AI model training and evaluation methods for identifying risk of intestinal dysfunction

Abstract

This application belongs to the field of medical perioperative monitoring technology, specifically an AI model training and evaluation method for identifying the risk of bowel dysfunction in colorectal cancer patients during the perioperative period. The method involves: obtaining a trained AI model or an AI model feature dataset; collecting patient medical record data; obtaining positive or negative data on perioperative bowel dysfunction in colorectal patients; calculating the intestinal inflammatory stress index IISi, IISi=ln[(MME×T1) / (LCR+0.1)]; where LCR=K1 / corrected preoperative CRP; setting independent variable data based on medical record data; the independent variable data includes parameter 1 and parameter 2; parameter 1 corresponds to the intestinal inflammatory stress index IISi; parameter 2 corresponds to the surgical procedure; setting positive or negative data on perioperative bowel dysfunction in colorectal patients as dependent variable data; and training the AI ​​model using the above independent and dependent variable data to obtain the AI ​​model feature dataset or the trained AI model.

Description

Technical Field

[0001] This application belongs to the field of medical technology, specifically the field of perioperative monitoring technology, and particularly relates to an AI model training and evaluation method for identifying the risk of bowel dysfunction in colorectal cancer patients during the perioperative period. Background Technology

[0002] Glossary: ​​CRP (C-reactive protein) is an acute-phase protein produced by the liver during infection, inflammation, or tissue damage. It rises rapidly within 6-8 hours, peaking at 24-48 hours. Its level is positively correlated with the severity of inflammation, but it cannot distinguish between infection types (e.g., bacterial / viral). Acute inflammation: CRP levels are significantly elevated during bacterial infections (e.g., pneumonia, postoperative infections). Chronic inflammation: hs-CRP (high-sensitivity CRP) is used to assess cardiovascular disease risk.

[0003] The perioperative period referred to in this application is radical right / left hemicolectomy or total mesorectal excision. The perioperative period encompasses the entire surgical procedure, specifically including the following three consecutive phases: Preoperative phase: from the patient's admission and preoperative examinations until the induction of anesthesia. Intraoperative phase: from the induction of anesthesia until the end of the surgery (skin suturing or the patient leaving the operating room). Postoperative phase: from the end of the surgery until the patient's discharge or complete recovery of bowel function.

[0004] Radical surgery for colorectal cancer is widely used globally as a core treatment to improve patient survival rates; however, the high incidence of postoperative bowel dysfunction remains a significant clinical challenge. These complications include postoperative bowel dysfunction, anastomotic edema, and bacterial translocation infection. Among these, postoperative bowel dysfunction is a common complication.

[0005] Its main manifestations include delayed flatus and defecation, abdominal distension, nausea and vomiting, and decreased bowel sounds. In severe cases, intestinal paralysis or anastomotic edema may occur, and even bacterial translocation infection may occur. It significantly prolongs the patient's hospital stay, increases the 30-day readmission rate, and increases medical costs, and has become an important challenge for perioperative management.

[0006] Currently, methods for predicting postoperative bowel dysfunction in patients mainly rely on clinical observation and laboratory indicators. Traditional methods, such as the time to first flatus, require subjective reports from patients, which can lead to significant inter-observer variability; bowel sound auscultation has insufficient sensitivity and often low accuracy; and while abdominal X-rays can be used for differentiation, they are not suitable for routine high-frequency monitoring due to radiation exposure and insufficient specificity.

[0007] Existing methods generally suffer from a recognition delay of more than 12 hours, causing the golden intervention window to be missed and leading to irreversible damage to the intestinal barrier. However, premature drug intervention, such as using dopamine receptor agonists based on subjective symptoms, can increase the risk of cardiac arrhythmias in patients. On the other hand, delayed intervention also increases the rate of irreversible intestinal barrier damage several times over. Accurately detecting or predicting postoperative bowel dysfunction is crucial for successful postoperative recovery.

[0008] Currently, there is a lack of methods to comprehensively evaluate patients' intestinal function during surgery and postoperative care. If patients miss the golden intervention window after surgery, it will lead to irreversible damage to the intestinal barrier, and there is an urgent need for effective prediction or detection and evaluation methods. Summary of the Invention

[0009] The technical problem this application aims to solve is predicting the probability of postoperative bowel dysfunction in patients. Probabilistic data helps physicians seize the optimal intervention time, reduce the risk of complications, and increase the probability of successful postoperative recovery for patients.

[0010] In this application, the applicant designed a method for obtaining parameters for risk assessment of intestinal dysfunction, which obtains the preoperative lymphocyte count value K1, the intraoperative opioid dosage MME and the operation duration T1; calculates the intestinal inflammatory stress index IISi, and assesses the risk of intestinal dysfunction based on the intestinal inflammatory stress index IISi.

[0011] The Intestinal Inflammatory Stress Index (IISi) is a pioneering intestinal dysfunction risk prediction system and its core innovative index based on routine preoperative indicators and some intraoperative indicators. By integrating multi-dimensional data such as preoperative systemic inflammatory status and intraoperative anesthesia and surgical stress, it can output individualized intestinal dysfunction risk at the end of the operation. After experimental verification, the accuracy rate is over 85%.

[0012] In this application, the applicant also designed an AI model training method for identifying the risk of bowel dysfunction, which is used to obtain a trained AI model or an AI model feature dataset to further improve the accuracy of identifying the risk of bowel dysfunction, raising the accuracy rate to 90%.

[0013] In this application, the applicant also designed an assessment method for identifying the risk of bowel dysfunction, which identifies the risk of bowel dysfunction based on the aforementioned AI model or training dataset and outputs the probability of occurrence of the risk of bowel dysfunction.

[0014] The technical solution proposed in this application to solve the above-mentioned technical problems is a method for obtaining parameters for risk assessment of intestinal dysfunction, which includes obtaining the preoperative lymphocyte count value K1; obtaining the intraoperative opioid dosage MME; obtaining the operation duration T1; and calculating the intestinal inflammatory stress index IISi, IISi=ln[(MME×T1) / (LCR+0.1)]; where LCR=K1 / corrected preoperative CRP; CRP is the C-reactive protein content. MME is the morphine milligram equivalent per unit body weight.

[0015] The operation duration T1 is calculated in hours; a preset risk value for the intestinal inflammatory stress index IISi is set; when the calculated intestinal inflammatory stress index IISi is greater than the preset risk value, a high-risk assessment result for intestinal dysfunction is output; when the calculated intestinal inflammatory stress index IISi is less than the preset risk value, a low-risk assessment result for intestinal dysfunction is output.

[0016] The technical solution of this application to solve the above-mentioned technical problems can also be an AI model training method for identifying the risk of intestinal dysfunction, used to obtain a trained AI model or an AI model feature dataset; collect patient medical record data; the medical record data includes surgical method, preoperative CRP, preoperative lymphocyte count K1, intraoperative opioid dosage MME, and surgical duration T1; CRP is the C-reactive protein content value; obtain positive or negative data of perioperative intestinal dysfunction in colorectal patients; calculate the intestinal inflammatory stress index IISi, IISi=ln[(MME×T1) / (LCR+0.1)]; where LCR=K1 / corrected preoperative CRP; set independent variable data according to the medical record data; the independent variable data includes parameter 1 and parameter 2; parameter 1 corresponds to the intestinal inflammatory stress index IISi; parameter 2 corresponds to the surgical method; set the positive or negative data of perioperative intestinal dysfunction in colorectal patients as dependent variable data; train the AI ​​model with the above independent variable data and dependent variable data to obtain an AI model feature dataset or a trained AI model.

[0017] When the intestinal inflammatory stress index (IISi) is greater than 2.355, parameter 1 is assigned a value of 1; when the intestinal inflammatory stress index (IISi) is less than or equal to 2.355, parameter 1 is assigned a value of 0. Surgical procedures include laparoscopic surgery and open surgery; if it is laparoscopic surgery, parameter 2 is assigned a value of 0; if it is open surgery, parameter 2 is assigned a value of 1. Parameter 1 is the independent variable data of the intestinal inflammatory stress index (IISi). Parameter 2 is the independent variable data of the surgical procedure.

[0018] The independent variable data also includes parameter 3, which corresponds to the patient's stage N. Stage N includes N0, N1, and N2. When the stage N is N0, parameter 3 has a value of 0; when the stage N is N1, parameter 3 has a value of 1; and when the stage N is N2, parameter 3 has a value of 2. Parameter 3 is the independent variable data for stage N.

[0019] The independent variable data also includes parameter group 3, which corresponds to the N-stage of the patient. The N-stage includes N0, N1, and N2. Parameter group 3 includes parameters x31 and x32. When the patient is in stage N0: x31=0, x32=0; when the patient is in stage N1: x31=1, x32=0; when the patient is in stage N2: x31=0, x32=1. Parameter group 3 is the independent variable data for the N-stage.

[0020] The independent variable data also includes parameter 4, which corresponds to the patient's T stage. The T stage includes non-T4 stage and T4 stage. When the T stage is non-T4 stage, parameter 4 is assigned a value of 0; when the T stage is T4 stage, parameter 4 is assigned a value of 1. Parameter 4 is the T stage independent variable data.

[0021] The T-period can be divided into T1, T2, T3, and T4 periods. When the T-period is T1, parameter 4 is assigned a value of 0; when the T-period is T2, parameter 4 is assigned a value of 0; when the T-period is T3, parameter 4 is assigned a value of 1; and when the T-period is T4, parameter 4 is assigned a value of 2. Parameter 4 is the independent variable data for the T-period.

[0022] The operation duration T1 is calculated in hours; the definition of positive or negative data for bowel dysfunction is as follows: During the operation period of colorectal patients, within 72 hours after the operation, the patient is evaluated. If two or more of the following conditions are present, it is considered positive; if less than two are present, it is considered negative: Condition 1, abdominal distension with VAS-A score ≥6; Condition 2, bowel sounds <2 times / minute; Condition 3, need for pharmacological prokinetic or mechanical decompression intervention; the dependent variable data for positive patients is assigned a value of 1; the dependent variable data for negative patients is assigned a value of 0.

[0023] The AI ​​model is a machine learning classification model, which includes any one of the following: multifactor logistic regression model, random forest model, support vector machine model, or artificial neural network model.

[0024] The AI ​​model is a multifactor logistic regression model;

[0025] P=1 / (1+Exp{-(β0+β1 · x1+β2 · x2+β31 · x31+β32 · x32+β4 ·x4)}); The model coefficients obtained after training are: β0=−10.631; β1=4.164; β2=1.177; β31=0.577; β32=1.112; β4=0.849; where: x1 is parameter 1; x2 is parameter 2; x31: parameter x31 corresponding to N stage N1; x32: parameter x32 corresponding to N stage N2; x4 is parameter 4 corresponding to T stage. When T stage is not T4 stage, parameter 4 is assigned a value of 0; when T stage is T4 stage, parameter 4 is assigned a value of 1.

[0026] The technical solution of this application to solve the above-mentioned technical problems can also be an AI assessment method for identifying the risk of intestinal dysfunction. This method involves collecting patient medical record data, setting independent variable data based on the medical record data, using the AI ​​model feature dataset obtained by the above method in conjunction with the AI ​​calculation model, outputting the predicted probability of intestinal dysfunction, and identifying the risk of colorectal patients' intestinal dysfunction during surgery based on the predicted probability.

[0027] Set a preset risk value for the intestinal inflammatory stress index IISi; when the calculated intestinal inflammatory stress index IISi is greater than the preset risk value, output a high-risk assessment result for intestinal dysfunction; when the calculated intestinal inflammatory stress index IISi is less than the preset risk value, output a low-risk assessment result for intestinal dysfunction.

[0028] The technical solution of this application to solve the above-mentioned technical problems can also be an electronic device, the electronic device comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the executable instructions in the memory to implement the method as described above.

[0029] The technical solution of this application to solve the above-mentioned technical problems can also be a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above-mentioned method is implemented.

[0030] The technical effects of the above-mentioned technical solution include: the creation of the intestinal inflammatory stress index IISi and its acquisition method, which can calculate the risk assessment parameters of intestinal dysfunction, namely the intestinal inflammatory stress index IISi, through several key factors.

[0031] The technical benefits of the above-mentioned solution include: timely assessment of bowel dysfunction risk post-surgery; assessment and prediction of bowel dysfunction risk based on output bowel dysfunction risk assessment parameters; and convenient implementation of appropriate interventions during the golden intervention period post-surgery. This avoids the problem of delayed bowel dysfunction identification leading to missed golden intervention periods.

[0032] The technical effects of the above-mentioned technical solution include: setting a preset risk value for the intestinal inflammatory stress index IISi, facilitating risk assessment, and providing intuitive risk warnings.

[0033] The technical effects of the above-mentioned technical solution include: an AI model training method for identifying the risk of intestinal dysfunction, which can convert case data into models and training datasets, facilitating subsequent evolution and updates, and improving the efficiency and accuracy of risk identification.

[0034] The technical effects of the above-mentioned technical solution include: accurate definition of positive or negative data for intestinal dysfunction, which improves the applicability of subsequent models.

[0035] The technical effects of the above solution include: multiple independent variable data can be used, either one or multiple simultaneously for training, improving training accuracy. The accuracy of independent variable data acquisition is high.

[0036] The technical effects of the above-mentioned technical solution include: the setting of the intestinal inflammatory stress index IISi independent variable data is simple and clear, and easy to calculate.

[0037] The technical effects of the above-mentioned technical solution include: the data on surgical method, T-stage, and N-stage are simple, clear, and easy to calculate.

[0038] The technical effects of the above-mentioned technical solution include: the AI ​​model includes a multi-factor logistic regression model, which is very suitable for the application scenario of this application, has good model matching, and has high recognition accuracy after training.

[0039] The technical effects of the above-mentioned technical solution include: the model parameters are very suitable for assessing the risk of intestinal dysfunction during the operation of colorectal patients, the parameter matching is good, and the recognition accuracy after training is high.

[0040] The technical effects of the above-mentioned technical solution include: an AI assessment method for identifying the risk of intestinal dysfunction; the trained model and parameters are easy to apply; it is convenient for continuous updates and iterations; and it improves the accuracy of risk assessment. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of a method for obtaining parameters for risk assessment of bowel dysfunction, as described in one of the embodiments. Figure 1 ;

[0042] Figure 2 This is a schematic diagram of an AI assessment method for identifying the risk of bowel dysfunction, one of the embodiments.

[0043] Figure 3 This is a schematic diagram of an AI model training method as one of the embodiments. Figure 1 ;

[0044] Figure 4 This is a partial schematic diagram of one embodiment of the AI ​​model training method;

[0045] Figure 5 This is a partial schematic diagram of one embodiment of the AI ​​model training method;

[0046] Figure 6 This is a partial schematic diagram of one embodiment of the AI ​​model training method;

[0047] Figure 7 This is a partial schematic diagram of one embodiment of the AI ​​model training method;

[0048] Figure 8 This is a partial schematic diagram of one embodiment of the AI ​​model training method;

[0049] Figure 9 This is a partial schematic diagram of one embodiment of the AI ​​model training method;

[0050] Figure 10 This is a schematic diagram of one embodiment of the AI ​​evaluation method;

[0051] Figure 11 This is one example of the basic case information for the training and validation queues;

[0052] Figure 12 This is a schematic diagram of the PIP and stable selection frequency of one of the candidate variables in the embodiment.

[0053] Figure 13 This is a schematic diagram of the results of a multivariate logistic regression analysis, one of the embodiments.

[0054] Figure 14 This is a predictive nomograph constructed as one of the embodiments;

[0055] Figure 15 This is the ROC curve of the training queue in one of the embodiments;

[0056] Figure 16 This is a comparison of the area under the ROC curve, the p-value, and the 95% confidence interval calculated from the training queue in one of the embodiments.

[0057] Figure 17 This is the ROC curve of the verification queue in one of the embodiments;

[0058] Figure 18 This is a comparison of the area under the ROC curve, the p-value, and the 95% confidence interval calculated by the verification queue in one of the embodiments.

[0059] Figure 19 This is the calibration curve for the training queue in one of the embodiments;

[0060] Figure 20This is a calibration curve for a verification queue, as shown in one of the embodiments.

[0061] Figure 21 This is a clinical decision curve for a training cohort, one of the implementation examples;

[0062] Figure 22 This is a clinical decision curve for a validation cohort, one of the implementation examples;

[0063] Figure 23 It is a comparison of the sensitivity curves of training queues for three different algorithms;

[0064] Figure 24 It is a comparison of the sensitivity curves of three different algorithms for verifying queues;

[0065] Figure 25 This is a flowchart illustrating an AI-based assessment method for identifying the risk of bowel dysfunction. Detailed Implementation

[0066] The content of this application will be further described in detail below with reference to the accompanying drawings. It should be noted that the following description is of preferred embodiments of this application and does not constitute any limitation on this application. The description of preferred embodiments is merely an explanation of the general principles of this application. The use of terms such as "first," "second," "A," "B," "parameter 1," "parameter 2," and "parameter 3" in this application is for ease of explanation only and does not represent a temporal or spatial order. The combinations of letters and numbers "TA," "TB," "H," and "parameter 1" in this application are for ease of explanation only, and their specific meanings are determined by the specific content they represent.

[0067] like Figure 1 In one embodiment of a method for obtaining parameters for risk assessment of intestinal dysfunction, the preoperative lymphocyte count K1 is obtained; the intraoperative opioid dosage MME is obtained; the operation duration T1 is obtained; and the intestinal inflammatory stress index IISi is calculated, IISi=ln[(MME×T1) / (LCR+0.1)]; where LCR=K1 / corrected preoperative CRP (mg / L).

[0068] The method for obtaining the corrected preoperative C-reactive protein (CRP) test value is as follows: set a lower limit threshold, and when the CRP test value is lower than the lower limit threshold, such as 1 mg / L, calculate it as 1 mg / L.

[0069] Specific preoperative CRP correction rules: When the CRP test value is <1 mg / L, it is forced to be calculated as 1 mg / L in the formula; when the CRP test value is ≥1 mg / L, the actual measured value is used directly.

[0070] Avoid CRP values ​​of 0 or extremely low values ​​(e.g., 0.05), as these will result in an abnormally large LCR. Without correction, patients with very low CRP (representing extremely low inflammation and should be low-risk) will have extremely high LCRs, leading to an extremely large denominator (LCR + 0.1) in the IISi formula. Consequently, the calculated IISi value will become a very small negative number, thus affecting the model's predictive ability.

[0071] Preoperative blood tests are defined as the most recent blood test results taken after the patient's admission and before the induction of surgical anesthesia. The specific time frame is limited to 24-48 hours before surgery.

[0072] The preoperative lymphocyte count K1 is the absolute lymphocyte count (LYM) in a complete blood count (CBC), measured in units of 10^9 / L. Similar to CRP, it refers to the result of the most recent venous blood sample taken within 24-48 hours before surgery. Clinically, CRP and CBC are performed using the same batch of blood.

[0073] Intraoperative opioid dosage (MME) refers to morphine milligram equivalents per kilogram (MME / kg). First, the dosage of each drug is converted into morphine milligram equivalents (mg) using an equivalent conversion factor. Then, it is divided by the patient's preoperative weight (kg) to obtain the final parameter included in the formula (unit: mg / kg).

[0074] Assume a patient weighs 60kg and uses 40μg sufentanil and 0.8mg remifentanil during surgery. The equivalence factor for sufentanil is 1000, and for remifentanil it is 100. The surgery duration T1 is 3 hours. Then: Calculate the total MME (mg): (0.04mg × 1000) + (0.8mg × 100) = 40 + 80 = 120mg. Calculate the dose per unit body weight (MME / kg): 120mg / 60kg = 2.0mg / kg. Then substitute into the IISi formula for calculation.

[0075] Example 1: Since sufentanil is currently the most commonly used potent analgesic in radical resection of colorectal cancer, this example sets its conversion factor to 1000. That is: 1 μg sufentanil = 1 mg MME.

[0076] Example 2: Remifentanil is commonly used for continuous intraoperative infusion to counteract the stress of skin incision and exploration. In this example, the conversion factor is set to 100. That is: 1 mg remifentanil = 100 mg MME.

[0077] Example 3: Fentanyl is a classic analgesic, and in this example, its conversion factor is set to 100. That is: 0.1 mg fentanyl = 10 mg MME.

[0078] Example 4: If other opioids are used during the procedure, the following conversion standards should be used: Morphine: conversion factor is 1. Hydromorone: conversion factor is 7 (i.e., 1mg ≈ 7mg MME). Oxycodone: conversion factor for injection is 1.

[0079] like Figure 2 In one embodiment of a method for obtaining parameters for risk assessment of intestinal dysfunction, the operation duration T1 is calculated in hours; a preset risk value for the intestinal inflammatory stress index IISi is set; when the calculated intestinal inflammatory stress index IISi is greater than the preset risk value, a high-risk assessment result for intestinal dysfunction is output; when the calculated intestinal inflammatory stress index IISi is less than or equal to the preset risk value, a low-risk assessment result for intestinal dysfunction is output.

[0080] Another option is a perioperative bowel dysfunction risk assessment device, comprising a data input module, a calculation module, and an assessment module. The data input module is used to input the preoperative lymphocyte count K1, intraoperative opioid dosage MME, and surgical duration T1. The calculation module is used to calculate the intestinal inflammatory stress index IISi; the intestinal inflammatory stress index IISi = ln[(MME × T1) / (LCR + 0.1)]; where LCR = K1 / corrected preoperative CRP (mg / L). The assessment module is used to compare a preset risk value with the intestinal inflammatory stress index IISi to obtain the bowel dysfunction risk assessment result.

[0081] The method to correct the preoperative C-reactive protein (CRP) value is as follows: when the CRP test value is below 1 mg / L, adjust it to 1 mg / L.

[0082] like Figure 3 A method for training an AI model to identify the risk of perioperative bowel dysfunction, used to obtain a trained AI model or an AI model feature dataset, includes the following steps: collecting patient medical record data and setting independent variable data based on the medical record data; obtaining positive or negative data on perioperative bowel dysfunction in colorectal patients and setting the positive or negative data as dependent variable data; training the AI ​​model with the above independent variable data and dependent variable data to obtain an AI model feature dataset or a trained AI model.

[0083] The definition of positive or negative data for colorectal patients with postoperative bowel dysfunction is as follows: within 72 hours after surgery, the patient is assessed and two or more of the following conditions are present, which is considered positive; less than two conditions are considered negative.

[0084] Condition 1: Abdominal distension with a VAS-A score ≥ 6. Condition 2: Bowel sounds < 2 times / minute. Condition 3: Requires pharmacological prokinetic intervention or mechanical decompression.

[0085] The dependent variable data for positive patients was assigned a value of 1; the dependent variable data for negative patients was assigned a value of 0.

[0086] like Figure 4 This paper presents an AI model training method for identifying the risk of perioperative bowel dysfunction. The independent variables include parameter 1 and parameter 2; parameter 1 corresponds to the intestinal inflammatory stress index IISi, and parameter 2 corresponds to the surgical procedure. Specifically, parameter 1 is the independent variable data for the intestinal inflammatory stress index IISi, and parameter 2 is the independent variable data for the surgical procedure.

[0087] like Figure 5 When the intestinal inflammatory stress index IISi is greater than 2.355, parameter 1 is assigned a value of 1; when the intestinal inflammatory stress index IISi is less than or equal to 2.355, parameter 1 is assigned a value of 0.

[0088] like Figure 6 A method for training an AI model to identify the risk of perioperative bowel dysfunction, wherein the surgical procedure includes laparoscopic surgery and open surgery; if it is laparoscopic surgery, parameter 2 is assigned a value of 0; if it is open surgery, parameter 2 is assigned a value of 1.

[0089] like Figure 7 This paper describes an AI model training method for identifying the risk of perioperative bowel dysfunction. The independent variable data includes parameter 4, which corresponds to the patient's T stage. The T stage includes non-T4 and T4 stages. When the T stage is non-T4, parameter 4 is assigned a value of 0; when the T stage is T4, parameter 4 is assigned a value of 1. Parameter 4 is the T stage independent variable data.

[0090] like Figure 8 The independent variable data also includes parameter 4, which corresponds to the patient's stage, T stage. T stage includes T1, T2, T3, and T4 stages. When the T stage is T1, parameter 4 is assigned a value of 0; when the T stage is T2, parameter 4 is assigned a value of 0; when the T stage is T3, parameter 4 is assigned a value of 1; and when the T stage is T4, parameter 4 is assigned a value of 2.

[0091] like Figure 9 This paper presents an AI model training method for identifying the risk of perioperative bowel dysfunction. The independent variable data also includes parameter 3, which corresponds to the patient stage N. Stage N includes N0, N1, and N2. When stage N is N0, parameter 3 has a value of 0; when stage N is N1, parameter 3 has a value of 1; and when stage N is N2, parameter 3 has a value of 2. Parameter 3 is the independent variable data for stage N.

[0092] The training data can include the following four independent predictors: IISi (Intestinal Inflammatory Stress Index) is the core continuous variable. Surgery Type is a categorical variable (laparoscopic / open surgery). T Stage: a binary variable (non-T4 / T4). N Stage: a categorical variable (N0 / N1 / N2).

[0093] A method for training an AI model to identify the risk of perioperative bowel dysfunction is disclosed. The AI ​​model is a machine learning classification model, which includes any one of a multivariate logistic regression model, a random forest model, a support vector machine model, or an artificial neural network model. Preferably, a multivariate logistic regression model is used. This embodiment demonstrates the best-performing multivariate logistic regression model, but this prediction is essentially a binary classification supervised learning task. Therefore, other mainstream machine learning algorithms for processing this type of structured data are also applicable.

[0094] The AI ​​model includes a multifactor logistic regression model;

[0095] P=1 / (1+Exp{-(β0+β1 · x1+β2 · x2+β31 · x31+β32 · x32+β4 · x4)}); The model coefficients obtained after training are: β0=−10.631; β1=4.164; β2=1.177; β31=0.577; β32=1.112; β4=0.849.

[0096] Where: x1 is the intestinal inflammatory stress index IISi independent variable data, i.e., parameter 1; x2 is the surgical method independent variable data, i.e., parameter 2; x31: parameter x31 corresponding to N stage N1; x32: parameter x32 corresponding to N stage N2; x4 is parameter 4 corresponding to T stage. When T stage is not T4, parameter 4 is assigned a value of 0; when T stage is T4, parameter 4 is assigned a value of 1.

[0097] like Figure 10 In one embodiment of an AI assessment method for identifying the risk of perioperative bowel dysfunction, patient medical record data is collected, and independent variable data is set based on the medical record data; the AI ​​model feature dataset obtained by the above method is used in conjunction with the AI ​​calculation model to identify the risk assessment data of perioperative bowel dysfunction in colorectal patients.

[0098] An embodiment of an electronic device includes a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the executable instructions in the memory to implement the method described above.

[0099] In one embodiment of a computer-readable storage medium, computer program instructions are stored on the computer-readable storage medium, which, when executed by a processor, implement the method as described above.

[0100] The research process involved in this application is described below. First, step 1: medical record data acquisition; medical records of 1580 patients with colorectal cancer who underwent surgical treatment at the Second Affiliated Hospital of Nanchang University from July 2015 to June 2025 were retrieved. They were divided into a training cohort (n=1106) and a validation cohort (n=474) in a 7:3 ratio.

[0101] like Figure 11 The figure compares the basic information of the training queue (n=1106) and the validation queue (n=474).

[0102] Inclusion criteria for cases: 1) undergoing radical right / left hemicolectomy or total mesorectal excision; 2) age ≥ 18 years; 3) having complete preoperative examination and postoperative 6-hour monitoring data; 4) colorectal cancer being a single primary malignant tumor.

[0103] Exclusion criteria: 1) Emergency surgery or palliative resection; 2) Preoperative bowel obstruction; 3) Data missing >20%.

[0104] Step 2. Data grouping and variable study: The selected cases were further grouped as follows: (1) Age was regrouped into <60 years, 60-74 years and ≥75 years; (2) Gender was divided into male and female; (3) BMI was grouped into <25kg / m² and ≥25kg / m²; (4) Tumor size was divided into two groups: ≤5cm and >5cm; (5) Surgical method was divided into two groups: open surgery group and laparoscopic surgery group; (6) All T stages and N stages (radiological stages) were recorded in detail according to the 8th edition of the American Joint Committee on Cancer (AJCC) staging.

[0105] In this application, T staging refers to the depth of invasion of the intestinal wall by the primary tumor. According to the AJCC 8th edition colorectal cancer criteria: T1: Tumor invades the submucosa. T2: Tumor invades the muscularis propria. T3: Tumor penetrates the muscularis propria and reaches the pericolorectal tissues. T4: Tumor penetrates the visceral peritoneum (T4a) or directly invades or adheres to other organs / structures (T4b).

[0106] This model treats T-stage as a binary variable: Non-T4 stage: defined as the low-risk group. This includes T1 (invasion of the submucosa), T2 (invasion of the muscularis propria), and T3 (penetration of the muscularis propria but without invasion of the peritoneum / organs). T4 stage: defined as the high-risk group. Specifically refers to T4a (penetration of the visceral peritoneum) and T4b (direct invasion of adjacent organs or structures).

[0107] In this application, the N-stage (Node Stage) reflects the involvement of mesenteric lymph nodes. The N-stage data in this system are based on the clinical staging (cN Stage) results assessed by preoperative enhanced CT or MRI of the abdominal and pelvic region. N0: No enlarged or structurally abnormal regional lymph nodes seen on imaging. N1: 1 to 3 regional lymph node metastases detected on imaging. N2: ≥4 regional lymph node metastases detected on imaging.

[0108] (7) Calculation of Intestinal Inflammatory Stress Index (IISi): The calculation method is as follows: IISi = ln[(Intraoperative opioid dosage MME × Operation duration T1) / (LCR + 0.1)], where LCR = Preoperative lymphocyte count (×10) 9 / L) / Corrected preoperative CRP (mg / L); Corrected preoperative CRP: When the CRP value is below 1 mg / L, it is calculated as 1 mg / L; According to the X-tile cutoff value analysis, when IISi is greater than 2.355, the patient is judged as a high-risk patient for bowel dysfunction; when IISi is less than or equal to 2.355, the patient is judged as a low-risk patient for bowel dysfunction.

[0109] In this application, X-tile Cut-off Analysis is a statistical method and visualization software specifically designed to find the optimal classification threshold for biomarkers.

[0110] In this application, the specific meaning is as follows: Since the core indicator of this invention, IISi (Intestinal Inflammatory Stress Index), is a continuous value, in order to facilitate doctors' use in clinical practice, we need to find a specific dividing value to divide patients into two groups: "high risk" and "low risk".

[0111] X-tile analysis iterates through all possible numerical cut-off points, calculates the statistical difference (X2 value) between the two groups at each cut-off point, and finally selects the value with the most significant difference (smallest P value), namely 2.355, as the best cut-off value for determining the risk of bowel dysfunction.

[0112] Step 3. Definition of positive endpoint: Bowel dysfunction is considered to have occurred if ≥2 of the following criteria are met within 72 hours after surgery: ① Abdominal distension VAS-A score ≥6; ② Bowel sounds <2 times / minute; ③ Requires pharmacological prokinetic or mechanical decompression intervention.

[0113] Step 4: Quantitative assignment, screening, and predictive model construction. Dependent variable (outcome variable) assignment: Based on the definition of the positive endpoint above, the outcome variable is assigned a value of 1 for patients who achieved the final outcome; and a value of 0 for patients who did not achieve the final outcome.

[0114] Independent variable (risk factor) assignment: (1) Age: <60 years old (assigned value 0), 60-74 years old (assigned value 1), ≥75 years old (assigned value 2); (2) Gender: male (assigned value 0), female (assigned value 1); (3) BMI: <25kg / m² (assigned value 0), ≥25kg / m² (assigned value 1); (4) Tumor size: ≤5cm (assigned value 0), >5cm (assigned value 1); (5) Surgical method: laparoscopic surgery (assigned value 0), open surgery (6) According to the AJCC 8th edition staging: T1 / T2 (0), T3 (1), T4 (2); N0 (0), N1 (1), N2 (2); (7) IISi index: when IISi is greater than 2.355, the patient is judged as a high-risk patient for bowel dysfunction (1); when IISi is less than or equal to 2.355, the patient is judged as a low-risk patient for bowel dysfunction (0).

[0115] based on Figure 11 The training queue data shown uses a dual-channel method to filter candidate variables.

[0116] In this application, the dual-channel method refers to a dual verification and screening strategy of "Bayesian probability + frequency stability" adopted by the present invention to ensure the robustness of the predictor. Only variables that pass the screening of both of these independent statistical channels can enter the final AI model.

[0117] First, a variable selection method based on the Bayesian posterior inclusion probability (PIP) was used to calculate the probability of each variable being included in the model within the Bayesian model averaging framework, with a threshold of PIP ≥ 0.80. Then, a stability selection method was employed, using repeated sampling and stepwise regression to statistically determine the inclusion frequency of each variable, with a threshold of stab_freq ≥ 0.70.

[0118] like Figure 12 The results showed that the PIP values ​​for surgical type (Surgery Type), Intestinal Inflammatory Stress Index (IISi), T stage, and N stage were close to or reached 1, with inclusion frequencies of 1.00, 1.00, 0.8875, and 0.825, respectively. All met the dual threshold and were thus identified as robust candidate predictive factors. Figure 12 Although age had a high inclusion frequency (stab_freq=0.7875), its PIP was 0.446, which did not reach the threshold, so it was not included in the modeling. Sex, body mass index (BMI), and tumor size group all did not reach the threshold and were not included in the modeling.

[0119] Therefore, surgical type, intestinal inflammatory stress index (IISi), T stage, and N stage were selected as retained independent variables for subsequent model training.

[0120] The variables retained above were incorporated into a multivariate logistic regression model to screen for independent risk factors of postoperative bowel dysfunction in colorectal cancer and to construct the model.

[0121] like Figure 13 The analysis showed that IISi (per unit) was an independent risk factor (OR = 64.361, 95% CI 36.449–119.835, P < 0.001); OR (Odds Ratio): an indicator quantifying the intensity of risk. OR > 1 indicates that the factor is a risk factor; the larger the value, the stronger the risk effect. In this application, it reflects how many times higher the probability of postoperative bowel dysfunction is in the high-risk group compared to the low-risk group.

[0122] 95% CI (95% Confidence Interval): The range of possible values ​​for the estimated true effect size. An interval not containing 1 indicates statistical significance. In this application, it refers to the range of values ​​that the population parameter (here referring to the true OR value) may fall into under a certain probability guarantee (i.e., 95%).

[0123] P (Probability value, also known as significance level / probability value): Used to determine whether a statistical difference is significant. It represents the probability that the observed difference is caused by random error.

[0124] Surgical category (open surgery: OR=3.247, 95%CI 2.125–4.991, P<0.001; laparoscopic surgery as a reference) was an independent risk factor; heterogeneity existed within the T stage, with T4 being significantly different from T1–T2 (OR=2.321, 95%CI 1.244–4.413, P=0.009), while T3 was not statistically significant relative to T1–T2 (OR=0.991, 95%CI 0.547–1.818, P=0.975); N stage N1 (OR=1.780, 95%CI 1.111–2.869, P=0.017) and N2 (OR=3.038, 95%CI 1.780–5.253, P<0.001) were both independent risk factors.

[0125] Given that the T stage is a multilevel variable and the T3 level was not significant, the final model classified the T stage into "non-T4 / T4" categories. After the binary classification, the effect of T4 relative to non-T4 was (OR=2.34, 95%CI1.51–3.63, P<0.001); and the likelihood ratio test showed that adding this binary variable significantly improved the model fit (LRT: P=0.0001201), so it was retained in the final model.

[0126] Further comprehensive analysis of all independent risk factors for postoperative bowel dysfunction in colorectal cancer patients was conducted, and the multivariate regression formula was updated as follows:

[0127] P=1 / (1+Exp{-(β0+β1 · x1+β2 · x2+β31 · x31+β32 · x32+β4 ·x4)}); where the model coefficients are: β0 = −10.631; β1 = 4.164; β2 = 1.177; β31 = 0.577; β32 = 1.112; β4 = 0.849. Where: x1: IISi; x2: surgical category; x31: N stage N1; x32: N stage N2; x4: T stage (T4 = 1, not T4 = 0). In the Logistic regression model, since N stage is a categorical variable containing three categories (N0, N1, N2), using N0 stage (no lymph node metastasis) as the reference group, N stage is decomposed into two indicator variables: x31: representing the N1 stage effect (i.e., the additional risk brought by N1 relative to N0). x32: representing the N2 stage effect (i.e., the additional risk brought by N2 relative to N0).

[0128] The corresponding relationship matrix is ​​as follows: When the patient is in stage N0: x31=0, x32=0 (baseline state). When the patient is in stage N1: x31=1, x32=0. When the patient is in stage N2: x31=0, x32=1.

[0129] It should be noted that the multivariate logistic regression model and its specific coefficients (β0 to β4) given in the above embodiments are only one of the preferred prediction models constructed by this invention based on training cohort data. The core technical concept of this invention lies in selecting a specific combination of feature variables (i.e., intestinal inflammatory stress index IISi, surgical method, T stage, and N stage) to assess the risk of intestinal dysfunction. Therefore, those skilled in the art can achieve the purpose of this invention by using other artificial intelligence or machine learning algorithms for training based on the same combination of feature variables.

[0130] As a parallel embodiment of the present invention, the AI ​​prediction model can also be constructed using the following algorithms, all of which fall within the scope of protection of the present invention: Decision trees and ensemble learning algorithms: such as Random Forest, Gradient Boosting Decision Tree (GBDT), XGBoost, Light GBM, etc. These algorithms can effectively handle nonlinear relationships between features and output comprehensive prediction results by integrating multiple weak classifiers. Support Vector Machine (SVM): Uses kernel functions to map features such as IISi to a high-dimensional space to find the optimal classification hyperplane. Artificial Neural Network (ANN) or Deep Learning Model: Constructs a Multilayer Perceptron (MLP) containing an input layer, hidden layers, and an output layer, optimizes the weights through the backpropagation algorithm, and outputs the probability of intestinal dysfunction. Bayesian Classifier: Calculates the posterior probability of each feature combination based on Bayes' theorem. When using the above alternative algorithms, although the mathematical expression of the final model is no longer the Logit linear equation mentioned above, its logic for risk prediction using IISi and clinicopathological features is completely consistent with this embodiment.

[0131] Since Random Forest, Support Vector Machine (SVM), and other similar models are non-linear "black box" models, their mathematical principles dictate that they cannot be written out as a single-line explicit analytical formula like Logistic Regression. This is an inherent characteristic of the algorithms themselves.

[0132] like Figure 23 and Figure 24 The chart presents a comparison of ROC curves for the training and validation queues of the Random Forest, Support Vector Machine (SVM), and Logistic Regression methods. The Logistic Regression method is labeled as the original model in the chart. The horizontal axis represents 1-specificity, and the vertical axis represents sensitivity.

[0133] Depend on Figure 23 and Figure 24 The data shows that the AUC values ​​of the aforementioned alternative models in the validation queue are consistently above 0.90, highly consistent with the original model. This demonstrates that the core feature combination of the present invention has universality.

[0134] A predictive nomograph is constructed based on the results of a multivariate logistic regression model.

[0135] The nomograph in this application is essentially a visualization scale of the aforementioned multivariate logistic regression equation. Its specific logical relationship with the risk of bowel dysfunction and the steps for its use are as follows:

[0136] Individual scoring: For each patient, first find the actual value on the corresponding variable scale (scales 2-5) in the graph (e.g., IISi=2.5, surgical method=open abdomen), and project it vertically upwards to the top score scale (Points) to obtain the individual risk score corresponding to that indicator.

[0137] Total score calculation: The individual scores of all four variables (IISi, surgical method, T stage, N stage) of the patient are directly added together to obtain the total risk score.

[0138] Risk Conversion: Locate the total score on the 6th scale and project it vertically downwards to the Diagnostic Probability scale at the bottom.

[0139] Result determination: The value read on the probability scale at this time (range 0-1) is the predicted probability (i.e., risk value) of the patient developing postoperative bowel dysfunction.

[0140] For example, if the reading is 0.85, it means that the patient has an 85% probability of postoperative bowel dysfunction. Clinicians should classify the patient as a high-risk group and initiate preventive intervention.

[0141] like Figure 14 The diagram illustrates a predictive nomograph, a nomogram containing seven scales: The first scale is a fractional scale, representing the risk score corresponding to each scale mark from the second to the fifth scale; its scale value ranges from 0 to 140, with 0 at the leftmost end and 140 at the rightmost end, and the scale is equally divided. The second scale represents the patient's IISi (a continuous variable); the third scale represents the patient's surgical procedure; the fourth scale represents the patient's N stage; the fifth scale represents the patient's T stage (T4 / non-T4); the sixth scale is the total risk score, which is calculated as: total score = risk score corresponding to IISi + risk score corresponding to surgical procedure + risk score corresponding to N stage + risk score corresponding to T stage, i.e., the sum of the scores corresponding to the scale marks from the second to the fifth scale; the seventh scale is the diagnostic probability scale, representing the predicted probability of postoperative bowel dysfunction, its value obtained by mapping the total score from the sixth scale to the seventh scale.

[0142] The feature represented on the second scale is IISi (continuous); the IISi scale range is approximately -0.5 to 4.0. The higher the IISi value, the higher the score on the first scale: when the IISi value is low (close to -0.5), the score on the first scale is close to 0, and when the IISi value is high (close to 4.0), the score on the first scale is close to full marks.

[0143] The features represented on the third scale include laparoscopic surgery and open surgery; laparoscopic surgery corresponds to a score of 0 on the first scale, while open surgery corresponds to a higher score on the first scale (which can be read from the score scale in the figure).

[0144] The features represented on the fourth scale include N0, N1, and N2; N0 has a score of 0 on the first scale, and the scores of N1 and N2 on the first scale increase progressively (with N2 being higher than N1).

[0145] The features represented on the 5th scale include non-T4 and T4; non-T4 corresponds to a score of 0 on the 1st scale, while T4 corresponds to a higher score on the 1st scale.

[0146] The sixth scale represents the total score, with a scale of 0 to 160, where 0 is at the far left and 160 is at the far right, and the scale is divided equally. When the score on the sixth scale is close to the upper limit (approximately ≥150 points), the corresponding value on the seventh scale is located at the right end of 0.95, and the predicted diagnostic probability is close to >0.95. When the score on the sixth scale is close to the lower limit (approximately ≤20 points), the corresponding value on the seventh scale is located at the left end of 0.05, and the predicted diagnostic probability is close to <0.05.

[0147] In the research related to this application, three different verification methods were used to verify the trained AI model or AI model feature dataset obtained in this application.

[0148] In the studies related to this application, the sensitivity and specificity of the model were assessed using ROC curves to distinguish between patients who experienced and those who did not. The ROC curves for the training cohort are shown below. Figure 15 As shown; the ROC curve of the verification queue is as follows. Figure 17 . Figure 16 and Figure 18 To compare the predictive ability of the postoperative bowel function prediction system calculated for the training and validation cohorts, the area under the receiver operating characteristic (AUC) was used for quantification, and the 95% confidence interval for each AUC was calculated. Generally, an AUC above 0.9 indicates high accuracy, an AUC between 0.7 and 0.9 indicates some accuracy, an AUC between 0.5 and 0.7 indicates low accuracy, and an AUC of 0.5 indicates that the diagnostic method is completely ineffective and has no diagnostic value.

[0149] In this application, the area under the ROC curve (AUC), the 95% confidence interval (95% CI), and the p-value are used together to evaluate the predictive performance of the AI ​​model.

[0150] The area under the ROC curve (AUC) represents accuracy. The closer the value is to 1, the stronger the model's ability to distinguish between "sick" and "not sick". This model has an AUC > 0.94, which is considered extremely high accuracy.

[0151] 95% confidence interval (95% CI): Represents stability. The narrower the interval, the more stable the model is across different samples. The compact CI range of this model (e.g., 0.929-0.957) demonstrates minimal performance fluctuation.

[0152] P-value: Represents significance. P < 0.001 proves that the model's predictive ability is significantly better than random guessing (AUC = 0.5), which is statistically significant. Core conclusion of the validation results: Comparing the training cohort (AUC 0.943) and the validation cohort (AUC 0.942), the values ​​of the two are almost identical, strongly demonstrating that the model in this application has excellent generalization predictive ability, no overfitting, and can be accurately applied to new patients who have not participated in the training.

[0153] In the related research of this application, the performance of the model was further evaluated using calibration curves. First, the constructed nomograph was used to predict the probability of each research subject, and the subjects were arranged in ascending order of predicted probability. Without grouping, the relationship between "predicted probability and actual probability of occurrence" was smoothed using Loess and offset correction was performed using a 1000-time bootstrap method to obtain the calibration results. The horizontal axis represents the predicted probability of occurrence, and the vertical axis represents the actual probability of occurrence.

[0154] Figure 19 and Figure 20 In the graph, the horizontal axis represents the predicted probability, and the vertical axis represents the actual probability. Figure 19 and Figure 20 In the graph, the Ideal curve is at the bottom, represented by a wide, light gray solid line; it represents a perfect calibration baseline where the predicted probability perfectly matches the actual observed probability (i.e., y=x). The Apparent curve is in the middle layer, represented by a dark gray dashed line. It represents the model's direct fit performance on the current dataset. The Bias-corrected curve is at the top layer, represented by a thin black solid line (with data points). It represents the model's actual predictive performance after correcting for optimism bias using 1000 bootstrap sampling iterations.

[0155] like Figure 19 , is the calibration curve of the training queue, i.e., the training set; such as Figure 20 , is the calibration curve of the verification queue, i.e., the verification set; Figure 19 and Figure 20The diagonal dashed line represents the calibration curve under ideal conditions, where the model's predicted probability equals the actual probability of occurrence; the dashed curve is the result obtained by Loess fitting on all data; the solid curve is the curve obtained by performing 1000 bootstrapping, estimation, and deduction of optimistic bias on all data. Figure 19 and Figure 20 These are key pieces of evidence used to validate model calibration. They demonstrate the consistency between the probability values ​​predicted by the model and the observed occurrences in the real world.

[0156] exist Figure 19 and Figure 20 In the model, the black calibration curve, representing actual performance, consistently and closely follows the gray ideal curve, representing the perfect baseline. Even at the extremes of the predicted probabilities with small sample sizes, the calibration curves show no significant deviation. This high degree of visual overlap indicates that the predicted probabilities output by this model have extremely high reliability. For example, when the model predicted a 40% risk of bowel dysfunction in a certain group of patients, actual observation confirmed that approximately 40% of patients in that group did indeed experience the outcome event, indicating no systematic overestimation or underestimation. This is especially true in validation cohorts not involved in the modeling, such as... Figure 20 In the calibration curve, the model maintained excellent fit. This strongly demonstrates that the model successfully overcame the risk of overfitting, exhibits good robustness, and is capable of accurately quantifying the risk of new samples.

[0157] Analysis of validation results for the calibration curve: Figure 19 and Figure 20 The results demonstrate a high degree of consistency in the calibration performance of the prediction model across both the training and independent validation queues. In both figures, the bias-corrected curve (solid line), representing actual prediction performance, closely matches the ideal curve (dial dashed line), representing perfect prediction. Across the entire X-axis range (prediction probability) from 0 to 1, the solid line consistently fluctuates slightly around the dashed line without significant deviation.

[0158] This indicates that the predicted probability output by this model accurately reflects the actual observed probability of patients experiencing bowel dysfunction. For example, when the model predicted a 60% risk for a group of patients, actual observation confirmed that approximately 60% of the patients in that group did indeed experience the outcome event.

[0159] The comparative advantages of training versus validation: Of particular note is... Figure 20In the validation cohort, the calibration curve still maintained excellent goodness of fit. This strongly demonstrates from the dimension of "prediction accuracy" that the model has good robustness and generalization ability.

[0160] Comparison and explanation: If the model has overfitting, there will usually be a significant deviation of the solid line from the diagonal in the validation cohort (such as an S-shaped deviation). The results shown in this application confirm that the model has successfully overcome this common technical defect and can accurately quantitatively evaluate the risks of new samples that have not participated in training.

[0161] In the research related to this application, the clinical decision-making ability of the model was also evaluated using the DCA curve. As Figure 21 , is the DCA curve of the training cohort; as Figure 22 , is the DCA curve of the validation cohort.

[0162] DCA curve (Decision Curve Analysis, clinical decision curve analysis): DCA is a method used to evaluate whether a prediction model can bring net benefit in actual clinical applications.

[0163] The abscissa of the DCA curve is the threshold probability, and the ordinate of the DCA curve is the net benefit rate. When various evaluation methods reach a certain value, the risk probability of postoperative inflammatory reaction in patients is recorded as Pi; when Pi reaches a certain threshold (recorded as Pt), it is defined as positive, and a certain intervention measure (such as enhancing anti-inflammatory treatment measures) is taken. Then, enhancing anti-inflammatory treatment measures naturally changes the balance between the anti-inflammatory effect and the side effects of anti-inflammatory drugs, and the ordinate is the net benefit rate (Net Benefit, NB) after subtracting the disadvantages from the advantages. NB = A×P - B×L. A is the proportion of true positives; P is the benefit value of applying the intervention to true positive patients. B is the proportion of false positives; L is the loss value of applying the intervention to false positive patients.

[0164] As Figure 21 and Figure 22 shown: The DCA curve values of the modeling group and the validation group are in the range of 0.1 - 1.0, above the NONE and ALL lines. The results of this model are good. The ALL value line, that is, the gray line, represents the full intervention strategy, that is, assuming that all samples are positive and all receive intervention. It presents as a curve with a negative slope, indicating the clinical net benefit of the sample population when adopting this strategy at different risk thresholds. The NONE value line, that is, the black straight line, represents the no-intervention strategy, that is, assuming that all samples are negative (PI < PT), and thus no intervention measures are taken, and its clinical net benefit is constantly zero.

[0165] consistency: Figure 22 The shape of the verification queue curve and Figure 21 The training cohorts showed high consistency, proving that the model was not overfitting. Superiority: Within a wide threshold range of 0.1 to 0.9, the net benefit of this model (Example 2) was consistently higher than that of the "comparative model (excluding IISi)". The comparative model refers to the Logistic regression prediction model constructed solely based on three conventional clinicopathological features: surgical method, T stage, and N stage. This model represents the predictive level achievable using traditional clinical methods without using the core indicator of this invention—the Intestinal Inflammatory Stress Index (IISi). It is significantly higher than both the ALL and NONE value lines. Conclusion: This demonstrates that the evaluation system of this application can effectively improve the accuracy and benefit of clinical decision-making when dealing with new patients.

[0166] This application uses R software (version 4.3.0), i.e., R language and its plugin packages are open source and free software under the GPL license, for statistical analysis. The following packages are used: data.table, BMA, stabs, glmnet, broom, rms, pROC, ggplot2, and rmda.

[0167] like Figure 25 The method for obtaining risk assessment parameters for bowel dysfunction involved in this application involves collecting patients' preoperative clinical information and surgical plan information; calculating the IISi index by combining preoperative information and intraoperative data; and compiling information on surgical method, N stage, T stage, and other information to calculate the total risk score for postoperative bowel dysfunction. Based on the calculated IISi index, at-risk populations are evaluated and corresponding interventions are implemented.

[0168] In this application, the preoperative lymphocyte count K1, intraoperative opioid dosage MME, and operation duration T1 are obtained; the intestinal inflammatory stress index IISi is calculated, IISi=ln[(MME×T1) / (LCR+0.1)]. In the method for training and evaluating an AI model to identify the risk of intestinal dysfunction, patient medical record data is collected and set as independent variables to obtain the trained AI model or AI model feature dataset; positive or negative data on perioperative intestinal dysfunction in colorectal patients are obtained and set as dependent variables; the AI ​​model is trained using the above independent and dependent variable data to obtain the AI ​​model feature dataset or the trained AI model, which is used to identify the risk of perioperative intestinal dysfunction in colorectal patients.

[0169] While this application has been described and illustrated with reference to preferred embodiments and several alternatives, it is not intended to be limited to the specific descriptions herein. Other alternatives or equivalent components may also be used to practice this application.

Claims

1. A method for training an AI model to identify the risk of bowel dysfunction, characterized in that: Obtain the trained AI model or the AI ​​model feature dataset; Collect patient medical record data; the medical record data includes surgical procedure, preoperative CRP, preoperative lymphocyte count K1, intraoperative opioid dosage MME, and operation duration T1. CRP is a numerical value for C-reactive protein content; Obtain positive or negative data on perioperative bowel dysfunction in colorectal patients; Calculate the intestinal inflammatory stress index IISi, IISi=ln[(MME×T1) / (LCR+0.1)]; where LCR=K1 / corrected preoperative CRP; Independent variables are set based on medical record data; independent variables include parameter 1 and parameter 2. Parameter 1 corresponds to the intestinal inflammatory stress index IISi; Parameter 2 corresponds to the surgical procedure. Positive or negative data on perioperative bowel dysfunction in colorectal patients were set as dependent variable data. The AI ​​model is trained using the independent and dependent variable data to obtain the AI ​​model feature dataset or the trained AI model. The definition of positive or negative data for intestinal dysfunction is as follows: During the surgical period of colorectal patients, within 72 hours after surgery, patients are assessed, and the presence of two or more of the following conditions is considered positive; less than two conditions are considered negative. Condition 1: Abdominal distension occurs, VAS-A score ≥ 6; Condition 2: Bowel sounds <2 times / minute; Condition 3 requires pharmacological prokinetic intervention or device-based decompression intervention; The dependent variable data for positive patients was assigned a value of 1; the dependent variable data for negative patients was assigned a value of 0.

2. The method according to claim 1, characterized in that, When the intestinal inflammatory stress index IISi is greater than 2.355, parameter 1 is assigned a value of 1; When the intestinal inflammatory stress index IISi is less than or equal to 2.355, parameter 1 is assigned a value of 0; Surgical procedures include laparoscopic surgery and open surgery; If it is a laparoscopic surgery, then parameter 2 is set to 0; If it is an open abdominal surgery, then parameter 2 is assigned a value of 1.

3. The method according to claim 1, characterized in that, The independent variable data also includes parameter 3, which corresponds to patient stage N. The N-stage includes N0, N1, and N2 stages; When N is N0, the value of parameter 3 is 0; When N is N1, parameter 3 is assigned the value 1; When N is N2, parameter 3 is assigned the value 2.

4. The method according to claim 1, characterized in that, The independent variable data also includes parameter group 3, which corresponds to the N stage of the patient. N-stages include N0, N1, and N2; parameter group 3 includes parameter x31 and parameter x32. When the patient is in stage N0: x31=0, x32=0; When the patient is in stage N1: x31=1, x32=0; When the patient is in stage N2: x31=0, x32=1.

5. The method according to claim 1, characterized in that, The independent variable data also includes parameter 4, which corresponds to the patient stage, T stage. The T-period includes non-T4 period and T4 period; when the T-period is non-T4 period, parameter 4 is assigned a value of 0; when the T-period is T4 period, parameter 4 is assigned a value of 1. The T-period can be divided into T1, T2, T3, and T4 periods. When the T-period is T1, parameter 4 is assigned a value of 0; when the T-period is T2, parameter 4 is assigned a value of 0; when the T-period is T3, parameter 4 is assigned a value of 1; and when the T-period is T4, parameter 4 is assigned a value of 2.

6. The method according to claim 1, characterized in that, The operation duration T1 is calculated in hours.

7. The method of claim 1, wherein, The AI ​​model is a machine learning classification model, which includes any one of the following: multifactor logistic regression model, random forest model, support vector machine model, or artificial neural network model.

8. An Al evaluation method for identifying the risk of intestinal dysfunction, characterized by: Collect patient medical record data, set independent variable data based on the medical record data; use the AI ​​model feature dataset obtained by the method of any one of claims 1 to 7 in conjunction with the AI ​​calculation model to output the predicted probability of bowel dysfunction, and identify the risk of colorectal patients with bowel dysfunction during surgery based on the predicted probability.

9. The method according to claim 8, characterized in that, Set a preset risk value for the intestinal inflammatory stress index IISi; when the calculated intestinal inflammatory stress index IISi is greater than the preset risk value, output a high-risk assessment result for intestinal dysfunction; When the calculated intestinal inflammatory stress index IISi is less than the preset risk value, a low-risk assessment result for intestinal dysfunction is output.

10. An electronic device, characterized in that, The electronic device includes: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the executable instructions in the memory to implement the method as described in any one of claims 1 to 9.

Images