Colorectal cancer whole-course management system based on medical agent

Abstract

The invention discloses a colorectal cancer whole-course management system based on a medical agent, relates to the technical field of medical management platforms, and solves the technical problem that an existing medical platform cannot perform accurate and efficient auxiliary diagnosis and treatment and follow-up visit management when many colorectal cancer patients are diagnosed. A patient end is used for uploading information, receiving a diagnosis result and carrying out medical question and answer dialogue; the doctor auxiliary end is used for receiving the information and providing auxiliary diagnosis and treatment suggestions and follow-up visit plans; the colorectal cancer medical agent is used for extracting key information from patient information, recommending auxiliary diagnosis and treatment suggestions according to medical knowledge base data and performing medical question and answer dialogues; the colorectal cancer data platform is used for carrying out data exchange and sharing with the HI system and the scientific research system and carrying out intelligent management on patient information; individualized precise service is achieved, differentiated schemes are automatically generated through the system, key node reminding is triggered, and prognosis experience of a patient is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical management platform, and particularly relates to a colorectal cancer whole-process management system based on a medical agent. BACKGROUND

[0002] In the face of colorectal cancer patient management, most medical structures use existing patient whole-process management platforms for unified management, but due to the characteristics of colorectal cancer, this management method is not effective, and the main deficiencies include:

[0003] I. High dependence on manual management, imbalance between service quality and cost

[0004] Colorectal cancer management involves screening, treatment, follow-up and other links. The existing system relies too much on manual operation, lacks service refinement, and cannot meet the individual needs of patients. For example, different patients have different postoperative rehabilitation needs, but manual management mostly uses standardized processes and cannot be targeted. And manual processes are prone to miss key nodes, such as colorectal cancer patients who need regular colonoscopy after surgery, relying on manual reminders may lead to loss of follow-up. As the number of patients increases, the demand for manpower rises, costs rise but efficiency does not improve, forming a vicious cycle.

[0005] II. Disconnected from multiple systems, data integration and utilization hindered

[0006] The system cannot effectively link with HIS system, research database and special disease database, forming a data island. When making clinical decisions, doctors need to switch between multiple systems for queries, affecting efficiency. For example, to determine whether a colorectal cancer patient is suitable for targeted therapy, multiple system information needs to be integrated, but this is difficult to achieve. Research data collection relies on manual input, which is time-consuming and prone to errors, reducing credibility. Special disease data is disconnected from the management system, such as Lynch syndrome information for colorectal cancer patients, which is not synchronized, which may lead to inappropriate follow-up plans and increase the risk of missed diagnosis.

[0007] III. Limited service content, insufficient research value

[0008] The system is limited to the basic management of colorectal cancer patients in the treatment stage, and does not cover the whole life cycle from prevention to end-of-life care. For example, prevention management of high-risk groups and postoperative psychological intervention of patients are not included. Data lacks long-term, multi-dimensional records, making it difficult to support in-depth research, such as the impact of different treatment plans on the long-term survival rate of colorectal cancer patients, which is limited, hindering the progress of diagnosis and treatment technology. SUMMARY

[0009] In order to solve the problems existing in the prior art, the present application provides a colorectal cancer whole-process management system based on a medical agent, which solves the technical problem that the existing medical platform cannot accurately and efficiently assist in diagnosis and treatment and follow-up management when there are many colorectal cancer patients.

[0010] A colorectal cancer whole-process management system based on medical agent, comprising: a patient end, a doctor assistant end, a colorectal cancer medical agent, a colorectal cancer data platform and a medical knowledge base;

[0011] The patient end is used for uploading information, receiving diagnosis results and conducting medical question and answer dialogue;

[0012] The doctor assistant end is used for receiving information, providing auxiliary diagnosis and treatment suggestions and follow-up plans;

[0013] The colorectal cancer medical agent is used for extracting key information from patient information, recommending auxiliary diagnosis and treatment suggestions according to medical knowledge base data and conducting medical question and answer dialogue;

[0014] The colorectal cancer data platform is used for collecting, sorting and analyzing patient information and surgery information, exchanging and sharing data with HIS and scientific research systems, and intelligently managing patient information;

[0015] The medical knowledge base is used for storing medical knowledge related to colorectal cancer, including colorectal cancer medical textbooks, guidelines, expert consensus, high-quality journal literature, desensitization hospital electronic medical record data and image data (patient examination report pictures).

[0016] Further, the follow-up plan is determined by the colorectal cancer medical agent according to patient information and surgery information, the follow-up plan includes follow-up time and corresponding reexamination items, the doctor assistant end sends reexamination notification to the patient end according to the follow-up plan to remind the patient to reexamine, the patient operates the patient end to confirm in-hospital reexamination or out-of-hospital reexamination, if the doctor assistant end does not receive feedback from the patient end for the reexamination notification, the doctor is reminded to confirm offline, at this time, the doctor confirms whether the patient has reexamined in hospital in the HIS, if not, the hospital contacts the patient offline to let the patient confirm in the patient end, if the patient confirms to reexamine in hospital, the doctor is reminded, the doctor registers and arranges reexamination in the in-hospital system offline, after the reexamination, the doctor uploads the examination report through the doctor assistant end; if the patient confirms to reexamine out of hospital, a page for uploading out-of-hospital examination report is generated, and the patient uploads the examination report through the patient end;

[0017] The colorectal cancer medical agent analyzes and reasons the received medical report. If the reasoning is that the re-visit result is normal, the follow-up process is continued according to the follow-up plan until the follow-up plan is completed. If the reasoning is that the re-visit result is abnormal, the re-visit result and the re-visit basis are pushed to the doctor assistant end to remind the doctor to intervene, and the doctor confirms whether it is recurrence. If it is not recurrence, it is transferred to the re-visit result normal process, and if it is recurrence, the follow-up process is ended. The colorectal cancer medical agent performs pathological analysis to generate a new adjuvant treatment plan or surgery arrangement for the patient, and the doctor confirms it on the doctor assistant end. The doctor assistant end notifies the patient end. If the patient end confirms, the doctor is notified in advance to arrange treatment and registration matters.

[0018] Further, the colorectal cancer medical agent includes a scene agent, an information fusion agent, a knowledge base agent, and a clinical reasoning agent.

[0019] The scene agent is used for input and output information, and judges the subsequent agent link according to the input information, including task decomposition and directional path selection.

[0020] The information fusion agent is used for converting the data input by the scene agent into structured data through multi-modal text and image information recognition, structured understanding, and context integration.

[0021] The knowledge base agent is used for connecting the model output with structured and unstructured medical knowledge resources, evidence support, and knowledge verification.

[0022] The clinical reasoning agent adopts a thinking model and outputs a model inference link, which is used for determining disease assessment, diagnosis and treatment suggestions, and follow-up plans according to structured data and medical knowledge base data.

[0023] Further, the scene agent is connected with an FAQ system. When the patient initiates a question and answer dialogue, the consultation question is first searched in the FAQ system. If the same question is searched, the corresponding answer is directly fed back. If the same question is not searched, the question is sent to the clinical reasoning agent.

[0024] Further, the information fusion agent performs structured processing on the medical report uploaded by the patient during re-visit, including:

[0025] The source of the medical report is identified, and the information about the report source in the medical report is extracted, including the hospital name, the visit time, and the person visited. If the visit time is not within the follow-up time or the person visited is not the same person as the patient himself, the client is fed back that the report is wrong and please re-verify.

[0026] If the visit time and the patient verification pass, it is determined whether the visit report comes from inside or outside the hospital according to the hospital name. If the visit report comes from inside the hospital, the corresponding tabular in-hospital report is extracted from the HIS through the scene agent, and the visit report information is directly extracted from the tabular in-hospital report. If the visit report comes from outside the hospital, the structured data is extracted from the out-of-hospital report image through the multi-modal large model.

[0027] Further, the clinical reasoning agent analyzes the pathology reasoning in the patient visit stage according to the disease information and basic information of the patient in combination with the data in the medical knowledge base, outputs the corresponding recommended treatment scheme, reasoning drug and treatment plan, and determines the reasoning basis by the knowledge base agent. The scene agent arranges and outputs the pathology reasoning analysis, recommended treatment scheme, reasoning drug, treatment plan and reasoning basis to the doctor auxiliary end for assisting the doctor to make a judgment.

[0028] Further, the clinical reasoning agent judges whether the patient re-visit result is normal in the patient re-visit stage. The judgment process is as follows:

[0029] The structured information in the re-visit report is extracted, and the extracted content is divided into four types of indexes of tumor markers, imaging examination, pathology / endoscopy and clinical symptoms;

[0030] The specific description of each index is extracted through keyword matching combined with regular expressions. If there is ambiguous expression, it is marked separately to facilitate the subsequent doctor to focus on judgment. Similar expressions are converted into the same expressions.

[0031] The extracted index-description word pair is matched with abnormal, possibly abnormal, existence of abnormal hidden danger, normal level. Abnormal refers to the existence of clear recurrence evidence, which is matched to the confirmed keyword. Possibly abnormal refers to the existence of relatively clear but non-confirmed recurrence clues, which is matched to the suspected keyword. Existence of abnormal hidden danger refers to the existence of slight abnormality, but it is not enough to point to recurrence, which is matched to the slight abnormality keyword. Normal refers to all indexes without abnormal description, which is matched to the normal keyword.

[0032] After the judgment is completed, the re-visit report, judgment result and judgment basis are uploaded to the doctor auxiliary end to assist the doctor to judge whether the re-visit result is normal.

[0033] Further, the medical knowledge base includes reviews, guidelines related to rectal cancer, and other general medical knowledge content.

[0034] Further, the medical knowledge base also includes a management module for regularly fetching the latest colorectal cancer data to update the medical knowledge base, and constructing a knowledge graph for the content in the medical knowledge base, and setting a frequency label, the frequency label involves three attribute fields, including: a frequency level field, an access counter, and a last update date; each time the knowledge content is retrieved, the access counter value is automatically increased by 1; a daily timing task is executed to recalculate the frequency level: obtain the latest access count of all entries; calculate the percentile distribution of the global access amount and update the frequency label and the last update date of the corresponding entry.

[0035] Further, when the large model calls the medical knowledge base content, hierarchical retrieval is performed according to the frequency label, including:

[0036] High-frequency priority retrieval stage: receiving a user query request; initiating a retrieval instruction to the medical knowledge base: limiting the retrieval range to only include entries with high frequency labels; setting the number of returned results to n times the actual demand, then sorting all results by relevance, and selecting the most relevant specified number of results to return to the user, and the process terminates; if the highest score does not exceed the threshold, then enter the medium-frequency extended retrieval stage;

[0037] Medium-frequency extended retrieval stage: initiating an extended retrieval to the medical knowledge base: expanding the retrieval range to high and medium label entries; setting the number of returned results to m times the demand; merging the results obtained in the high-frequency stage; calculating the highest relevance score of the merged results; if the highest score exceeds the threshold, then reordering all merged results, and selecting the most relevant specified number of results to return, and the process terminates; if the highest score does not exceed the threshold, then enter the full library bottom-up retrieval stage; wherein m>n;

[0038] The full library bottom-up retrieval stage includes: initiating a global retrieval to the medical knowledge base: the retrieval range covers all knowledge entries, and the returned results are set to the actual demand, and the most relevant specified number of results are directly returned.

[0039] The beneficial effects of the present application include:

[0040] 1. Individualized precision service is realized, differential schemes are automatically generated by the system, key node reminders are triggered, patient prognosis experience is improved, a large number of repetitive manual operations are reduced, human cost is reduced, and the vicious cycle of cost and efficiency imbalance is broken; through automatic triggering of reminders by the system, combined with multi-channel reach such as SMS and APP push, the loss rate is greatly reduced, ensuring standardized management throughout the patient process and reducing the risk of recurrence; improving patient management efficiency and effect: the system automatically completes follow-up plan generation, process triggering, data aggregation, and the like, medical personnel can focus on diagnosis and treatment decisions and patient communication, and the work efficiency of medical personnel is improved. Through automatic management, even if the patient scale expands, the increase in human input is controllable, and the unit service cost decreases while the service volume increases.

[0041] 2. A highly efficient tumor follow-up report processing system was constructed through "hierarchical processing + graded retrieval," which reduces system burden while ensuring follow-up efficiency. Specifically, a rough judgment without reasoning is first made on follow-up reports: using structured extraction technology to break down key information such as tumor markers, imaging, and pathology, the reports are quickly classified into three categories: "abnormal," "potentially abnormal," and "potentially abnormal" through keyword matching and rule mapping. This stage does not require inference capabilities, relying only on string matching and preset rules, which significantly reduces the computational cost of real-time processing and enables the system to efficiently handle the accumulated number of follow-up patients. Once the doctor confirms recurrence, targeted reasoning analysis is initiated, and the knowledge base content called during the reasoning process is tagged with frequency—high-frequency core knowledge (such as treatment pathways after recurrence and treatment plans for common metastatic sites in the NCCN guidelines) is marked as "high frequency," medium-frequency auxiliary knowledge (such as differential diagnosis of rare metastatic lesions) is marked as "medium frequency," and low-frequency special cases or cutting-edge research is marked as "low frequency." Through graded retrieval, high-frequency knowledge is prioritized, reducing redundant information queries and significantly accelerating the reasoning response speed. The core advantage of this design lies in the "separation of priorities": the preliminary judgment stage filters out reports with no risk of recurrence and focuses on cases that require attention; the reasoning stage improves efficiency through knowledge-based hierarchical retrieval. This dual mechanism ensures that the system can still operate smoothly even when the number of patients followed up accumulates to a large scale, while reducing the repetitive work of doctors in the initial screening stage, allowing them to concentrate on handling confirmed recurrence cases, ultimately achieving the dual goals of reducing the burden of follow-up treatment and improving efficiency.

[0042] 3. It breaks down data silos, enabling one-stop access to all-dimensional data and real-time synchronization of key information, significantly improving the efficiency and accuracy of clinical decision-making. Scientific data collection has also become more automated and of higher quality, helping to achieve precise diagnosis and treatment and reducing the risk of missed diagnoses.

[0043] 4. It covers the entire life cycle management of colorectal cancer, including prevention, psychological intervention and other services. At the same time, the long-term and multi-dimensional data accumulation can support in-depth scientific research and accelerate the progress of diagnosis and treatment technology and scientific research transformation. Attached Figure Description

[0044] Figure 1 This is a flowchart of a colorectal cancer end-to-end management system based on a medical agent, which is an embodiment of this application.

[0045] Figure 2 This is a schematic diagram illustrating patient-related information involved in an embodiment of this application.

[0046] Figure 3 This is a schematic diagram of surgical information involved in an embodiment of this application.

[0047] Figure 4is a colorectal cancer medical agent output result diagram involved in the embodiments of the present application.

[0048] Figure 5 is a follow-up plan automatic execution flowchart involved in the embodiments of the present application. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0050] Embodiment 1

[0051] A colorectal cancer whole-process management system based on a medical agent, as shown in Figure 1 includes a patient end, a doctor assistant end, a colorectal cancer medical agent, a colorectal cancer data platform and a medical knowledge base;

[0052] The patient end is used for uploading information, receiving diagnosis results and conducting medical question and answer dialogue;

[0053] The doctor assistant end is used for receiving information, providing assistant diagnosis and treatment suggestions and follow-up plans; the received information involves patient information such as Figure 2 as shown: including patient basic information, diagnosis record, medical history, medical order, label and treatment plan; follow-up record includes follow-up plan, follow-up confirmation, follow-up record and follow-up completion progress;

[0054] The colorectal cancer medical agent is used for extracting key information from patient information, recommending assistant diagnosis and treatment suggestions according to medical knowledge base data and conducting medical question and answer dialogue;

[0055] The colorectal cancer data platform is used for collecting, sorting and analyzing patient information and surgery information, exchanging and sharing data with HIS and scientific research system, intelligently managing patient information; as shown in Figure 3 surgery information includes preoperative information, intraoperative information and pathological information of the patient;

[0056] The medical knowledge base is used for storing medical knowledge related to colorectal cancer.

[0057] Specifically, the information uploaded by the patient end includes:

[0058] 1) Inspection results, fecal occult blood, carcinoembryonic antigen; 2) examination results, image report (chest CT, whole abdominal CT, liver ultrasound, etc.), gastrointestinal endoscopy report; 3) pathological report.

[0059] The doctor-assisted end receives information including:

[0060] 1) The patient uploads the preoperative / postoperative examination results on the patient end, and the preliminary answer is obtained by the colorectal agent analysis. The doctor end receives the reminder to be processed, and the doctor confirms the result and pushes it back to the patient end. At the same time, a specific diagnosis and treatment plan is made, and the plan time is set;

[0061] 2) The patient initiates a health knowledge question on the patient end, and the colorectal agent answers it preliminarily. After the doctor confirms it on the doctor end, it is pushed back to the patient end;

[0062] 3) Patients with diagnosis and treatment plans remind doctors to arrange according to the plan and time sequence on the doctor end, such as when to return to hospital for chemotherapy and when to return to hospital for surgery;

[0063] 4) The data platform links the hospital medical record system and synchronizes the examination results uploaded by the patient end. The patient's medical data form a data chain and are displayed on the patient view, which can also be viewed by the doctor on the doctor end.

[0064] The data of the medical knowledge base includes: multi-modal data covering the fields of epidemiology, pathology, diagnosis, treatment, prognosis, prevention, etc. of colorectal cancer, which provides data support for the colorectal cancer medical agent.

[0065] Specifically: ① Structured data, including diagnosis and treatment guideline flowchart and table, structured clinical case data, etc.; ② Unstructured data, including medical literature, expert consensus, unstructured clinical case data, etc.; ③ Picture data, patient uploaded picture format examination report.

[0066] The patient end uploads colorectal cancer related health knowledge questions and disease examination reports. After the colorectal medical agent obtains the information, it outputs possible differential diagnosis and diagnosis and treatment opinions based on guideline recommendations and historical case records.

[0067] In another embodiment, the follow-up plan is determined by the colorectal cancer medical agent according to patient information and surgery information, such as Figure 5As shown, the follow-up plan includes follow-up time and corresponding reexamination items. The doctor assistant end sends a reexamination notification to the patient end according to the follow-up plan to remind the patient of reexamination. The patient operates the patient end to confirm in-hospital reexamination or out-of-hospital reexamination. If the doctor assistant end does not receive feedback from the patient end for the reexamination notification, the doctor is reminded to confirm offline. At this time, the doctor confirms whether the patient has reexamined in hospital in the HIS. If not, the hospital contacts the patient offline to let the patient confirm in the patient end. If the patient confirms to reexamine in hospital, the doctor is reminded. The doctor registers and arranges reexamination in the in-hospital system offline. After the reexamination, the doctor uploads the treatment report through the doctor assistant end. If the patient confirms to reexamine out of hospital, a page for uploading an out-of-hospital treatment report is generated, and the patient uploads the treatment report through the patient end.

[0068] The designation process of the follow-up plan is as follows: a basic follow-up template is set according to the patient's surgery time, reexamination is performed every 3-6 months after 1 month, CEA is monitored every 3 months, imaging such as chest / abdominal CT is performed every 6 months, and colonoscopy is performed every year after surgery.

[0069] The colorectal cancer medical agent analyzes and reasons the received treatment report. If the reasoning is that the reexamination result is normal, the follow-up process is continued according to the follow-up plan until the follow-up plan is completed. If the reasoning is that the reexamination result is abnormal, the reexamination result and the reexamination basis are pushed to the doctor assistant end to remind the doctor to intervene, and the doctor confirms whether it is recurrence. If it is not recurrence, the process of the reexamination result being normal is entered. If it is recurrence, the follow-up process is ended. The colorectal cancer medical agent performs pathological analysis to generate a new adjuvant therapy plan or surgery arrangement for the patient, which is confirmed by the doctor in the doctor assistant end. The doctor assistant end notifies the patient end. If the patient end confirms, the doctor is notified in advance to arrange treatment and registration matters.

[0070] In another embodiment, the colorectal cancer medical agent analyzes and reasons the received treatment report. Whether the reasoning is that the reexamination result is normal or abnormal, the doctor needs to be reminded to intervene, and the doctor confirms whether it is recurrence.

[0071] In another embodiment, the colorectal cancer medical agent includes a scene agent, an information fusion agent, a knowledge base agent, and a clinical reasoning agent.

[0072] The scene agent is used for input and output information, and judges the subsequent agent link according to the input information, including task decomposition and directional path selection.

[0073] The information fusion agent is used for converting the data input by the scene agent into structured data through multi-modal image-text information recognition, structured understanding, and context integration.

[0074] The knowledge base agent is used to connect the model output with structured and unstructured medical knowledge resources, to support evidence and verify knowledge; on the one hand, when the model generates an answer, the knowledge base agent automatically retrieves structured medical resources such as diagnosis and treatment guideline items, drug database fields, and unstructured medical resources such as literature abstracts and clinical path texts, to match the corresponding authoritative basis for the answer, ensuring that each piece of information can be traced back to a specific knowledge source, and avoiding the generation of answers without basis; on the other hand, the agent verifies the model output content in reverse, compares the core conclusions in the knowledge resources such as disease diagnosis standards and drug contraindications, identifies and corrects possible biases or errors such as confusing symptom descriptions of similar diseases, to ensure the accuracy of the answer from the source.

[0075] The clinical reasoning agent adopts a thinking model and outputs a model inference link, which is used to determine disease assessment, diagnosis and treatment suggestions, and follow-up plans according to structured data and medical knowledge base data;

[0076] In another embodiment, the scene agent is connected with an FAQ system, and when a patient initiates a question and answer conversation, the consultation question is first searched in the FAQ system, and if the same question is searched, the corresponding answer is directly fed back, and if the same question is not searched, the question is sent to the clinical reasoning agent.

[0077] In another embodiment, the information fusion agent performs structured processing on the medical report uploaded by the patient during the reexamination, including:

[0078] Identifying the source of the medical report, extracting information about the source of the medical report in the medical report, including the hospital name, the medical time, and the medical person; if the medical time is not within the follow-up time or the medical person is not the same person as the patient, the client is fed back that the report is wrong and please recheck;

[0079] If the medical time and the medical person are verified, it is determined whether the medical report comes from the hospital or the outside of the hospital according to the hospital name, if the medical report comes from the hospital, the corresponding tabular hospital report is extracted from the HIS through the scene agent, and the medical report information is directly extracted from the tabular hospital report; if the medical report comes from the outside of the hospital, structured data is extracted from the outside report image through the multi-modal large model.

[0080] In another embodiment, the clinical reasoning agent, during the patient's medical stage, performs pathological reasoning analysis according to the patient's disease information and basic information combined with the data in the medical knowledge base, outputs the corresponding recommended treatment plan, reasoning drug, and treatment plan, and determines the reasoning basis by the knowledge base agent, and arranges and outputs the pathological reasoning analysis, recommended treatment plan, reasoning drug, treatment plan, and reasoning basis to the doctor's auxiliary end by the scene agent, to assist the doctor in making judgments.

[0081] Regarding the picture information used for model diagnosis, the clinical reasoning agent stage mainly collects information in three directions. The first is the patient's HIS system information, the second is the knowledge base data, and the third is the picture report information uploaded by the patient. The types of pictures provided by the hospital include patient examination reports such as blood tests, urine tests, gastrointestinal endoscopy, and pathological examinations. The clinical reasoning agent mainly extracts text information and table information as the basis for judgment.

[0082] For the extraction of picture report information, first, the text in the picture report is recognized, and then the information is extracted according to the keywords / key indicators in the report, sorted by examination time, and stored in the patient view. For example, in the case of pathological report structuring, key indicators such as differentiation degree, nerve invasion status, and tumor bud count are identified and the information data of these indicators is stored.

[0083] The clinical reasoning agent currently uses two types of large models, the first being a language large model and the second being a multi-modal large model. The language large model is first pre-trained based on the collected basic data, and then fine-tuned for colorectal cancer by constructing corresponding SFT question and answer pairs. Regarding the fine-tuning of the multi-modal large model, more than 10,000 multi-modal SFT question and answer pairs are constructed using the 3,000 picture report information provided by the hospital. The fine-tuning of the multi-modal model is mainly for question and answer based on the text information in the picture report.

[0084] Specifically, the scene agent output effect is as shown in Figure 4 , including data acquisition time, data source, data content, AI pathological reasoning analysis (based on the patient's re-examination report in a certain city hospital on October 10, 2024, and the follow-up report submitted by the patient on October 12, 2024, deep analysis based on model reasoning abnormal information as follows: (10-15 cm from the anus) chronic active inflammation of the rectal membrane, erosion, ulcer formation, local area Ki-67 proliferation index high, nuclear division easy to see, tend to malignant cancer, not ruled out sarcoma-like, biopsy tissue is limited, please combine with clinical, suggest re-examination or further consultation for diagnosis. Immunohistochemistry 1: CK(), CK8 118(), CDX-2(), CD34(vessel·), CD117(-), Dog-1(-), SMA(focal-), Desmin(), CD31(), IV8(), CD10(), K1-67(>40%*)) AI reasoning basis (Chinese Society of Clinical Oncology (CSCO) Colorectal Cancer Diagnosis and Treatment Guidelines Chapter 4.1.2 CT1-2N0 Rectal Cancer Treatment Principles), AI recommended diagnosis and treatment plan (neoadjuvant therapy), AI reasoning for drug use and treatment plan (scheme number, drug name, dosage, administration time and cycle, clinical application);

[0085] In another embodiment, the clinical reasoning agent determines whether the patient recheck result is normal at the patient recheck stage, and the determination process is:

[0086] Structured information in the recheck report is extracted, and the extracted content is divided into four types of indexes, namely tumor markers, imaging examinations, pathology / endoscopy, and clinical symptoms. The specific division can be achieved by matching the title with keywords, such as paragraphs containing “CT” and “MRI” being classified as imaging, and paragraphs containing “CEA” being classified as tumor markers.

[0087] The specific description of each index is extracted by keyword matching combined with regular expressions, such as extracting the “CEA” index and the corresponding description word “significant increase” from “CEA: 56 ng / mL (reference value 0-5 ng / mL), significantly higher than the last time (20 ng / mL)”. From “pelvic CT: anastomotic site shows a 2.3 cm soft tissue shadow, considering metastasis”, the “pelvic CT” index and the corresponding description word “considering metastasis” are extracted. If there is a vague expression, it is marked separately for subsequent doctor's attention; synonyms / paronyms are converted to the same expression, such as “increase” “increase” “increase” being unified as “increase”.

[0088] The extracted index-description word pairs are matched with multiple levels of abnormalities, possible abnormalities, potential abnormal risks, and normality. Specifically: abnormality refers to the presence of clear evidence of recurrence, matching to diagnostic keywords, such as pathology / endoscopy reports containing “cancer cells” “cancer tissue”; imaging reports containing “metastasis” “recurrence”; multiple indexes pointing to clear recurrence, such as CEA significantly increased + CT found metastasis;

[0089] Possible abnormality refers to the presence of relatively clear but non-diagnostic recurrence clues, matching to suspected keywords, such as imaging reports containing “suspected metastasis” “nodule (nature to be determined, not excluding recurrence)”; tumor markers “significantly increased” accompanied by warning symptoms, such as “CEA significantly increased + blood in stool aggravated”;

[0090] Potential abnormality refers to the presence of slight abnormalities, but not enough to point to recurrence, matching to slight abnormality keywords, such as tumor markers “mildly elevated”, such as CEA slightly higher than the reference value, no other abnormalities; single warning symptom, such as “mild weight loss”, no other index abnormalities

[0091] Normality refers to all indexes without abnormal description, matching to normal keywords, such as tumor markers “normal”, imaging “no recurrence”, and no warning symptoms.

[0092] After the determination is completed, the recheck report, the determination result, and the determination basis are uploaded to the doctor's auxiliary end to assist the doctor in determining whether the recheck result is normal.

[0093] In another embodiment, the medical knowledge base also includes a management module for regularly fetching the latest colorectal cancer data to update the medical knowledge base, and building a knowledge graph for the content in the medical knowledge base and setting a frequency tag. Specifically:

[0094] Storage and dynamic updating of the frequency tag:

[0095] 1. The frequency tag involves three attribute fields, including:

[0096] Frequency level field: stores high / medium / low three types of identification; access counter: records the cumulative number of times the content is retrieved; last update date: records the last update time of the tag.

[0097] 2. The dynamic updating process of the tag includes: each time the knowledge content is retrieved, the access count value is automatically increased by 1; a timing task is executed daily to recalculate the frequency level: obtain the latest access count of all entries; calculate the percentile distribution of global access volume; mark the top 20% high access volume entries as high; mark the middle 30% access volume entries as medium; mark the remaining 50% entries as low; update the frequency tag and the last update date of the corresponding entries.

[0098] Execution process of hierarchical retrieval:

[0099] 1. High-frequency priority retrieval stage includes: receiving a user query request; initiating a retrieval instruction to the medical knowledge base: limiting the retrieval range to only include entries with high frequency tags; setting the number of returned results to be 2 times the actual demand (such as 5 entries, take 10 entries); calculating the highest relevance score of the returned results; if the highest score exceeds the threshold (such as 0.7), then sorting all results by relevance, and selecting the most relevant specified number of results to return to the user, and the process terminates; if the highest score does not exceed the threshold, then enter the medium-frequency extended retrieval stage.

[0100] 2. Medium-frequency extended retrieval stage includes: initiating an extended retrieval to the medical knowledge base: expanding the retrieval range to high and medium tag entries; setting the number of returned results to be 3 times the demand; merging the results obtained in the high-frequency stage; calculating the highest relevance score of the merged results; if the highest score exceeds the threshold, then re-sorting all merged results, and selecting the most relevant specified number of results to return, and the process terminates; if the highest score does not exceed the threshold, then enter the full library bottom-up retrieval stage.

[0101] 3. Full library bottom-up retrieval stage includes: initiating a global retrieval to the medical knowledge base: the retrieval range covers all knowledge entries, and the returned results are set to the actual demand, and the most relevant specified number of results are directly returned.

[0102] Other aspects

[0103] 1. The missed detection compensation mechanism includes: in the high-frequency retrieval stage, two retrieval modes are executed synchronously: semantic retrieval based on vector similarity and traditional retrieval based on keyword matching, and then the result sets of the two modes are combined and the repeated entries are eliminated to enter the subsequent processing.

[0104] 2. The cold start processing scheme includes: pre-allocating labels to newly entered warehouse content according to text features: if it contains keywords such as "common", "basic", "important", etc., the label is pre-set to high, if it contains keywords such as "advanced", "reference", "technology", etc., the label is pre-set to medium, and if it is a professional term-intensive content, the label is pre-set to low; and after accumulating enough access data, switch to dynamic labels.

[0105] In another embodiment, the colorectal cancer medical agent combines the medical knowledge base to provide the implementation process of medical question and answer for the patient, specifically as follows:

[0106] Step 1: According to the patient's question, the answer prompt word is clear, the model is clear through instruction constraint, only based on knowledge base information answer, do not fabricate content; for the question beyond the scope of knowledge base, uniformly reply "no relevant authoritative information is found, it is suggested to consult the attending physician".

[0107] Step 2: In the face of patients asking questions in spoken language, the large model first performs normalization processing and converts it into keywords that can be searched in the knowledge base, while identifying potential risk intentions; based on the pre-processed keywords, the most relevant information is matched from the structured knowledge base, and the latest and highest level basis is preferentially selected; the large model converts the retrieved professional information into language that is easy for patients to understand, while forcibly embedding safety prompts.

[0108] Step 3: Set up a rule engine to filter high-risk content; perform keyword audit on the output content, and automatically intercept and prompt for medical treatment if relevant content is involved; collect patient feedback and doctor's review opinions, update the knowledge base and optimize the model fine-tuning data regularly, forming a positive cycle of "data-model-answer".

[0109] Thus, efficient and accurate medical question and answer dialogue is realized, and the patient's problem is solved in a timely manner.

[0110] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A colorectal cancer end-to-end management system based on a medical agent, characterized in that, include: Patient-side, doctor-assistant-side, colorectal cancer medical agent, colorectal cancer data platform, and medical knowledge base; The patient terminal is used to upload information, receive diagnostic results, and conduct medical Q&A dialogues. The doctor's auxiliary terminal is used to receive information, provide auxiliary diagnosis and treatment suggestions, and follow-up plans; The colorectal cancer medical agent is used to extract key information from patient information, recommend auxiliary diagnosis and treatment suggestions based on medical knowledge base data, and conduct medical question-and-answer dialogues. The colorectal cancer data platform is used to collect, organize, and analyze patient and surgical information, exchange and share data with HIS and research systems, and intelligently manage patient information. The medical knowledge base is used to store medical knowledge related to colorectal cancer.

2. The colorectal cancer end-to-end management system based on a medical agent according to claim 1, characterized in that, The follow-up plan is determined by the colorectal cancer medical agent based on patient and surgical information. The plan includes follow-up time and corresponding re-examination items. The doctor's assistant periodically sends re-examination notifications to the patient's end according to the follow-up plan, reminding the patient to attend re-examinations. The patient confirms on their end whether the re-examination will be in-hospital or out-of-hospital. If the doctor's assistant does not receive feedback from the patient regarding the re-examination notification, it prompts the doctor to confirm offline. The doctor then checks with the HIS (Hospital Information System) to confirm whether the patient has already had an in-hospital re-examination. If not, the hospital contacts the patient offline to ask them to confirm on their end. If the patient confirms an in-hospital re-examination, the doctor is notified, and the doctor registers and arranges the re-examination offline in the hospital system. After the re-examination, the doctor uploads the medical report through the doctor's assistant. If the patient confirms an out-of-hospital re-examination, a page for uploading out-of-hospital medical reports is generated, and the patient uploads the medical report through their end. The colorectal cancer medical agent analyzes and infers from the received medical reports. If the inference is that the follow-up results are normal, the follow-up process continues according to the follow-up plan until the follow-up plan is completed. If the reasoning indicates an abnormal follow-up result, the follow-up result and supporting evidence are pushed to the doctor's auxiliary terminal to remind the doctor to intervene and confirm whether there is a recurrence. If there is no recurrence, the process proceeds to the normal follow-up result stage. If there is a recurrence, the follow-up process ends, and the colorectal cancer medical agent performs pathological analysis to generate a new adjuvant treatment plan or surgical arrangement for the patient. The doctor then confirms the plan on the doctor's auxiliary terminal, which notifies the patient's terminal. If the patient confirms the plan, the doctor is notified in advance to arrange treatment and registration.

3. The colorectal cancer end-to-end management system based on a medical agent according to claim 1, characterized in that, The colorectal cancer medical agent includes a scenario agent, an information fusion agent, a knowledge base agent, and a clinical reasoning agent; Scene agents are used to input and output information and determine subsequent agent links based on the input information, including task decomposition and directional path selection; Information fusion agents are used to transform data input from scene agents into structured data through the recognition, structured understanding, and context integration of multimodal graphic information. Knowledge base agents are used to connect model outputs with structured and unstructured medical knowledge resources for evidence support and knowledge verification. The clinical reasoning agent employs a thinking model and outputs model inference links to determine disease assessment, treatment recommendations, and follow-up plans based on structured data and medical knowledge base data.

4. The colorectal cancer end-to-end management system based on a medical agent according to claim 3, characterized in that, The scenario agent is connected to an FAQ system. When a patient initiates a question-and-answer dialogue, the question is first searched in the FAQ system. If the same question is found, the corresponding answer is directly provided. If the same question is not found, the question is forwarded to the clinical reasoning agent.

5. The colorectal cancer end-to-end management system based on a medical agent according to claim 1, characterized in that, The information fusion agent performs structured processing on the medical reports uploaded by patients during follow-up visits, including: Identify the source of the medical report and extract information about the report source, including the hospital name, visit time, and patient. If the visit time is not within the follow-up period or the patient is not the same person as the patient, report an error to the client and ask for re-verification. If the appointment time and patient are verified, the system determines whether the medical report comes from within or outside the hospital based on the hospital name. If the medical report comes from within the hospital, the system extracts the corresponding tabular internal report from the HIS through a scene agent and directly extracts the medical report information from the tabular internal report. If the medical report comes from outside the hospital, the system extracts structured data from the external report image through a multimodal large model.

6. The colorectal cancer end-to-end management system based on a medical agent according to claim 3, characterized in that, During the patient's consultation, the clinical reasoning agent performs pathological reasoning analysis based on the patient's disease information and basic information, combined with data from the medical knowledge base. It outputs corresponding recommended treatment plans, inferred medications, and treatment plans. The knowledge base agent determines the reasoning basis, and the scenario agent organizes and outputs the pathological reasoning analysis, recommended treatment plans, inferred medications, treatment plans, and reasoning basis to the doctor's auxiliary terminal to assist the doctor in making judgments.

7. A colorectal cancer end-to-end management system based on a medical agent according to claim 3, characterized in that, The clinical reasoning agent determines whether the patient's follow-up examination results are normal during the follow-up visit. The determination process is as follows: Extract the structured information from the follow-up visit report and break down the extracted content into four categories of indicators: tumor markers, imaging examinations, pathology / endoscopy, and clinical symptoms. The specific description of each indicator is extracted by combining keyword matching with regular expressions. If there are ambiguous descriptions, they are marked separately to facilitate doctors' subsequent judgment. Then, similar descriptions are converted into identical descriptions. The extracted indicators and descriptive terms were matched with abnormal, possibly abnormal, potentially abnormal, and normal levels. Abnormal means there is clear evidence of recurrence and is matched with diagnostic keywords. Possibly abnormal means there are relatively clear but not confirmed clues to recurrence and is matched with suspected keywords. Potentially abnormal means there is a slight abnormality but not enough to indicate a recurrence and is matched with slightly abnormal keywords. Normal means there are no abnormal descriptions for any indicators and is matched with normal keywords.

8. The colorectal cancer end-to-end management system based on a medical agent according to claim 1, characterized in that, The medical knowledge base includes reviews and guidelines related to rectal cancer, as well as other general medical knowledge.

9. A colorectal cancer end-to-end management system based on a medical agent according to claim 1, characterized in that, The medical knowledge base also includes a management module for regularly crawling the latest colorectal cancer data to update the medical knowledge base, constructing a knowledge graph for the content in the medical knowledge base, and setting frequency tags. The frequency tags involve three attribute fields: frequency level field, access counter, and last update date. Each time knowledge content is retrieved, its access count is automatically incremented by 1. A scheduled task is executed daily to recalculate the frequency level: obtain the latest access count for all entries; calculate the percentile distribution of global access volume and update the frequency tag and last update date of the corresponding entry.

10. A colorectal cancer end-to-end management system based on a medical agent according to claim 9, characterized in that... When the large model calls upon medical knowledge base content, it performs hierarchical retrieval based on frequency tags, including: High-frequency priority retrieval phase: Receive user query request; send retrieval instruction to medical knowledge base: limit the retrieval scope to include only items with high frequency tags; set the number of returned results to n times the actual demand, then sort all results by relevance, select the most relevant specified number of results to return to the user, and the process terminates; if the highest score does not exceed the threshold, proceed to the mid-frequency extended retrieval phase; Mid-frequency extended search phase: Initiate an extended search to the medical knowledge base: expand the search scope to high and mid-level tag entries; set the number of returned results to m times the required amount; merge the results obtained in the high-frequency phase; calculate the highest relevance score of the merged results; if the highest score exceeds the threshold, reorder all merged results and select the most relevant specified number of results to return, and the process terminates; if the highest score does not exceed the threshold, proceed to the full database catch-all search phase; where m>n; The comprehensive database search phase includes: initiating a global search of the medical knowledge base: the search scope covers all knowledge items, the returned results are set to the actual demand, and the most relevant specified number of results are returned directly.

Images