CN-121983340-A - Clinical special disease prediction method and device based on large model instruction prompt

CN121983340ACN 121983340 ACN121983340 ACN 121983340ACN-121983340-A

Abstract

The application relates to the technical field of medical big data processing, and provides a method and a device for predicting clinical special diseases based on big model instruction prompt. According to the method, firstly, a triplet event unit is extracted from clinical special disease EHR data, semantic coding is conducted on the triplet event unit to obtain clinical event coding features, secondly, a QA problem pair is built and a natural language prompt instruction is constructed based on the triplet event unit, the natural language prompt instruction is subjected to embedded coding to obtain instruction embedded expression, then the clinical event coding features and the instruction embedded expression are aligned in dimension and semantic alignment, finally, a bimodal mixed embedded sequence is obtained by splicing the clinical event coding features and the instruction embedded expression, a first large model is input to obtain a clinical special disease prediction result, accuracy of extracting clinical special disease related information from the EHR data is improved, a high-quality instruction sample adapting to a clinical special disease scene can be generated, a high-quality bimodal mixed embedded sequence is obtained, and classification prediction accuracy is further improved.

Inventors

WU YAJING
TANG YONGQIANG
ZHONG YUQI
ZHANG WENSHENG

Assignees

中国科学院自动化研究所

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The method for predicting the clinical special disease based on the large model instruction prompt is characterized by comprising the following steps of: Acquiring EHR data of a structured electronic health record of a clinical specific disease; extracting a triple event unit from the clinical special disease EHR data, wherein the triple event unit comprises a unique identifier, an event type and an event value; Carrying out semantic coding on the triplet event unit to obtain clinical event coding characteristics; constructing a QA question pair based on the triplet event unit, and constructing a natural language prompt instruction by utilizing the QA question pair; Performing embedded coding on the natural language prompt instruction to obtain an instruction embedded expression; Projecting the clinical event coding features by using a linear projector so as to conduct dimension alignment and semantic alignment on the projected clinical event coding features and the instruction embedded expression; Orderly splicing projected clinical event coding features and instruction embedded expressions to obtain a bimodal mixed embedded sequence; And predicting based on the bimodal mixed embedded sequence by using a first large model to obtain a clinical disease prediction result.
2. The method of claim 1, wherein extracting the triplet event element from the clinical suites EHR data comprises: Preprocessing the EHR data of the clinical special diseases, wherein the preprocessing at least comprises time sequence alignment and outlier rejection; performing multi-table association integration on the EHR data of the clinical special diseases through the unique identifier, and splitting the data in each table word by word segment to obtain the triplet event unit; wherein the unique identifier comprises at least one of a patient unique identifier and an admission unique identifier.
3. The large model instruction hint based clinical suites prediction method of claim 1, wherein constructing QA problem pairs based on the triplet event unit comprises: Acquiring a preset QA problem pair template; taking the event type and the event value in the target triplet event unit as the problem of a preset QA problem pair template to obtain an initial question-answer pair; invoking a second large model to perform semantic rewriting on the answers in the initial question-answer pair to obtain rewritten answers, wherein the rewritten answers have medical semantic consistency with the answers in the initial question-answer pair; Combining the target triplet event unit and the rewritten answer to obtain a QA question pair of the target triplet event unit; the target triplet event unit is any triplet event unit.
4. The method for predicting clinical suites based on large model instruction hints according to claim 1, wherein semantically encoding the triplet event unit to obtain clinical event encoding features comprises: combining all triad event units of the patient to obtain a clinical event stream; and inputting each triplet event unit in the clinical event stream into a pre-training encoder one by one, and modeling and integrating the internal semantics of the event through a multi-layer attention mechanism to obtain high-dimensional continuous clinical event coding characteristics.
5. The method for predicting clinical suites based on large model instruction cues as set forth in claim 1, wherein constructing natural language cues using the QA problem pair comprises: determining role definitions based on the target QA question pairs, wherein the role definitions comprise artificial intelligent reasoning assistant identity definitions with clinical special disease medical expertise; determining an input information guide, wherein the input information guide at least comprises a guide language and a placeholder, and the content in the placeholder is determined based on the event type and the event value of a target QA question pair; Determining a clinical target definition, wherein the clinical target definition comprises a clinical feature extraction target constraint, and the target constraint at least comprises a clinical index type and a time node; Determining an output format constraint; combining role definition, input information guidance, clinical target definition and output format constraint to obtain a natural language prompt instruction of a target QA problem pair; wherein the target QA question pair is any QA question pair.
6. The method of claim 1, wherein projecting the clinical event encoded features using a linear projector comprises: Inputting clinical event encoding features into a linear projector; transforming the clinical event coding feature by using a weight matrix of the linear projector to obtain a clinical event coding feature projection value; updating the weight matrix in response to determining that the clinical event encoded feature projection values and the instruction embedded expression do not meet the alignment target; iteratively executing the steps of transforming the clinical event coding feature again by using the updated weight matrix to obtain an updated clinical event coding feature projection value, and updating the weight matrix again in response to determining that the updated clinical event coding feature projection value and the instruction embedded expression do not meet the alignment target until the updated clinical event coding feature projection value and the instruction embedded expression meet the alignment target; Wherein the alignment targets include dimension alignment and semantic alignment.
7. The method of claim 1, wherein the first large model includes at least a classification head that determines a clinical suites prediction outcome based on a bimodal mixed embedded sequence; The first training stage uses natural language prompt instructions in the clinical disease field to carry out field self-adaptive training on the general large model, and the second training stage uses EHR data in the clinical disease field to carry out task-oriented training on the classification head; the first large model adopts cross entropy loss as a loss function in a training stage, model parameters are updated by using back propagation until convergence, and the adjustment parameters when the model parameters are updated comprise at least one of learning rate, training round number, characteristic dimension of the model, encoder layer number or attention head number.
8. A clinical specific disease prediction device based on large model instruction prompt, characterized by comprising: The acquisition module is configured to acquire the EHR data of the structured electronic health record of the clinical specific disease; An extraction target configured to extract a triplet event unit from clinical special disease EHR data, the triplet event unit comprising a unique identifier, an event type, and an event value; The semantic coding module is configured to carry out semantic coding on the triplet event unit to obtain clinical event coding characteristics; a construction module configured to construct a QA problem pair based on the triplet event unit, and construct a natural language prompt instruction using the QA problem pair; The embedded coding module is configured to carry out embedded coding on the natural language prompt instruction to obtain an instruction embedded expression; A bimodal alignment module configured to project the clinical event coding features using a linear projector to dimension align and semantically align the projected clinical event coding features and the instruction embedded representation; the bimodal alignment module is further configured to splice the projected clinical event coding features and the command embedded expressions in order to obtain a bimodal mixed embedded sequence; And the prediction module is configured to predict based on the bimodal mixed embedded sequence by using a first large model to obtain a clinical specific disease prediction result.
9. An electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of the method for clinical indicated disease prediction based on large model instruction cues as claimed in any one of claims 1 to 7.
10. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the method for predicting clinical comic based on large model instruction cues as claimed in any one of claims 1 to 7.

Description

Clinical special disease prediction method and device based on large model instruction prompt Technical Field The application relates to the technical field of medical big data processing, in particular to a method and a device for predicting clinical special diseases based on big model instruction prompt. Background Clinical specialization is a standardized full-period diagnosis and treatment system constructed by modern medical institutions around specific disease integration multidisciplinary resources on the basis of traditional clinical specialization subdivision. Along with the increase of the accurate medical demands, the clinical special disease mode with diseases as cores is widely applied. In this context, the early prediction of disease progression and adverse outcome based on structured electronic health records (Electronic Health Record, EHR) is of great importance for achieving early risk stratification and improving patient prognosis. Real world EHRs are typically composed of tables, sequences, and multi-source clinical events, have high dimensions, strong heterogeneity, and complex time dependence, and are rich in disease evolution information. In the related art, a support vector machine, a random forest and other traditional machine learning methods are often used for disease diagnosis and risk prediction. However, the method has obvious defects in processing high-dimensional and multi-source heterogeneous real world EHR data, such as the need of a great deal of expert knowledge for manual feature extraction, low efficiency, difficulty in capturing complex nonlinear correlations in long-sequence clinical events, lack of reasoning capability on medical deep logic, and difficulty in adapting to complex clinical special disease scenes. Some solutions attempt to introduce large language models (Large Language Model, LLM) for EHR data processing. Compared with the traditional machine learning method, LLM can capture complex association between clinical events under the condition of less artificial feature engineering, so that greater application potential is shown in clinical disease prediction and risk assessment. However, when LLM is used for HER data processing, there are problems that because LLM has a context length limitation, it is difficult to fully utilize long-sequence EHR, possibly causing key information to be hidden or lost, general LLM often lacks deep adaptation to specialized medical knowledge, and it is difficult to accurately capture complex association and development rules specific to a disease in clinical specific diseases, so that suboptimal prediction performance is shown in a real-world clinical scene, and labeling data of clinical specific diseases in the real world requires a large amount of manpower and material resources, so that labeling data is generally limited, and difficulty in training and generalizing a prediction model is further increased. Therefore, there is a need for an efficient and accurate LLM-based prediction method for clinical specific diseases. Disclosure of Invention In view of the above, the embodiment of the application provides a method and a device for predicting a clinical specific disease based on large model instruction prompt, which are used for solving the problems that in the prior art, high-dimensional heterogeneous characteristics of a structured electronic health record are difficult to effectively model, and labeling data in a clinical specific disease scene are scarce and are not suitable for the field of general LLM, so that specific disease prediction is difficult to accurately perform. In a first aspect of an embodiment of the present application, a method for predicting a clinical specific disease based on large model instruction hint is provided, including: Acquiring EHR data of a structured electronic health record of a clinical specific disease; The method comprises the steps of extracting a triplet event unit from clinical special disease EHR data, wherein the triplet event unit comprises a unique identifier, an event type and an event value; carrying out semantic coding on the triplet event units to obtain clinical event coding characteristics; Constructing a QA (Question Answer) Question pair based on the triplet event unit, and constructing a natural language prompt instruction by utilizing the QA Question pair; embedding and encoding the natural language prompt instruction to obtain an instruction embedded expression; Projecting the clinical event coding features by using a linear projector so as to carry out dimension alignment and semantic alignment on the projected clinical event coding features and the embedded expression of instructions; Orderly splicing projected clinical event coding features and instruction embedded expressions to obtain a bimodal mixed embedded sequence; and predicting based on the bimodal mixed embedded sequence by using the first large model to obtain a clinical disease prediction result. In a