Methods, devices, equipment, and media for assisted diagnosis based on large language models

Abstract

This disclosure relates to an auxiliary diagnostic processing method, apparatus, device, and medium based on a large language model, comprising: acquiring a trained large language model trained on basic training data samples and incremental training data; adjusting the trained large language model based on the incremental training data to obtain a candidate large language model; simultaneously processing the question text to be processed based on the trained large language model and the candidate large language model to obtain an inference answer text, a confidence score, and a candidate answer text; triggering a preset backoff strategy when the confidence score is less than a confidence threshold; and replacing the trained large language model with the candidate large language model when the evaluation result of the candidate answer text on the candidate large language model meets preset model switching conditions for processing real-time question text. Thus, incremental training ensures model effectiveness, and the human-machine collaboration method using subgroup dynamic confidence thresholds avoids decision interference and effectively reduces the negative effects of human-machine collaboration.

Description

Technical Field

[0001] This disclosure relates to the fields of medical artificial intelligence and natural language processing technology, and in particular to an auxiliary diagnostic processing method, device, equipment and medium based on a large language model. Background Technology

[0002] Typically, the emergency department is the core department of a hospital for handling acute and critical illnesses, characterized by a large number of patients, complex conditions, incomplete information, and a short decision-making window; neurological emergencies (such as acute ischemic stroke, epilepsy, vertigo, headache, etc.) have extremely high requirements for diagnostic timeliness.

[0003] Among related technologies, existing deep learning models based on Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformers, as well as large language models for medical question answering in recent years, have been used in scenarios such as stroke identification, automatic analysis of neuroimaging, and prediction of stroke recurrence risk. However, existing methods have technical problems such as the lack of closed-loop model version evolution and insufficient stability of human-machine collaboration in subgroups. Summary of the Invention

[0004] To address, or at least partially address, the aforementioned technical problems, this disclosure provides an auxiliary diagnostic processing method, apparatus, device, and medium based on a large language model. This addresses technical issues faced by existing medical large language models in real-world clinical deployments, such as unsustainable evolution of model versions and reduced accuracy due to human-machine collaboration in specific subgroups.

[0005] This disclosure provides an auxiliary diagnostic processing method based on a large language model. The method includes: acquiring a trained large language model trained on basic training data samples and incremental training data; wherein the incremental training data is generated based on quintuple structure data, the quintuple structure data including initial diagnosis information, model output diagnostic information, AI-assisted diagnostic information, confirmed diagnosis information, and scoring information; adjusting the trained large language model based on the incremental training data to obtain a candidate large language model; acquiring a problem text to be processed, and simultaneously processing the problem text based on the trained large language model and the candidate large language model to obtain inference results. The system includes answer text, corresponding credibility score, and candidate answer text; the credibility score is determined based on symptom entropy and model confidence estimation; when the credibility score is less than the confidence threshold corresponding to a preset disease seed group category, a preset fallback strategy is triggered; the preset fallback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision; the candidate large language model is evaluated based on the candidate answer text, and when the evaluation result meets the preset model switching conditions, the candidate large language model replaces the trained large language model for processing real-time question text.

[0006] Optionally, obtaining the trained large language model based on basic training data samples includes: obtaining a triplet aligned dataset consisting of a question, the original answer, and the corrected answer as the basic training data sample; performing supervised fine-tuning on a preset basic model based on the basic training data sample; and performing group-relative strategy optimization on the basis of supervised fine-tuning to obtain the trained large language model; wherein, the group-relative strategy optimization includes generating multiple candidate responses within each symptom cluster and ranking them within the group and awarding step rewards based on logical coherence, consistency with the target score, and misdiagnosis penalty score, wherein the misdiagnosis penalty score includes a disease grading weighted misdiagnosis penalty item.

[0007] Optionally, the base model is deployed using a hybrid low-bit quantization of W4A8 or W4A16; where W4 indicates that the weights are quantized with 4 bits, A8 indicates that the activations are quantized with 8 bits, and A16 indicates that the activations are quantized with 16 bits. The multi-head attention triple projection of the base model is fused with the activation function and distributed in blocks according to the computation kernel. The key-value cache is managed at the page level to support multi-request concurrency, and when the length of the input sequence is less than a preset length threshold, it switches to a small window mode that disables the page-based key-value cache.

[0008] Optionally, acquiring the incremental training data includes: after each clinical interaction, collecting the initial diagnosis information, the model output diagnosis information, the AI-assisted diagnosis information, the confirmed diagnosis information, and the scoring information to construct the five-tuple structure data as feedback samples; scoring and classifying the feedback samples according to preset screening rules as incremental training data to store in a training sample database; wherein, the preset screening rules include scoring thresholds, target diagnosis consistency, and misdiagnosis information.

[0009] Optionally, obtaining the text of the question to be processed includes: obtaining the original dialogue text by performing streaming speech recognition based on the doctor-patient spoken dialogue; extracting multiple structured fields from the original dialogue text based on an entity recognition model; filling the multiple structured fields into predefined clinical template slots, and repeatedly asking the user questions to fill in the missing key slots, thereby obtaining the text of the question to be processed.

[0010] Optionally, obtaining the credibility score includes: obtaining the model confidence estimate and symptom entropy corresponding to the trained large language model; and estimating the initial credibility score based on the model confidence estimate and the symptom entropy.

[0011] Optionally, when the confidence score is less than the confidence threshold corresponding to the preset disease seed group category, a preset fallback strategy is triggered, including: obtaining the preset disease seed group category corresponding to the reasoning answer text, and obtaining the confidence threshold corresponding to the preset disease seed group category; when the confidence score is less than the confidence threshold, triggering one or more of the following as the preset fallback strategy: only outputting differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision.

[0012] This disclosure also provides an auxiliary diagnostic processing device based on a large language model. The device includes: a first acquisition module for acquiring a trained large language model trained based on basic training data samples; a second acquisition module for acquiring incremental training data, wherein the incremental training data is generated based on quintuple structure data, the quintuple structure data including initial diagnosis information, model output diagnostic information, artificial intelligence-assisted diagnostic information, confirmed diagnosis information, and scoring information; an adjustment module for adjusting the trained large language model based on the incremental training data to obtain a candidate large language model; a third acquisition module for acquiring the problem text to be processed; and a first processing module for simultaneously processing the problem text based on the trained large language model and the candidate large language model. The system processes the question text to obtain the reasoning answer text, the corresponding credibility score, and the candidate answer text. The credibility score is determined based on symptom entropy and model confidence estimation. A second processing module triggers a preset fallback strategy when the credibility score is less than a preset confidence threshold corresponding to a disease seed group category. This preset fallback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make an independent judgment, and reducing the display weight of AI suggestions in the target decision. A third processing module evaluates the candidate large language model based on the candidate answer text and, when the evaluation result meets preset model switching conditions, replaces the trained large language model with the candidate large language model for processing real-time question text.

[0013] This disclosure also provides an electronic device, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the auxiliary diagnostic processing method based on a large language model as provided in this disclosure.

[0014] This disclosure also provides a computer-readable storage medium storing a computer program for executing the large language model-based auxiliary diagnostic processing method provided in this disclosure.

[0015] This disclosure also provides a computer program product, including a computer program, wherein the computer program, when executed by a processor, implements the auxiliary diagnostic processing method based on a large language model as described in one aspect above.

[0016] The technical solution provided in this disclosure has the following advantages compared with the prior art: The auxiliary diagnostic processing method based on a large language model provided in this disclosure includes: acquiring a trained large language model trained based on basic training data samples and incremental training data; wherein, incremental training data is generated based on quintuple structure data, the quintuple structure data including initial diagnosis information, model output diagnostic information, artificial intelligence-assisted diagnostic information, confirmed diagnosis information, and scoring information; adjusting the trained large language model based on the incremental training data to obtain a candidate large language model; acquiring the problem text to be processed, and simultaneously processing the problem text based on the trained large language model and the candidate large language model. This process yields the inference answer text, its corresponding credibility score, and candidate answer text. The credibility score is determined based on symptom entropy and model confidence estimation. When the credibility score is less than the confidence threshold corresponding to a preset disease seed group category, a preset fallback strategy is triggered. This preset fallback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of AI suggestions in the target decision. The candidate large language model is evaluated based on the candidate answer text, and when the evaluation result meets the preset model switching conditions, the candidate large language model replaces the trained large language model for processing real-time question text. Therefore, through clinical closed-loop feedback learning, incremental training of the alignment adaptation layer can be triggered at predetermined cycles without updating the backbone parameters of the basic model. The candidate large language model and the trained large language model can be run in parallel, and switching or rolling back can be performed after dual threshold judgment. This achieves sustainable evolution and observable model improvement, alleviates the problems of data drift and performance degradation in long-term deployment, and, through the human-machine collaboration method of subgroup dynamic confidence threshold, defaults to the fallback state in subgroups that are prone to negative effects of artificial intelligence assistance (such as the peripheral neuropathy subgroup and other neurological disease subgroups), only outputting differential diagnosis and examination suggestions without outputting the final conclusion, avoiding decision interference caused by conveying low-confidence conclusions to users, and effectively reducing the negative effects of human-machine collaboration. Attached Figure Description

[0017] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0018] Figure 1 A flowchart illustrating an auxiliary diagnostic processing method based on a large language model provided in this embodiment of the present disclosure; Figure 2 This is a schematic diagram of training data processing provided in an embodiment of the present disclosure; Figure 3A schematic diagram of GRPO intra-group ranking and step reward training provided in an embodiment of this disclosure; Figure 4 A schematic diagram illustrating the updating of high-freeze LoRA structure parameters provided in an embodiment of this disclosure; Figure 5 A schematic diagram of the hybrid quantization engineering deployment process provided in the embodiments of this disclosure; Figure 6 A schematic diagram illustrating clinical closed-loop feedback learning data reflux and model switching provided in the embodiments of this disclosure; Figure 7 This is a schematic diagram of an end-to-end clinical voice input pipeline provided in an embodiment of the present disclosure; Figure 8 A schematic diagram of the four-step thinking chain output structure and risk labeling provided in the embodiments of this disclosure; Figure 9 A schematic diagram illustrating the clinical decision support and four-stage diagnostic comparison provided in the embodiments of this disclosure; Figure 10 A schematic diagram of the human-machine collaborative backoff strategy for subgroup dynamic confidence threshold provided in this embodiment of the disclosure; Figure 11 A schematic diagram illustrating the overall architecture and clinical decision-making process provided for embodiments of this disclosure; Figure 12 This is a schematic diagram of the structure of an auxiliary diagnostic device based on a large language model, provided in an embodiment of this disclosure. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0021] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0022] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0023] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0024] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0025] Specifically, existing processing methods typically have the following drawbacks: (1) They are highly dependent on high-quality training data and lack generalization ability; (2) They consume a lot of computing resources and are difficult to deploy in real time under hospital-level computing power conditions; (3) They are mainly based on structured inputs and static tasks and lack adaptability to multimodal, unstructured, and long-tailed cases in real-world emergency scenarios; (4) The output lacks a transparent reasoning chain, making it difficult for doctors to assess its credibility and resulting in low clinical acceptance; (5) They do not make responsive adjustments to sociodemographic factors (such as education level and medical insurance type) and ignore the individualized needs of health education and rehabilitation recommendations.

[0026] Specifically, in the long-term operation of real-world clinical practice, the evolution of model versions lacks a closed loop, and there is a standard answer between diagnosis and final diagnosis. Once deployed, the model becomes static and faces long-term data drift and performance degradation. In some disease subgroups (such as peripheral neurology and other neurological disease subgroups), blindly providing doctors with AI-assisted suggestions may actually reduce the accuracy of doctors' diagnoses, exhibiting a "negative effect of AI assistance." In addition, existing large language models generally suffer from high memory consumption, large first token latency, insufficient concurrency, and inability to continue inference in scenarios where edge nodes (such as ambulances and primary hospitals) are offline or have weak networks, under hospital-level computing power conditions. Existing deployment solutions lack collaborative optimization for emergency neurology scenarios in terms of low bit quantization, attention operator fusion, KV-cache paging management, and edge network disaster recovery inference, affecting practical applications in high-time-efficiency emergency scenarios.

[0027] This disclosure proposes an auxiliary diagnostic processing method based on a large language model, addressing issues encountered in the real-world clinical deployment of large language models, such as "unsustainable evolution of model versions," "human-machine collaboration in specific subgroups actually reducing accuracy," and "insufficient engineering deployment of high-throughput, low-latency methods." Details are provided below. Figure 1 Provide a detailed description.

[0028] Figure 1 This is a flowchart illustrating an auxiliary diagnostic processing method based on a large language model, provided in an embodiment of this disclosure. This method can be executed by an auxiliary diagnostic processing device based on a large language model, wherein the device can be implemented in software and / or hardware, and is generally integrated into an electronic device. Figure 1 As shown, the method includes: Step 101: Obtain the trained large language model and incremental training data based on the basic training data samples; among them, incremental training data is generated based on the five-tuple structure data, which includes initial diagnosis information, model output diagnosis information, artificial intelligence-assisted diagnosis information, confirmed diagnosis information and scoring information.

[0029] In some embodiments, a triplet-aligned dataset consisting of a question, the original answer, and the corrected answer is obtained as a basic training data sample. The preset basic model is then fine-tuned under supervision based on the basic training data sample. Based on the supervised fine-tuning, a group-relative strategy optimization is performed to obtain a trained large language model. The group-relative strategy optimization includes generating multiple candidate responses within each symptom cluster and ranking them within the group and awarding step rewards based on logical coherence, consistency with the target score, and misdiagnosis penalty score. The misdiagnosis penalty score includes a disease grade weighted misdiagnosis penalty term.

[0030] For example, such as Figure 2The data processing flow shown includes training corpus processing, triples, and the alignment module (Neuro-Aligner). Specifically, it involves acquiring a corpus of dialogues between physicians and patients from multiple specialties, approximately 300,000 entries. The data processing flow includes: semantic noise detection based on pre-trained language models such as BERT (Bidirectional Encoder Representations from Transformers); regularization and de-identification to remove personally identifiable information such as names, ID numbers, and contact information; multi-round question-answer rewriting to form high-quality question-answer pairs, approximately 90,000 entries; and the alignment dataset using triples composed of questions, original answers, and corrected answers compiled by medical professionals, with a lightweight alignment module deployed to integrate existing question-answer corpora, interdisciplinary clinical contexts, and domain terminology databases. This module refines the original answers and simultaneously generates a supervised fine-tuning dataset (positive sample set) and a reinforcement learning preference dataset (preference sample set) as basic training data samples.

[0031] In some embodiments, a hybrid low-bit quantization of W4A8 or W4A16 is used to deploy the base model; where W4 indicates that the weights are quantized with 4 bits, A8 indicates that the activations are quantized with 8 bits, and A16 indicates that the activations are quantized with 16 bits. The multi-head attention triple projection of the base model is fused with the activation function and distributed in blocks according to the computation kernel. The key-value cache is managed at the page granularity to support multi-request concurrency, and when the length of the input sequence is less than a preset length threshold, it switches to a small window mode that disables the page-based key-value cache.

[0032] Specifically, the basic model is obtained through knowledge distillation, structured pruning, and parameter sharing, based on a general large language model with no less than a certain number of parameters, such as 70B.

[0033] For example, the base model is obtained from a general large language model through structured pruning and parameter sharing distillation. The parameter size is, for example, about 70 bytes. Through the above distillation strategy, while retaining the original model's preset percentage, for example, more than 90% of the task performance, the inference resource consumption can be reduced to a certain value, for example, the size of 34 bytes. The base model supports a maximum context length of no less than a certain number, for example, 4096 tokens, and a maximum number, for example, 6144 tokens. The inference stage adopts a key-value caching preheating mechanism and a zero-redundancy optimizer distributed optimization. The deployment stage adopts quantization strategies such as 4-bit weight quantization, which can realize multi-path concurrent inference on a single GPU (Graphics Processing Unit) with an average response latency of about 1 second.

[0034] In this embodiment, the model training process includes the following three stages: (1) Supervised Fine-Tuning (SFT) to align medical semantics; (2) Group Relative Policy Optimization (GRPO) incorporating expert feedback, such as... Figure 3 As shown, inputting a symptom cluster or the same query generates N candidate responses (N is a positive integer greater than 0) within each symptom cluster. The total reward is determined based on logical coherence reward, consistency with expert scores, and misdiagnosis penalty score (misdiagnosis penalty item for disease grading weight). Candidate responses for the same query are sorted within the group according to the total reward. A step reward strategy is adopted to update the model parameters, resulting in a trained large language model.

[0035] In this embodiment of the disclosure, a penalty weighting function for misdiagnosis of critical diseases is used to assign significantly higher penalty weights to missed or misdiagnosed critical diseases such as cerebral hemorrhage, acute large artery occlusion ischemic stroke, intracranial infection, and status epilepticus in the above-mentioned GRPO reward function, thereby further suppressing missed or misdiagnosis of critical diseases.

[0036] For example, a disease severity grading table: In one implementation, diseases are divided into three levels: CriticalSet, including CriticalSet1={cerebral hemorrhage, acute ischemic stroke with large artery occlusion, intracranial infection, status epilepticus}, with weight w1≥8; CriticalSet2={transient ischemic attack, acute peripheral neuropathy exacerbation, complex migraine, first-onset epilepsy}, with weight w2∈[3,6]; CriticalSet3={general headache, benign paroxysmal positional vertigo}, with weight w3=1; The above grades and weight values ​​can be selected and set according to the actual application scenario, but the grading structure and the inequality w1>w2>w3 remain unchanged.

[0037] Furthermore, a graded weighted misdiagnosis penalty term R is added to the aforementioned GRPO reward term. m R m = Σ i w i ·1[ ≠y i And y i ∈CriticalSet i ];in For model prediction and diagnosis, y i For the true diagnosis of this sample, 1[·] is the indicator function; the overall reward R = R_logic + R_expert + R mWhere R_logic is the logical coherence reward and R_expert is the consistency reward with expert rating.

[0038] Therefore, when introducing R m Subsequently, the missed diagnosis rate and misdiagnosis rate of CriticalSet1 diseases were significantly suppressed, and the overall accuracy and recall rate did not significantly degrade.

[0039] Furthermore, sensitivity analysis was conducted on several values ​​of w1 within the range of [6, 10]. It was observed that when w1=8, a good balance was achieved between the suppression of missed diagnoses of key diseases and the overall performance.

[0040] Therefore, the misdiagnosis rate can be greatly reduced by introducing the GRPO training mechanism based on misdiagnosis penalty.

[0041] For example, the neurology specialty fine-tuning employs a high-freeze LoRA (low-rank adaptive) strategy, updating only the input embedding layer and selected intermediate layers. This allows for specialty adaptation to be completed within 400 GPU hours, on, for example, approximately 2000-3000 neurology question-answer pairs. Figure 4 As shown, the question-and-answer text of the neurology specialty is input into the embedding layer (LoRA can be updated), then passes through the frozen Transformer Block 1 and Transformer Block 2, and then enters Transformer Block k (the selected intermediate layer that can be updated by LoRA). Then, Transformer Block k+1 to Transformer Block N are all in a frozen state. Finally, the projection layer (frozen state) is output, and the neurology specialty fine-tuned model is output as the supervised fine-tuned model.

[0042] It is understandable that a hybrid quantization approach is used to deploy the underlying model to meet the requirements of hospital-level computing power, including lower memory usage, lower first-token latency, higher concurrency, and continuous inference capabilities for edge scenarios. For example, such as... Figure 5 As shown, the domain foundation model is deployed using a hybrid low-bit quantization strategy of W4A8 (4-bit weights, 8-bit activation) or W4A16 (4-bit weights, 16-bit activation). In one implementation, W4 quantization is used for the linear layer weights, A8 quantization is used for the intermediate activations, and FP16 (16-bit half-precision floating-point) is retained for the multi-head attention Softmax (activation function) and Layer Norm (layer normalization).

[0043] Furthermore, operator fusion is performed by combining the QKV (Query, Key, Value) projections of multi-head attention triples into a single GEMM (General Matrix Multiplication) operator and fusing it with RoPE (Rotary Position Embedding) and Softmax, and then distributing it in blocks according to the computation kernel; FFN (Feed Forward Network) is fused with residual connections; and RMSNorm is fused with downstream linear layers to reduce memory access round trips.

[0044] Furthermore, KV-cache implements paging management. KV-cache stores data in high-bandwidth memory in pages of 16 or 32 tokens, supporting concurrent requests. When the request queue length exceeds the threshold Q0 (e.g., 8), a page-granular LRU (Least Recently Used) eviction policy is enabled.

[0045] Furthermore, when the length of the input sequence is less than the threshold L0 (e.g., 1024 tokens), the small window mode is switched to: page-based KV-cache is disabled, and continuous KV-cache and a more aggressive operator fusion strategy are used to reduce the latency of the first token.

[0046] Furthermore, for edge network outage disaster recovery inference, in pre-hospital emergency care (e.g., ambulance terminals) and primary hospitals (e.g., tiered diagnosis and treatment nodes), a lightweight version that has undergone secondary distillation (e.g., 7B or 13B distillation version) is deployed; in network outage or weak network conditions, local caching and local inference services are used to continuously provide preliminary diagnosis and treatment suggestions; after the network is restored, the case data generated by local inference is entered into the feedback pool and synchronization is triggered.

[0047] Furthermore, in terms of hardware applicability, the implementation of linear layer GEMM, attention block computation, and KV-cache paging management is not dependent on the hardware instruction set of any specific vendor. It can be adapted to various inference acceleration hardware that supports data types such as INT4 (4-bit integer) / INT8 (8-bit integer) / FP16, including but not limited to general-purpose or special-purpose inference accelerators such as GPU, NPU (Neural Processing Unit), and TPU (Tensor Processing Unit).

[0048] Therefore, through hybrid quantization engineering deployment, the basic model of the field can be deployed with hybrid low-bit quantization such as W4A8 or W4A16 under hospital-level computing power conditions. Attention operator fusion, KV-cache paging management and small window mode switching can achieve multi-request concurrency and low first token latency. Among them, the edge network outage disaster recovery inference submodule further supports continuous inference in pre-hospital emergency care and hierarchical diagnosis and treatment scenarios under weak network or network outage conditions.

[0049] Step 102: Adjust the trained large language model based on incremental training data to obtain candidate large language models.

[0050] Step 103: Obtain the text of the question to be processed. Process the text of the question to be processed simultaneously based on the trained large language model and the candidate large language model to obtain the inference answer text, the corresponding credibility score, and the candidate answer text; wherein, the credibility score is determined based on the symptom entropy and the model confidence estimation.

[0051] Step 104: When the confidence score is less than the confidence threshold corresponding to the preset disease seed group category, a preset fallback strategy is triggered; wherein, the preset fallback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision.

[0052] Step 105: Evaluate the candidate large language model based on the candidate answer text, and when the evaluation result meets the preset model switching conditions, replace the trained large language model with the candidate large language model to process the real-time question text.

[0053] In some embodiments, after each clinical interaction, initial diagnosis information, model output diagnosis information, AI-assisted diagnosis information, confirmed diagnosis information, and scoring information are collected to construct a five-tuple structured data as feedback samples. The feedback samples are scored and classified according to preset screening rules and used as incremental training data to store the training sample database. The preset screening rules include scoring thresholds, target diagnostic consistency, and misdiagnosis information.

[0054] Specifically, such as Figure 6 As shown, after each clinical interaction, initial diagnosis information, model-output diagnostic information, AI-assisted diagnostic information, confirmed diagnosis information, and scoring information are automatically collected. Pre-defined structured fields are serialized into feedback samples, and all fields undergo de-identification processing before being written into the feedback pool. Each feedback sample is scored according to a scoring threshold, target diagnostic consistency, and misdiagnosis information. For example, score = α·ExpertScore (scoring threshold) + β·1[ModelDx=FinalDx] (target diagnostic consistency). γ·MisdiagSeverityWeight (misdiagnosis information), where α, β, and γ are adjustable weights, such as α=0.5, β=1.0, and γ=1.5; MisdiagSeverityWeight can be used as the weight table for disease grading. i Samples with a score greater than the threshold θ_pos are entered into the positive sample pool, samples with a score less than the threshold θ_neg are entered into the adversarial sample pool, and other samples are treated as regular samples.

[0055] Furthermore, incremental training is triggered according to a predetermined cycle (e.g., once every two weeks) or a positive sample accumulation threshold (e.g., greater than or equal to 500 cases). Incremental training only updates the input embedding layer of Neuro-Aligner and high-freeze LoRA and selected intermediate layers, without updating the backbone parameters of the basic model. A single incremental training session takes no more than 100 GPU hours to obtain a candidate large language model.

[0056] Furthermore, the candidate large language model (V) obtained through incremental training n+1 ) and trained large language models (V n It runs in parallel in shadow mode; it performs inference on the same problem text separately, but only displays V to clinicians. n The reasoning answer text, V n+1 The candidate answer texts were recorded blinded and included in the controlled evaluation pool; when V n+1 When the preset model switching conditions, such as "diagnostic accuracy improvement ≥ 0.5% and misdiagnosis rate decrease ≥ 0.3%", are simultaneously met on both the validation set and the shadow control evaluation pool, the model will automatically switch to V. n+1 Replace V n Used for processing real-time problem text; otherwise, retain V. n V n+1 Rollback and retain data for use in the next cycle; if the evaluation window fails to meet the standard for N consecutive (e.g. 3) consecutive times, an adaptive adjustment of the training over-participation screening threshold is triggered; in addition, all model switching, rollback events and training sample sources are written to the audit log to support regulatory review.

[0057] In some embodiments, obtaining the text of the question to be processed includes: obtaining the original dialogue text by performing streaming speech recognition based on the doctor-patient spoken dialogue; extracting multiple structured fields from the original dialogue text based on an entity recognition model; filling the multiple structured fields into predefined clinical template slots; and repeatedly asking the user questions to fill in the missing key slots to obtain the text of the question to be processed.

[0058] Specifically, such as Figure 7As shown, streaming speech recognition is performed on the spoken dialogue between doctors and patients to obtain the original dialogue text. In one implementation, the streaming ASR (Automatic Speech Recognition) adopts an endpoint detection and acoustic model that is robust to hospital environmental noise, and supports Mandarin and common local accents. The speech transcription delay is no greater than, for example, 500ms. The following structured fields are extracted from the original dialogue text through an entity recognition model for neurological symptoms, such as symptoms (including location, nature, and onset), signs, onset time, accompanying symptoms, medication history, past medical history, and allergy history. The entity recognition model is trained on a subset of the aforementioned de-identified corpus that has been manually annotated at the entity level.

[0059] Furthermore, the extracted fields are filled into predefined clinical template slots; when key slots (such as onset time, accompanying symptoms, and history of cerebrovascular disease) are missing, a follow-up question is initiated to the physician via interface or voice; the completed clinical template is used as the text of the question to be processed.

[0060] It should be noted that the voice clinical input pipeline supports fully localized deployment; the raw voice data does not leave the local node after transcription, and the transcribed text and extracted fields undergo de-identification processing before being written to the feedback pool, further meeting user needs.

[0061] Thus, through the end-to-end clinical voice input pipeline, the system enables streaming input of emergency room doctor-patient spoken dialogues, extraction of neurological symptom entities, template slot filling, and reverse questioning of missing slots. This reduces the input burden on physicians and completes the necessary fields for decision-making, further improving the accuracy of assisted diagnostic processing based on a large language model.

[0062] In this embodiment of the disclosure, the text to be processed is processed simultaneously based on the trained large language model and the candidate large language model, and the resulting inference answer text and candidate answer text both include symptom analysis, differential diagnosis, further examination suggestions and conclusions.

[0063] Specifically, such as Figure 8 As shown, the model follows the explicit four-step chain of thought (CoT) framework during the decoding phase, sequentially outputting symptom analysis (key symptom extraction, onset time and progression pattern), differential diagnosis (candidate diagnosis list and probability of each candidate), further examination suggestions (impact, laboratory tests, neuroelectrophysiology, etc.), and conclusion (final diagnosis and risk label, such as low, medium, and high). The output is temperature-calibrated, and the diagnosis is quantified based on attention-weighted uncertainty and misdiagnosis risk, providing clinicians with diagnostic suggestions with explicit risk labels.

[0064] It should be noted that the inputs to this embodiment may include transcripts of medical records, physical examination findings, laboratory tests, imaging reports, and demographic information. The output includes text containing independent reasoning answers as a preliminary diagnosis, along with the reasoning process, examination suggestions, and treatment recommendations. It supports a four-stage diagnostic comparison: initial physician diagnosis; independent model diagnosis; AI-assisted diagnosis after the physician reads the model's reasoning; final diagnosis by an expert panel; strict isolation between model reasoning and the final diagnosis, pairwise consistency testing, and scoring; and finally, a comparison between the model and the final diagnosis information. Figure 9 As shown.

[0065] In some embodiments, the model confidence estimate and symptom entropy corresponding to the trained large language model are obtained, and an initial confidence score is obtained based on the model confidence estimate and symptom entropy.

[0066] In some embodiments, a preset disease seed group category corresponding to the reasoning answer text is obtained, and a confidence threshold corresponding to the preset disease seed group category is obtained. When the confidence score is less than the confidence threshold, one or more of the following preset fallback strategies are triggered: only outputting differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision.

[0067] Specifically, such as Figure 10 As shown, diseases are divided into predefined subgroups. ={C1, C2, ..., C n In one embodiment, it includes: C1 acute ischemic stroke, C2 transient ischemic attack, C3 cerebral hemorrhage, C4 epilepsy, C5 vertigo, C6 headache, C7 intracranial infection, C8 peripheral neuropathy, and C9 other neurological diseases, totaling 9 categories; each subgroup C i Associate a confidence threshold τ i The initial values ​​were derived by back-calculating the physician independent diagnosis accuracy and the model independent diagnosis accuracy of this subgroup in the prospective study. Based on the observation of the "AI-assisted negative effect" phenomenon in subgroups C8 and C9 in historical studies, the initial values ​​τ8 and τ9 of these two subgroups were set to be higher than those of other subgroups.

[0068] Specifically, for each inference, the model outputs the raw confidence score (model confidence estimate) p (a normalized value from the last layer's Softmax probability) and the symptom entropy H (a bag-of-words entropy estimate of the patient's subject feature vector); the credibility score is estimated using the following formula: CredibilityScore = p λ·H, where λ is a learnable parameter with a value range of [0, 1], such as 0.3; in actual inference scenarios, when CredibilityScore < τ i Time trigger for C i rollback strategy R(C)i R1 only outputs differential diagnosis and examination suggestions, not the final conclusion; R2 explicitly prompts "It is recommended that physicians make independent judgments" in the AI-assisted interface; R3 reduces the display weight of AI suggestions in the final decision suggestion (e.g., presenting them in a light color or in a collapsed manner); R1, R2, and R3 can be used individually or in combination; for subgroups C8 and C9, the default is to enter the R1+R2 combination fallback state; each fallback event, the physician's final decision, and the comparison results of that decision with the expert diagnosis are all fed back into the aforementioned feedback pool for online learning and updating λ and τ. i .

[0069] It should be noted that the embodiments disclosed herein support two input modes: a standard version (containing only medical history, physical examination, and auxiliary examinations) and an extended version (additionally including gender, age, education level, and medical insurance type). The differences in recommendations across the four dimensions of examination, treatment, rehabilitation, and health education under the two input modes are compared to achieve personalized recommendations based on sociodemographic characteristics.

[0070] It should be noted that the model evaluation disclosed herein can be based on automated benchmarking and real-world double-blind human scoring; a minimum number of neurologists will conduct a pre-set five-level scoring (comprehensiveness, accuracy, clinical applicability) of the model and physician outputs; statistical analysis will employ the McNemar test, Friedman rank test, and Wilcoxon signed-rank test with Bonferroni correction, and multivariate logistic regression will be used to evaluate the interaction effects of education level, health insurance type, and disease category on changes in recommendations; the training set and evaluation set will be strictly segregated.

[0071] As an example of a scenario, such as Figure 11 As shown, the end-to-end voice input pipeline acquires doctor-patient spoken dialogues, structured clinical input (medical history / physical examination / laboratory tests / imaging / demographics), and the basic model includes a certain size, such as 70B distillation + structured pruning; triple alignment / Neuro-Aligner; GRPO enhanced alignment (including k-level misdiagnosis penalty); high-freeze LoRA specialist fine-tuning; four-step explicit CoT + risk labeling in the inference and expression modules; subgroup dynamic confidence threshold human-machine collaborative backtracking; structured output diagnosis / examination / treatment / health education; clinical closed-loop feedback learning, five-tuple collection (physician initial diagnosis / model independent / AI-assisted / expert diagnosis / score); screening, scoring and feedback pool, incremental training with a shadow version of the model; parallel operation of two models and dual threshold switching; hybrid quantization engineering deployment, W4A8 / W4A16 hybrid quantization, QKV+RoPE+Softmax operator fusion, KV-cache pagination, small window mode and edge network outage disaster recovery inference.

[0072] Therefore, in addition to known technologies such as basic model (domain-specific) distillation, triplet alignment of question, original answer and expert-corrected answer, relative strategy optimization of misdiagnosis penalty group, high-freeze low-rank adaptation, explicit four-step thinking chain output and risk labeling, and sociodemographic responsive recommendation, it further includes differentiated technology modules such as clinical closed-loop feedback learning module, subgroup dynamic confidence threshold human-machine collaboration module, key disease misdiagnosis penalty graded weight function, hybrid quantization engineering deployment module, and end-to-end speech clinical input pipeline. This enables sustainable evolution, reliable human-machine collaboration, high-throughput low-latency deployment and end-to-end speech input engineering closed loop under real-world clinical conditions.

[0073] Specifically, the output diagnosis (inference answer text) and recommendation are presented in the form of an interpretable structured report. The report includes at least: a collapsible segmented display of the four-step thinking chain, color-coded and differentiated display of risk labels, visual prompts of the triggering of subgroup dynamic confidence thresholds, and a link to the closed-loop feedback learning module for the subsequent follow-up of the case. The report can be presented to clinicians via mobile terminals.

[0074] It should be noted that the base model can be replaced with any open-source or self-developed large language model with ≥7B parameters; the quantization scheme can be replaced with INT8 (8-bit integer quantization) / FP8 (8-bit floating-point quantization) / AWQ (activation-aware weighted quantization), etc.; the acceleration framework can be replaced with TensorRT-LLM (a dedicated inference acceleration framework for large models), SGLang (Structured Generation Language), etc.; the number of CoT steps can be extended to five or six steps; the specialty data can be extended to other emergency scenarios such as cardiovascular emergencies, respiratory emergencies, and digestive emergencies; the sociodemographic dimension can be extended to income, language, region, and cultural background; and the assessment scenario can be extended to telemedicine, hierarchical diagnosis and treatment, and pre-hospital emergency care.

[0075] In summary, this disclosed embodiment, building upon known improvements such as basic model distillation, alignment of triples (question, original answer, and expert-corrected answer), group-relative strategy optimization with misdiagnosis penalty, high-freeze low-rank adaptation, explicit four-step thinking chain output and risk labeling, and sociodemographic responsive recommendation, further utilizes clinical closed-loop feedback learning to structurally backflow the five-tuple of "physician initial diagnosis, model independence, AI assistance, expert diagnosis, and expert scoring" and perform parallel evaluation and dual threshold switching on the trained large language model using a candidate large language model; the subgroup dynamic confidence threshold human-machine collaboration module addresses the potential for AI assistance to cause negative consequences. The system includes a default fallback strategy for subgroups of effects; a weighted function for grading penalties for misdiagnosis of key diseases, which classifies diseases into different levels according to severity and reflects them with differentiated weights in the GRPO reward function; a hybrid quantization engineering deployment model, which provides hybrid low-bit quantization such as W4A8 or W4A16, attention operator fusion, KV-cache paging management, small window mode switching, and edge network outage disaster recovery inference; and an end-to-end speech clinical input pipeline. Ultimately, this system achieves a closed-loop engineering system for the sustainable evolution of the medical large language model under real-world clinical conditions, controllable subgroups, efficient engineering deployment, and end-to-end speech input.

[0076] Corresponding to the aforementioned auxiliary diagnostic method based on a large language model, this disclosure provides an auxiliary diagnostic device based on a large language model. Figure 12 This is a schematic diagram of a large language model-based auxiliary diagnostic device provided in an embodiment of this disclosure. The device can be implemented by software and / or hardware, and is generally integrated into an electronic device, such as... Figure 12 As shown, the large language model-based auxiliary diagnostic device includes: The first acquisition module 1201 is used to acquire a trained large language model trained based on basic training data samples; The second acquisition module 1202 is used to acquire incremental training data; wherein, the incremental training data is generated based on the five-tuple structure data, and the five-tuple structure data includes initial diagnosis information, model output diagnosis information, artificial intelligence-assisted diagnosis information, confirmed diagnosis information and scoring information. The adjustment module 1203 is used to adjust the trained large language model based on the incremental training data to obtain a candidate large language model; The third acquisition module 1204 is used to acquire the text of the problem to be processed; The first processing module 1205 is used to process the text of the question to be processed simultaneously based on the trained large language model and the candidate large language model to obtain the inference answer text, the corresponding credibility score, and the candidate answer text; wherein the credibility score is determined based on symptom entropy and model confidence estimation. The second processing module 1206 is used to trigger a preset rollback strategy when the confidence score is less than the confidence threshold corresponding to the preset disease seed group category; wherein, the preset rollback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision. The third processing module 1207 is used to evaluate the candidate large language model based on the candidate answer text, and when the evaluation result meets the preset model switching conditions, replace the trained large language model with the candidate large language model to process the real-time question text.

[0077] Optionally, the first acquisition module 1201 is specifically used for: acquiring a triplet aligned dataset consisting of a question, the original answer, and the corrected answer as the basic training data sample; performing supervised fine-tuning on a preset basic model based on the basic training data sample, and then performing group-relative strategy optimization on the basis of supervised fine-tuning to obtain the trained large language model; wherein, the group-relative strategy optimization includes generating multiple candidate responses within each symptom cluster and ranking them within the group and awarding step rewards based on logical coherence, consistency with the target score, and misdiagnosis penalty score, wherein the misdiagnosis penalty score includes a disease grading weight misdiagnosis penalty item.

[0078] Optionally, the base model is deployed using a hybrid low-bit quantization of W4A8 or W4A16; where W4 indicates that the weights are quantized with 4 bits, A8 indicates that the activations are quantized with 8 bits, and A16 indicates that the activations are quantized with 16 bits. The multi-head attention triple projection of the base model is fused with the activation function and distributed in blocks according to the computation kernel. The key-value cache is managed at the page level to support multi-request concurrency, and when the length of the input sequence is less than a preset length threshold, it switches to a small window mode that disables the page-based key-value cache.

[0079] Optionally, the second acquisition module 1202 is used to: after each clinical interaction, collect the initial diagnosis information, the model output diagnosis information, the artificial intelligence-assisted diagnosis information, the confirmed diagnosis information, and the scoring information to construct the five-tuple structure data as feedback samples; score and classify the feedback samples according to preset screening rules as incremental training data to store the training sample database; wherein, the preset screening rules include scoring threshold, target diagnosis consistency, and misdiagnosis information.

[0080] Optionally, the third acquisition module 1204 is specifically used for: obtaining the original dialogue text by performing streaming speech recognition based on the doctor-patient spoken dialogue; extracting multiple structured fields from the original dialogue text based on an entity recognition model; filling the multiple structured fields into predefined clinical template slots, and repeatedly asking the user questions to fill in the missing key slots, thereby obtaining the text of the question to be processed.

[0081] Optionally, the first processing module 1205 is further configured to: obtain the model confidence estimate and symptom entropy corresponding to the trained large language model; and estimate based on the model confidence estimate and the symptom entropy to obtain an initial confidence score.

[0082] Optionally, the second processing module 1206 is specifically used to: obtain the preset disease seed group category corresponding to the reasoning answer text, and obtain the confidence threshold corresponding to the preset disease seed group category; when the confidence score is less than the confidence threshold, trigger one or more of the following as the preset fallback strategy: only output differential diagnosis and examination suggestions, explicitly prompt the user to make independent judgments, and reduce the display weight of artificial intelligence suggestions in the target decision.

[0083] The auxiliary diagnostic processing device based on a large language model provided in this disclosure can execute the auxiliary diagnostic processing method based on a large language model provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects of the method execution.

[0084] This disclosure also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implements the auxiliary diagnostic processing method based on a large language model provided in any embodiment of this disclosure.

[0085] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0086] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0087] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0088] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the auxiliary diagnostic processing method based on a large language model as described in the foregoing embodiments.

[0089] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including but not limited to object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0090] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0091] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0092] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0093] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0094] According to one or more embodiments of this disclosure, this disclosure provides an electronic device, including: processor; Memory used to store the processor's executable instructions; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the auxiliary diagnostic processing method based on a large language model as provided in this disclosure.

[0095] According to one or more embodiments of the present disclosure, the present disclosure provides a computer-readable storage medium storing a computer program for performing an auxiliary diagnostic processing method based on a large language model as described in any of the present disclosure.

[0096] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0097] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0098] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for auxiliary diagnostic processing based on a large language model, characterized in that, include: Obtain a trained large language model and incremental training data based on basic training data samples; wherein, the incremental training data is generated based on five-tuple structure data, which includes initial diagnosis information, model output diagnosis information, artificial intelligence-assisted diagnosis information, confirmed diagnosis information, and scoring information; The trained large language model is adjusted based on the incremental training data to obtain a candidate large language model; The text of the question to be processed is obtained, and the text of the question to be processed is processed simultaneously based on the trained large language model and the candidate large language model to obtain the inference answer text, the corresponding credibility score, and the candidate answer text; wherein, the credibility score is determined based on symptom entropy and model confidence estimation. When the confidence score is less than the confidence threshold corresponding to the preset disease seed group category, a preset rollback strategy is triggered; wherein, the preset rollback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision; The candidate large language model is evaluated based on the candidate answer text, and when the evaluation result meets the preset model switching conditions, the candidate large language model replaces the trained large language model for processing real-time question text.

2. The method according to claim 1, characterized in that, The process of obtaining a trained large language model based on basic training data samples includes: The aligned dataset consisting of a triplet of question, original answer, and corrected answer is used as the basic training data sample. Based on the aforementioned basic training data samples, the preset basic model is subjected to supervised fine-tuning. On the basis of supervised fine-tuning, a group-relative strategy optimization is performed to obtain the trained large language model. The group-relative strategy optimization includes generating multiple candidate responses within each symptom cluster and ranking them within the group and awarding step rewards based on logical coherence, consistency with the target score, and misdiagnosis penalty score. The misdiagnosis penalty score includes a disease grading weighted misdiagnosis penalty item.

3. The method according to claim 2, characterized in that, The basic model is deployed using a hybrid low-bit quantization of W4A8 or W4A16; where W4 indicates that the weights are quantized with 4 bits, A8 indicates that the activations are quantized with 8 bits, and A16 indicates that the activations are quantized with 16 bits. The basic model uses multi-head attention to fuse attention triplet projections with activation functions and distribute them in blocks according to computational kernels. It manages key-value cache at the page level to support multiple concurrent requests, and switches to a small window mode that disables page-based key-value caching when the length of the input sequence is less than a preset length threshold.

4. The method according to claim 1, characterized in that, Obtaining the incremental training data includes: After each clinical interaction, the initial diagnosis information, the model output diagnosis information, the artificial intelligence-assisted diagnosis information, the confirmed diagnosis information, and the scoring information are collected to construct the five-tuple structure data as feedback samples; The feedback samples are scored and categorized according to preset screening rules and used as incremental training data to store the training sample database; wherein, the preset screening rules include scoring thresholds, target diagnostic consistency, and misdiagnosis information.

5. The method according to claim 1, characterized in that, The process of obtaining the text of the problem to be processed includes: The original dialogue text is obtained by streaming speech recognition based on the spoken dialogue between doctors and patients; Multiple structured fields are extracted from the original dialogue text based on an entity recognition model; The multiple structured fields are filled into predefined clinical template slots, and the missing key slots are repeatedly filled by asking the user questions to obtain the text of the problem to be processed.

6. The method according to claim 1, characterized in that, Obtaining the credibility score includes: Obtain the model confidence estimate and symptom entropy corresponding to the trained large language model; An initial confidence score is obtained by estimating the model confidence level and the symptom entropy.

7. The method according to claim 1, characterized in that, When the confidence score is less than the confidence threshold corresponding to the preset disease seed group category, a preset fallback strategy is triggered, including: Obtain the preset disease seed group category corresponding to the reasoning answer text, and obtain the confidence threshold corresponding to the preset disease seed group category; When the confidence score is less than the confidence threshold, one or more of the following are triggered as the preset fallback strategy: only output differential diagnosis and examination suggestions, explicitly prompt the user to make independent judgments, and reduce the display weight of artificial intelligence suggestions in the target decision.

8. An auxiliary diagnostic processing device based on a large language model, characterized in that, The device includes: The first acquisition module is used to acquire a trained large language model based on basic training data samples; The second acquisition module is used to acquire incremental training data; wherein, the incremental training data is generated based on the five-tuple structure data, and the five-tuple structure data includes initial diagnosis information, model output diagnosis information, artificial intelligence-assisted diagnosis information, confirmed diagnosis information and scoring information. An adjustment module is used to adjust the trained large language model based on the incremental training data to obtain a candidate large language model; The third acquisition module is used to acquire the text of the problem to be processed; The first processing module is used to process the text of the question to be processed simultaneously based on the trained large language model and the candidate large language model to obtain the inference answer text, the corresponding credibility score, and the candidate answer text; wherein, the credibility score is determined based on symptom entropy and model confidence estimation. The second processing module is used to trigger a preset rollback strategy when the confidence score is less than the confidence threshold corresponding to the preset disease seed group category; wherein, the preset rollback strategy includes one or more of the following: outputting only differential diagnosis and examination suggestions, explicitly prompting the user to make independent judgments, and reducing the display weight of artificial intelligence suggestions in the target decision; The third processing module is used to evaluate the candidate large language model based on the candidate answer text, and when the evaluation result meets the preset model switching conditions, replace the trained large language model with the candidate large language model to process the real-time question text.

9. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store the processor's executable instructions; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the auxiliary diagnostic processing method based on a large language model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program for executing the auxiliary diagnostic processing method based on a large language model as described in any one of claims 1-7.

Images