a pre-set error coding library, it can automatically and accurately locate the specific types of deficiencies in trainees' clinical reasoning processes, rather than just providing general evaluations. The assessment report generated by the system clearly identifies the skill gaps and maps them to targeted training courses. This closed-loop mechanism of "assessment-location-suggestion" completely changes the situation of vague feedback and blind improvement in traditional training, enabling trainees to conduct targeted reinforcement training, significantly shortening the time to master core skills (tests show a 56% reduction) and reducing the clinical error rate (tests show a 72% decrease).

[0041] 3. This invention constructs a highly realistic and continuously optimized intelligent training environment, reducing teaching costs and risks. By utilizing deep learning models to generate patient responses and combining reinforcement learning with human feedback loops for continuous optimization, the simulated patient's dialogue behavior more closely resembles the complexity and dynamism of real clinical scenarios. Simultaneously, the system possesses self-learning capabilities, continuously expanding its knowledge base through unknown question clustering and human annotation, maintaining the timeliness of the content. This provides trainees with a risk-free, high-standard, and repeatable "standardized patient" for 24 / 7 training, significantly reducing reliance on scarce clinical teaching resources (real patients, experienced physicians) and addressing a core pain point in psychiatric education.

[0042] 4. This invention promotes the digitalization and intelligentization of teaching management, providing data support for individualized instruction. The system can automatically record and statistically analyze the number of outpatient visits, number of dialogue rounds, diagnostic accuracy, and multi-dimensional ability scores of all students, and generate a visual ability radar chart. This enables teachers and teaching administrators to accurately grasp the teaching effectiveness and students' weaknesses from both group and individual levels, thereby developing more scientific and personalized teaching plans, realizing the transformation from experience-based teaching to data-driven teaching, and comprehensively improving the quality and level of psychiatric medical education. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the AI-based clinical thinking training simulation method for schizophrenia in Embodiment 1 of the present invention.

[0044] Figure 2 This is a module structure diagram of the AI-based clinical thinking training simulation system for schizophrenia in Embodiment 2 of the present invention. Detailed Implementation

[0045] The following is in conjunction with the appendix Figures 1-2 The present invention will be described in further detail below.

[0046] Example 1: AI-based Clinical Thinking Training Simulation Method for Schizophrenia

[0047] This embodiment is based on human-computer dialogue robot technology. It uses an intelligent question-and-answer robot to simulate a schizophrenic patient and engage in dialogue with the user. The user needs to take the role of a clinical doctor to conduct a consultation with the schizophrenic patient simulated by the intelligent question-and-answer robot to obtain effective information. After analysis, a diagnosis is made. The system can record data such as the number of consultations and accuracy rate of the user.

[0048] After a user's question is received by the robot, it will search a mental illness type knowledge base configured in the background. Questions with matching keywords will return answers from the knowledge base. For unknown questions, automatic clustering will match similar questions and answers, assisting human staff in continuously expanding the knowledge base. The knowledge base includes a mental illness knowledge base and a public knowledge base. The knowledge base's diagnostic description of schizophrenia covers core symptoms and includes main question types and a large number of similar questions, along with diverse matching answers, for negative and positive symptoms within the core symptoms. The question-answering robot is also equipped with deep learning capabilities to help train and optimize the question-answering model, improving the quality of question answers. Specifically, this includes the following steps:

[0049] S1. Construct a dynamic knowledge base: Input the ICD-10 diagnostic criteria for schizophrenia, clinical operation guidelines, real desensitized medical records, and standardized patient scripts into a dynamic knowledge base with a graph-vector-inverted triple index structure.

[0050] S2. Multimodal input reception and preprocessing: Receives keyboard, voice, or touch input, performs medical synonym normalization, sensitive word filtering, and pinyin error correction, and generates standardized text.

[0051] S3. Keyword Extraction and Standardization Mapping: Using a three-level strategy of AC automata, edit distance, and SBERT vector cosine similarity, symptom keywords are extracted from the standardized text and mapped to ICD-10 standardized symptom vectors Vkey. The ICD-10 standardized evaluation matrix is ​​shown in Table 1.

[0052] Table 1 ICD-10 Standardization Evaluation Matrix

[0053]

[0054] Keyword extraction includes: constructing a four-level tree of "symptom-keyword-synonym-score", using precise, fuzzy and semantic three-level recall, and standardizing it through a medical synonym mapping table to output a 768-dimensional symptom vector Vkey, whose elements take scores of 0, 1, 2 and 5 to represent four intensities: not mentioned, possible, clear and typical.

[0055] S4. Two-layer matching judgment: First, perform rule-layer matching, then perform semantic-layer vector cosine matching, triggering unknown question clustering for questions that do not match; the two-layer matching judgment includes: the rule layer directly checks whether there is a suicide keyword and no follow-up question plan, if so, immediately mark it with error code E305; the semantic layer calculates the cosine of the mean of Vkey and the question library vector, if it is lower than 0.78, it is marked as an unknown question and sent to HDBSCAN clustering.

[0056] S5. Deep Learning Response Generation: The student's question and the standard answer from the dynamic knowledge base are fed into the domain-trained and distilled Chinese-TinyBERT model to generate a simulated patient response and return it. The deep learning response generation includes: using the Chinese-BERT-wwm-ext model trained after the MLM domain as the teacher model, distilling to obtain a 6-layer Chinese-TinyBERT student model, jointly training the matching task and the generation task, using the weighted sum of BCELoss and CrossEntropy loss function, and using Redis caching and Triton batch inference in the inference stage to make the single-sentence latency ≤120 ms.

[0057] S6. Four-dimensional quantitative scoring: Based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions, the scores for symptoms, course of illness, severity, and exclusion are automatically calculated. The ICD-10 four-dimensional quantitative scoring includes: symptom criteria are calculated by summing the positive / negative symptom vector elements and multiplying by 5 points; course of illness criteria are calculated by extracting the number of months in the "how long" response using regular expressions, with ≥1 month receiving 20 points; severity is calculated by keywords related to unemployment and lack of insight, each receiving 10 points; and exclusion criteria are calculated by keywords related to substance abuse and alcohol abuse, each receiving 10 points. The maximum score for all four items is 100 points.

[0058] S7. Multi-dimensional scoring of clinical operation: quantitative scoring of the coverage of the seven domains of mental health examination, the frequency of humanistic care and empathy, the keyword recognition rate of risk assessment, and the proportion of open-ended questions; among which, the multi-dimensional scoring of clinical operation includes: the completeness of mental health examination is calculated as 3 points for the number of keywords in the seven domains hit, with a maximum of 15 points.

[0059] Example of a clinical performance evaluation system program:

[0060] Completeness of mental examination (0-15 points):

[0061] Python

[0062] def check_exam_completeness(dialog):

[0063] # 7 Inspection Dimensions Required by ICD-10

[0064] domains = ["General Performance","Emotional Response","Thought Content","Cognitive Function","Insight"]

[0065] score = 0

[0066] for domain in domains:

[0067] if contains_keywords(dialog, domain_keywords[domain]):

[0068] score += 3 # 3 points per item

[0069] return score

[0070] Humanistic care index is accumulated based on the number of times the preset empathy dictionary is hit, with a maximum of 10 points; risk assessment ability is scored from 0 to 15 points based on the depth of follow-up questions in the suicide / harm dictionary; communication skills are scored by linear interpolation based on the proportion of open-ended questions, with a proportion >60% earning 10 points.

[0071] Humanistic Care Index (0-10 points): Counts the frequency of empathetic statements (such as "I understand your pain");

[0072] Risk assessment ability (0-15 points): Suicidal / harmful tendencies keyword recognition rate

[0073] Communication skills assessment (0-10 points): Percentage of open-ended questions (>60% for full marks).

[0074] S8. Defect Localization and Error Coding: Map dimensions below the threshold and unaddressed key information to a pre-defined error coding library, outputting error codes and their location sentences. Defect localization and error coding include: establishing 18 error codes across three categories: E101 Symptom Omission, E203 Disease Course / Severity, and E305 Risk / Humanistic Care. If a dimension score is less than the threshold and the corresponding element of Vkey is 0, the corresponding error code is triggered and the location sentence position is recorded. The error type coding table is shown in Table 2.

[0075] Table 2 Error Type Coding Table

[0076]

[0077] The following is a program example of the feedback report generation algorithm:

[0078] def generate_report(scores, errors):

[0079] weakness = []

[0080] if scores["symptom"] < 30:

[0081] weakness.append("Insufficient recognition of negative symptoms")

[0082] If "E305" appears in errors:

[0083] weakness.append("Suicide risk assessment missing")

[0084] return {

[0085] "Overall Score": f"{sum(scores.values())} / 100",

[0086] "Weakness"

[0087] Training suggestion: map_weakness_to_course(weakness)

[0088] S9. Targeted Training Course Recommendation: Based on the error code-course mapping matrix and historical improvement rate, personalized training courses are recommended to students. The targeted training course recommendation includes: constructing an 18×24 error-course Boolean matrix, using Top-N collaborative filtering to recommend micro-lessons, simulation exercises and quizzes with a historical improvement rate >30%, N=3, and generating personalized training paths.

[0089] S10, Incremental Learning and Knowledge Base Update: After clustering unknown problems and undergoing double-blind manual review, the data is hot-updated to the dynamic knowledge base to achieve continuous model optimization.

[0090] Incremental learning and knowledge base updates include: clustering unknown problems into minimum clusters of ≥5, and after double-blind consistency verification of ≥0.85, synchronizing them to Neo4j graph nodes, Milvus vector indexes, and Elasticsearch inverted indexes through a zero-downtime hot update mechanism, and periodically re-distilling the model.

[0091] The complete training and evaluation process in this embodiment is as follows:

[0092] 1. Trainee consultation:

[0093] Student: "Did you hear a sound that doesn't exist?"

[0094] System record: Positive symptom (auditory hallucination) successfully identified → Symptom criteria +5 points

[0095] 2. Key omissions:

[0096] The student did not ask, "How long have the symptoms lasted?"

[0097] System check: Missing disease course criteria → Error code E203

[0098] 3. AI patients proactively disclose:

[0099] AI: "Life is meaningless..."

[0100] Trainees did not follow up with suicide risk assessment → Error code E305

[0101] 4. Generate an evaluation report:

[0102] {

[0103] "ICD-10 rating": {

[0104] Symptom criteria: 32 / 40

[0105] "Course duration criteria": "5 / 20",

[0106] Severity: 15 / 20

[0107] Exclusion criteria: 18 / 20

[0108] },

[0109] "Clinical proficiency score": {

[0110] Mental health examination: 12 / 15

[0111] "Humanistic care": "8 / 10"

[0112] Risk Assessment: 3 / 15

[0113] Communication Skills: 7 / 10

[0114] },

[0115] "Major Deficiencies": ["E203 - Unverified Disease Course", "E305 - Lack of Suicide Risk Assessment"],

[0116] Training Recommendations: ["Learning Module M03: Disease Inquiry Techniques", "Exercise P12: Risk Assessment Intensive Training"]

[0117] }

[0118] This embodiment equips the intelligent question-answering robot with a relatively detailed knowledge base and public knowledge base for schizophrenia, covering questions and answers related to clinical thinking training for schizophrenia. After extensive testing by several users over a period of six months, the direct hit rate for questions simulating schizophrenia patients reached 90%. The backend can also collect and analyze the question-answering records of all users, including the number of consultations, number of dialogue rounds, and diagnostic accuracy, effectively achieving more comprehensive training of users' clinical thinking. The accuracy of its assessment is shown in Table 3.

[0119] Table 3. Accuracy Assessment Table (Data from a test conducted on 120 students at a medical school)

[0120]

[0121] This embodiment specifically transforms the ICD-10 diagnostic criteria into a quantifiable assessment algorithm. By establishing a multi-dimensional scoring system for clinical operation standards and developing a closed-loop mechanism of "defect localization-targeted training," it ultimately solves three major pain points in psychiatric clinical teaching: subjective assessment, inefficient training, and blind improvement. It has now been applied to the National Standardized Training Base for Psychiatrists.

[0122] Example 2: An AI-based clinical thinking training simulation system for schizophrenia, including a dynamic knowledge base module, a multimodal input preprocessing module, a keyword processing module, a two-layer matching judgment module, a deep learning response generation module, a quantitative scoring module, a defect analysis module, a targeted recommendation module, and an incremental learning module.

[0123] The dynamic knowledge base module stores the ICD-10 diagnostic criteria for schizophrenia, clinical operating procedures, real desensitized medical records, and standardized patient scripts, and uses a graph-vector-inverted triple index structure for data organization and management. The multimodal input preprocessing module receives keyboard, voice, or touch input, and generates standardized text data through medical synonym normalization, sensitive word filtering, and pinyin error correction, which is then transmitted to the keyword processing module. The keyword processing module receives the standardized text data from the multimodal input preprocessing module and uses a three-level strategy—AC automata, edit distance, and SBERT vector cosine similarity—to extract symptom keywords from the standardized text and map them to ICD-10 standardized symptom vectors Vkey. The two-layer matching judgment module communicates with both the keyword processing module and the dynamic knowledge base module, performing rule-level matching followed by semantic-level vector cosine matching, triggering unknown question clustering for missed questions. The deep learning response generation module communicates with the two-layer matching judgment module, inputting student questions and standard answers from the dynamic knowledge base into a domain-trained and distilled C... The system employs the hinese-TinyBERT model to generate simulated patient responses and return them. A quantitative scoring module, connected to the deep learning response generation module, includes a four-dimensional scoring submodule and a clinical operation scoring submodule. The four-dimensional scoring submodule automatically calculates scores for symptoms, course of illness, severity, and exclusion based on the symptom vector Vkey, dialogue timeline, social function, insight, and exclusion questions. The clinical operation scoring submodule quantifies and scores the coverage of the seven domains of mental health examination, the frequency of empathy in humanistic care, the recognition rate of risk assessment keywords, and the proportion of open-ended questions. A defect analysis module, connected to the quantitative scoring module, maps scoring dimensions below a threshold and key information not followed up to a pre-set error coding library, outputting error codes and their location sentences. A targeted recommendation module, also connected to the defect analysis module, recommends personalized training courses to students based on the error code-course mapping matrix and historical improvement rate. An incremental learning module, connected to both the two-layer matching judgment module and the dynamic knowledge base module, hot-updates the clustering results of unknown questions to the dynamic knowledge base module after double-blind manual review, enabling continuous system optimization.

[0124] Example 3: An AI-based clinical thinking training simulation device for schizophrenia, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement an AI-based clinical thinking training simulation method for schizophrenia, the method comprising: a dynamic knowledge base construction step: inputting the ICD-10 diagnostic criteria for schizophrenia, clinical operation specifications, real desensitized medical records, and standardized patient scripts into a dynamic knowledge base with a graph-vector-inverted triple index structure; a multimodal input reception and preprocessing step: receiving keyboard, voice, or touch input, performing medical synonym normalization, sensitive word filtering, and pinyin error correction to generate standardized text; a keyword extraction and standardized mapping step: using a three-level strategy of AC automata, edit distance, and SBERT vector cosine similarity to extract symptom keywords from the standardized text and map them to ICD-10 standardized symptom vectors Vkey; a two-layer matching judgment step: first performing rule-layer matching, then performing semantic-layer vector cosine similarity. The process includes: matching, triggering clustering of unknown questions for missed questions; deep learning response generation step: feeding student questions and standard answers from the dynamic knowledge base into the domain-trained and distilled Chinese-TinyBERT model to generate simulated patient answers and return them; four-dimensional quantitative scoring step: automatically calculating scores for symptoms, course of disease, severity, and exclusion based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions; multi-dimensional clinical operation scoring step: quantifying and scoring the coverage of the seven domains of mental health examination, the frequency of humanistic care and empathy, the keyword recognition rate of risk assessment, and the proportion of open-ended questions; defect localization and error coding step: mapping dimensions below the threshold and key information not followed up to a pre-set error coding library, outputting error codes and their localization sentences; targeted training course recommendation step: recommending personalized training courses to students based on the error code-course mapping matrix and historical improvement rate; incremental learning and knowledge base update step: clustering unknown questions and, after double-blind manual review, hot-updating to the dynamic knowledge base to achieve continuous model optimization.

[0125] Example 4: A computer-readable storage medium storing computer instructions for execution by a computer to implement an AI-based clinical thinking training simulation method for schizophrenia. This method includes: a dynamic knowledge base construction step: inputting the ICD-10 diagnostic criteria for schizophrenia, clinical operating procedures, real desensitized medical records, and standardized patient scripts into a dynamic knowledge base with a graph-vector-inverted triple index structure; a multimodal input reception and preprocessing step: receiving keyboard, voice, or touch input, performing medical synonym normalization, sensitive word filtering, and pinyin correction to generate standardized text; a keyword extraction and standardized mapping step: using a three-level strategy of AC automata, edit distance, and SBERT vector cosine similarity to extract symptom keywords from the standardized text and map them to ICD-10 standardized symptom vectors Vkey; a two-layer matching judgment step: first performing rule-layer matching, then performing semantic-layer vector cosine matching, triggering unknown question clustering for missed questions; deep... The learning response generation steps are as follows: Student questions and standard answers from the dynamic knowledge base are fed into the domain-trained and distilled Chinese-TinyBERT model to generate simulated patient responses and return them; Four-dimensional quantitative scoring steps: Based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions, scores for symptoms, course of illness, severity, and exclusion are automatically calculated; Multi-dimensional clinical operation scoring steps: Quantitative scoring is performed on the coverage of the seven domains of mental health examination, the frequency of empathy in humanistic care, the keyword recognition rate of risk assessment, and the proportion of open-ended questions; Defect localization and error coding steps: Dimensions below the threshold and key information not followed up are mapped to a pre-set error coding library, and error codes and their localization sentences are output; Targeted training course recommendation steps: Based on the error code-course mapping matrix and historical improvement rate, personalized training courses are recommended to students; Incremental learning and knowledge base update steps: Unknown questions are clustered and, after double-blind manual review, are hot-updated to the dynamic knowledge base to achieve continuous model optimization.

[0126] Example 5: A computer program product, including a computer program, which, when executed by a processor, implements an AI-based clinical thinking training simulation method for schizophrenia. This method includes: a dynamic knowledge base construction step: inputting the ICD-10 diagnostic criteria for schizophrenia, clinical operating procedures, real desensitized medical records, and standardized patient scripts into a dynamic knowledge base with a graph-vector-inverted triple index structure; a multimodal input reception and preprocessing step: receiving keyboard, voice, or touch input, performing medical synonym normalization, sensitive word filtering, and pinyin error correction to generate standardized text; a keyword extraction and standardized mapping step: using a three-level strategy of AC automata, edit distance, and SBERT vector cosine similarity to extract symptom keywords from the standardized text and map them to ICD-10 standardized symptom vectors Vkey; a two-layer matching judgment step: first performing rule-layer matching, then performing semantic-layer vector cosine matching, triggering unknown question clustering for missed questions; and deep learning response generation. Steps: 1. Feed the student's questions and standard answers from the dynamic knowledge base into the domain-trained and distilled Chinese-TinyBERT model to generate simulated patient responses and return them; 2. Four-dimensional quantitative scoring: Automatically calculate scores for symptoms, course of illness, severity, and exclusion based on the symptom vector Vkey, dialogue timeline, social function, insight, and exclusion questions; 3. Multi-dimensional clinical operation scoring: Quantify and score the coverage of the seven domains of mental health examination, frequency of empathy in humanistic care, keyword recognition rate of risk assessment, and proportion of open-ended questions; 4. Deficiency localization and error coding: Map dimensions below the threshold and key information not followed up to a pre-set error coding library, outputting error codes and their localization sentences; 5. Targeted training course recommendation: Recommend personalized training courses to students based on the error code-course mapping matrix and historical improvement rate; 6. Incremental learning and knowledge base update: Cluster unknown questions and, after double-blind manual review, hot-update them to the dynamic knowledge base to achieve continuous model optimization.

[0127] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. An AI-based clinical thinking training simulation method for schizophrenia, characterized by: Includes the following steps: S1. Construct a dynamic knowledge base: Input the ICD-10 diagnostic criteria for schizophrenia, clinical operation guidelines, real desensitized medical records, and standardized patient scripts into a dynamic knowledge base with a graph-vector-inverted triple index structure; S2. Multimodal input reception and preprocessing: Receives keyboard, voice, or touch input, performs medical synonym normalization, sensitive word filtering, and pinyin error correction, and generates standardized text; S3. Keyword extraction and standardized mapping: Using a three-level strategy of AC automaton, edit distance and SBERT vector cosine similarity, symptom keywords are extracted from the standardized text and mapped to ICD-10 standardized symptom vectors Vkey; S4. Two-layer matching judgment: First, perform rule layer matching, then perform semantic layer vector cosine matching, and trigger unknown problem clustering for missed problems; S5. Deep Learning Response Generation: The student's question and the standard answer from the dynamic knowledge base are fed into the domain-trained and distilled Chinese-TinyBERT model to generate a simulated patient response and return it. S6. Four-dimensional quantitative scoring: Based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions, the scores of the four items of symptoms, course of disease, severity, and exclusion are automatically calculated. S7. Multi-dimensional scoring of clinical operation: quantitative scoring of the coverage of the seven domains of mental health examination, the frequency of humanistic care and empathy, the keyword recognition rate of risk assessment, and the proportion of open-ended questions. S8. Defect localization and error coding: Map dimensions below the threshold and key information that has not been followed up to a preset error coding library, and output the error code and its localization sentence; S9. Targeted training course recommendation: Based on the error code-course mapping matrix and historical improvement rate, recommend personalized training courses to students; S10, Incremental Learning and Knowledge Base Update: After clustering unknown problems and undergoing double-blind manual review, the data is hot-updated to the dynamic knowledge base to achieve continuous model optimization.

2. The AI-based clinical thinking training simulation method for schizophrenia according to claim 1, characterized in that: The keyword extraction and standardization mapping steps include: constructing a four-level tree of symptoms-keywords-synonyms-scores, using precise, fuzzy, and semantic three-level recall, and standardizing it through a medical synonym mapping table to output a 768-dimensional symptom vector Vkey, whose elements take scores of 0, 1, 2, and 5 to represent four intensities: not mentioned, possible, clear, and typical.

3. The AI-based clinical thinking training simulation method for schizophrenia according to claim 1, characterized in that: The two-layer matching judgment step includes: the rule layer directly checks whether the keyword "suicide" is contained and whether there is no follow-up question plan. If so, it immediately marks the error code E305; the semantic layer calculates the cosine of the mean of Vkey and the question library vector. If it is lower than 0.78, it is marked as an unknown question and sent to HDBSCAN clustering.

4. The AI-based clinical thinking training simulation method for schizophrenia according to claim 1, characterized in that: The deep learning response generation step includes: using Chinese-BERT-wwm-ext, which is trained after continuing the MLM domain, as the teacher model; distilling to obtain a 6-layer Chinese-TinyBERT student model; jointly training the matching task and the generation task; using the loss function as a weighted sum of BCELoss and CrossEntropy; and using Redis caching and Triton batch inference in the inference stage to make the single sentence latency ≤120 ms.

5. The AI-based clinical thinking training simulation method for schizophrenia according to claim 1, characterized in that: The four-dimensional quantitative scoring steps include: symptom criteria are calculated by summing the elements of the positive / negative symptom vector and multiplying by 5 points; the course of the disease is determined by extracting the number of months in the response using regular expressions, with ≥1 month receiving 20 points; the severity is determined by keywords related to unemployment and lack of insight, each receiving 10 points; and the exclusion criteria are determined by keywords related to drug abuse and alcohol abuse, each receiving 10 points. The total score for all four items is 100 points.

6. The AI-based clinical thinking training simulation method for schizophrenia according to claim 1, characterized in that: The multi-dimensional scoring steps of the clinical operation include: completeness of mental examination is calculated as 3 points based on the number of keywords hit in the 7 major domains, with a maximum of 15 points; humanistic care index is accumulated based on the number of hits in the pre-set empathy dictionary, with a maximum of 10 points; risk assessment ability is scored from 0 to 15 points based on the depth of follow-up questions in the suicide / harm dictionary; communication skills are calculated by linear interpolation based on the proportion of open-ended questions, with a proportion >60% earning 10 points.

7. An AI-based clinical thinking training simulation system for schizophrenia, applied to the AI-based clinical thinking training simulation method for schizophrenia as described in any one of claims 1-6, characterized in that: It includes a dynamic knowledge base module, a multimodal input preprocessing module, a keyword processing module, a two-layer matching and judgment module, a deep learning response generation module, a quantitative scoring module, a defect analysis module, a targeted recommendation module, and an incremental learning module; The dynamic knowledge base module is used to store the ICD-10 diagnostic criteria for schizophrenia, clinical operation guidelines, real desensitized medical records, and standardized patient scripts, and uses a graph-vector-inverted triple index structure for data organization and management. The multimodal input preprocessing module is used to receive multimodal input from keyboard, voice or touch, and generate standardized text data through medical synonym normalization, sensitive word filtering and pinyin error correction, and transmit it to the keyword processing module. The keyword processing module receives standardized text data transmitted by the multimodal input preprocessing module and uses a three-level strategy of AC automaton, edit distance and SBERT vector cosine similarity to extract symptom keywords from the standardized text and map them into ICD-10 standardized symptom vectors Vkey. The dual-layer matching judgment module is communicatively connected to the keyword processing module and the dynamic knowledge base module, respectively. It is used to first perform rule-layer matching, then perform semantic-layer vector cosine matching, and trigger unknown question clustering for questions that are not matched. The deep learning response generation module is communicatively connected to the two-layer matching judgment module, and is used to input the student's question and the standard answer in the dynamic knowledge base into the Chinese-TinyBERT model that has been trained and distilled by the domain, generate the patient's simulated answer and return it; The quantitative scoring module is communicatively connected to the deep learning response generation module and includes a four-dimensional scoring submodule and a clinical operation scoring submodule. The four-dimensional scoring submodule is used to automatically calculate the scores for symptoms, course of disease, severity, and exclusion based on the symptom vector Vkey and the dialogue timeline, social function, insight, and exclusion questions. The clinical operation scoring submodule is used to quantitatively score the coverage of the seven domains of mental health examination, the frequency of empathy in humanistic care, the recognition rate of risk assessment keywords, and the proportion of open-ended questions. The defect analysis module is communicatively connected to the quantitative scoring module and is used to map scoring dimensions below the threshold and key information that has not been followed up to a preset error coding library, and output error codes and their location sentences. The targeted recommendation module is communicatively connected to the defect analysis module and is used to recommend personalized training courses to students based on the error code-course mapping matrix and historical improvement rate. The incremental learning module is communicatively connected to the two-layer matching judgment module and the dynamic knowledge base module, respectively. It is used to hot update the clustering results of unknown problems to the dynamic knowledge base module after double-blind manual review, so as to realize continuous system optimization.

8. An AI-based clinical thinking training simulation device for schizophrenia, characterized in that: It includes at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor to implement the AI-based clinical thinking training simulation method for schizophrenia as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions that are executed by the computer to implement the AI-based clinical thinking training simulation method for schizophrenia as described in any one of claims 1-6.

10. A computer program product, characterized in that: The invention includes a computer program that, when executed by a processor, implements the AI-based clinical thinking training simulation method for schizophrenia as described in any one of claims 1-6.

Images