A respiratory disease clinical decision making method, system, device, and medium

Abstract

The application discloses a respiratory system disease clinical decision method, system, equipment and medium, belongs to the artificial intelligence technical field, aims at solving the problem of poor accuracy and accessibility of respiratory system disease clinical decision, and comprises the following steps: acquiring general medical sample data, electronic medical record sample data and label data; using the general medical sample data to pre-train a diagnosis sub-model and an inference decision sub-model; using the electronic medical record sample data and the label data to fine-tune the diagnosis sub-model and the inference decision sub-model; acquiring to-be-tested medical data and inputting the diagnosis sub-model to obtain a diagnosis result; taking the diagnosis conclusion as a query anchor point, searching in a structured database according to the medical data, and returning medical knowledge context; constructing a final prompt word by using the medical data, the diagnosis conclusion and the medical knowledge context, and inputting the final prompt word, the diagnosis conclusion and the medical knowledge context into the inference decision sub-model together, and outputting an explanation text by the inference decision sub-model.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to clinical decision-making for diseases, particularly a method, system, device, and medium for clinical decision-making in respiratory diseases. Background Technology

[0002] Respiratory diseases are common and frequently occurring illnesses, primarily affecting the trachea, bronchi, lungs, and pleural cavity. Mild cases often present with cough, chest pain, and impaired breathing, while severe cases can lead to respiratory distress, hypoxia, and even respiratory failure and death. The diagnosis and treatment of respiratory diseases involve complex clinical decisions. Traditional clinical decision-making heavily relies on physicians' professional knowledge and experience; however, in areas with uneven distribution of medical resources, primary healthcare institutions often lack experienced respiratory specialists. Furthermore, the diagnosis of respiratory diseases requires the integration of multimodal data, including imaging reports, laboratory tests, and pathological results. The decision-making process involves extensive professional knowledge and complex logical reasoning, placing extremely high demands on physicians' comprehensive abilities. Therefore, the processing of medical texts is particularly important in auxiliary diagnostic tasks.

[0003] There are two main approaches to processing medical text: 1. Approaches based on rule-based or discriminative models (such as BERT). These techniques excel at extracting entities and classifying them (such as determining whether someone has pneumonia), but they cannot generate fluent descriptions of symptoms or treatment suggestions and lack interactivity; 2. Approaches based on large language models (such as LLM). In recent years, large language models have shown potential in auxiliary diagnostic tasks such as medical question answering and medical text processing, thanks to their powerful natural language processing capabilities and knowledge reserves.

[0004] Large Language Models (LLMs) are natural language processing models based on the transformer architecture. They have a large number of parameters and are trained on a vast amount of text, thus closely resembling human language cognition and generation processes. Compared to traditional NLP models, LLMs can better understand and generate natural text, and possess a certain degree of logic and reasoning ability, demonstrating great potential in tasks such as text classification, dialogue, and generation. Designing and training LLMs from scratch is very costly. With the open-sourcing of a number of high-performing pre-trained models, such as LLaMa2, ChatGLM, Bloom, and Ziya, fine-tuning based on open-source models has become a research hotspot. To adapt to various specific scenarios, industry-specific large models are typically created by fine-tuning a base LLM using data from vertical domains.

[0005] Patent application number 202511396056.8 discloses a method, system, and medium for clinical decision support in liver cancer based on a large language model. The method includes: S1, collecting professional corpus data in the field of liver cancer, using incremental pre-training under a federated learning framework combined with efficient parameter fine-tuning technology to inject domain knowledge into a large language model base to obtain a domain-enhanced model; S2, based on real clinical data, using the large language model to extract structured features from the model output by fusing image features, synthesizing a high-quality set of clinical reasoning instructions for liver cancer containing intermediate reasoning basis, and using the high-quality set of clinical reasoning instructions to perform supervised instruction fine-tuning of the domain-enhanced model to obtain an instruction fine-tuning model; S3, constructing positive and negative sample pairs, introducing a real-time doctor feedback mechanism, and using a grouping relative strategy optimization algorithm and Monte Carlo tree search to train the instruction fine-tuning model, aligning the model output with human expert preferences to obtain the final large language model for auxiliary diagnosis of liver cancer; S4, receiving a decision request containing clinical data of the target patient, inputting the decision request into the large language model for auxiliary diagnosis of liver cancer, and obtaining a structured result from the model output containing auxiliary decision suggestions and reasoning basis. By efficiently incorporating medical guidelines, clinical data, and expert experience in the field of liver cancer into the model, it acquires precise diagnostic logic and terminology. The model not only provides diagnostic suggestions but also gradually explains the basis for its judgments, greatly enhancing doctors' trust in and verifiability of the model's output.

[0006] While solutions based on large language models (such as LLM) offer good interactivity, they typically employ a single "end-to-end" model to address all issues, failing to separate diagnostic and interpretative parameters. Consequently, they are prone to factual errors when dealing with complex medical decisions and struggle to perform targeted parameter fine-tuning. Summary of the Invention

[0007] The purpose of this invention is to address the technical problem that existing large language models are prone to factual errors and are difficult to fine-tune, resulting in poor accuracy and accessibility of clinical decision-making for diseases, by providing a clinical decision-making method, system, device, and medium for respiratory diseases.

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution: A clinical decision-making approach for respiratory diseases includes the following steps: Step S1: Obtain sample and label data; Acquire sample data, including general medical sample data and electronic medical record sample data; label the ICD-10 diagnosis, diagnostic basis, treatment plan, and treatment decision in the electronic medical record data to obtain labeled data; Step S2: Pre-training and fine-tuning the diagnostic decision model; The diagnostic decision model includes a diagnostic sub-model Dx and an inference decision sub-model Ex; the diagnostic sub-model Dx and the inference decision sub-model Ex are pre-trained using general medical sample data; the diagnostic sub-model Dx and the inference decision sub-model Ex are fine-tuned using electronic medical record sample data and corresponding label data. Step S3: Real-time clinical decision-making; Acquiring medical data to be tested Input the diagnostic sub-model Dx, and the diagnostic sub-model Dx will output the diagnostic results. ; the diagnosis results As an anchor point for queries, based on medical data Retrieve from structured databases and return medical knowledge context. Using medical data Diagnostic results and medical knowledge context Build the final prompt words ; Final prompt Input the inference decision sub-model Ex, and output the explanatory text. Explain the text It is a structured result that includes suggestions for decision support and reasoning.

[0009] Further, in step S2, the diagnostic sub-model Dx includes an input layer, a decoder layer, an intermediate layer, and an output layer. The input layer is configured with a structured EHR parsing module and word embeddings, the decoder layer is configured with 12 layers of Transformer decoder blocks, the intermediate layer is configured with a clinical feature attention enhancement module, and the output layer is configured with an ICD-10 dedicated classification head and a confidence scoring module. The reasoning decision sub-model Ex includes an input layer, a decoder layer, an intermediate layer, and an output layer. The input layer is configured with a multi-source data fusion module and word embedding. The decoder layer is configured with a 32-layer Transformer decoder block. The intermediate layer is configured with a diagnostic anchoring constraint module and a knowledge fusion module. The output layer is configured with a clinical text standardization module and an autoregressive generation head.

[0010] Furthermore, in step S2, when fine-tuning the diagnostic sub-model Dx and the inference decision sub-model Ex, the total loss... Represented as:

[0011] Fine-tuning loss of the diagnostic sub-model Dx Represented as: ; The inference decision sub-model Ex undergoes supervised fine-tuning of the negative log-likelihood loss in the SFT. Represented as: ; The reasoning decision sub-model Ex performs policy gradient loss in human feedback reinforcement learning (RLHF). Represented as: ; in, , , Each represents the weight of the loss; Indicates the first ICD-10 diagnostic categories for respiratory diseases This indicates the total number of ICD-10 diagnostic categories for respiratory diseases. This represents the category label value in the ICD-10 labeled diagnostic set for respiratory diseases. This represents the set of all trainable parameters for the diagnostic sub-model Dx; Represented as medical data for a given patient and model parameters At that time, the patient was diagnosed with the first Conditional probability of ICD-10-labeled diagnostic categories for each respiratory disease; This represents all trainable parameters of the inference decision sub-model Ex. This represents the maximum sequence length of the text generated by the inference decision submodel Ex. This represents the word vector generated in step t. This represents the historical word sequence generated up to step t. Indicates medical data, Indicates the diagnosis result. Indicates the context of medical knowledge. For hyperparameters; Indicates regular expression processing; This represents the parameters of the reasoning and decision-making sub-model Ex during the reinforcement learning phase. This represents the set of parameters for the reasoning and decision-making sub-model Ex. Indicating in strategy Expected calculation under the following conditions This represents the sequence of clinical explanatory texts generated by the reasoning and decision-making submodel Ex. This represents the rating given by clinicians to the explanatory text generated by the model. Representation Strategy Generate sequence The probability of; Indicates hyperparameters; Denotes KL divergence, This represents the baseline model strategy during the supervised fine-tuning phase; This represents the KL divergence between the policy distribution and the SFT pre-trained distribution.

[0012] Furthermore, in step S3, the medical data to be tested... After inputting the diagnostic submodel Dx, the explanation text continues until the inference decision submodel Ex outputs it. The specific process is as follows: 1. Preliminary diagnosis and anchor point extraction; The diagnostic sub-model Dx processes the input medical data. It outputs diagnostic results including ICD-10 diagnostic code and diagnostic category label. Diagnostic results Represented as: ; 2. Knowledge retrieval based on chained components; Diagnostic results As a query anchor, it performs retrieval in the structured database through a chained interface; retrieval function Return to relevant medical knowledge context ; Medical knowledge context Represented as: ; 3. Dynamic construction and modification of prompt words; Construct the final prompt words as the final input to the inference decision sub-model Ex. Final prompt It is based on medical data Diagnostic results Medical knowledge context The composite tensor formed by these elements is represented as: ; 4. Controlled interpretation generation; The reasoning decision sub-model Ex is based on the final prompt word. Generate the final explanatory text. Explain the text Represented as: ; in, Medical data representing a given patient and model parameters At that time, the patient was diagnosed with a respiratory disease, and the ICD-10 diagnostic category was marked. The conditional probability; This represents the set of all trainable parameters for the diagnostic sub-model Dx; Represents a structured database; This represents the set of all trainable parameters for the inference decision sub-model Ex.

[0013] Furthermore, in step S3, the inference decision sub-model Ex generates the explanatory text. The specific process is as follows: Step E1, Final Prompt Structured construction; Based on medical data Diagnostic results and medical knowledge context A standardized prompt word template is constructed, and the three types of data are integrated into a parsable input text for the reasoning and decision-making sub-model Ex according to a fixed format to obtain the final prompt word. ; Step E2: Hierarchical analysis of prompt words by the reasoning decision sub-model Ex; The reasoning decision sub-model Ex receives the final prompt word. Then, layered parsing and data interaction are completed through internal modules; Step E3: The inference decision sub-model Ex generates explanatory text through autoregression; The reasoning decision sub-model Ex outputs explanatory text based on the autoregressive generation capability of the large model.

[0014] Furthermore, in step E2, the specific steps are as follows: Step E2-1, Anchoring layer analysis; Analysis of the final prompt words Diagnostic results The core diagnostic direction is locked through the diagnostic anchoring constraint module in the intermediate layer, and redundant information unrelated to diagnosis is filtered out. Step E2-2: Knowledge Layer Integration; Medical data and medical knowledge context Feature fusion is performed to screen for suitable treatment recommendations; Step E2-3: Context alignment; Verify medical data Clinical data and diagnostic results Medical knowledge context Logical consistency.

[0015] Furthermore, in step E3, the specific steps are as follows: Step E3-1: The reasoning decision sub-model Ex generates the first sentence of the diagnostic basis; Step E3-2: Based on the previously generated lexical units, combine them with the fused medical data. Diagnostic results Medical knowledge context Generate the next word element; Step E3-3: Repeat the above autoregressive process until a complete diagnostic basis and treatment plan are generated; and the clinical text standardization module verifies the generated content in real time to ensure that it conforms to the clinical document format. Step E3-4: After generation, output the natural language interpretation text. .

[0016] A clinical decision-making system for respiratory diseases, comprising: The sample and label data acquisition module is used to acquire sample data, including general medical sample data and electronic medical record sample data; it also marks the ICD-10 diagnosis, diagnostic basis, treatment plan, and treatment decision in the electronic medical record data to obtain label data. The diagnostic decision model pre-training and fine-tuning module is used to pre-train and fine-tune the diagnostic decision model, which includes a diagnostic sub-model Dx and an inference decision sub-model Ex. The module uses general medical sample data to pre-train the diagnostic sub-model Dx and the inference decision sub-model Ex respectively. The module uses electronic medical record sample data and corresponding label data to fine-tune the diagnostic sub-model Dx and the inference decision sub-model Ex respectively. The real-time clinical decision module is used to acquire the medical data to be tested. Input the diagnostic sub-model Dx, and the diagnostic sub-model Dx will output the diagnostic results. ; the diagnosis results As a query anchor, based on medical data Retrieve from structured databases and return medical knowledge context. Using medical data Diagnostic results and medical knowledge context Build the final prompt words ; Final prompt Input the inference decision sub-model Ex, and output the explanatory text. Explain the text It is a structured result that includes suggestions for decision support and reasoning.

[0017] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0018] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method.

[0019] The beneficial effects of this invention are as follows: 1. In this invention, a dual-channel parameter separation design is adopted, consisting of Dx (diagnosis) and Ex (interpretation). The Dx channel focuses on outputting high-confidence diagnostic results, while the Ex channel focuses on generating professional text. Each channel performs its own function and is precisely adapted to clinical needs. This effectively solves the problem that large language models are prone to factual errors and are difficult to fine-tune, resulting in poor accuracy and accessibility of clinical decision-making for diseases. It is less prone to factual errors when dealing with complex medical decisions and allows for targeted parameter fine-tuning of the two models, Dx (diagnosis) and Ex (interpretation), thereby improving the accuracy and accessibility of clinical decision-making for respiratory diseases.

[0020] 2. In this invention, when the diagnostic results are... When used as a query anchor and retrieved from a structured database via a chained interface, this database contains not only text but also medical image feature vectors (ImageEmbeddings). It can retrieve not only text guides but also CT image features of similar cases, thereby constructing multimodal prompts and inputting them into the inference decision sub-model Ex to improve the accuracy of the explanatory text generated by the inference decision sub-model Ex.

[0021] 3. In this invention, a knowledge-based dynamic prompt word looping mechanism is constructed when constructing prompt words. Standard medical knowledge is retrieved using the diagnostic results as anchor points, and prompt word input is dynamically reconstructed to achieve evidence-based AI generation and ensure the reliability of suggestions.

[0022] 4. The method of the present invention has flexible expansion capabilities. It can improve diagnostic accuracy, expand the multimodal knowledge base, adapt to various fine-tuning strategies, and incorporate thought chain reasoning through a two-way feedback verification mechanism. This makes the system not only conform to human medical professional logic, but also adapt to different clinical scenarios, ultimately achieving synergistic optimization of diagnostic accuracy, interpretive professionalism, and interactive reliability. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a schematic diagram of the pre-training of the diagnostic decision model in this invention; Figure 3 This is a schematic diagram of the fine-tuning of the diagnostic decision model in this invention; Figure 4 This is a schematic diagram of real-time clinical decision-making in this invention; Figure 5 This is a schematic diagram of the specific application interface of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] ICD-10 is an internationally unified classification standard for diseases developed by the World Health Organization (WHO). Its full name is the International Statistical Classification of Diseases and Related Health Problems. SFT, short for Supervised Fine-tuning, refers to the technical process of performing secondary training on a pre-trained large model using high-quality manually labeled data. RLHF, or Reinforcement Learning Based on Human Feedback, is an AI training technique that combines reinforcement learning with human preferences. It optimizes model behavior through human feedback and is widely used in fields such as dialogue systems and text generation. KL divergence, also known as relative entropy, is a core metric in information theory and machine learning used to measure the difference between two probability distributions. It quantifies the information loss caused when using a distribution Q to approximate another true distribution P. HER stands for Electronic Health Record, which is an individual's official health record that can be shared across devices and institutions. It covers medical data such as medical insurance information, allergy history, and disease information, and is mainly used to record patients' health. Qwen-7B is an open-source large language model launched by Alibaba Cloud's Tongyi Qianwen series. It is developed based on the Transformer architecture and uses more than 2 trillion tokens of multilingual data for pre-training. Token embedding is a technique that converts the smallest unit of text segmentation (Token, i.e., word) into a fixed-dimensional vector, enabling computers to understand and compute language semantics. It is a core component of large models for processing text. Qwen-110B, also known as Qwen1.5-110B, is a large language model that was open-sourced by the Tongyi Qianwen team in April 2024.

[0027] Example 1 This embodiment provides a clinical decision-making method for respiratory diseases. By pre-training and fine-tuning the model before application and inference, the model can output structured results containing auxiliary decision-making suggestions and reasoning basis, enhancing the trust and verifiability of the model's output. For example... Figure 1 As shown, the specific steps include: Step S1: Obtain sample and label data; Acquire sample data, including general medical sample data and electronic medical record sample data; label the ICD-10 diagnosis, diagnostic basis, treatment plan, and treatment decision in the electronic medical record data to obtain labeled data.

[0028] The general medical sample data comes from the open-source datasets WanJuan CC (English) and WuDaoCorpora (Chinese), which together include 5.8 billion words of open-source human dialogue data. In addition, to ensure that the model has a solid foundation in medical knowledge, a dedicated medical corpus of 1.8 billion words was also constructed. The corpus is derived from peer-reviewed research papers, clinical guidelines, and textbooks indexed by PubMed, MEDLINE, and CNKI. This general medical corpus covers clinical research, specialty textbooks, and practice guidelines related to the diagnosis and treatment of respiratory diseases over the past five years.

[0029] The electronic medical record (EHR) sample data is domain-specific EHR data, derived from clinical medical records (de-identified respiratory cases) of patients with respiratory diseases collected by West China Hospital of Sichuan University. The inclusion criteria for EHR data were: a clear diagnosis of the target disease, complete EHR data, and at least one hospitalization record during the study period. For patients with multiple visits during the observation period, two records were randomly retained for each patient to maintain data balance and avoid over-enrichment of a single individual sample. Exclusion criteria included: missing data ratio ≥30%, unclear diagnostic information or suspected misdiagnosis, inability to match a specific respiratory disease, and rare disease cases with only a single occurrence. ICD-10 diagnoses, diagnostic criteria, treatment plans, and treatment decisions in the EHR data were labeled to obtain tagged data. To avoid data leakage, stratified sampling was performed by patient ID during dataset partitioning to ensure that all records of the same patient were uniquely assigned to either the training or test set.

[0030] Step S2: Pre-training and fine-tuning the diagnostic decision model; The diagnostic decision model includes a diagnostic sub-model Dx and an inference decision sub-model Ex. The diagnostic sub-model Dx and the inference decision sub-model Ex are pre-trained using general medical sample data. The diagnostic sub-model Dx and the inference decision sub-model Ex are fine-tuned using electronic medical record sample data and corresponding label data.

[0031] The diagnostic sub-model Dx is the LungGPT-Dx(7B) model, built upon the Qwen-7B architecture. Its core objective is "high-accuracy ICD-10 classification and fast inference," and it includes an input layer, a decoder layer, an intermediate layer, and an output layer. Specifically: The input layer is configured with a structured EHR parsing module and 768-dimensional word embeddings to convert unstructured EHRs (text / numerical data) into structured tensors, adapting them to the input dimensions of subsequent models. Specifically, the structured EHR parsing module converts unstructured text such as chief complaints, present medical history, and examination results into a structured tensor of "feature dimension × 768," retaining only diagnosis-related features (filtering out irrelevant information, such as non-medical fields in the patient's basic information). The structured EHR parsing module is based on existing technology and can be BioBERT (Lee J, Yoon S, Kim S, et al. BioBERT: a pre-trained biomedical language representation model for bio), which is the most commonly used 768-dimensional BERT-like baseline model in the clinical text domain; or it can be ClinicalBERT (Alsentzer E, Murphy WR, Boag W, et al. Publicly Available Clinical BERT Embeddings. NAACL Clinical NLP Workshop, 2019), which is specifically pre-trained on massive medical records in EHR such as MIMIC, with a standard 768-dimensional output adapted to medical record text encoding.

[0032] The decoder layer is configured with a 12-layer Transformer decoder block, which simplifies the original Qwen-7B's 32-layer decoder to 12 layers, eliminates redundant computations in general dialogue, and improves inference speed. This Transformer decoder block can be any existing Transformer decoder block, and those skilled in the art can choose and apply it according to their needs without any inventive effort.

[0033] The intermediate layer includes a clinical feature attention enhancement module, which increases the attention weight (weight coefficient 1.5) of key diagnostic features in the EHR (such as CT images, blood gas analysis, and physical signs). Specifically, this module assigns higher attention weights to a pre-defined "diagnostic key feature vocabulary" (such as "patchy shadows," "wet rales," and "pH value") within the multi-head attention layer, enhancing the model's ability to capture core diagnostic information. This clinical feature attention enhancement module is an existing module and a common enhancement method for medical NLP and HER modeling (Automated classification of clinical diagnostics in electronic health records using transformer (PLOS ONE, 2025)). However, the innovation of the clinical feature attention enhancement module in the intermediate layer of this embodiment is as follows: 1. Fixed weight coefficient of 1.5, non-adaptive, non-learning, explicit fixed multiplier enhancement (different from most dynamic / learnable weights); 2. Pre-set "diagnostic key feature vocabulary", a closed-source / expert-defined vocabulary for diagnostic tasks (patchy shadows, wet rales, pH value, etc.); 3. Direct weighting within the multi-head attention layer, with explicit multiplication of coefficients during the attention score calculation stage, rather than post-processing / gating.

[0034] The output layer is configured with an ICD-10-specific classification header and a confidence scoring module to output ICD-10 codes (one-hot codes) for 86 respiratory diseases and calculate the confidence score for each code (i.e., diagnosis). The ICD-10-specific classification header replaces the original Qwen-7B text generation header, employs a fully connected layer with a Softmax activation function, and has an output dimension of 86 (the predefined number of ICD-10 diagnostic categories), directly mapping to standard diagnostic codes. Both the ICD-10-specific classification head and the confidence scoring module are existing technologies. For the ICD-10-specific classification head, evlin et al. (2019) BERT: token embedding → fully connected → Softmax can be used for classification; Liu et al. (2023) LLaMA / Qwen medical fine-tuning can be used: discarding the generated head and adding a classification layer for disease / ICD prediction; medical NLP standards: ClinicalBERT and BioBERT both use 768 dimensions → FC → Softmax for diagnostic classification; the confidence scoring module can use Leveraging LLMs for ICD Coding and Uncertainty Estimation (DiVA, 2025).

[0035] The details of each layer in the diagnostic sub-model Dx are shown in the table below:

[0036] The diagnostic sub-model Dx uses a 12-layer simplified decoder, which improves the inference speed by 60% compared to the original Qwen-7B. The time taken for a single EHR diagnosis is less than 0.5 seconds, which is suitable for the needs of real-time clinical diagnosis. In addition, the classification head directly outputs ICD-10 encoding without the need for additional text parsing steps, and the diagnostic results can be directly connected to the hospital information system (HIS).

[0037] The reasoning and decision-making sub-model Ex is the LungGPT-Ex (110B) model, which is built on the Qwen-110B (i.e., Qwen1.5-110B, a large language model open-sourced by the Tongyi Qianwen team in April 2024) architecture. Its core objectives are "evidence-based text generation, logical coherence, and adherence to clinical norms," ​​and it includes an input layer, decoder layer, intermediate layer, and output layer. Specifically: The input layer is configured with a multi-source data fusion module and 768-dimensional word embedding, used to integrate medical data. Diagnostic results and medical knowledge context The data is fused into a unified tensor. Specifically, this multi-source data fusion module is used to integrate medical data. Diagnostic results and medical knowledge context Perform weighted fusion (weight ratio: 40% 30% (30%), generating a unified input tensor. This multi-source data fusion module is an existing technology, Multi-SourceFusion Transformer for Diagnosis Prediction Using EHR Data (IEEE Journal of Biomedical and Health Informatics, 2024.), which discloses the weighted feature fusion of three types of information: electronic medical record data, diagnostic results, and medical knowledge graphs, for disease diagnosis prediction.

[0038] The decoder layer is configured with 32 layers of Transformer decoder blocks to fully preserve the original Qwen-110B decoder structure, ensuring the fluency of natural language generation and logical reasoning capabilities. This Transformer decoder block can be any existing Transformer decoder block; those skilled in the art can choose and apply it according to their needs without any inventive effort.

[0039] The intermediate layer includes a diagnostic anchoring constraint module and a knowledge fusion module. These modules are used to force the module to anchor to the diagnostic results of the diagnostic sub-model Dx and to fuse retrieved medical knowledge, preventing the generation of content that deviates from the diagnosis. Specifically, the diagnostic anchoring constraint module adds a "diagnostic feature verification" to the output of each layer of the decoder. If the generated content deviates from the diagnostic results... If the core diagnostic direction (e.g., the diagnostic sub-model Dx diagnoses "severe pneumonia," and the generated content focuses on "asthma"), then the attention weights are adjusted in reverse to force a regression to the diagnostic topic; this knowledge fusion module is used to integrate the retrieved medical knowledge context. (For example, the guidelines for the treatment of severe pneumonia) are broken down into sub-features such as "diagnostic criteria", "treatment principles", and "contraindications", and are linked to medical data. The patient data is matched field by field (e.g., "elderly patients" matches "antibiotic dosage adjustment"). Both the diagnostic anchoring constraint module and the knowledge fusion module are existing technologies. The diagnostic anchoring constraint module can use topic anchoring and generation constraints (general NLP), such as the long-published technologies of Controllable Text Generation with Constrained Decoding ICML / NeurIPS series, which verify the consistency of generated content with the target topic and dynamically adjust attention or probability distribution during the decoding stage. Alternatively, it can use diagnostic anchoring and attention correction in medical scenarios, such as Diagnosis-Guided Medical NoteGeneration with Constrained Transformer AMIA / JMIR, which uses the diagnostic result as an anchor point and penalizes or reweights attention for text generation that deviates from the diagnosis, forcing the generated content to be consistent with the diagnosis. It can also use successive verification of the decoding layer + reverse correction weights from multiple large-scale medical model patents, which performs diagnostic consistency verification on the generated content at each layer of the decoder, adjusting attention weights if deviations are found, so that the generation returns to the diagnostic topic. The knowledge fusion module can be derived from the knowledge fusion module in invention patent application 202610052171.1.

[0040] The output layer is configured with a clinical text standardization module and an autoregressive generation head, used to generate explanatory text according to the clinical document format, and to control the uniformity of terminology and logical progression. Among them, the clinical text standardization module has a built-in respiratory medical record writing template (divided into three levels of headings: "Diagnostic Basis", "Treatment Plan" and "Precautions"). When generating text, it forces the output to follow the template format, and at the same time verifies the standardization of terminology (such as uniformly using "ICD-10 code" instead of "diagnostic code"). Both the clinical text standardization module and the autoregressive generation head are existing technologies. The clinical text standardization module can use the general large model generation paradigm (basic), Devlin J, et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding [C] / / NAACL-HLT, 2019. (Autoregressive decoding is a core capability of Transformer-type models, common knowledge); the autoregressive generation head can use autoregressive generation (best match) in medical scenarios, Diagnosis-Guided Medical Note Generation with Constrained Transformer (AMIA, 2023): publicly adopts an autoregressive generation head to generate diagnostic interpretations and treatment recommendations that conform to clinical standards.

[0041] The details of each layer in the diagnostic sub-model Dx are shown in the table below:

[0042] The reasoning and decision-making sub-model Ex, through a complete 32-layer decoder and relying on the large model capabilities of Qwen-110B with a large number of parameters, enables reasoning of complex clinical logic (such as causal analysis of "adjustment of antibiotic dosage for elderly patients"). In addition, through multi-module constraints, it ensures that the explanatory text anchors the diagnostic results, fits the specific patient data, and conforms to clinical documentation standards, thus significantly reducing the probability of "illusion".

[0043] When pre-training the diagnostic sub-model Dx and the inference decision-making sub-model Ex using general medical sample data, the pre-training process is an autoregressive process that does not require labels and can utilize existing pre-training methods. Those skilled in the art can choose and apply existing training methods (including loss functions, such as cross-entropy loss) according to actual circumstances and needs, without requiring creative effort. Figure 2 As shown.

[0044] When fine-tuning the diagnostic sub-model Dx and the inference decision-making sub-model Ex using electronic medical record sample data and corresponding label data, such as Figure 3 As shown, specifically: When fine-tuning the diagnostic sub-model Dx, the cross-entropy loss function is used. To achieve the core optimization goal of full-scale fine-tuning, full-scale parameter fine-tuning after 4-bit quantization is implemented based on QLoRA technology. The core steps are as follows: ① Based on the Qwen-7B architecture initialization model, low-power, long-distance wireless communication technology with 4-bit quantization reduces memory usage; ② Construct an ICD-10 labeled dataset for respiratory diseases and complete the data cleaning and structuring of electronic medical record data; ③ Fine-tune the loss using Dx To optimize the full fine-tuning of the target and enhance the ICD-10 classification capability; among which, the Dx fine-tuning loss... Represented as: ; in, Indicates the first ICD-10 diagnostic categories for respiratory diseases This indicates the total number of ICD-10 diagnostic categories for respiratory diseases. This represents the category label value (unique thermal code, 1 for a match, 0 otherwise) in the ICD-10 labeled diagnostic set for respiratory diseases. This represents the set of all trainable parameters of the diagnostic sub-model Dx (including all parameters of the word embedding layer, attention layer, and ICD-10 dedicated classification head). Represented as medical data for a given patient and model parameters At that time, the patient was diagnosed with the first Conditional probability of ICD-10-labeled diagnostic categories for respiratory diseases.

[0045] ④ Validate on the clinical test set, optimize the classification threshold, and obtain the final diagnostic model.

[0046] When fine-tuning the inference decision sub-model Ex, it includes supervised fine-tuning of SFT and human feedback reinforcement learning RLHF.

[0047] Supervised fine-tuned SFT, which is a negative log-likelihood loss with diagnostic anchoring. The aim is to enable the model to learn the clinical logic of "diagnosis-knowledge-interpretation." Negative log-likelihood loss. Specifically: ; in, This represents all trainable parameters of the inference decision sub-model Ex. This represents the maximum sequence length of the text generated by the inference decision submodel Ex. This represents the word vector generated in step t. This represents the historical word sequence generated up to step t. Indicates medical data, Indicates the diagnosis result. Indicates the context of medical knowledge; This indicates regularization to prevent overfitting and ensure that the parameters do not deviate from the basic capabilities of Qwen-110B.

[0048] Human Feedback Reinforcement Learning (RLHF) is based on policy gradient loss using a reward model. Its purpose is to enable the model to generate clinical explanatory texts that align with human preferences; policy gradient loss Specifically: ; in, This represents the parameters of the reasoning and decision-making sub-model Ex during the reinforcement learning phase. This represents the set of parameters for the reasoning and decision-making sub-model Ex. Indicating in strategy Expected calculation under the following conditions This represents the sequence of clinical explanatory texts generated by the reasoning and decision-making submodel Ex. This represents the clinician's rating of the explanatory text generated by the model (award value, 1-5 points). Representation Strategy Generate sequence The probability of; This represents the hyperparameter, namely the KL divergence penalty coefficient, with a value of 0.05; Denotes KL divergence, This represents the baseline model strategy during the supervised fine-tuning phase; This represents the KL divergence between the policy distribution and the SFT pre-training distribution, i.e., the inter-policy bias constraint, used to ensure that the generated content does not deviate from medical professional knowledge.

[0049] Step S3: Real-time clinical decision-making; Acquiring medical data to be tested Input the diagnostic sub-model Dx, and the diagnostic sub-model Dx will output the diagnostic results. ; the diagnosis results As an anchor point for queries, based on medical data Retrieve from structured databases and return medical knowledge context. Using medical data Diagnostic results and medical knowledge context Build the final prompt words (Medical data during construction) It is used as the basic context and incorporated into the reasoning and decision-making sub-model Ex); the final prompt word Input the inference decision sub-model Ex, and output the explanatory text. Explain the text It is a structured result that includes suggestions for decision support and reasoning.

[0050] Medical data After inputting into the clinical decision model, such as Figure 4 As shown, the specific process is as follows: 1. Preliminary diagnosis and anchor point extraction; The diagnostic sub-model Dx first processes the input structured medical data. It outputs diagnostic results including ICD-10 diagnostic code and diagnostic category label. Diagnostic results Represented as: ; in, Medical data representing a given patient and model parameters At that time, the patient was diagnosed with a respiratory disease, and the ICD-10 diagnostic category was marked. The conditional probability; This represents the set of all trainable parameters for the diagnostic sub-model Dx.

[0051] 2. Knowledge retrieval based on chained components; Diagnostic results As a query anchor, it is used in structured databases via chained interfaces (such as LangChain). Searching in ); Search function Return to relevant medical knowledge context (e.g., the pathophysiological mechanisms of the disease, standard treatment regimens, contraindications); medical knowledge context Represented as: ; Advantages: This step ensures that subsequent interpretations are based not only on patient data but also on standard medical guidelines, achieving "evidence-based" AI generation that complements the medical expertise of the reasoning and decision-making sub-model Ex.

[0052] Furthermore, it should be noted that the structured database not only contains text, but also medical image feature vectors. When searching through the chain interface, not only text guides are retrieved, but also CT image features of similar cases are retrieved. This allows for the construction of multimodal prompts, which are then input into the inference decision sub-model Ex.

[0053] 3. Dynamic construction and modification of prompt words; Construct the final prompt words as the final input to the inference decision sub-model Ex. This is a dynamic correction process; the prompts are no longer static templates, but rather based on medical data. ICD-10 diagnostic results Medical knowledge context The composite tensor formed by these elements is represented as: .

[0054] 4. Controlled interpretation generation; Based on the prompts constructed above, the reasoning decision sub-model Ex generates the final explanatory text. (Includes diagnostic explanation and treatment plan); Explanatory text Represented as: ; in, This represents the set of all trainable parameters for the inference decision sub-model Ex; The generation probability can be expressed as: ; in, This represents the maximum sequence length of the text generated by the inference decision submodel Ex. This represents the word vector generated in step t. This represents the historical word sequence generated up to step t. Indicates medical data, Indicates the diagnosis result. This indicates the context of medical knowledge.

[0055] By decomposing the traditional end-to-end clinical decision generation probability P(E|X) into a two-step chain probability structure, namely: first, from medical data... Received diagnosis results Then, from medical data Diagnostic results and medical knowledge context Generate explanatory text together Thus forming the overall joint probability. .

[0056] In the inference decision sub-model Ex, it does not directly receive medical data. Instead, it completes data interaction and text output through a three-step logic of "dynamic construction of prompt words → hierarchical model parsing → autoregressive generation". The specific process is as follows: Step E1, Final Prompt Structured construction (data integration); Based on medical data Diagnostic results and medical knowledge context A standardized prompt word template is constructed, and the three types of data are integrated into a parsable input text for the reasoning and decision-making sub-model Ex according to a fixed format to obtain the final prompt word. Final prompt Template examples (actually structured tensors, illustrated in natural language below): [Patient Basic Data] { Structured analysis results: such as "Gender: Male, Age: 90 years old, Chief complaint: Recurrent cough and sputum for more than 2 years, worsening with shortness of breath for 10 days; Chest CT: Moderate pleural effusion on the left side; Blood gas analysis: pH 7.437, PCO2 32.3 mmHg" [Confirmed Case Result] { Standardized expression: such as "ICD-10 code: J18.9, primary diagnosis: severe pneumonia; secondary diagnosis: acute exacerbation of chronic obstructive pulmonary disease (J44.1)" [Evidence-Based Medicine Knowledge] { Structured content: such as "Diagnostic criteria for severe pneumonia: acute onset + patchy shadows on lung imaging + moist rales in the lungs; Treatment principles: broad-spectrum antibiotics for infection control (piperacillin-tazobactam) + monitoring of vital signs; Triggers for acute exacerbations of COPD: infection / fatigue; Treatment: bronchodilator nebulization" The template will be converted into a tensor form that the Ex model can recognize, containing eigenvectors, Encoding vector, The knowledge vectors, along with the other two dimensions, are unified to the input dimensions of the Ex model (adapted to the 768-dimensional word embedding of Qwen-110B).

[0057] Step E2: The reasoning and decision-making sub-model Ex performs hierarchical parsing of prompt words (the core of data interaction). The reasoning decision sub-model Ex receives the final prompt word. Then, layered parsing and data interaction are completed through internal modules; the specific steps are as follows: Step E2-1, Anchoring layer analysis; Analysis of the final prompt words Diagnostic results The core diagnostic direction is locked through the "diagnostic anchoring constraint module" in the middle layer (such as prioritizing the parsing of "severe pneumonia" related content) and filtering out redundant information that is not related to the diagnosis. Step E2-2: Knowledge Layer Integration; Medical data and medical knowledge context Feature fusion can be performed, for example, by combining the "treatment principles for severe pneumonia" with the characteristics of "patients being 90 years old and having pleural effusion" to screen for appropriate treatment recommendations (such as adjusting antibiotic dosage and strengthening monitoring). Step E2-3: Context alignment; Verify medical data Clinical data and diagnostic results Medical knowledge context Logical consistency; for example, if medical data The diagnosis was made without "moist rales in the lungs". For cases classified as "severe pneumonia," the explanatory text will specify "diagnosed by combining imaging features and patient symptoms" to avoid logical contradictions.

[0058] Step E3: The inference decision sub-model Ex generates explanatory text through autoregression (final output); After completing the cue word parsing and data fusion, the inference decision sub-model Ex outputs explanatory text based on the autoregressive generation capability of Qwen-110B; the specific steps are as follows: Step E3-1, the reasoning decision sub-model Ex starts from the final prompt word. Starting from the "generation requirements", first generate the first sentence of the diagnostic basis (e.g., "The patient is diagnosed with severe pneumonia, based on the following:"). Step E3-2: Based on the previously generated lexical units (e.g., "Based on the following:"), combine them with the fused medical data. Diagnostic results Medical knowledge context Generate the next word element (e.g., "The patient had an acute onset of illness, accompanied by recurrent fever and cough"). Step E3-3: Repeat the above autoregressive process until a complete diagnostic basis and treatment plan are generated. The model's built-in "clinical text standardization module" will verify the generated content in real time to ensure that it conforms to the clinical document format (points, terminology consistency). Step E3-4: After generation, output the natural language interpretation text. Explain the text This is a structured result that includes suggestions for decision support and the rationale behind them; an example is: [Diagnostic Basis] (i.e., Reasoning Basis): 1. Severe pneumonia (ICD-10: J18.9, primary diagnosis): The patient had an acute onset, accompanied by recurrent fever, cough, and yellow sputum. Chest CT showed scattered patchy shadows in both lungs. Combined with blood gas analysis results, the diagnosis met the criteria for severe pneumonia. 2. Acute exacerbation of chronic obstructive pulmonary disease (ICD-10: J44.1, secondary diagnosis): The patient has a history of cough and sputum production for more than 2 years. Physical examination revealed barrel chest, and CT showed signs of emphysema. The symptoms have recently worsened, suggesting an acute exacerbation.

[0059] [Treatment Plan] (i.e., decision support recommendations): 1. Level 1 nursing care + ECG monitoring, closely monitoring vital signs; 2. Administer piperacillin-tazobactam for anti-infective treatment (considering the patient's advanced age, adjust the dose to 75% of the usual dose); 3. Inhalation of bronchodilators via nebulization to relieve airway spasm; 4. A follow-up chest CT scan and blood gas analysis should be performed within 3 days to assess the treatment effect.

[0060] like Figure 5 As shown, the specific application steps of this method are as follows: 1) Information input and retrieval; The process begins with entering a unique patient ID. Based on this, the system automatically retrieves and displays a complete medical history in a structured format from the hospital information system, including the chief complaint, present illness, past medical history, and physical examination results. Simultaneously, the system links and displays relevant auxiliary examination results, such as crucial chest CT reports, blood gas analysis, and coagulation function tests, providing objective evidence for subsequent analysis.

[0061] 2) Intelligent analysis and suggestions; After integrating the above information, the work begins according to the method steps of Embodiment 1 described above.

[0062] Diagnosis generation: Based on all information, automatically generate preliminary diagnoses (such as "severe pneumonia" or "acute exacerbation of chronic obstructive pulmonary disease") and determine their category (such as "preliminary diagnosis" or "minor diagnosis"). Interactive reasoning: Users can click "Expand Diagnostic Basis" and "Expand Treatment Plan" to request the model to provide specific reasoning processes and clinical decision suggestions. For example, for the diagnosis of "severe pneumonia," the model will combine the patient's acute symptoms, CT imaging features (scattered patchy and nodular shadows in both lungs), and auscultation results (moist rales in both lower lung fields) to generate evidence.

[0063] 3) Clinical decision-making and review; All diagnoses, justifications, and treatment plans automatically generated using the above methods ultimately require review and decision-making by clinicians. The interface allows users to "double-click to accept or edit modifications." Figure 5 The prompt (not explicitly shown) clarifies the role of AI in assisted positioning, providing doctors with structured references while retaining the final decision-making and modification rights.

[0064] Example 2 This embodiment provides a clinical decision-making system for respiratory diseases, including: The sample and label data acquisition module is used to acquire sample data, including general medical sample data and electronic medical record sample data; it also marks the ICD-10 diagnosis, diagnostic basis, treatment plan, and treatment decision in the electronic medical record data to obtain label data. The diagnostic decision model pre-training and fine-tuning module is used to pre-train and fine-tune the diagnostic decision model, which includes a diagnostic sub-model Dx and an inference decision sub-model Ex. The module uses general medical sample data to pre-train the diagnostic sub-model Dx and the inference decision sub-model Ex respectively. The module uses electronic medical record sample data and corresponding label data to fine-tune the diagnostic sub-model Dx and the inference decision sub-model Ex respectively. The real-time clinical decision module is used to acquire the medical data to be tested. Input the diagnostic sub-model Dx, and the diagnostic sub-model Dx will output the diagnostic results. ; the diagnosis results As an anchor point for queries, based on medical data Retrieve from structured databases and return medical knowledge context. Using medical data Diagnostic results and medical knowledge context Build the final prompt words ; Final prompt Input the inference decision sub-model Ex, and output the explanatory text. Explain the text It is a structured result that includes suggestions for decision support and reasoning.

[0065] Example 3 A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform steps of a clinical decision-making method for respiratory diseases.

[0066] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0067] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the memory may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the clinical decision-making method for respiratory diseases. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0068] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run program code for a clinical decision-making method for respiratory diseases.

[0069] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform steps of a clinical decision-making method for respiratory diseases.

[0070] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the clinical decision-making method for respiratory diseases as described above.

[0071] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the clinical decision-making method for respiratory diseases described in the embodiments of this application.

Claims

1. A clinical decision-making method for respiratory diseases, characterized in that, Includes the following steps: Step S1: Obtain sample and label data; Acquire sample data, including general medical sample data and electronic medical record sample data; label the ICD-10 diagnosis, diagnostic basis, treatment plan, and treatment decision in the electronic medical record data to obtain labeled data; Step S2: Pre-training and fine-tuning the diagnostic decision model; The diagnostic decision model includes a diagnostic sub-model Dx and an inference decision sub-model Ex; the diagnostic sub-model Dx and the inference decision sub-model Ex are pre-trained using general medical sample data; the diagnostic sub-model Dx and the inference decision sub-model Ex are fine-tuned using electronic medical record sample data and corresponding label data. Step S3: Real-time clinical decision-making; Acquiring medical data to be tested Input the diagnostic sub-model Dx, and the diagnostic sub-model Dx will output the diagnostic results. ; the diagnosis results As a query anchor, based on medical data Retrieve from structured databases and return medical knowledge context. Using medical data Diagnostic results and medical knowledge context Build the final prompt words ; Final prompt Input the inference decision sub-model Ex, and output the explanatory text. Explain the text The results are structured and include suggestions for decision support and reasoning. In step S2, the diagnostic sub-model Dx includes an input layer, a decoder layer, an intermediate layer, and an output layer. The input layer is configured with a structured EHR parsing module and word embeddings, the decoder layer is configured with 12 layers of Transformer decoder blocks, the intermediate layer is configured with a clinical feature attention enhancement module, and the output layer is configured with an ICD-10 dedicated classification head and a confidence scoring module. The reasoning decision sub-model Ex includes an input layer, a decoder layer, an intermediate layer, and an output layer. The input layer is configured with a multi-source data fusion module and word embedding. The decoder layer is configured with a 32-layer Transformer decoder block. The intermediate layer is configured with a diagnostic anchoring constraint module and a knowledge fusion module. The output layer is configured with a clinical text standardization module and an autoregressive generation head.

2. The clinical decision-making method for respiratory diseases as described in claim 1, characterized in that, In step S2, when fine-tuning the diagnostic sub-model Dx and the inference decision sub-model Ex, the total loss is... Represented as: Fine-tuning loss of the diagnostic sub-model Dx Represented as: ； Supervised fine-tuning of the negative log-likelihood loss in the inference decision sub-model Ex Represented as: ； Policy gradient loss for the inference decision sub-model Ex in reinforcement learning with human feedback Represented as: ； in, , , Each represents the weight of the loss; Indicates the first ICD-10 diagnostic categories for respiratory diseases This indicates the total number of ICD-10 diagnostic categories for respiratory diseases. This represents the category label value in the ICD-10 labeled diagnostic set for respiratory diseases. This represents the set of all trainable parameters for the diagnostic sub-model Dx; Represented as medical data for a given patient and model parameters At that time, the patient was diagnosed with the first Conditional probability of ICD-10-labeled diagnostic categories for each respiratory disease; This represents all trainable parameters of the inference decision sub-model Ex. This represents the maximum sequence length of the text generated by the inference decision submodel Ex. This represents the word vector generated in step t. This represents the historical word sequence generated up to step t. Indicates medical data, Indicates the diagnosis result. Indicates the context of medical knowledge. For hyperparameters; Indicates regular expression processing; This represents the parameters of the reasoning and decision-making sub-model Ex during the reinforcement learning phase. This represents the set of parameters for the reasoning and decision-making sub-model Ex. Indicating in strategy Expected calculation under the following conditions This represents the sequence of clinical explanatory texts generated by the reasoning and decision-making submodel Ex. This represents the rating given by clinicians to the explanatory text generated by the model. Representation Strategy Generate sequence The probability of; Indicates hyperparameters; Denotes KL divergence, This represents the baseline model strategy during the supervised fine-tuning phase; This represents the KL divergence between the policy distribution and the SFT pre-trained distribution.

3. The clinical decision-making method for respiratory diseases as described in claim 1, characterized in that, In step S3, the medical data to be tested After inputting the diagnostic submodel Dx, the explanation text continues until the inference decision submodel Ex outputs it. The specific process is as follows: I. Preliminary diagnosis and anchor point extraction; The diagnostic sub-model Dx processes the input medical data. It outputs diagnostic results including ICD-10 diagnostic code and diagnostic category label. Diagnostic results Represented as: ； II. Knowledge retrieval based on chain components; Diagnostic results As a query anchor, it performs retrieval in the structured database through a chained interface; retrieval function Return to the relevant medical knowledge context ; Medical knowledge context Represented as: ； III. Dynamic Construction and Correction of Prompt Words; Construct the final prompt words as the final input to the inference decision sub-model Ex. Final prompt It is based on medical data Diagnostic results Medical knowledge context The composite tensor formed by these elements is represented as: ； IV. Controlled interpretation generation; The reasoning decision sub-model Ex is based on the final prompt word. Generate the final explanatory text. Explain the text Represented as: ； in, Medical data representing a given patient and model parameters At that time, the patient was diagnosed with a respiratory disease, and the ICD-10 diagnostic category was marked. The conditional probability; This represents the set of all trainable parameters for the diagnostic sub-model Dx; Represents a structured database; This represents the set of all trainable parameters for the inference decision sub-model Ex.

4. The clinical decision-making method for respiratory diseases as described in claim 3, characterized in that, In step S3, the inference decision sub-model Ex generates the explanatory text. The specific process is as follows: Step E1, Final Prompt Structured construction; Based on medical data Diagnostic results and medical knowledge context A standardized prompt word template is constructed, and the three types of data are integrated into a parsable input text for the reasoning and decision-making sub-model Ex according to a fixed format to obtain the final prompt word. ; Step E2: Hierarchical analysis of prompt words by the reasoning decision sub-model Ex; The reasoning decision sub-model Ex receives the final prompt word. Then, layered parsing and data interaction are completed through internal modules; Step E3: The inference decision sub-model Ex generates explanatory text through autoregression; The reasoning decision sub-model Ex outputs explanatory text based on the autoregressive generation capability of the large model.

5. A clinical decision-making method for respiratory diseases as described in claim 4, characterized in that, In step E2, the specific steps are as follows: Step E2-1, Anchoring layer analysis; Analysis of the final prompt words Diagnostic results The core diagnostic direction is locked through the diagnostic anchoring constraint module in the intermediate layer, and redundant information unrelated to diagnosis is filtered out. Step E2-2: Knowledge Layer Integration; Medical data and medical knowledge context Feature fusion is performed to screen for suitable treatment recommendations; Step E2-3: Context alignment; Verify medical data Clinical data and diagnostic results Medical knowledge context Logical consistency.

6. A clinical decision-making method for respiratory diseases as described in claim 4, characterized in that, In step E3, the specific steps are as follows: Step E3-1: The reasoning decision sub-model Ex generates the first sentence of the diagnostic basis; Step E3-2: Based on the previously generated lexical units, combine them with the fused medical data. Diagnostic results Medical knowledge context Generate the next word element; Step E3-3: Repeat the above autoregressive process until a complete diagnostic basis and treatment plan are generated; and the clinical text standardization module verifies the generated content in real time to ensure that it conforms to the clinical document format. Step E3-4: After generation, output the natural language interpretation text. .

7. A clinical decision-making system for respiratory diseases, characterized in that, include: The sample and label data acquisition module is used to acquire sample data, including general medical sample data and electronic medical record sample data; it also marks the ICD-10 diagnosis, diagnostic basis, treatment plan, and treatment decision in the electronic medical record data to obtain label data. The diagnostic decision model pre-training and fine-tuning module is used to pre-train and fine-tune the diagnostic decision model, which includes a diagnostic sub-model Dx and an inference decision sub-model Ex. The module uses general medical sample data to pre-train the diagnostic sub-model Dx and the inference decision sub-model Ex respectively. The module uses electronic medical record sample data and corresponding label data to fine-tune the diagnostic sub-model Dx and the inference decision sub-model Ex respectively. The real-time clinical decision module is used to acquire the medical data to be tested. Input the diagnostic sub-model Dx, and the diagnostic sub-model Dx will output the diagnostic results. ; the diagnosis results As an anchor point for queries, based on medical data Retrieve from structured databases and return medical knowledge context. Using medical data Diagnostic results and medical knowledge context Build the final prompt words ; Final prompt Input the inference decision sub-model Ex, and output the explanatory text. Explain the text The results are structured and include suggestions for decision support and reasoning. In the diagnostic decision model pre-training and fine-tuning module, the diagnostic sub-model Dx includes an input layer, a decoder layer, an intermediate layer, and an output layer. The input layer is configured with a structured EHR parsing module and word embeddings, the decoder layer is configured with 12 layers of Transformer decoder blocks, the intermediate layer is configured with a clinical feature attention enhancement module, and the output layer is configured with an ICD-10 dedicated classification head and a confidence scoring module. The reasoning decision sub-model Ex includes an input layer, a decoder layer, an intermediate layer, and an output layer. The input layer is configured with a multi-source data fusion module and word embedding. The decoder layer is configured with a 32-layer Transformer decoder block. The intermediate layer is configured with a diagnostic anchoring constraint module and a knowledge fusion module. The output layer is configured with a clinical text standardization module and an autoregressive generation head.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

Images