An automated method for perioperative patient consultation note generation

Abstract

The application provides a method for automatic generation of perioperative patient consultation opinions, comprising: extracting key patient indicators from electronic medical records using natural language processing technology; cleaning and standardizing the key patient indicators to ensure data format consistency and obtaining preprocessed indicator data; mapping the preprocessed indicator data to the parameter format of a standardized risk assessment scale to obtain structured patient data; calling multiple clinically verified risk assessment scales to evaluate the structured patient data and generating a comprehensive risk assessment score based on the evaluation results; classifying the comprehensive risk of the patient according to the comprehensive risk assessment score to obtain a comprehensive risk assessment result; and converting the comprehensive risk assessment result into consultation opinion text using a generative large language model. The application significantly improves the efficiency and accuracy of perioperative patient risk assessment and consultation opinion generation by combining natural language processing, machine learning and generative large language models.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical informatization technology, in particular to an automatic generation method for perioperative patient consultation opinions. BACKGROUND

[0002] With the progress of medical technology and the intensification of population aging, the number of surgeries increases year by year, especially the proportion of elderly and high-risk patients significantly increases. This leads to a significant increase in perioperative risk assessment and consultation, and puts forward higher requirements for the quality of risk assessment and consultation. However, there are the following problems in the prior art:

[0003] Growth of surgical volume and increase of patient complexity: the growth of surgical volume and the increase of patient complexity lead to the increasing demand for preoperative anesthesia consultation, while the relative shortage of anesthesiologists leads to the shortage of medical resources.

[0004] Heavy workload of anesthesiologists: anesthesiologists need to handle a large number of preoperative assessments of patients, which leads to limited effective consultation time and insufficient consultation quality.

[0005] Risk assessment scales vary: there are many types of existing risk assessment scales, but doctors have different levels of mastery, which leads to insufficient consistency and accuracy of assessment results.

[0006] Heterogeneity of consultation opinions: the format and content of consultation opinions lack standardization, which affects the efficiency of shift handover and the accuracy of information transmission.

[0007] Therefore, there is an urgent need for a method for automatically generating perioperative patient consultation opinions to improve the efficiency and accuracy of risk assessment, reduce the workload of doctors, and ensure the standardization and consistency of consultation opinions. SUMMARY

[0008] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide an automatic generation method for perioperative patient consultation opinions, which aims to improve the efficiency and accuracy of perioperative patient risk assessment and consultation opinion generation through intelligent technology.

[0009] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0010] An automatic generation method for perioperative patient consultation opinions, comprising:

[0011] extracting key patient indicators from electronic medical records using natural language processing technology;

[0012] cleaning and standardizing the key patient indicators to ensure data format consistency, obtaining preprocessed indicator data;

[0013] mapping the pre-processed index data to a parameter format of a standardized risk assessment scale to obtain structured patient data;

[0014] calling a plurality of clinically validated risk assessment scales to evaluate the structured patient data, and generating a comprehensive risk assessment score according to the evaluation results;

[0015] based on a machine learning algorithm, the comprehensive risk assessment score is used to stratify the comprehensive risk of the patient, and a comprehensive risk assessment result is obtained;

[0016] using a generative large language model to convert the comprehensive risk assessment result into a standardized consultation opinion text.

[0017] Preferably, the key patient indicators include age, gender, height, weight, BMI, medical history and auxiliary examination results.

[0018] Preferably, the key patient indicators are extracted from the electronic medical record using natural language processing technology, including:

[0019] segmenting the free text in the electronic medical record into words;

[0020] annotating the part of speech for each word;

[0021] using a named entity recognition model in the medical field to identify medical entities in the free text, and restoring the words to root form;

[0022] text classification is performed on the root form in the free text to obtain classification data;

[0023] using a trained named entity recognition (NER) model to extract the key patient indicators in the classification data.

[0024] Preferably, the text classification is performed on the root form in the free text to obtain classification data, including:

[0025] inputting the root form in the free text into a text context relationship feature extraction layer to extract context relationship feature information;

[0026] inputting the root form in the free text into a global feature extraction layer to extract global feature information;

[0027] fuse the context relationship feature information and the global feature information to obtain fused text features;

[0028] use the fused text features to construct a loss function;

[0029] constantly optimize the loss function to obtain a text classification model;

[0030] The text classification model is used to complete text classification, and the classification data is obtained.

[0031] Preferably, the root form in the free text is input into a text context relationship feature extraction layer to extract context relationship feature information, including:

[0032] The pre-trained language model is used to extract initial feature information of the root form in the free text;

[0033] The initial feature information is input into a forward gate recurrent unit and a reverse gate recurrent unit;

[0034] The outputs of the forward gate recurrent unit and the reverse gate recurrent unit are spliced to obtain context relationship feature information; the expression of the context relationship feature information is: wherein H t ={H1,H2,…H l} represents the initial feature information input at time t, represents the output of the forward gate recurrent unit at time t, represents the output of the reverse gate recurrent unit at time t, and GRU represents a gate recurrent unit, and G represents context relationship feature information.

[0035] Preferably, the expression of the global feature information is:

[0036] c i =f(ω●H+b)

[0037]

[0038] In the formula, f is an activation function, ω is a convolution kernel, h is a convolution kernel size, b is a bias value, c i is the i-th extracted feature vector of the convolution layer, represents the value of the features extracted by three different convolution kernels after passing through a max-pooling layer, and C represents the extracted global feature information.

[0039] Preferably, the risk assessment scale includes: ASA grading table, Child-Pugh score table, NYHA cardiac function grading table, RCRI modified cardiac risk assessment table, ARISCAT postoperative pulmonary complication risk score table, and Essen stroke risk score table.

[0040] Preferably, a plurality of clinically verified risk assessment scales are called to evaluate the structured patient data, and a comprehensive risk assessment score is generated according to the evaluation results, including:

[0041] The scoring algorithm of each scale is called according to the structured patient data to calculate the scoring results of each scale;

[0042] generating the comprehensive risk assessment score according to the score results based on a weighted average algorithm.

[0043] Preferably, the comprehensive risk of the patient is classified according to the comprehensive risk assessment score based on a machine learning algorithm to obtain a comprehensive risk assessment result, including:

[0044] obtaining a preset comprehensive risk assessment score data set;

[0045] constructing an initial neural network deep learning model;

[0046] training the initial neural network deep learning model according to the comprehensive risk assessment score data set to obtain a trained classifier;

[0047] connecting a trained LSTM neural network after the classifier to obtain a comprehensive risk classification model;

[0048] inputting the comprehensive risk assessment score into the comprehensive risk classification model to obtain the comprehensive risk assessment result.

[0049] Preferably, the generative large language model is a GPT language model.

[0050] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0051] The present application provides an automatic generation method for perioperative patient consultation opinions, including: extracting key patient indicators from electronic medical records using natural language processing technology; cleaning and standardizing the key patient indicators to ensure data format consistency and obtain preprocessed indicator data; mapping the preprocessed indicator data to the parameter format of the standardized risk assessment scale to obtain structured patient data; calling multiple clinically verified risk assessment scales to evaluate the structured patient data, and generating a comprehensive risk assessment score according to the evaluation results; based on a machine learning algorithm, the comprehensive risk of the patient is classified according to the comprehensive risk assessment score to obtain a comprehensive risk assessment result; using a generative large language model to convert the comprehensive risk assessment result into a standardized consultation opinion text. The present application significantly improves the efficiency, accuracy and standardization of perioperative patient risk assessment and consultation opinion generation by combining advanced technologies such as natural language processing, machine learning and generative large language models, reduces the burden of medical workers, optimizes the medical service process, and provides higher quality personalized medical services for patients. The implementation of this system will bring significant efficiency improvement and technological progress to the medical industry. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0053] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0055] The purpose of the present application is to provide an automatic generation method for perioperative patient consultation opinions, which significantly improves the efficiency, accuracy and standardization of perioperative patient risk assessment and consultation opinion generation by combining advanced intelligent technologies such as natural language processing, machine learning and generative large language models.

[0056] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0057] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in Figure 1 The present application provides an automatic generation method for perioperative patient consultation opinions, which includes:

[0058] Step 100: extracting key patient indicators from electronic medical records using natural language processing technology;

[0059] Step 200: cleaning and standardizing the key patient indicators to ensure data format consistency, obtaining preprocessed indicator data;

[0060] Step 300: mapping the preprocessed indicator data to the parameter format of the standardized risk assessment scale, obtaining structured patient data;

[0061] Step 400: calling multiple clinically validated risk assessment scales to evaluate the structured patient data, and generating a comprehensive risk assessment score according to the evaluation results;

[0062] Step 500: Based on the machine learning algorithm, the comprehensive risk of the patient is risk stratified according to the comprehensive risk assessment score, and a comprehensive risk assessment result is obtained.

[0063] Step 600: The comprehensive risk assessment result is converted into a standardized consultation opinion text by using a generative large language model.

[0064] Preferably, the key patient indicators include age, gender, height, weight, BMI, medical history, and auxiliary examination results.

[0065] Preferably, the key patient indicators are extracted from the electronic medical record using natural language processing technology, including:

[0066] The free text in the electronic medical record is segmented into words;

[0067] Each of the words is tagged with a part of speech;

[0068] A named entity recognition model in the medical field is used to identify medical entities in the free text, and the words are restored to root forms;

[0069] The root forms in the free text are classified to obtain classification data;

[0070] The trained named entity recognition (NER) model is used to extract the key patient indicators in the classification data.

[0071] Specifically, step 100 of the present embodiment comprises:

[0072] First step: word segmentation and text segmentation

[0073] First, the free text in the electronic medical record is segmented into words or phrases. For Chinese text, a word segmentation tool such as jieba or HanLP can be used to cut sentences into words; for English text, spaCy or NLTK can be used for word segmentation. The purpose of word segmentation is to break down continuous text into basic language units (such as words or phrases) for subsequent processing. It should be noted that medical terms (such as "hypertension" and "diabetes") may be fixed phrases, so it is necessary to combine a medical dictionary (such as ICD-10 or SNOMED CT) for professional term recognition and segmentation to avoid incorrect splitting.

[0074] Second step: part-of-speech tagging

[0075] On the basis of word segmentation, each word is labeled with its part of speech (such as noun, verb, adjective, etc.). Part-of-speech tagging can help identify important information in the text, such as disease names are usually nouns, and test results may be numerical or adjectives. Tools such as spaCy, Stanza, etc. can be used for part-of-speech tagging, and combined with medical domain part-of-speech tagging models (such as BioBERT or ClinicalBERT) to improve the accuracy of tagging. For example, in "The patient's blood pressure is 120 / 80 mmHg", "blood pressure" is labeled as a noun, and "120 / 80" is labeled as a numerical value. Through part-of-speech tagging, context information can be provided for subsequent medical entity recognition.

[0076] Step 3: Named Entity Recognition and Lemmatization

[0077] Use a medical domain named entity recognition (NER) model to identify medical entities (such as diseases, drugs, test results, etc.) in free text. NER models can be based on deep learning (such as BioBERT, Med7) or rule matching. The model will identify the boundaries and categories of medical entities based on context information, such as "hypertension" is identified as a disease entity, and "aspirin" is identified as a drug entity. At the same time, the word is restored to the root form (lemmatization), such as "check" is restored to "check", to unify the expression form and reduce redundancy. Lemmatization can be achieved using the lemmatization tools of spaCy or NLTK.

[0078] Step 4: Text Classification

[0079] Text classification is performed on the root form of the free text to classify the text into different categories (such as medical history, test results, diagnosis information, etc.). Text classification can be achieved using machine learning models (such as support vector machine SVM, random forest) or deep learning models (such as BERT, TextCNN). When training the model, use the annotated electronic medical record data as the training set. For example, "The patient has a history of hypertension" is classified as "past medical history", and "blood pressure is 120 / 80 mmHg" is classified as "test results". The purpose of classification is to classify the information in the text and provide context support for subsequent extraction of key indicators.

[0080] Step 5: Extract Key Patient Indicators

[0081] The trained named entity recognition (NER) model is used to extract key patient indicators from the classified data. The NER model extracts the patient's age, gender, height, weight, BMI, medical history, and auxiliary examination results from the classified text content. For example, from "the patient is 65 years old, weighs 70 kg, and is 170 cm tall", the age = 65, weight = 70 kg, and height = 170 cm are extracted, and the BMI = 24.2 is calculated; from "the medical history includes hypertension and diabetes", the disease entities "hypertension" and "diabetes" are extracted. The extracted indicators are stored as structured data (such as JSON or database table) to provide standardized input for subsequent risk assessment and consultation opinion generation.

[0082] Preferably, the root forms in the free text are text classified to obtain classified data, including:

[0083] The root forms in the free text are input into a text context relationship feature extraction layer to extract context relationship feature information;

[0084] The root forms in the free text are input into a global feature extraction layer to extract global feature information;

[0085] The context relationship feature information and the global feature information are fused to obtain fused text features;

[0086] A loss function is constructed using the fused text features;

[0087] The loss function is continuously optimized to obtain a text classification model;

[0088] The text classification model is used to complete text classification to obtain the classified data.

[0089] In practical applications, the present embodiment can use an XLNet model to complete the initial feature extraction step. The XLNet model is a pre-training language model based on the combination of autoregression and auto-encoding. Its core part is the Transformer-XL architecture, which solves the defect of the BERT model (Bidirectional Encoder Representations from Transformers) in the length constraint of the input context, so that the pre-training model can learn more distant context information.

[0090] Preferably, the root forms in the free text are input into a text context relationship feature extraction layer to extract context relationship feature information, including:

[0091] The pre-training language model is used to extract the initial feature information of the root forms in the free text;

[0092] inputting the initial feature information into a forward gate recurrent unit and a backward gate recurrent unit;

[0093] splicing outputs of the forward gate recurrent unit and the backward gate recurrent unit to obtain context relationship feature information; an expression of the context relationship feature information is: wherein H t ={H1, H2, … H l} represents initial feature information input at a t time point, represents an output of the forward gate recurrent unit at the t time point, represents an output of the backward gate recurrent unit at the t time point, and GRU represents a gate recurrent unit, and G represents context relationship feature information.

[0094] Specifically, the embodiment can more fully learn text context relationship and obtain context information by splicing outputs of the forward gate recurrent unit and the backward gate recurrent unit.

[0095] Preferably, an expression of the global feature information is:

[0096] c i =f(ω·H+b)

[0097]

[0098] In the formula, f is an activation function, ω is a convolution kernel, h is a convolution kernel size, b is a bias value, c i is an i-th extracted feature vector of a convolution layer, represents values of features extracted by three different convolution kernels after passing through a max-pooling layer, and C represents extracted global feature information.

[0099] Specifically, the embodiment fuses the context relationship feature information and the global feature information to obtain fused text features, and specifically includes:

[0100] splicing the context relationship feature information and the global feature information to obtain first fused text features; wherein the first fused text features are:

[0101] performing max-pooling processing on the first fused text features to obtain second fused text features; max-pooling processing is used on the fused features O, and the purpose is to extract maximum features in the first fused text features O.

[0102] performing mean-pooling processing on the first fused text features to obtain third fused text features; mean-pooling processing is used on the fused features O, and the purpose is to extract mean features in the fused features O.

[0103] Further, a loss function is constructed using the fusion text features, including:

[0104] The first fusion text feature is input into a full connection layer and a Softmax layer to output a first probability value of a text category; wherein an expression of the full connection layer and the Softmax layer outputting the first probability value is:

[0105] Y1=tanh(W O ·O+b O )

[0106]

[0107] wherein W O represents a weight matrix of the full connection layer when the input is the first fusion text feature, O represents the first fusion text feature, b O represents a bias vector of the full connection layer when the input is the first fusion text feature, represents an output layer weight matrix when the input is the first fusion text feature, θ represents a network parameter, and P(y i |Y1, θ) represents a probability value of Y1 belonging to a y i category.

[0108] The second fusion text feature is input into a full connection layer and a Softmax layer to output a second probability value of a text category; wherein an expression of the full connection layer and the Softmax layer outputting the second probability value is:

[0109]

[0110] wherein, represents a weight matrix of the full connection layer when the input is the second fusion text feature, M p represents the second fusion text feature, represents a bias vector of the full connection layer when the input is the second fusion text feature, represents an output layer weight matrix when the input is the second fusion text feature, θ represents a network parameter, and P(y i |Y2, θ) represents a probability value of Y2 belonging to a y i category;

[0111] The third fusion text feature is input into a full connection layer and a Softmax layer to output a third probability value of a text category; wherein an expression of the full connection layer and the Softmax layer outputting the third probability value is:

[0112]

[0113] wherein, A represents a weight matrix of the full connection layer when the input is the second fused text feature p A represents a weight matrix of the full connection layer when the input is the third fused text feature A represents a bias vector of the full connection layer when the input is the third fused text feature A represents an output layer weight matrix when the input is the third fused text feature, and represents a network parameter, and P(y i |Y3,θ) represents a probability value that Y3 belongs to the y i th class.

[0114] The maximum value among the first probability value, the second probability value and the third probability value is taken as a prediction result.

[0115] A loss function is constructed by using the prediction result, and specifically, the expression of the loss function is as follows:

[0116]

[0117] Wherein, Y represents a real class of the text to be classified, m represents the number of total classes of the text, and represents a regularization parameter. k A represents a prediction result.

[0118] The present application can improve the classification performance and effect of the text by fusing the context relationship feature information and the global feature information of the text by using the gated recurrent unit, the convolution layer and the pooling layer to obtain the fused text feature, constructing and optimizing the loss function based on the fused text feature, and then completing the text classification.

[0119] Preferably, the risk assessment scale includes: ASA grading table, Child-Pugh score table, NYHA cardiac function grading table, RCRI modified cardiac risk assessment table, ARISCAT postoperative pulmonary complication risk score table, and Essen stroke risk score table.

[0120] Specifically, the step 300 of the embodiment includes:

[0121] Firstly, the parameter requirements and scoring rules of each risk assessment scale need to be clarified. For example, ASA classification requires information such as patient's medical history, activity tolerance; Child-Pugh score requires liver function-related indicators (such as bilirubin, albumin, prothrombin time, etc.); RCRI modified cardiac risk assessment requires cardiovascular-related risk factors (such as previous myocardial infarction, diabetes, etc.); ARISCAT score requires patient's age, preoperative oxygen saturation, preoperative infection, type of surgery, etc. According to the parameter requirements of these scales, a standardized parameter mapping table is established to map the preprocessed indicator data one by one with the scale parameters. For example, the patient's "age" is directly mapped to the "age" parameter of the ARISCAT score, and "hypertension" in the "medical history" is mapped to the "risk factors" parameter of the RCRI score.

[0122] After clarifying the parameter format, the preprocessed indicator data is converted according to the mapping table. For example, "patient age 87 years old" extracted from the electronic medical record is directly mapped to the "age" field of the ARISCAT score; "medical history: hypertension for 10 years" is mapped to the "risk factors" field of the RCRI score; "bilirubin 1.0 mg / dL" is mapped to the "bilirubin" field of the Child-Pugh score. For parameters that need to be calculated (such as BMI), calculate according to the formula (weight / height2) and then map to the parameter field of the related scale. In this way, the preprocessed indicator data is converted into a standardized format that meets the requirements of each risk assessment scale, ensuring the integrity and consistency of the data.

[0123] After completing the parameter mapping, all parameter data of the scales are integrated into a structured patient data format (such as JSON or database table). For example, the generated structured data may contain the following fields: {"age": 87, "BMI": 22, "ASA classification": 2, "Child-Pugh score": "A level", "RCRI risk factors": 1, "ARISCAT score": 47}. This structured data format facilitates subsequent risk assessment algorithm calls and processing, while ensuring data standardization and scalability. In this way, the system can quickly and accurately map the patient's preprocessed indicator data to the parameter format of each risk assessment scale, providing a reliable data foundation for comprehensive risk assessment and consultation opinion generation.

[0124] Preferably, a plurality of risk assessment scales that have been clinically verified are called to evaluate the structured patient data, and a comprehensive risk assessment score is generated according to the evaluation results, including:

[0125] The scoring algorithm of each scale is called according to the structured patient data to calculate the scoring result of each scale;

[0126] generate the comprehensive risk assessment score according to the score results based on a weighted average algorithm.

[0127] First paragraph: call the score algorithm of each scale to calculate the score result

[0128] Specifically, according to the structured patient data, the score algorithm of each risk assessment scale is called one by one to calculate the score result of each scale. The score algorithm of each scale is based on its specific parameter requirements and scoring rules. For example, the ASA classification directly outputs the classification result according to the patient's medical history and activity tolerance; the Child-Pugh score calculates the total score and determines the classification according to parameters such as bilirubin, albumin and prothrombin time; the RCRI score obtains the score by counting the number of cardiovascular-related risk factors (such as previous myocardial infarction, diabetes, etc.); the ARISCAT score calculates the total score according to the patient's age, preoperative oxygen saturation, preoperative infection, etc. By calling these algorithms, the system can generate independent score results for each scale, such as "ASA classification = 2", "Child-Pugh score = A", "RCRI score = 1", "ARISCAT score = 47".

[0129] After obtaining the score results of each scale, a weighted average algorithm is used to generate a comprehensive risk assessment score. First, according to the importance of each scale, weights are assigned, for example, the ASA classification weight is 0.3, the Child-Pugh score weight is 0.2, the RCRI score weight is 0.3, and the ARISCAT score weight is 0.2. Then, multiply the score result of each scale by the corresponding weight and accumulate to get the comprehensive risk assessment score. For example, if the ASA classification = 2 (corresponding to the score 0.4), the Child-Pugh score = A (corresponding to the score 0.2), the RCRI score = 1 (corresponding to the score 0.3), and the ARISCAT score = 47 (corresponding to the score 0.5), then the comprehensive risk assessment score is (0.4 x 0.3) + (0.2 x 0.2) + (0.3 x 0.3) + (0.5 x 0.2) = 0.35. The final comprehensive risk assessment score can be used for further classification (such as low risk, medium risk, high risk) or to generate consultation opinions.

[0130] Preferably, based on a machine learning algorithm, the comprehensive risk of the patient is classified according to the comprehensive risk assessment score to obtain a comprehensive risk assessment result, including:

[0131] Obtain a preset comprehensive risk assessment score data set;

[0132] Construct an initial neural network deep learning model;

[0133] Train the initial neural network deep learning model according to the comprehensive risk assessment score data set to obtain a trained classifier;

[0134] A trained LSTM neural network is connected after the classifier to obtain a comprehensive risk classification model.

[0135] The comprehensive risk assessment score is input into the comprehensive risk classification model to obtain the comprehensive risk assessment result.

[0136] Specifically, the present embodiment first collects and organizes a comprehensive risk assessment score dataset for training. The dataset should contain the comprehensive risk assessment scores of patients and their corresponding risk classification labels (such as low risk, medium risk, and high risk). The data sources can be historical electronic medical record data, surgical records, or clinical research data, ensuring that the data covers different patient populations and risk levels. In order to improve the generalization ability of the model, the dataset needs to be cleaned and standardized, such as removing outliers, filling missing values, and normalizing the comprehensive risk assessment scores (such as mapping the scores to the range of 0 to 1). In addition, the dataset is divided into training set, validation set and test set (such as 8:1:1 ratio) for subsequent model training and performance evaluation.

[0137] Based on the comprehensive risk assessment score dataset, an initial neural network (DNN) model is constructed for classifying comprehensive risk. The input layer of the model accepts comprehensive risk assessment scores, the hidden layer is composed of several fully connected layers, each layer contains a certain number of neurons (such as 64 or 128), and uses activation functions (such as ReLU) to introduce non-linear characteristics. The output layer uses the Softmax activation function to convert the model output into a probability distribution of risk classification (such as the probability of low risk, medium risk, and high risk). The loss function of the model selects the cross-entropy loss function (Cross-Entropy Loss), and the optimizer selects the Adam optimizer. The model is iteratively trained on the training set, and the performance (such as accuracy, F1 score) of the model is monitored on the validation set, and the hyperparameters (such as learning rate, number of hidden layers) are adjusted to optimize the model effect.

[0138] After the initial neural network model is trained, it is used as a classifier, and a trained LSTM (Long Short-Term Memory) neural network is connected after it to capture the time series features or potential complex patterns of comprehensive risk assessment scores. The input of the LSTM network can be the sequence of comprehensive risk assessment scores of patients (such as the trend of score changes in multiple assessments), and the output is the probability distribution of risk classification. The hidden layer of the LSTM network can contain a certain number of units (such as 64 or 128), and use Tanh and Sigmoid activation functions to process time series data. Through joint training or phased training, the classifier and LSTM network are integrated into a comprehensive risk classification model, further improving the model's ability to recognize complex risk patterns.

[0139] The comprehensive risk assessment score of the patient is input into the trained comprehensive risk classification model, and the model performs forward propagation according to the input data, outputting the probability distribution of each risk category (such as low risk = 0.7, medium risk = 0.2, and high risk = 0.1). According to the maximum value of the output probability, the comprehensive risk assessment result of the patient is determined (such as low risk). In addition, the model can also provide the confidence of the classification, helping doctors better understand the reliability of the assessment result. The final comprehensive risk assessment result can be used to generate standardized consultation opinion text, providing support for perioperative clinical decision-making.

[0140] Preferably, the generative large language model is a GPT language model.

[0141] Optionally, the present embodiment first arranges the comprehensive risk assessment result (such as low risk, medium risk, and high risk) and related patient information (such as age, gender, medical history, auxiliary examination results, etc.) into structured input data as the input of the generative large language model. For example, the input data can be in JSON format, containing fields such as {"age": 65, "gender": "male", "comprehensive risk assessment result": "medium risk", "medical history": ["hypertension", "diabetes"], "auxiliary examination results": {"blood pressure": "140 / 90 mmHg", "blood glucose": "8.5 mmol / L"}}. At the same time, a standardized consultation opinion template is designed to clearly define the structure and content of the generated text, such as including patient basic information, comprehensive risk assessment result, main risk factor analysis, and recommended preoperative preparation or intervention measures. This template design can provide clear guidance for the generative model, ensuring that the generated text meets clinical standards.

[0142] In order to ensure that the generated consultation opinion text meets the professionalism and standardization of the medical field, the GPT language model can be fine-tuned. Fine-tuning requires a high-quality medical text dataset, including a large number of consultation opinion instances covering different risk assessment results and patient conditions. Through supervised learning, these medical texts are used to train the GPT model, enabling it to generate text that conforms to the style and logic of medical language. During fine-tuning, specific training objectives can be set, such as the generated text must include the patient's comprehensive risk level, main risk factors, and specific preoperative recommendations, and other key content. The fine-tuned model will be more suitable for medical scenarios and can generate more accurate and standardized consultation opinions.

[0143] After fine-tuning is completed, the comprehensive risk assessment results and patient information are input into the GPT model to generate consultation opinion text. For example, the input prompt can be "generate consultation opinion based on the following patient information: age 65 years old, male, comprehensive risk assessment result is moderate risk, past medical history includes hypertension and diabetes, auxiliary examination results show blood pressure 140 / 90 mmHg, blood glucose 8.5 mmol / L." The model will output a standardized consultation opinion, such as "Patient, male, 65 years old, with a history of hypertension and diabetes. Comprehensive risk assessment result is moderate risk, main risk factors include poor blood pressure control (140 / 90 mmHg) and elevated blood glucose (8.5 mmol / L). It is recommended to strengthen blood pressure and blood glucose management before surgery, and adjust the dosage of antihypertensive drugs and insulin as needed." After generation, the text can be post-processed through rule verification or manual review to ensure the accuracy and standardization of the content, and the final consultation opinion text is used for clinical decision support.

[0144] As an optional implementation, the consultation opinion of the present embodiment is as follows:

[0145] Patient information: Zhang San 0005599888*** ward 01 bed

[0146] Chief complaint: abdominal pain for 3 days

[0147] Diagnosis: liver occupying lesion

[0148] Handover: The patient is an elderly female, 150 cm tall, weighing 49.5 kg, BMI = 22, ASA 2 grade. Child-Pugh liver function classification 0 points A grade. NYHA cardiac function 2 grade. Activity tolerance 4 METs, RCRI modified cardiac risk assessment risk factor 1, corresponding to the incidence of cardiac death / non-fatal myocardial infarction / non-fatal cardiac arrest about 1.0%. ARISCAT postoperative pulmonary complications risk prediction score 47 points, corresponding to high risk of postoperative pulmonary complications, Essen stroke risk score 3 points, perioperative stroke risk moderate risk, risk 7-9%. Non-cardiac surgery 30-day cardiovascular adverse event rate according to surgical type is moderate risk, incidence 1-5%.

[0149] History well understood, patient has visited.

[0150] The patient is an elderly female, 87 years old. Past "hypertension for 10 years, regular use" of "nifedipine" to control blood pressure. Planning to undergo laparoscopic liver lesion resection. Details are shown in the medical record.

[0151] Physical examination: mouth opening > 3 fingers, head and neck mobility normal, Mallampati classification grade 1, no removable dentures or loose teeth.

[0152] Auxiliary examination: electrocardiogram, chest X-ray showed no obvious abnormalities, and other examination results have been reviewed.

[0153] Proposal: 1. The patient is an elderly female, height 150 cm, weight 49.5 kg, BMI = 22, ASA 2.

[0154] Risk assessment: NYHA cardiac function 2. Activity tolerance 4 METs, RCRI modified cardiac risk assessment risk factor 1, corresponding to the incidence of cardiac death / non-fatal myocardial infarction / non-fatal cardiac arrest about 1.0%. ARISCAT postoperative pulmonary complication risk prediction score 47, corresponding to high risk of postoperative pulmonary complications, Essen stroke risk score 3, perioperative stroke risk moderate risk, risk 7-9%. Anesthesia risk is moderate to high risk.

[0155] Communicate with the patient and family members about the risks related to the operation, and obtain informed consent: There is a very high risk of postoperative pulmonary complications such as pulmonary infection, atelectasis, respiratory failure, and hypoxemia. Postoperative exclusion of further treatment in ICU. The risk of cerebrovascular events, embolism and infection during the perioperative period is correspondingly increased, such as cerebral hemorrhage, cerebral ischemia, cerebral infarction, delirium, cognitive dysfunction, etc.

[0156] Pay attention to the stability of circulation during the perioperative period, avoid sharp fluctuations in blood pressure and heart rate, use advanced circulation monitoring during the operation, and ensure the perfusion of important organs. Strengthen respiratory tract management, protective ventilation strategy, avoid excessive ventilation and carbon dioxide accumulation. Pay attention to maintain water and electrolyte balance and keep warm. Postoperative analgesia and antiemesis.

[0157] My department will follow up, thank you for your invitation!

[0158] Further, the patient information table of the collected data in this embodiment is as follows:

[0159]

[0160]

[0161]

[0162] The beneficial effects of the present application are as follows:

[0163] (1) The present application automatically extracts key patient indicators from electronic medical records by using natural language processing technology, and performs data cleaning and standardization processing, which greatly reduces the time of manual data collection and arrangement, and improves the efficiency of risk assessment; The automatic risk assessment process enables anesthesiologists to obtain comprehensive risk assessment results of patients more quickly, so as to focus on individualized assessment and decision-making.

[0164] (2) The application integrates multiple clinically validated risk assessment scales (such as ASA classification, Child-Pugh score, RCRI score, etc.), and generates a comprehensive risk assessment score through the "voting + weight" method, ensuring the comprehensiveness and accuracy of the evaluation results; based on the machine learning algorithm, the comprehensive risk assessment score is classified, further improving the accuracy and consistency of risk assessment, and reducing the influence of human subjective factors.

[0165] (3) The application reduces the repetitive work burden of anesthesiologists in preoperative consultation through an automated and standardized process, allowing them to more efficiently manage the growing number of patients; the automatically generated consultation opinion text reduces the time for doctors to write reports, optimizing the workflow.

[0166] (4) The consultation opinion text generated by the application uses a standardized and modular format, facilitating information sharing and shift handover between medical teams, improving communication efficiency and reducing errors in information transmission; the application strictly adheres to laws and regulations on medical data security and privacy protection, uses multi-level data encryption technology and access control mechanisms to ensure the security and privacy of patient data, protecting the interests of patients and medical institutions.

[0167] (5) The application uses a generative large language model to convert the comprehensive risk assessment results into a standardized consultation opinion text, ensuring the professionalism, consistency and readability of the report content, and reducing the understanding bias caused by individual expression differences among doctors.

[0168] (6) The application provides more accurate and personalized medical services for perioperative patients through comprehensive risk assessment and standardized consultation opinions, improving patient care quality and safety; the automatically generated consultation report must be reviewed by authorized anesthesiologists before it can be officially released, ensuring the accuracy and professionalism of the report, while reducing the repetitive labor of doctors.

[0169] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be mutually referred to.

[0170] The principles and implementation methods of the application are described using specific examples in this paper. The above description of the embodiments is only to help understand the method and core idea of the application; at the same time, for those skilled in the art, according to the idea of the application, the specific implementation method and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the application.

Claims

1. A method for automatically generating perioperative patient consultation opinions, characterized in that: include: Extract key patient metrics from electronic medical records using natural language processing technology; Cleaning and standardizing the key patient indicators to ensure data format consistency and obtain preprocessed indicator data; Mapping the pre-treatment indicator data to a parameter format of a standardized risk assessment scale to obtain structured patient data; Calling multiple clinically validated risk assessment scales to evaluate the structured patient data, and generating a comprehensive risk assessment score based on the evaluation results; Based on a machine learning algorithm, risk stratification is performed on the patient's comprehensive risk according to the comprehensive risk assessment score to obtain a comprehensive risk assessment result; Using a generative large language model to convert the comprehensive risk assessment results into standardized consultation opinion text; Leverage natural language processing to extract key patient metrics from electronic medical records, including: segmenting free text in said electronic medical record into words; Marking the part of speech for each of the words; Using a named entity recognition model in the medical field to identify medical entities in the free text and restore the words to their root forms; Performing text classification on the root forms in the free text to obtain classification data; Extracting the key patient indicators from the classified data using a trained named entity recognition model; Text classification is performed on the root forms in the free text to obtain classification data, including: Inputting the root form in the free text into the text context feature extraction layer to extract context feature information; Inputting the root form in the free text into the global feature extraction layer to extract global feature information; Fusing the contextual feature information and the global feature information to obtain a fused text feature; Constructing a loss function using the fused text features; Continuously optimizing the loss function to obtain a text classification model; Using the text classification model to complete text classification and obtain the classification data; Based on the machine learning algorithm, the patient's comprehensive risk is risk stratified according to the comprehensive risk assessment score to obtain a comprehensive risk assessment result, including: Obtain preset comprehensive risk assessment sub-datasets; Build an initial neural network deep learning model; Training the initial neural network deep learning model according to the comprehensive risk assessment sub-data set to obtain a trained classifier; Connecting a trained LSTM neural network after the classifier to obtain a comprehensive risk classification model; The comprehensive risk assessment score is input into the comprehensive risk classification model to obtain the comprehensive risk assessment result.

2. The method for automatically generating perioperative patient consultation opinions according to claim 1, characterized in that: The key patient indicators include: age, gender, height, weight, BMI, medical history and auxiliary examination results.

3. The method for automatically generating perioperative patient consultation opinions according to claim 1, characterized in that: The root form in the free text is input into the text context feature extraction layer to extract context feature information, including: Extracting initial feature information of word roots from the free text using a pre-trained language model; Inputting the initial feature information into a forward gated recurrent unit and a reverse gated recurrent unit; The outputs of the forward gated recurrent unit and the reverse gated recurrent unit are concatenated to obtain contextual feature information; the expression of the contextual feature information is: ;in, Indicates t The initial feature information input at each moment, Indicates t The output of the positive gated recurrent unit at all times, Indicates t The output of the moment-to-moment reverse gated recurrent unit, GRU represents a gated recurrent unit, G Represents contextual feature information.

4. The method for automatically generating perioperative patient consultation opinions according to claim 3, characterized in that: The expression of the global feature information is: ； ； Where, is the activation function, is the convolution kernel, is the convolution kernel size, is the bias value, is the feature vector extracted by the i-th convolutional layer, Represents the value of the features extracted by 3 different convolution kernels after passing through the maximum pooling layer, C Represents the extracted global feature information.

5. The method for automatically generating perioperative patient consultation opinions according to claim 1, characterized in that: The risk assessment scales include: ASA classification, Child-Pugh scoring table, NYHA cardiac function classification table, RCRI modified cardiac risk assessment table, ARISCAT postoperative pulmonary complications risk score table, and Essen stroke risk score table.

6. The method for automatically generating perioperative patient consultation opinions according to claim 1, characterized in that: Multiple clinically validated risk assessment scales are used to evaluate the structured patient data and generate a comprehensive risk assessment score based on the evaluation results, including: The scoring algorithm of each scale is called according to the structured patient data to calculate the scoring results of each scale; The comprehensive risk assessment score is generated according to the scoring results based on a weighted average algorithm.

7. The method for automatically generating perioperative patient consultation opinions according to claim 1, characterized in that: The generative large language model is a GPT language model.

Images