Postoperative symptom prediction system, device, and storage medium

Abstract

This application discloses a postoperative symptom prediction system, device, and storage medium. The postoperative symptom prediction system includes a terminal device and a server. The server is configured to perform the following operations: before surgery, using a preset feature extraction model to generate first feature information based on diagnostic information, where the diagnostic information includes a diagnostic report, surgical plan information, and the patient's preoperative physiological parameters; matching the first feature information with a knowledge base to determine context information matching the first feature information; generating prompt information based on the diagnostic information and context information using a preset prompt information template; and predicting the patient's postoperative symptoms based on the prompt information using a large model, generating a corresponding postoperative symptom report. The terminal device is configured to perform the following operations: sending the received diagnostic information to the server; and displaying the postoperative symptom report.

Description

Technical Field

[0001] This application relates to the field of medical instrument technology, and in particular to a postoperative symptom prediction system, device, and storage medium. Background Technology

[0002] For decades, surgical risk prediction has primarily relied on analyzing structured data such as patient age, medical history, and laboratory indicators to calculate the probability of known complications (such as infection and bleeding) and other known symptoms.

[0003] Patent CN120510482A discloses an intraoperative risk assessment system based on multimodal surgical image fusion. This system acquires multimodal surgical grayscale images; identifies difference images of the same type; obtains change regions on the difference images and projects them to obtain a modal variation sequence; determines anomaly coefficients based on the consistency of the modal variation sequence; performs dynamic time warping matching on the modal variation sequence, and determines a midpoint offset coefficient based on the positional difference between the center point and the corresponding warped point in the sequence; combines the anomaly coefficient and the midpoint offset coefficient to determine the operational impact; and combines the operational impact of all types of surgical grayscale images within the sampling period to achieve risk assessment. This invention integrates surgical data from different modalities, compressing complex and delicate surgical scenarios into simple multimodal sequence processing, enhancing the quantitative assessment capability of intraoperative risks, and improving the accuracy and timeliness of risk assessment.

[0004] Patent CN112949685A discloses a risk prediction method for aortic dissection surgery based on a boosting tree model. The method includes the following steps: Step 1: Data preprocessing; filling in missing data in surgical records to form a database containing surgical records; Step 2: Data mining; processing the data in the database based on a decision tree model; Step 3: Result analysis; analyzing the data mining results to evaluate the effectiveness of the data mining algorithm; Step 4: Knowledge application; applying the model to the data corresponding to the current aortic dissection surgery to predict the risk of the current surgery. The model of this invention has 100% accuracy in predicting the risk of death after aortic dissection surgery, which is higher than other prediction methods. The prediction results can provide key factors for postoperative mortality in aortic dissection surgery, providing decision-making basis for future doctors and patients, and helping medical policymakers to make full use of medical resources.

[0005] These methods provide a basic tool for preoperative risk stratification, but they are only applicable to specific surgical types and rely on fixed correlation patterns between disease and postoperative symptoms in historical data. However, dynamic parameters such as intraoperative blood loss and postoperative inflammatory marker changes are difficult to integrate in real time, leading to a lag in risk assessment.

[0006] With the increasing complexity of surgeries and the diversity of patients, simply predicting known complications is no longer sufficient. For example, known complications after thoracic surgery include long-term air leakage, lung collapse, limited lung volume, or irregular breathing. However, unknown symptoms such as neurological injury syndrome or drug reactions may also occur postoperatively. Traditional models are only applicable to complications that have been clearly recorded in historical cases and can only predict the probability of occurrence of known complications, but cannot identify new symptoms (such as metabolic disorders or neurological damage).

[0007] There is currently no effective solution to the technical problem of the existing technology, which only predicts the probability of occurrence of known postoperative symptoms and cannot identify unknown symptoms, resulting in a high degree of limitation in postoperative disease prediction. Summary of the Invention

[0008] The embodiments of this application provide a postoperative symptom prediction system, method, device, and storage medium to at least solve the technical problem in the prior art that only predicts the probability of occurrence of known postoperative symptoms and cannot identify unknown symptoms, resulting in a high degree of limitation in postoperative disease prediction.

[0009] According to one aspect of the embodiments of this application, a postoperative symptom prediction system is provided, including: a terminal device and a server, wherein the server is configured to perform the following operations: generating first feature information based on diagnostic information using a preset feature extraction model before surgery, wherein the diagnostic information includes a diagnostic report, surgical plan information, and the patient's preoperative physiological parameters; matching the first feature information using a knowledge base to determine context information matching the first feature information, wherein the context information is used to indicate symptoms related to the patient; generating prompt information based on the diagnostic information and the context information using a preset prompt information template; and predicting the patient's postoperative symptoms based on the prompt information using a large model, generating a corresponding postoperative symptom report, and the terminal device is configured to perform the following operations: sending the received diagnostic information to the server; and displaying the postoperative symptom report.

[0010] According to another aspect of the embodiments of this application, a method for predicting postoperative symptoms is also provided, comprising: acquiring diagnostic information related to the patient before surgery, wherein the diagnostic information includes a diagnostic report and surgical plan information; generating first feature information based on the diagnostic information using a preset feature extraction model; matching the first feature information using a knowledge base to determine context information matching the first feature information, wherein the context information is used to indicate symptoms related to the patient; generating prompt information based on the diagnostic information and the context information using a preset prompt information template; and predicting the patient's postoperative symptoms based on the prompt information using a large model to generate a corresponding postoperative symptom report.

[0011] According to another aspect of the embodiments of this application, a postoperative symptom prediction device is also provided, comprising: an information acquisition module for acquiring diagnostic information related to the patient before surgery, wherein the diagnostic information includes a diagnostic report and surgical plan information; an information generation module for generating first feature information based on the diagnostic information using a preset feature extraction model; an information determination module for matching the first feature information using a knowledge base to determine context information matching the first feature information, wherein the context information is used to indicate symptoms related to the patient; a prompt generation module for generating prompt information based on the diagnostic information and the context information using a preset prompt information template; and a symptom prediction module for predicting the patient's postoperative symptoms based on the prompt information using a large model and generating a corresponding postoperative symptom report.

[0012] According to another aspect of the embodiments of this application, a postoperative symptom prediction device is also provided, including: a processor; and a memory connected to the processor, for providing the processor with instructions to perform the following processing steps: acquiring patient-related diagnostic information before surgery, wherein the diagnostic information includes a diagnostic report and surgical plan information; generating first feature information based on the diagnostic information using a preset feature extraction model; matching the first feature information using a knowledge base to determine context information matching the first feature information, wherein the context information is used to indicate patient-related symptoms; generating prompt information based on the diagnostic information and the context information using a preset prompt information template; and predicting the patient's postoperative symptoms based on the prompt information using a large model, and generating a corresponding postoperative symptom report.

[0013] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above method.

[0014] In this embodiment, preoperative diagnostic reports, surgical plan information, and physiological parameter information are used to predict possible postoperative symptoms. Compared to existing technologies that predict postoperative complications solely based on postoperative monitoring patient information, this technology can determine postoperative symptoms before surgery, allowing for adjustments to the surgical plan and preventing symptom occurrence. Furthermore, traditional models rely on static historical data and can only identify known complications. This technology, however, uses a knowledge base and feature information matching to multi-dimensionally associate known diagnostic information with symptoms in the knowledge base, generating contextual information pointing to novel symptoms and reducing the limitations of postoperative disease prediction. Moreover, this technology uses a large model to filter and generate corresponding postoperative symptoms based on contextual and diagnostic information, thereby accurately predicting possible postoperative symptoms. This solves the technical problem of existing technologies that only predict the probability of known postoperative symptoms, failing to identify unknown symptoms and resulting in limited postoperative disease prediction. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 This is a schematic diagram of the postoperative symptom prediction system according to Embodiment 1 of this application;

[0017] Figure 2 This is a block diagram of the server according to the first aspect of Embodiment 1 of this application;

[0018] Figure 3 This is a schematic diagram of the server's prediction of postoperative symptoms according to the first aspect of Embodiment 1 of this application;

[0019] Figure 4 This is a schematic diagram of the process for generating first feature information according to the first aspect of Embodiment 1 of this application;

[0020] Figure 5 This is a schematic diagram of the process for generating second feature information according to the first aspect of Embodiment 1 of this application;

[0021] Figure 6 This is a schematic diagram of the process for generating third feature information according to the first aspect of Embodiment 1 of this application;

[0022] Figure 7 This is a flowchart illustrating the postoperative symptom prediction method according to the second aspect of Embodiment 1 of this application;

[0023] Figure 8 This is a schematic diagram of the postoperative symptom prediction device according to Embodiment 2 of this application; and

[0024] Figure 9 This is a schematic diagram of the postoperative symptom prediction device according to Embodiment 3 of this application. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] Example 1

[0028] Figure 1 A schematic diagram of the postoperative symptom prediction system according to Embodiment 1 of this application is shown. Figure 2 A block diagram of a server according to Embodiment 1 of this application is shown. (Reference) Figure 1 and Figure 2 As shown, this embodiment proposes a postoperative symptom prediction system, including: a terminal device 100 and a server 200, wherein the terminal device 100 is configured to send received diagnostic information to the server 200; and to display postoperative symptom reports.

[0029] also, Figure 3 A flowchart illustrating the operations performed by server 200 is further provided. (Reference) Figure 3 As shown, server 200 is configured to perform the following operations:

[0030] S302: Before surgery, the first feature information is generated based on the diagnostic information using a preset feature extraction model. The diagnostic information includes the diagnostic report, surgical plan information, and the patient's preoperative physiological parameters.

[0031] S304: Use a knowledge base to match the first feature information to determine the context information that matches the first feature information, wherein the context information is used to indicate symptoms related to the patient;

[0032] S306: Generate prompt information based on diagnostic information and context information using a preset prompt information template; and

[0033] S308: Based on prompts and information from a large model, predict the patient's postoperative symptoms and generate a corresponding postoperative symptom report.

[0034] Specifically, the doctor obtains the patient's diagnostic information through terminal device 100, and then sends the diagnostic information to server 200 through terminal device 100. The diagnostic information includes diagnostic reports, surgical plan information, and the patient's physiological parameter information.

[0035] This example uses a lung cancer patient.

[0036] (1) A diagnostic report may be, for example:

[0037] Chief complaint: Persistent dry cough for 3 months, accompanied by chest pain and blood in sputum for 1 week.

[0038] Past medical history: 40-year history of smoking (20 cigarettes / day), 10-year history of hypertension.

[0039] Chest CT: An irregular soft tissue nodule (2.8 cm × 2.1 cm × 3.0 cm) with spiculated margins and pleural traction was observed in the medial segment of the right middle lobe. FDG metabolism was abnormally elevated (SUVmax 8.7), suggesting peripheral lung cancer.

[0040] Whole-body PET / CT: Enlarged lymph nodes in the right hilar and mediastinal (groups 4R and 7), increased FDG metabolism, suspected metastasis; no distant metastasis was observed.

[0041] (2) Surgical plan information can be, for example:

[0042] Surgical type: Right middle lobectomy + systematic lymph node dissection.

[0043] The patient's tumor was located in the middle lobe of the right lung (2.8 cm × 2.1 cm × 3.0 cm), with no distant metastasis (M0) and regional lymph node metastasis (N1).

[0044] According to the AJCC 8th edition TNM staging and CSCO guidelines, stage IIB non-small cell lung cancer (N1) requires lobectomy combined with systematic lymph node dissection to ensure negative resection margins (R0 resection).

[0045] (3) Physiological parameter information, for example:

[0046] Age: 62 years old;

[0047] Gender: Male;

[0048] Height: 1.72 m;

[0049] Weight 72 kg;

[0050] Resting blood pressure: 138 / 86 mmHg;

[0051] Average heart rate: 82 beats / minute;

[0052] Resting oxygen saturation (SpO2): 95%;

[0053] Post-activity oxygen saturation (SpO2): 9%;

[0054] CEA (carcinoembryonic antigen): 8.7 ng / mL; and

[0055] Cyfra21-1 (cytokeratin 19 fragment): 4.5 ng / mL.

[0056] Furthermore, the server 200 receives the diagnostic information sent by the terminal device 100, and performs semantic analysis and feature extraction on the diagnostic report, surgical plan information, and patient physiological parameter information in the diagnostic information through the feature extraction model of the feature extraction module 210, generating the first feature information A=[a1,a2,...,a...]. m ] T Then, the feature extraction module 210 sends the first feature information to the feature retrieval module 220. And the server 200 sends diagnostic information to the feature enhancement module 230.

[0057] Furthermore, after receiving the first feature information sent by the feature extraction module 210, the feature retrieval module 220 of the server 200 matches the first feature information using a preset knowledge base to generate context information that matches the first feature information. The knowledge base stores context information that matches various feature vectors. For example, the context information could be possible symptoms that the patient may experience. The knowledge base is shown in Table 1.

[0058] Table 1

[0059]

[0060] The knowledge base contains n records of context information corresponding to feature information. The context information corresponding to feature information B1 is xxx1, the context information corresponding to feature information B2 is xxx2, ..., and so on. n The corresponding context information is xxx n Therefore, the feature retrieval module 220 compares the first feature information A with the feature information B1~B1 in the knowledge base as shown in Table 1. n The matching process is performed to determine the corresponding context information. For example, the feature retrieval module 220 matches the first feature information A with the feature information B1~B1 in the knowledge base. n The matching process is performed. When the first feature information A matches B1, the context information xxx1 corresponding to B1 is used as the context information matching the first feature information A. Thus, the feature retrieval module 220 determines the context information corresponding to the first feature information A. Then, the feature retrieval module 220 sends the context information corresponding to the first feature information A to the feature enhancement module 230.

[0061] Furthermore, the feature enhancement module 230 receives diagnostic information sent by the server 200 and context information sent by the feature retrieval module 220. Then, the feature enhancement module 230 obtains a preset prompt word template, which is shown below:

[0062] Patient basic information: [Patient basic information 1];

[0063] Imaging examination: [Imaging examination information 1];

[0064] Pathological results: [Pathological result 1];

[0065] Molecular detection: [Molecular detection 1];

[0066] Treatment goal: [Treatment goal 1];

[0067] Possible symptoms: [Symptom 1].

[0068] The prompt word template contains the prompt words that need to be filled in within [].

[0069] Therefore, the feature enhancement module 230 inputs the diagnostic information and contextual information into the prompt word template, generates prompt information, and sends the prompt information to the prediction module 240. For example, the prompt information is:

[0070] Patient basic information: [62-year-old male, BMI 24.3 kg / m²] 2 (Height 1.72 m, weight 72 kg), 40-year smoking history (20 cigarettes / day), quit smoking 1 year ago; 10-year history of hypertension, resting blood pressure 138 / 86 mmHg, average heart rate 82 beats / minute; resting SpO2 95%, which drops to 89% after activity (indicating a significant decrease in lung function reserve));

[0071] Imaging findings: [Chest CT: Irregular soft tissue nodule (2.8 cm × 2.1 cm × 3.0 cm) in the medial segment of the right middle lobe, with spiculated margins, pleural retraction, and abnormally elevated FDG metabolism (SUVmax 8.7), consistent with peripheral lung cancer. Whole-body PET / CT: Enlarged lymph nodes in the right hilar and mediastinal (groups 4R and 7) with elevated FDG metabolism, suggesting regional lymph node metastasis (N1), no distant metastasis (M0)];

[0072] Pathological results: [Pending postoperative pathological confirmation (intraoperatively planned right middle lobectomy + systematic lymph node dissection)];

[0073] Molecular testing: [Tumor markers: CEA 8.7 ng / mL↑, Cyfra21-1 4.5 ng / mL↑; Gene testing: It is recommended to perform EGFR, ALK, ROS1 and other driver gene testing after surgery to guide targeted therapy (not yet completed)];

[0074] Treatment goals: [Radical surgery: Right middle lobectomy + systematic lymph node dissection, ensuring R0 resection (negative margins); Staging-guided treatment: According to the AJCC 8th edition TNM staging (stage IIB, T1cN1M0), adjuvant chemotherapy will be initiated according to the CSCO guidelines after surgery, and radiotherapy may be considered in combination or not depending on the specific situation].

[0075] Possible symptoms: [xxx1, xxx2].

[0076] Based on the information above, please provide a postoperative symptom report for this patient, including:

[0077] Reasonable inferences about pathological results based on existing information (such as possible histological types, differentiation trends, and the possibility of pleural invasion).

[0078] Postoperative complication risks (such as bleeding, infection, atelectasis / pneumonia, arrhythmia, deep vein thrombosis / pulmonary embolism, bronchopleural fistula, pain, long-term effects on lung function, etc.);

[0079] Comorbidity management and perioperative risks (focus: hypertension management goals, risk and management strategies for respiratory complications caused by decreased lung function, such as breathing exercises and oxygen therapy needs assessment).

[0080] Postoperative follow-up recommendations (including follow-up timelines, follow-up content such as imaging examinations, tumor marker monitoring, complication monitoring, lifestyle guidance, and adjuvant therapy arrangements).

[0081] Furthermore, the prediction module 240 receives the prompt information sent by the feature enhancement module 230, and then inputs the prompt information into a preset large model. The large model can be any existing large model, such as ChatGPT. The prediction module 240 processes the prompt information using the large model to predict the patient's postoperative symptoms and outputs a corresponding postoperative symptom report. The postoperative symptom report may, for example, be:

[0082] "1. The probability of malignancy of lung nodules is >95%, and possible pathologies include glandular precursor lesions, microinvasive carcinoma, and invasive carcinoma."

[0083] 2. Follow-up recommendations: 1 month post-surgery, follow-up chest HRCT, elective surgery, and follow-up examination after anti-inflammatory treatment;

[0084] 3. Surgical procedure: Thoracoscopic right middle lobectomy + systematic lymph node dissection (4R, group 7);

[0085] 4. Postoperative lung function loss ratio (expected postoperative FEV1 decrease of 15%-18% (middle lobe accounts for the proportion of total lung function)).

[0086] 5. Considering the patient's underlying medical conditions, the risk of perioperative pneumonia / respiratory failure is approximately 25%; the risk of arrhythmia is approximately 18% (hypertension + intraoperative traction); and the risk of VTE is approximately 30% (Caprini score ≥ 5).

[0087] 6. Postoperative follow-up recommendations (contrast-enhanced chest CT + CEA / Cyfra21-1 every 3 months; chest CT + abdominal ultrasound + brain MRI every 6 months (if headache occurs)).

[0088] 7. Postoperative treatment recommendations (based on postoperative gene testing results and postoperative tumor stage, chemotherapy, immunotherapy, or targeted therapy are recommended)".

[0089] Furthermore, the server 200 sends the postoperative symptom report generated by the prediction module 240 to the terminal device 100 for display. The doctor can then view the report on the terminal device 100 and determine the various symptoms the patient may experience after surgery. Based on this, the doctor can revise the surgical plan to avoid the symptoms mentioned in the report. Additionally, the doctor can plan postoperative follow-up in advance based on the recommendations in the postoperative symptom report.

[0090] As described in the background section, with the increasing complexity of surgeries and the diversity of patients, simply predicting known complications is no longer sufficient. For example, known complications after thoracic surgery include long-term air leakage, lung collapse, limited lung volume, or irregular breathing. However, unknown symptoms such as neurological injury syndrome or drug reactions may also occur postoperatively. Traditional models are only applicable to complications that have been clearly recorded in historical cases and can only predict the probability of occurrence of known complications, but cannot identify new symptoms (such as metabolic disorders or neurological damage).

[0091] To address the aforementioned technical problems, the technical solution of this application predicts possible postoperative symptoms by utilizing diagnostic reports, surgical plan information, and physiological parameter information preoperatively. Compared to existing technologies that predict postoperative complications solely based on patient information from postoperative monitoring, this solution can determine postoperative symptoms before surgery, allowing for adjustments to the surgical plan and preventing symptom occurrence. Furthermore, traditional models rely on static historical data and can only identify known complications. This solution, however, uses a knowledge base and feature information matching to multi-dimensionally associate known diagnostic information with symptoms in the knowledge base, generating contextual information pointing to novel symptoms and reducing the limitations of postoperative disease prediction. Moreover, this solution uses a large model to filter and generate corresponding postoperative symptoms based on contextual and diagnostic information, accurately predicting possible postoperative symptoms. This solves the technical problem of existing technologies that only predict the probability of known postoperative symptoms, failing to identify unknown symptoms and resulting in limited postoperative disease prediction. Additionally, according to the technical solution of this embodiment, postoperative follow-up plans can be planned in advance based on recommendations in the postoperative symptom report.

[0092] Optionally, the operation of generating first feature information based on diagnostic information using a preset feature extraction model before surgery includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information.

[0093] Specifically, refer to Figure 4 As shown, the feature extraction module 210 is pre-configured with a first feature extraction model, a second feature extraction model, and a mapping layer. The first feature extraction model is used to extract features from the diagnostic report, the second feature extraction module is used to extract features from the surgical plan information, and the mapping layer is used to extract features from the physiological parameter information.

[0094] Therefore, firstly, the feature extraction module 210 inputs the diagnostic report into the first feature extraction model, extracts features from the diagnostic report through the first feature extraction model, and outputs the second feature information.

[0095] Furthermore, the feature extraction module 210 inputs the surgical plan information into the second feature extraction model, extracts features from the surgical plan information through the second feature extraction model, and outputs the third feature information.

[0096] Furthermore, the feature extraction module 210 inputs the physiological parameter information into the mapping layer, and performs feature extraction on the physiological parameter information through the mapping layer to generate the fourth feature information. Preferably, the mapping layer used in this embodiment can be, for example, a feedforward neural network layer.

[0097] Furthermore, the feature extraction module 210 fuses the second feature information, the third feature information, and the fourth feature information to generate the first feature information.

[0098] Therefore, this technical solution extracts features from diagnostic reports, surgical plans, and physiological parameters separately, thereby capturing key features of different diagnostic information. Then, the features corresponding to the diagnostic reports, surgical plans, and physiological parameters are fused and correlated, achieving multi-dimensional feature extraction.

[0099] Optionally, the operation of extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using the first word segmentation module of the first feature extraction model to generate first words; performing word embedding operations on the first words using the first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using the first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using the first fully connected layer of the first feature extraction model to generate second feature information.

[0100] Specifically, refer to Figure 5 As shown, the first feature extraction model includes a first word segmentation module, a first word embedding module, a first attention module, and a first fully connected layer.

[0101] Feature extraction module 210 inputs the diagnostic report into the first feature extraction model. The first feature extraction model segments the diagnostic report into words using the first word segmentation module, generating the first word segment. For example, the diagnostic report is:

[0102] Chief complaint: Persistent dry cough for 3 months, accompanied by chest pain and blood in sputum for 1 week.

[0103] Past medical history: 40-year history of smoking (20 cigarettes / day), 10-year history of hypertension.

[0104] Chest CT: An irregular soft tissue nodule (2.8 cm × 2.1 cm × 3.0 cm) with spiculated margins and pleural traction was observed in the medial segment of the right middle lobe. FDG metabolism was abnormally elevated (SUVmax 8.7), suggesting peripheral lung cancer.

[0105] Whole-body PET / CT: Enlarged lymph nodes in the right hilar and mediastinal (groups 4R and 7), increased FDG metabolism, suspected metastasis; no distant metastasis was observed.

[0106] The first word segmentation module generates the following words: "chief complaint", "persistent dry cough", "3 months", "chest pain", "blood in sputum", and "1 week", etc., and the generated word segmentation identifiers are oc1~oc. x1 (That is, the first word segmentation). x1 represents the number of words in the first word segmentation. Then, the first word segmentation module will segment the first words oc1~oc x1 Send to the first word embedding module.

[0107] Furthermore, the first word embedding module receives the first word segmentation oc1~oc x1 and the first segmentation oc1~oc x1 Perform word embedding operations to generate words with the first word segmentation oc1~oc x1 The corresponding first word vector od1~od x1 Then the first word embedding module will embed the first word vector od1~od x1 Send to the first attention module.

[0108] Furthermore, the first attention module receives the first word vector od1~od sent by the first word embedding module. x1 Then, the first word vector od1~od x1 Feature extraction is performed to generate first semantic feature information. Then, the first attention module sends the first semantic feature information to the first fully connected layer.

[0109] Furthermore, the first fully connected layer receives the first semantic feature information sent by the first attention module, processes the first semantic feature information, and generates the second feature information.

[0110] Therefore, this technical solution first segments the diagnostic report into words, and then uses an attention mechanism to extract features from it, thereby focusing on the key parts for predicting postoperative symptoms, thus improving the pertinence and accuracy of the prediction.

[0111] Optionally, the operation of extracting features from the first word vector through the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector through the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector through the enhanced attention module of the first feature extraction model to generate third semantic feature information, wherein the enhanced attention module is used to perform enhanced attention feature extraction on the first word vector based on the correlation between the first word segments determined by the global attention module; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information.

[0112] Specifically, the first attention module includes a global attention module and an enhanced attention module.

[0113] The first feature extraction model extracts the first word vector od1~od x1 Input the global attention module. The global attention module generates the first word vectors od1~od1 respectively. x1 The corresponding key vectors AK1~AK x1 Query vectors AQ1~AQ x1 and value vectors AV1~AV x1 The global attention module then calculates the sum of the first word vector od using the following formula. i The corresponding second semantic feature information AF i Where i = 1 ~ x1:

[0114] (1)

[0115] (2)

[0116] (3)

[0117] Where j = 1 ~ x1; Aw i,j The first word vector od j Compared to the first word vector od i The weight value; Ah i,j The first word vector od j Key vector AK j With the first word vector od i Query vector AQ i The correlation coefficient between them; d represents the dimension of the key vector.

[0118] Therefore, in order to determine the first word vector od i Second semantic feature information AF i Referring to the above formula, firstly, the first word vector od is calculated using formula (1). j Compared to the first word vector od i correlation coefficient Ah i,j .

[0119] Then, using formula (2) to calculate Ah i,j The softmax function is calculated to obtain the correlation coefficient Ah. i,j The corresponding weight Aw i,j .

[0120] Then, the first word vector od is calculated using formula (3). i Second semantic feature information AF i .

[0121] Therefore, by using the above methods, we can obtain the first word vector od. i The corresponding semantic feature information AF depends on long-distance context. i .

[0122] Furthermore, the first feature extraction model extracts the first word vectors od1~od x1 The input is an enhanced attention module, which is then used to extract semantic features for enhanced attention. Specifically, for each first word segment and its corresponding first word vector, the enhanced attention module performs attention-based feature extraction based on first words that are semantically highly correlated with that first word segment, thereby achieving enhanced semantic feature extraction.

[0123] More specifically, the attention enhancement module generates word vectors od1~od1 respectively. x1 The corresponding key vectors BK1~BK x1 Query vectors BQ1~BQ x1 and value vectors BV1~BV x1 The attention enhancement module then calculates each first word vector od1~od using the following formula. x1 With the first word vector od i (i.e., the first participle oa) i The word vectors (corresponding to the first target word segmentation) and their corresponding correlation coefficients Bh i,j Where i,j=1~x1:

[0124] (4)

[0125] Then, the enhanced attention module extracts the relevant information Bh i,1 ~Bh i,x1 A predetermined number of related information items (e.g., L related information items) are selected from the information Bh in descending order. i,1 ~Bh i,x1 Select L of the largest correlation information Bh' i,1 ~Bh' i,L (Bh' i,1 ~Bh' i,L ∈Bh i,1 ~Bh i,x1 ), and determine the correlation information Bh' i,1 ~Bh' i,L The corresponding first segmentation oc'1~oc' L (oc'1~oc') L ∈oc1~oc x1 (This corresponds to the multiple segments that are most strongly associated with the first target segment).

[0126] Then, determine the corresponding segments oc'1~oc'. L There are L windows of width W. After deduplication of the first word segment within each of these L windows, K first word segments oc''1~oc'' are obtained. K (where oc''1~oc'') K ∈oc1~oc x1 And oc i ∈oc''1~oc'' K The first word vector corresponding to it is od''1~od'' K (where od''1~od'') K ∈od1~od x1 And od i ∈od''1~od'' K ).

[0127] The attention enhancement module then works from key vectors BK1 to BK. x1 The first segment oc''1~oc'' is determined in the middle. K The corresponding key vectors are BK''1~BK'' K (where BK''1~BK'') K ∈BK1~BK x1 Then, calculate the first word segment oc''1~oc'' using the following formula. K With the first participle oc i The corresponding correlation coefficient Bh'' i,k Where k = 1 ~ K:

[0128] (5)

[0129] Among them, BK'' k For the first participle oc'' k The corresponding key vector, where k = 1 to K. Then, the enhanced attention module recalculates the first word segmentation oc''1~oc''. K Compared to the first word segmentation oc i The weight value Bw'' i,k :

[0130] (6)

[0131] Then, the enhanced attention module works from the value vectors BV1 to BV x1 The middle part is determined to be related to the first segmentation oc''1~oc'' respectively. K The corresponding value vector is BV''1~BV'' K (where BV''1~BV'') K ∈BV1~BV x1Then, the enhanced attention module calculates the objective function (OC) of the first word segmentation. i The corresponding third semantic feature information BF with enhanced attention i :

[0132] (7)

[0133] Therefore, by using the above methods, we can target each first-segment word objective. i Identify the L first segmented words oc'1~oc' with high relevance to it. L And utilize the L first segmentation words oc'1~oc' with high relevance. L The corresponding window further determines the first word segmentation oc''1~oc'' K Then use the first segmentation oc''1~oc'' K For the first segmentation oc i Feature extraction for enhanced attention.

[0134] Furthermore, the first feature extraction model extracts the second semantic feature information AF1~AF2. x1 and the corresponding third semantic feature information BF1~BF x1 The features are concatenated to generate the first semantic feature information EF1~EF. x1 That is, EF1 is formed by splicing AF1 and BF1; EF2 is formed by splicing AF2 and BF2; ...; and so on, EF x1 It is by AF x1 and BF x1 It was pieced together.

[0135] Therefore, this technical solution utilizes a global attention mechanism to comprehensively analyze the correlation between all first word vectors, identifying potential long-distance dependencies between different parts of the diagnostic report. Furthermore, it employs an enhanced attention mechanism, building upon highly correlated word segments with high relevance to the target word segmentation, and using the segments defined by the windows corresponding to these highly correlated segments to extract enhanced attention features from the target word segmentation. Thus, compared to the global attention mechanism, the enhanced attention mechanism can identify potential local dependencies between different parts of the diagnostic report; while compared to the traditional local attention mechanism, it can identify dependencies within highly correlated local areas on a global scale. Therefore, combining global attention and enhanced attention allows for a more comprehensive extraction of feature information.

[0136] Optionally, the operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate the third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate the third feature information.

[0137] Specifically, refer to Figure 6 As shown, the second feature extraction model includes a second word segmentation module, a second word embedding module, a second attention module, and a second fully connected layer.

[0138] The feature extraction module 210 inputs the surgical plan information into the second feature extraction model. The second feature extraction model then segments the surgical plan information using the second word segmentation module to generate second word segments. For example, the surgical plan information is as follows:

[0139] Surgical type: Right middle lobectomy + systematic lymph node dissection.

[0140] The patient's tumor was located in the middle lobe of the right lung (2.8 cm × 2.1 cm × 3.0 cm), with no distant metastasis (M0) and regional lymph node metastasis (N1).

[0141] According to the AJCC 8th edition TNM staging and CSCO guidelines, stage IIB non-small cell lung cancer (N1) requires lobectomy combined with systematic lymph node dissection to ensure negative resection margins (R0 resection).

[0142] Therefore, the second word segmentation module generates word segments including: "right middle lobe of lung", "resection", "systemic", "lymph node", "dissection", "patient", "tumor", "located in", and "right middle lobe of lung", etc., and the generated word segmentation identifiers are tc1~tc. x2 (That is, the second word segmentation). Here, x2 represents the number of second word segments. Then, the second word segmentation module will segment the second words tc1~tc... x2 Send to the second word embedding module.

[0143] Furthermore, the second word embedding module receives the second word segmentation tc1~tc x2 And the second segment tc1~tc x2 Perform word embedding operations to generate words from the second word segmentation tc1~tc. x2 The corresponding second word vector td1~td x2 Then the second word embedding module will embed the second word vectors td1~tdx2 Send to the second attention module.

[0144] Furthermore, the second attention module receives the second word vectors td1~td sent by the second word embedding module. x2 Then, for the second word vector td1~td x2 Feature extraction is performed to generate fourth semantic feature information.

[0145] The second attention module then sends the fourth semantic feature information to the second fully connected layer.

[0146] Furthermore, the second fully connected layer receives the fourth semantic feature information sent by the second attention module, processes the fourth semantic feature information, and generates the third feature information.

[0147] Therefore, this technical solution first segments the surgical plan information into words, and then uses an attention mechanism to extract features from it, thereby focusing on the key parts for predicting postoperative symptoms, thus improving the targeting and accuracy of the prediction.

[0148] Optionally, the operation of extracting features from the second word vector using the second attention module of the second feature extraction model to generate fourth semantic feature information includes: extracting features from the second word vector using the global attention module of the second feature extraction model to generate fifth semantic feature information; extracting features from the second word vector using the enhanced attention module of the second feature extraction model to generate sixth semantic feature information, wherein the enhanced attention module is used to perform enhanced attention feature extraction on the second word vector based on the correlation between the second word segments determined by the global attention module; and concatenating the fifth and sixth semantic feature information to generate fourth semantic feature information.

[0149] Specifically, the second attention module also includes a global attention module and an enhanced attention module.

[0150] The second feature extraction model will extract the second word vector td1~td x2 Input the global attention module. The global attention module then generates the second word vectors td1~td1 respectively. x2 The corresponding bond vectors CK1~CK x2 Query vectors CQ1~CQ x2 and value vectors CV1~CV x2 The global attention module then calculates the expression with the second word vector td using the following formula. v The corresponding fourth semantic feature information CF v Where v = 1 ~ x2:

[0151] (8)

[0152] (9)

[0153] (10)

[0154] Where u = 1 ~ x2; Cw v,u Represents the second word vector td v Relative to the second word vector td u The weight value; Ah v,u Represents the second word vector td u The key vector CK u With the second word vector td v Query vector CQ v The correlation coefficient between them; d represents the dimension of the key vector.

[0155] Therefore, in order to determine the second word vector td v The fourth semantic feature information CF v Referring to the above formula, firstly, use formula (8) to calculate the second word vector td. u Relative to the second word vector td v Correlation coefficient Ch v,u .

[0156] Then, using formula (9) to apply Ch v,u The softmax function is calculated to obtain the correlation coefficient Ch. v,u The corresponding weight Cw v,u .

[0157] Then, the second word vector td is calculated using formula (10). v The fifth semantic feature information CF v .

[0158] Therefore, by using the above method, we can obtain the second word vector td. v Corresponding semantic feature information CF that depends on long-range context v .

[0159] Furthermore, the second feature extraction model extracts the second word vectors td1~td x2 The input is an enhanced attention module, which is then used to extract semantic features for enhanced attention. Specifically, for each second word segment and its corresponding second word vector, the enhanced attention module performs attention-based feature extraction based on second words that are semantically highly related to that second word segment, thereby achieving enhanced semantic feature extraction.

[0160] More specifically, the enhanced attention module generates the second word vectors td1~td respectively. x2The corresponding key vectors DK1~DK x2 Query vectors DQ1~DQ x2 and value vectors DV1~DV x2 The attention enhancement module then calculates each second word vector td1~td using the following formula. x2 With the second word vector td v (That is, the second participle tc) v The word vectors (corresponding to the relevance coefficient Dh of the second target word segmentation) v,u Where v, u = 1 ~ x2:

[0161] (11)

[0162] Then, the enhanced attention module extracts the relevant information Dh v,1 ~Dh v,x2 A predetermined number of related information items (e.g., L related information items) are selected from the information Dh in descending order. v,1 ~Dh v,x2 Select L of the largest correlation information Dh' v,1 ~Dh' v,L (Dh' v,1 ~Dh' v,L ∈Dh v,1 ~Dh v,x2 ), and determine the correlation information Dh' v,1 ~Dh' v,L The corresponding second segmentation tc'1~tc' L (tc'1~tc') L ∈tc1~tc x2 (This corresponds to the multiple segments most strongly associated with the second target segment)

[0163] Then, determine the corresponding segments tc'1~tc' of the second word segment. L There are L windows of width W. After deduplication of the second word segmentation within these L windows, Y second word segments tc''1~tc'' are obtained. Y (where tc''1~tc'') Y ∈tc1~tc x2 And tc v ∈tc''1~tc'' Y The corresponding second word vector is td''1~td'' Y (where td''1~td'') Y ∈td1~td x2 And td v ∈td''1~td'' Y ).

[0164] The attention enhancement module then works from key vectors DK1 to DK. x2 The middle determines the second segment tc''1~tc'' Y The corresponding key vectors are DK''1~DK'' Y (where DK''1~DK'') Y ∈DK1~DK x2 Then, calculate each second segment tc''1~tc'' using the following formula. Y With the second segmentation tc v The corresponding correlation coefficient Dh'' v,y Where y = 1 ~ Y:

[0165] (12)

[0166] Among them, DK'' y For the second segmentation tc'' y The corresponding key vector, where y = 1 to Y. Then, the enhanced attention module recalculates the second word segmentation tc''1~tc''. Y Compared to the second word segmentation tc v The weight value Dw'' v,y :

[0167] (13)

[0168] Then, the enhanced attention module works from the value vectors DV1~DV x2 The middle part is determined to be related to the second segment tc''1~tc'' respectively. Y The corresponding value vector DV''1~DV'' Y (where DV''1~DV'') Y ∈DV1~DV x2 Then, the enhanced attention module calculates the tc of the second word segmentation. v The corresponding sixth semantic feature information DF with enhanced attention v :

[0169] (14)

[0170] Therefore, through the above methods, we can target each second word segmentation tc. v Identify the L second-segment words tc'1~tc' that have a high correlation with it. L And utilize the L highly relevant second segmentations tc'1~tc' L The corresponding window further determines the second word segment tc''1~tc'' Y Then, using the second segmentation tc''1~tc'' Y For the second segment tc vFeature extraction for enhanced attention.

[0171] Furthermore, the second feature extraction model incorporates the fifth semantic feature information CF 1~ CF x2 and the corresponding sixth semantic feature information DF 1~ DF x2 The data is concatenated to generate the fourth semantic feature information GF. 1~ GF x2 That is, GF1 is composed of CF1 and DF1; GF2 is composed of CF2 and DF2; ...; and so on, GF x2 It is CF x2 and DF x2 It was pieced together.

[0172] Therefore, this technical solution utilizes a global attention mechanism to comprehensively analyze the correlation between all second word vectors, identifying potential long-distance dependencies between different parts of the surgical plan information. Furthermore, it employs an enhanced attention mechanism, building upon highly correlated word segments with high relevance to the target word segmentation, and using the window-defined segments corresponding to these highly correlated segments to extract enhanced attention features from the target word segmentation. Thus, compared to the global attention mechanism, the enhanced attention mechanism can identify potential local dependencies between different parts of the surgical plan information; while compared to the traditional local attention mechanism, it can identify dependencies within highly correlated local areas on a global scale. Therefore, combining global attention and enhanced attention allows for a more comprehensive extraction of feature information.

[0173] Optionally, the operation of extracting features from the physiological parameter information of the diagnostic information through the mapping layer to generate the fourth feature information includes: constructing a physiological parameter vector reflecting the physiological parameter information based on the physiological parameter information; and inputting the physiological parameter vector into the mapping layer to obtain the fourth feature information.

[0174] Specifically, in this embodiment, physiological parameter information can be encoded to construct a physiological parameter vector corresponding to the physiological parameter information, as shown below:

[0175]

[0176] Based on the above encoding, a physiological parameter vector corresponding to the patient's physiological parameters is constructed. This embodiment then inputs this physiological parameter vector into a mapping layer based on a feedforward neural network for processing, thereby obtaining the fourth feature information. Thus, through the above method, feature information that can be fused with diagnostic reports and surgical plan information based on physiological parameters can be generated, thereby improving the accuracy of subsequent feature retrieval and matching.

[0177] Optionally, the operation of generating the first feature information based on the second feature information, the third feature information, and the fourth feature information includes: concatenating the second feature information, the third feature information, and the fourth feature information to generate fused feature information; and processing the fused feature information through a neural network model to generate the first feature information.

[0178] Specifically, the feature extraction module 210 acquires the second feature information, the third feature information, and the fourth feature information, and then concatenates the second feature information, the third feature information, and the fourth feature information to generate fused feature information.

[0179] Furthermore, the feature extraction module is pre-configured with a neural network model. This neural network model can be an MLP (Multi-Level Processing). The feature extraction module 210 inputs the fused feature information into the neural network model, processes the feature information through the neural network model, and outputs the first feature information.

[0180] Therefore, this technical solution extracts features from the fused feature information composed of feature vectors extracted from different diagnostic information (i.e., diagnostic reports, surgical plan information, and physiological parameter information). This allows for the simultaneous consideration of multiple influencing factors such as the patient's clinical manifestations, surgical interventions, and underlying health conditions, thus providing more comprehensive information support for subsequent predictions.

[0181] Furthermore, according to a second aspect of this embodiment, a method for predicting postoperative symptoms is provided. Figure 7 A flowchart illustrating the method is shown below. (Refer to...) Figure 7 As shown, the method includes:

[0182] S702: Obtain patient-related diagnostic information before surgery, including diagnostic reports and surgical plan information;

[0183] S704: Generate first feature information based on diagnostic information using a preset feature extraction model;

[0184] S706: Use a knowledge base to match the first feature information and determine the context information that matches the first feature information, wherein the context information is used to indicate symptoms related to the patient;

[0185] S708: Generate prompt messages based on diagnostic information and context information using a preset prompt message template; and

[0186] S710: Based on prompts and information from a large model, predict the patient's postoperative symptoms and generate a corresponding postoperative symptom report.

[0187] Optionally, the operation of generating first feature information based on diagnostic information using a preset feature extraction model includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information.

[0188] Optionally, the operation of extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using the first word segmentation module of the first feature extraction model to generate first words; performing word embedding operations on the first words using the first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using the first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using the first fully connected layer of the first feature extraction model to generate second feature information.

[0189] Optionally, the operation of extracting features from the first word vector through the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector through the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector through the enhanced attention module of the first feature extraction model to generate third semantic feature information; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information.

[0190] Optionally, the operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate the third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate the third feature information.

[0191] Optionally, the operation of generating the first feature information based on the second feature information, the third feature information, and the fourth feature information includes: concatenating the second feature information, the third feature information, and the fourth feature information to generate fused feature information; and processing the fused feature information through a neural network model to generate the first feature information.

[0192] Furthermore, according to a third aspect of this embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above method.

[0193] Therefore, according to this embodiment, by utilizing diagnostic reports, surgical plan information, and physiological parameter information preoperatively, possible postoperative symptoms can be predicted. Compared to existing technologies that predict postoperative complications solely based on patient information from postoperative monitoring, this technology can determine postoperative symptoms before surgery, allowing for adjustments to the surgical plan and preventing postoperative symptoms. Furthermore, traditional models rely on static historical data and can only identify known complications. This technology, however, uses a knowledge base and feature information matching to multi-dimensionally associate the patient's known diagnostic information with symptoms in the knowledge base, generating contextual information pointing to novel symptoms and reducing the limitations of postoperative disease prediction. Moreover, this technology uses a large model to filter and generate corresponding postoperative symptoms based on contextual and diagnostic information, thereby accurately predicting possible postoperative symptoms. This solves the technical problem of existing technologies that only predict the probability of known postoperative symptoms and cannot identify unknown symptoms, resulting in highly limited postoperative disease prediction.

[0194] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0195] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0196] Example 2

[0197] Figure 8A postoperative symptom prediction device 800 according to a first aspect of this embodiment is shown, which corresponds to the method described according to the first aspect of Embodiment 1. Reference Figure 8 As shown, the device 800 includes: an information acquisition module 810, used to acquire diagnostic information related to the patient before surgery, wherein the diagnostic information includes a diagnostic report and surgical plan information; an information generation module 820, used to generate first feature information based on the diagnostic information using a preset feature extraction model; an information determination module 830, used to match the first feature information using a knowledge base to determine context information matching the first feature information, wherein the context information is used to indicate symptoms related to the patient; a prompt generation module 840, used to generate prompt information based on the diagnostic information and context information using a preset prompt information template; and a symptom prediction module 850, used to predict the patient's postoperative symptoms based on the prompt information using a large model and generate a corresponding postoperative symptom report.

[0198] Optionally, the information generation module 820 includes: a first generation submodule, used to extract features from the diagnostic report of the diagnostic information using a first feature extraction model to generate second feature information; a second generation submodule, used to extract features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; a third generation submodule, used to extract features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and a fourth generation submodule, used to generate first feature information based on the second feature information, the third feature information, and the fourth feature information.

[0199] Optionally, the first generation submodule includes: a first generation unit, used to perform word segmentation processing on the diagnostic report through the first word segmentation module of the first feature extraction model to generate first word segments; a second generation unit, used to perform word embedding operation on the first word segments through the first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first word segments; a third generation unit, used to perform feature extraction on the first word vectors through the first attention module of the first feature extraction model to generate first semantic feature information; and a fourth generation unit, used to process the first semantic feature information through the first fully connected layer of the first feature extraction model to generate second feature information.

[0200] Optionally, the third generation unit includes: extracting features from the first word vector using the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector using the enhanced attention module of the first feature extraction model to generate third semantic feature information; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information.

[0201] Optionally, the second generation submodule includes: a fifth generation unit, used to segment the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; a sixth generation unit, used to perform word embedding operations on the second words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; a seventh generation unit, used to extract features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and an eighth generation unit, used to process the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate third feature information.

[0202] Optionally, the fourth generation submodule includes: a ninth generation unit, used to concatenate the second feature information, the third feature information, and the fourth feature information to generate fused feature information; and a tenth generation unit, used to process the fused feature information through a neural network model to generate first feature information.

[0203] Therefore, according to this embodiment, by utilizing diagnostic reports, surgical plan information, and physiological parameter information preoperatively, possible postoperative symptoms can be predicted. Compared to existing technologies that predict postoperative complications solely based on patient information from postoperative monitoring, this technology can determine postoperative symptoms before surgery, allowing for adjustments to the surgical plan and preventing postoperative symptoms. Furthermore, traditional models rely on static historical data and can only identify known complications. This technology, however, uses a knowledge base and feature information matching to multi-dimensionally associate the patient's known diagnostic information with symptoms in the knowledge base, generating contextual information pointing to novel symptoms and reducing the limitations of postoperative disease prediction. Moreover, this technology uses a large model to filter and generate corresponding postoperative symptoms based on contextual and diagnostic information, thereby accurately predicting possible postoperative symptoms. This solves the technical problem of existing technologies that only predict the probability of known postoperative symptoms and cannot identify unknown symptoms, resulting in highly limited postoperative disease prediction.

[0204] Example 3

[0205] Figure 9 A postoperative symptom prediction device 900 according to a first aspect of this embodiment is shown, which corresponds to the method described according to the first aspect of Embodiment 1. Reference Figure 9As shown, the device 900 includes: a processor 910; and a memory 920 connected to the processor 910, used to provide instructions to the processor 910 to process the following steps: acquiring patient-related diagnostic information before surgery, wherein the diagnostic information includes a diagnostic report and surgical plan information; generating first feature information based on the diagnostic information using a preset feature extraction model; matching the first feature information using a knowledge base to determine context information matching the first feature information, wherein the context information is used to indicate patient-related symptoms; generating prompt information based on the diagnostic information and context information using a preset prompt information template; and predicting the patient's postoperative symptoms based on the prompt information using a large model, and generating a corresponding postoperative symptom report.

[0206] Optionally, the operation of generating first feature information based on diagnostic information using a preset feature extraction model before surgery includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information.

[0207] Optionally, the operation of extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using the first word segmentation module of the first feature extraction model to generate first words; performing word embedding operations on the first words using the first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using the first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using the first fully connected layer of the first feature extraction model to generate second feature information.

[0208] Optionally, the operation of extracting features from the first word vector through the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector through the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector through the enhanced attention module of the first feature extraction model to generate third semantic feature information; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information.

[0209] Optionally, the operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate the third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate the third feature information.

[0210] Optionally, the operation of generating the first feature information based on the second feature information, the third feature information, and the fourth feature information includes: concatenating the second feature information, the third feature information, and the fourth feature information to generate fused feature information; and processing the fused feature information through a neural network model to generate the first feature information.

[0211] Therefore, according to this embodiment, by utilizing diagnostic reports, surgical plan information, and physiological parameter information preoperatively, possible postoperative symptoms can be predicted. Compared to existing technologies that predict postoperative complications solely based on patient information from postoperative monitoring, this technology can determine postoperative symptoms before surgery, allowing for adjustments to the surgical plan and preventing postoperative symptoms. Furthermore, traditional models rely on static historical data and can only identify known complications. This technology, however, uses a knowledge base and feature information matching to multi-dimensionally associate the patient's known diagnostic information with symptoms in the knowledge base, generating contextual information pointing to novel symptoms and reducing the limitations of postoperative disease prediction. Moreover, this technology uses a large model to filter and generate corresponding postoperative symptoms based on contextual and diagnostic information, thereby accurately predicting possible postoperative symptoms. This solves the technical problem of existing technologies that only predict the probability of known postoperative symptoms and cannot identify unknown symptoms, resulting in highly limited postoperative disease prediction.

[0212] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0213] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0214] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0215] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0216] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0217] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0218] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A postoperative symptom prediction system, characterized by, include: Terminal devices and servers, among which The server configuration is used to perform the following operations: Before the operation, a first feature information is generated based on the diagnostic information using a preset feature extraction model. The diagnostic information includes the diagnostic report, surgical plan information, and the patient's preoperative physiological parameters. In the knowledge base, feature information that matches the first feature information is determined, and context information corresponding to the matched feature information is determined as context information that matches the first feature information, wherein the context information is used to indicate symptoms related to the patient; Based on the diagnostic information and the context information, a prompt message is generated using a preset prompt message template; as well as Based on the provided information, the large model predicts the patient's postoperative symptoms and generates a corresponding postoperative symptom report. The terminal device is configured to perform the following operations: The received diagnostic information is sent to the server; as well as The postoperative symptom report is displayed, and in which, The operation of generating first feature information based on diagnostic information using a preset feature extraction model before surgery includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information, wherein... The operation of extracting features from the diagnostic report of the diagnostic information using a first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using a first word segmentation module of the first feature extraction model to generate first words; embedding the first words into words using a first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using a first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using a first fully connected layer of the first feature extraction model to generate second feature information, wherein... The operation of extracting features from the first word vector using the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector using the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector using the enhanced attention module of the first feature extraction model to generate third semantic feature information, wherein the enhanced attention module is used to determine the correlation between the first word segments based on the first word vector, and for a first target word in the first word segments, to perform enhanced attention feature extraction on the first target word using multiple word segments in the first word segments that are most strongly correlated with the first target word; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information, wherein... The operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate third feature information, wherein... The operation of extracting features from the second word vector using the second attention module of the second feature extraction model to generate fourth semantic feature information includes: extracting features from the second word vector using the global attention module of the second feature extraction model to generate fifth semantic feature information; extracting features from the second word vector using the enhanced attention module of the second feature extraction model to generate sixth semantic feature information, wherein the enhanced attention module is used to determine the correlation between the second word segments based on the second word vector, and for the second target word in the second word segment, to perform enhanced attention feature extraction on the second target word using multiple word segments in the second word segment that are most correlated with the second target word; and concatenating the fifth semantic feature information and the sixth semantic feature information to generate the fourth semantic feature information.

2. The system of claim 1, wherein, The operation of generating the first feature information based on the second feature information, the third feature information, and the fourth feature information includes: The second feature information, the third feature information, and the fourth feature information are concatenated to generate fused feature information; and The fused feature information is processed by a neural network model to generate the first feature information.

3. A postoperative symptom prediction device, characterized in that, include: The information acquisition module is used to acquire diagnostic information related to the patient before surgery, including diagnostic reports and surgical plan information. The information generation module is used to generate first feature information based on the diagnostic information using a preset feature extraction model; The information determination module is used to determine feature information that matches the first feature information in the knowledge base, and to determine the context information corresponding to the matched feature information as the context information that matches the first feature information, wherein the context information is used to indicate symptoms related to the patient; The prompt generation module is used to generate prompt information based on the diagnostic information and the context information using a preset prompt information template; as well as The symptom prediction module is used to predict the patient's postoperative symptoms based on the provided prompts using a large model, and to generate a corresponding postoperative symptom report. The operation of generating first feature information based on diagnostic information using a preset feature extraction model before surgery includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information, wherein... The operation of extracting features from the diagnostic report of the diagnostic information using a first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using a first word segmentation module of the first feature extraction model to generate first words; embedding the first words into words using a first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using a first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using a first fully connected layer of the first feature extraction model to generate second feature information, wherein... The operation of extracting features from the first word vector using the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector using the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector using the enhanced attention module of the first feature extraction model to generate third semantic feature information, wherein the enhanced attention module is used to determine the correlation between the first word segments based on the first word vector, and for a first target word in the first word segments, to perform enhanced attention feature extraction on the first target word using multiple word segments in the first word segments that are most strongly correlated with the first target word; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information, wherein... The operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate third feature information, wherein... The operation of extracting features from the second word vector using the second attention module of the second feature extraction model to generate fourth semantic feature information includes: extracting features from the second word vector using the global attention module of the second feature extraction model to generate fifth semantic feature information; extracting features from the second word vector using the enhanced attention module of the second feature extraction model to generate sixth semantic feature information, wherein the enhanced attention module is used to determine the correlation between the second word segments based on the second word vector, and for the second target word in the second word segment, to perform enhanced attention feature extraction on the second target word using multiple word segments in the second word segment that are most correlated with the second target word; and concatenating the fifth semantic feature information and the sixth semantic feature information to generate the fourth semantic feature information.

4. A postoperative symptom prediction device, characterized in that, include: processor; as well as A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: Preoperatively, obtain patient-related diagnostic information, including diagnostic reports and surgical plan information; First feature information is generated based on the diagnostic information using a preset feature extraction model; In the knowledge base, feature information that matches the first feature information is determined, and context information corresponding to the matched feature information is determined as context information that matches the first feature information, wherein the context information is used to indicate symptoms related to the patient; Based on the diagnostic information and the context information, a prompt message is generated using a preset prompt message template; as well as Based on the provided information, a large model predicts the patient's postoperative symptoms and generates a corresponding postoperative symptom report. The operation of generating first feature information based on diagnostic information using a preset feature extraction model before surgery includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information, wherein... The operation of extracting features from the diagnostic report of the diagnostic information using a first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using a first word segmentation module of the first feature extraction model to generate first words; embedding the first words into words using a first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using a first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using a first fully connected layer of the first feature extraction model to generate second feature information, wherein... The operation of extracting features from the first word vector using the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector using the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector using the enhanced attention module of the first feature extraction model to generate third semantic feature information, wherein the enhanced attention module is used to determine the correlation between the first word segments based on the first word vector, and for a first target word in the first word segments, to perform enhanced attention feature extraction on the first target word using multiple word segments in the first word segments that are most strongly correlated with the first target word; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information, wherein... The operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate third feature information, wherein... The operation of extracting features from the second word vector using the second attention module of the second feature extraction model to generate fourth semantic feature information includes: extracting features from the second word vector using the global attention module of the second feature extraction model to generate fifth semantic feature information; extracting features from the second word vector using the enhanced attention module of the second feature extraction model to generate sixth semantic feature information, wherein the enhanced attention module is used to determine the correlation between the second word segments based on the second word vector, and for the second target word in the second word segment, to perform enhanced attention feature extraction on the second target word using multiple word segments in the second word segment that are most correlated with the second target word; and concatenating the fifth semantic feature information and the sixth semantic feature information to generate the fourth semantic feature information.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it performs the following steps of the method: Preoperatively, obtain patient-related diagnostic information, including diagnostic reports and surgical plan information; First feature information is generated based on the diagnostic information using a preset feature extraction model; In the knowledge base, feature information that matches the first feature information is determined, and context information corresponding to the matched feature information is determined as context information that matches the first feature information, wherein the context information is used to indicate symptoms related to the patient; Based on the diagnostic information and the context information, a prompt message is generated using a preset prompt message template; as well as Based on the provided information, a large model predicts the patient's postoperative symptoms and generates a corresponding postoperative symptom report. The operation of generating first feature information based on diagnostic information using a preset feature extraction model before surgery includes: extracting features from the diagnostic report of the diagnostic information using the first feature extraction model to generate second feature information; extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information; extracting features from the physiological parameter information of the diagnostic information using a mapping layer to generate fourth feature information; and generating first feature information based on the second feature information, the third feature information, and the fourth feature information, wherein... The operation of extracting features from the diagnostic report of the diagnostic information using a first feature extraction model to generate second feature information includes: segmenting the diagnostic report into words using a first word segmentation module of the first feature extraction model to generate first words; embedding the first words into words using a first word embedding module of the first feature extraction model to generate first word vectors corresponding to the first words; extracting features from the first word vectors using a first attention module of the first feature extraction model to generate first semantic feature information; and processing the first semantic feature information using a first fully connected layer of the first feature extraction model to generate second feature information, wherein... The operation of extracting features from the first word vector using the first attention module of the first feature extraction model to generate first semantic feature information includes: extracting features from the first word vector using the global attention module of the first feature extraction model to generate second semantic feature information; extracting features from the first word vector using the enhanced attention module of the first feature extraction model to generate third semantic feature information, wherein the enhanced attention module is used to determine the correlation between the first word segments based on the first word vector, and for a first target word in the first word segments, to perform enhanced attention feature extraction on the first target word using multiple word segments in the first word segments that are most strongly correlated with the first target word; and concatenating the second semantic feature information and the third semantic feature information to generate first semantic feature information, wherein... The operation of extracting features from the surgical plan information of the diagnostic information using the second feature extraction model to generate third feature information includes: segmenting the surgical plan information into words using the second word segmentation module of the second feature extraction model to generate second words; embedding the second words into words using the second word embedding module of the second feature extraction model to generate second word vectors corresponding to the second words; extracting features from the second word vectors using the second attention module of the second feature extraction model to generate fourth semantic feature information; and processing the fourth semantic feature information using the second fully connected layer of the second feature extraction model to generate third feature information, wherein the second word vectors are processed by the second attention module of the second feature extraction model to extract features from the second word vectors. The operation of extracting features from the vectors to generate fourth semantic feature information includes: extracting features from the second word vectors using the global attention module of the second feature extraction model to generate fifth semantic feature information; extracting features from the second word vectors using the enhanced attention module of the second feature extraction model to generate sixth semantic feature information, wherein the enhanced attention module is used to determine the correlation between the second word segments based on the second word vectors, and for the second target word in the second word segment, to perform enhanced attention feature extraction on the second target word using multiple word segments in the second word segment that are most correlated with the second target word; and concatenating the fifth semantic feature information and the sixth semantic feature information to generate the fourth semantic feature information.

Images