Artificial intelligence-based post-endoscopy risk prediction management system and method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANXI MEDICAL UNIV
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-21
Smart Images
Figure CN122436216A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence postoperative management technology, and in particular to an artificial intelligence-based postoperative risk prediction and management system and method for digestive endoscopy. Background Technology
[0002] Digestive endoscopy has become a core method for the diagnosis and treatment of digestive tract diseases. The timing of postoperative eating needs to be confirmed by doctors or nurses. Eating too early may induce nausea, vomiting, wound bleeding or perforation, but fasting too late will increase patients' discomfort such as hypoglycemia and hunger anxiety, and reduce the turnover efficiency of day wards. Current clinical practice mainly relies on the surgeon's experience or fixed time to implement uniform medical orders, and lacks quantitative prediction tools for individual patient physiological status and the degree of intraoperative trauma.
[0003] Most current management systems utilize machine learning models to predict the risk of postoperative complications. They construct bleeding risk scoring models based on intraoperative variables and preoperative medical history, and process endoscopic images and vital sign data through deep learning to give an approximate time for eating after surgery. However, the prediction model relies on the average risk characteristics of the population and does not incorporate the patient's own physiological sensitivity to fasting stress baseline, resulting in significant differences in the predicted eating time for patients with different physical conditions. At the same time, the degree of intraoperative trauma varies, and the model cannot distinguish the differentiated recovery trajectories of different patients under the same surgical procedure. Summary of the Invention
[0004] The purpose of this invention is to provide an artificial intelligence-based postoperative risk prediction management system and method for digestive endoscopy, which solves the problems of the lack of individualized fasting programs for patients and insufficient quantification of the coupling effect between intraoperative trauma and individual physiological response in existing technologies.
[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention is an artificial intelligence-based postoperative risk prediction and management system for digestive endoscopy, comprising: The preoperative baseline construction module is used to acquire physiological indicator sets collected from patients under normal eating and fasting conditions before surgery. Based on the changing trend of the physiological indicator set over time, the dynamic change slope of each dimension of physiological indicator is calculated, and the fasting tolerance inflection point is determined. Based on the dynamic change slope and the tolerance inflection point time, a multidimensional individualized fasting stress feature vector is constructed. The intraoperative trauma coupling calibration module is used to acquire the patient's intraoperative operation data, determine the trauma level coefficient based on the intraoperative operation data, construct a nonlinear mapping function from the trauma level coefficient to the feature weights, and redistribute the weights of each dimension in the multidimensional individualized fasting stress feature vector based on the nonlinear mapping function to obtain the calibrated stress features. The redistribution of the weights of each dimension is used to characterize the degree of differential dependence of patients on preoperative metabolic reserves under different trauma loads. The postoperative risk prediction module is used to establish an eating risk prediction model and input calibrated stress characteristics to obtain the risk level of the patient's first postoperative food intolerance. The eating risk prediction model includes a general backbone network trained based on multi-center data and a lightweight adaptation layer deployed in the target hospital. The lightweight adaptation layer is used to adjust the bias parameters of the output layer of the general backbone network according to the local case data of the target hospital. The plan generation module is used to match the corresponding individualized eating guidance plan from a preset eating plan library based on the risk level.
[0006] Furthermore, during the patient's preoperative fasting process, a set of physiological indicators was collected sequentially at preset time intervals. The set of physiological indicators included blood glucose level, visual analog scale (VAS) score for hunger, and heart rate. The slope of the dynamic change was obtained by linearly fitting the sequence of changes in the physiological indicators over time. The fasting tolerance inflection point was the time point at which the VAS score for hunger first exceeded the preset threshold and the blood glucose level first fell below the preset threshold.
[0007] Furthermore, a trauma level coefficient is determined based on the surgical procedure type, resection depth level, and sensitivity level of the resection site; the nonlinear mapping function is: in, The original weights of the i-th dimension in the multidimensional individualized fasting stress feature vector are: The weights are calibrated, and TLC represents the trauma grade coefficient. For the preset scaling factor, is the surgical procedure sensitivity coefficient corresponding to the i-th dimension.
[0008] Furthermore, the lightweight adaptation layer consists of an output bias vector and a scaling factor. The output of the general backbone network is scaled by the scaling factor and then superimposed with the output bias vector to obtain the final risk prediction value. The parameters of the lightweight adaptation layer are updated using an exponential moving average method based on local case data from the target hospital, while the parameters of the general backbone network remain frozen during the update process. The training process of the general backbone network adopts a multi-task learning framework, simultaneously predicting the postoperative food intolerance risk level, postoperative infection risk probability, and postoperative hospital stay regression value. Among these, the tasks are jointly optimized by sharing underlying feature representations. The parameter initialization of the lightweight adaptation layer adopts a meta-learning-based strategy, specifically including the following: a set of basic parameters is pre-trained based on multi-center historical data, and then the basic parameters are rapidly gradient-adapted based on the initial local case data from the target hospital to obtain initial parameters adapted to the target hospital.
[0009] Furthermore, in the protocol generation module, when the risk level is low, the individualized eating guidance plan includes the first postoperative fasting period and the first liquid diet; when the risk level is medium, the individualized eating guidance plan includes the second fasting period of the same length as the first, and a sugary electrolyte drink as the first food intake; when the risk level is high, the individualized eating guidance plan includes the third fasting period of the same length as the first, and recommendations for sugary drinks and preventive antiemetics; the individualized eating guidance plan output by the protocol generation module also includes a visual timeline diagram, which marks the recommended eating content, expected eating time, and risk warning thresholds for each stage.
[0010] Furthermore, it also includes a closed-loop update module, used to collect the patient's actual eating response data after implementing the individualized eating guidance plan, and calculate the gradient loss based on the actual eating response data; wherein, during backpropagation, the parameter gradient of the general backbone network is locked, and only the parameters of the lightweight adaptation layer are updated online incrementally; during the update process, the closed-loop update module uses a differential privacy mechanism to process the noise perturbation of the actual eating response data, and when the prediction accuracy of the updated lightweight adaptation layer on the validation set improves by more than a preset threshold, the original parameters are replaced.
[0011] Furthermore, the preoperative baseline construction module also includes a process deviation correction unit; the process deviation correction unit is used to obtain the steady-state drift value of the scaling factor in the lightweight adaptation layer within a preset statistical period, and determine the process deviation factor vector of the target hospital based on the steady-state drift value; the process deviation factor vector is used to perform affine transformation correction on the dynamic change slope and tolerance inflection point time to obtain the corrected dynamic change slope and the corrected tolerance inflection point time, and a multidimensional individualized fasting stress feature vector is constructed using the corrected dynamic change slope and the corrected tolerance inflection point time.
[0012] An AI-based approach to postoperative risk prediction and management in gastrointestinal endoscopy includes the following steps: S1. Before surgery, blood glucose, hunger score and heart rate were collected from patients under normal eating and fasting conditions at time intervals. The slope of dynamic change of each indicator and tolerance inflection point were calculated to construct a multidimensional individualized fasting stress feature vector. S2. Determine the process deviation factor based on the steady-state drift value of the scaling factor of the target hospital adaptation layer, and reconstruct the feature vector after correcting the slope and inflection point; S3. Based on the surgical procedure type, resection depth and location during the operation, the trauma level coefficient is determined, and the weights of each dimension of the feature vector are redistributed through a nonlinear mapping function to obtain the calibrated stress characteristics. S4. Input the calibrated stress features into the prediction model, extract features through the frozen general backbone network, adjust the bias through the local lightweight adaptation layer, and output the postoperative food intolerance risk level. S5. Match and output individualized eating plans based on risk levels, with different levels corresponding to different fasting durations and initial food intake; S6. Collect actual patient eating response data to calculate gradient loss. During backpropagation, lock the general backbone network and perform incremental updates on the adaptation layer parameters under differential privacy noise perturbation. Replace parameters when the validation set accuracy improvement exceeds the threshold.
[0013] The present invention has the following beneficial effects: This invention establishes a multidimensional individualized fasting stress feature vector that includes dynamic change slope and tolerance inflection point by preoperative self-comparison. It shifts the prediction benchmark for postoperative eating risk from the statistical mean of the population to the actual measured value of the patient's own physiological stress, which significantly improves the accuracy of individualized prediction and avoids the problems of excessive fasting for some patients or insufficient warning for others.
[0014] This invention redistributes the weights of each dimension of the feature vector by using a nonlinear mapping function to simulate the differentiated dependence of patients on metabolic reserves under different trauma loads. This enables the prediction model to dynamically adapt to the coupling effect of surgical trauma and individual stress. It adopts a two-layer model architecture of a general skeleton network and a lightweight adaptation layer. The general skeleton is frozen and only the adaptation layer is updated locally. It takes into account both multi-center knowledge sharing and hospital local privacy protection, and has low deployment cost and strong generalization ability.
[0015] In this invention, the process deviation correction unit automatically corrects cross-center process deviations by utilizing the steady-state drift value of the adaptation layer parameters, eliminating preoperative measurement noise and further enhancing the consistency and reliability of multi-center deployment. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a system block diagram of the present invention.
[0018] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation
[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0020] See Figure 1This invention is an artificial intelligence-based postoperative risk prediction and management system for digestive endoscopy, comprising: The preoperative baseline construction module is used to acquire physiological indicator sets collected from patients under normal eating and fasting conditions before surgery. Based on the changing trend of the physiological indicator set over time, the dynamic change slope of each dimension of physiological indicator is calculated, and the fasting tolerance inflection point is determined. Based on the dynamic change slope and the tolerance inflection point time, a multidimensional individualized fasting stress feature vector is constructed. During the patient's preoperative fasting process, physiological indicators were collected sequentially at preset time intervals. The physiological indicators included blood glucose level, visual analog scale (VAS) score for hunger, and heart rate. The slope of the dynamic change was obtained by linearly fitting the change sequence of the physiological indicators over time. The fasting tolerance inflection point was the time point at which the VAS score for hunger first exceeded the preset threshold and the blood glucose level first fell below the preset threshold.
[0021] The intraoperative trauma coupling calibration module is used to acquire the patient's intraoperative operation data, determine the trauma level coefficient based on the intraoperative operation data, construct a nonlinear mapping function from the trauma level coefficient to the feature weights, and redistribute the weights of each dimension in the multidimensional individualized fasting stress feature vector based on the nonlinear mapping function to obtain the calibrated stress features. The redistribution of the weights of each dimension is used to characterize the degree of differential dependence of patients on preoperative metabolic reserves under different trauma loads. The trauma grade coefficient is determined based on the surgical procedure type, resection depth grade, and sensitivity grade of the resection site; the nonlinear mapping function is: in, The original weights of the i-th dimension in the multidimensional individualized fasting stress feature vector are: The weights are calibrated, and TLC represents the trauma grade coefficient. For the preset scaling factor, is the surgical procedure sensitivity coefficient corresponding to the i-th dimension.
[0022] The postoperative risk prediction module is used to establish an eating risk prediction model and input calibrated stress characteristics to obtain the risk level of the patient's first postoperative food intolerance. The eating risk prediction model includes a general backbone network trained based on multi-center data and a lightweight adaptation layer deployed in the target hospital. The lightweight adaptation layer is used to adjust the bias parameters of the output layer of the general backbone network according to the local case data of the target hospital. The lightweight adaptation layer consists of an output bias vector and a scaling factor. The output of the general backbone network is scaled by the scaling factor and then superimposed with the output bias vector to obtain the final risk prediction value. The parameters of the lightweight adaptation layer are updated using an exponential moving average method based on local case data from the target hospital, while the parameters of the general backbone network remain frozen during the update process. The training process of the general backbone network adopts a multi-task learning framework, simultaneously predicting the postoperative food intolerance risk level, postoperative infection risk probability, and postoperative hospital stay regression value. Among these, the tasks are jointly optimized by sharing the underlying feature representation. The parameter initialization of the lightweight adaptation layer adopts a meta-learning-based strategy, which includes the following: a set of basic parameters is pre-trained based on multi-center historical data, and then the basic parameters are rapidly gradient-adapted based on the initial local case data from the target hospital to obtain initial parameters adapted to the target hospital.
[0023] The plan generation module is used to match the corresponding individualized eating guidance plan from a preset eating plan library based on the risk level.
[0024] In the protocol generation module, when the risk level is low, the individualized feeding guidance plan includes the first postoperative fasting period and the first liquid diet; when the risk level is medium, the individualized feeding guidance plan includes the second fasting period of the same duration as the first, and a sugary electrolyte drink as the first food intake; when the risk level is high, the individualized feeding guidance plan includes the third fasting period of the same duration as the first, and recommendations for sugary drinks and preventive antiemetics; the individualized feeding guidance plan output by the protocol generation module also includes a visual timeline chart, which marks the recommended food intake content, expected feeding time, and risk warning thresholds for each stage.
[0025] The closed-loop update module is used to collect the patient's actual eating response data after implementing the individualized eating guidance plan, and calculate the gradient loss based on the actual eating response data. During backpropagation, the parameter gradient of the general backbone network is locked, and only the parameters of the lightweight adaptation layer are updated online incrementally. During the update process, the closed-loop update module uses a differential privacy mechanism to process the noise perturbation of the actual eating response data, and when the prediction accuracy of the updated lightweight adaptation layer on the validation set improves by more than a preset threshold, the original parameters are replaced.
[0026] The working principle of this invention is as follows: After the system is started, the preoperative baseline construction module first performs an individualized fasting stress test on patients who are about to undergo digestive endoscopy. The specific calculation steps are as follows: During the preoperative preparation phase, baseline values of physiological indicators were collected from the patient under normal eating conditions, as well as stress values of physiological indicators collected at multiple time points during fasting at preset time intervals. These physiological indicators included: fingertip blood glucose level (G), visual analog scale (H) for hunger, and resting heart rate (R). The collected data constituted a time-varying sequence of physiological indicators for subsequent dynamic feature calculations. To quantify the rate of physiological change in patients under hunger stress, this module used a weighted linear least squares method to calculate the dynamic slope of each indicator, with the blood glucose slope as an example. For example, the calculation formula is: in, This represents the time value corresponding to the j-th physiological indicator collection point after the start of fasting. This represents the baseline blood glucose level under normal eating conditions. This represents the average of the measurements taken at the time points following fasting. The j-th measurement point after the start of fasting The corresponding fingertip blood glucose measurement value was used to calculate the slope of blood glucose change using the method described above. Slope of change in hunger and the slope of heart rate change The dimensions are respectively Rating / h, bpm / h; The fasting tolerance inflection point is the critical moment when the patient's physiological reserves are significantly depleted. This embodiment uses a dual-condition trigger mechanism to determine the inflection point time. : or That is, the earliest time when a patient's self-rated hunger reaches 5 points or blood glucose level drops to the warning line of 4.0 mmol / L is the inflection point of the patient's fasting tolerance. The aforementioned dynamic slope and tolerance inflection point time together constitute the patient's multidimensional individualized fasting stress feature vector. This feature vector differs from the static single-point values in traditional methods and can characterize the entire process of the patient's physiological response under fasting stress.
[0027] Subsequently, the intraoperative trauma coupling calibration module acquires intraoperative operation data in real time from the endoscope host and electronic medical record, including surgical procedure type, resection depth grade, and resection site sensitivity grade. Based on the preset lookup table rules, it determines the trauma grade coefficient (TLC) for this operation, as shown in Table 1: Table 1 The three-dimensional feature vector constructed before surgery The initial weights are set to The weights are redistributed using a nonlinear mapping function. Taking blood glucose as an example, the weights are calibrated. The calculation example is as follows: Set global scaling factor For endoscopic submucosal dissection, the blood glucose sensitivity coefficient Take 0.9, heart rate sensitivity coefficient Take 0.3; The calculation yields: This shows that the greater the trauma, the greater the weight of the preoperative glucose metabolic reserve dimension.
[0028] The postoperative risk prediction module inputs post-calibrated stress features into the feeding risk prediction model. The general backbone network is pre-trained based on desensitized historical data from multiple medical centers. It adopts a multi-task learning framework to simultaneously predict three related targets on the basis of shared underlying feature representations: postoperative feeding intolerance risk level, postoperative infection risk probability, and postoperative hospitalization duration regression value. Multi-task joint optimization enables the underlying feature representation to capture general physiological patterns related to postoperative recovery and improves generalization ability. The lightweight adaptation layer is deployed locally in the target hospital and contains only two adjustable components: output bias vector and scaling factor. After receiving the deep features output by the general backbone network, it obtains the final risk prediction value by scaling and bias superposition. The initial parameters of the adaptation layer are obtained using a meta-learning strategy: first, a set of basic parameters are trained from multi-center data, and then a small number of gradient steps are used for rapid adaptation on the first batch of local cases in the target hospital, so that the model has basic adaptability to the patient population of the hospital in the early stage of deployment.
[0029] After receiving the risk level, the protocol generation module retrieves the corresponding individualized eating guidance protocol from the preset eating protocol library. The basic fasting time is consistent for the three risk levels. Low-risk patients are given ordinary liquids; medium-risk patients are given sugary electrolyte drinks in addition to liquids to prevent hypoglycemia; high-risk patients are given additional recommendations for preventive antiemetic medication in addition to sugary drinks. At the same time, the module generates a visual timeline diagram, marking the recommended eating content, expected eating time and risk warning threshold for each stage in the form of a horizontal timeline, so that nurses and patients can intuitively understand the execution process.
[0030] Finally, the closed-loop update module collects actual eating response data after the patient executes the feeding plan, including whether vomiting occurred, abdominal distension and pain scores, hypoglycemia events, and the patient's subjective comfort score. Based on the difference between the predicted risk level and the actual outcome, the gradient loss is calculated and the data enters the backpropagation process. During backpropagation, the parameter gradients of the general backbone network are forcibly locked, and the gradient signal is only transmitted to the bias vector and scaling factor of the lightweight adaptation layer, realizing online incremental updates only to the adaptation layer parameters. A differential privacy mechanism is used during the update process, adding Gaussian noise perturbation to the gradient to ensure that the eating response information of a single patient cannot be inferred in reverse, thus protecting patient privacy. After each update, the system evaluates the prediction accuracy of the new parameters on the local validation set. The original parameters are only replaced when the accuracy improvement exceeds a preset threshold, avoiding model performance fluctuations caused by individual abnormal samples. As the target hospital uses the system for an extended period, the adaptation layer parameters gradually converge to steady-state values that reflect the characteristics of the patient population in that hospital, and the system's prediction accuracy is continuously optimized.
[0031] In the above embodiments, the preoperative preparation procedures of different medical institutions vary systematically. For example, the control of water intake during fasting, the precision of blood collection timing, and the calibration standards of blood glucose meters may not be entirely consistent. This difference in procedures introduces measurement bias, resulting in a systematic shift in the values measured for the same patient in different hospitals. This shift is not due to the patient's actual physiological differences, but rather the superposition of noise in the hospital procedures. If this noisy feature vector is directly input into a general backbone network, even if the lightweight adaptation layer has a certain output correction capability, it will still cause cross-center attenuation of prediction accuracy, and the adaptation layer requires additional... The convergence time consumed by external factors can be used to digest this deviation. Therefore, to address this issue, in this embodiment, the preoperative baseline construction module further includes a process deviation correction unit. The process deviation correction unit is used to obtain the steady-state drift value of the scaling factor in the lightweight adaptation layer within a preset statistical period, and to determine the process deviation factor vector of the target hospital based on the steady-state drift value. The process deviation factor vector is used to perform affine transformation correction on the dynamic change slope and tolerance inflection point time to obtain the corrected dynamic change slope and the corrected tolerance inflection point time, and to construct a multidimensional individualized fasting stress feature vector based on the corrected dynamic change slope and the corrected tolerance inflection point time.
[0032] See Figure 2 An AI-based method for predicting and managing postoperative risks after gastrointestinal endoscopy includes the following steps: S1. Before surgery, blood glucose, hunger score and heart rate were collected from patients under normal eating and fasting conditions at time intervals. The slope of dynamic change of each indicator and tolerance inflection point were calculated to construct a multidimensional individualized fasting stress feature vector. S2. Determine the process deviation factor based on the steady-state drift value of the scaling factor of the target hospital adaptation layer, and reconstruct the feature vector after correcting the slope and inflection point; S3. Based on the surgical procedure type, resection depth and location during the operation, the trauma level coefficient is determined, and the weights of each dimension of the feature vector are redistributed through a nonlinear mapping function to obtain the calibrated stress characteristics. S4. Input the calibrated stress features into the prediction model, extract features through the frozen general backbone network, adjust the bias through the local lightweight adaptation layer, and output the postoperative food intolerance risk level. S5. Match and output individualized eating plans based on risk levels, with different levels corresponding to different fasting durations and initial food intake; S6. Collect actual patient eating response data to calculate gradient loss. During backpropagation, lock the general backbone network and perform incremental updates on the adaptation layer parameters under differential privacy noise perturbation. Replace parameters when the validation set accuracy improvement exceeds the threshold.
[0033] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. An artificial intelligence-based postoperative risk prediction and management system for digestive endoscopy, characterized in that, include: The preoperative baseline construction module is used to acquire physiological indicator sets collected by patients under normal eating and fasting conditions before surgery. Based on the changing trend of the physiological indicator set over time, the dynamic change slope of each dimension of physiological indicator is calculated, and the fasting tolerance inflection point is determined. Based on the dynamic change slope and the tolerance inflection point time, a multidimensional individualized fasting stress feature vector is constructed. The intraoperative trauma coupling calibration module is used to acquire the patient's intraoperative operation data, determine the trauma level coefficient based on the intraoperative operation data, construct a nonlinear mapping function from the trauma level coefficient to the feature weights, and redistribute the weights of each dimension in the multidimensional individualized fasting stress feature vector based on the nonlinear mapping function to obtain the calibrated stress features. The redistribution of the weights of each dimension is used to characterize the degree of differential dependence of patients on preoperative metabolic reserves under different trauma loads. The postoperative risk prediction module is used to establish an eating risk prediction model and input calibrated stress characteristics to obtain the risk level of the patient's first postoperative food intolerance. The eating risk prediction model includes a general backbone network trained based on multi-center data and a lightweight adaptation layer deployed in the target hospital. The lightweight adaptation layer is used to adjust the bias parameters of the output layer of the general backbone network according to the local case data of the target hospital. The plan generation module is used to match the corresponding individualized eating guidance plan from a preset eating plan library based on the risk level.
2. The postoperative risk prediction and management system for digestive endoscopy based on artificial intelligence according to claim 1, characterized in that, During the patient's preoperative fasting process, a set of physiological indicators was collected sequentially at preset time intervals. The set of physiological indicators included blood glucose level, visual analog scale (VAS) score for hunger, and heart rate. The slope of the dynamic change was obtained by linearly fitting the change sequence of the physiological indicators over time. The fasting tolerance inflection point was the time point corresponding to the first time when the VAS score for hunger exceeded a preset threshold and the first time when the blood glucose level fell below a preset threshold.
3. The postoperative risk prediction and management system for digestive endoscopy based on artificial intelligence according to claim 2, characterized in that, The trauma level coefficient is determined based on the surgical procedure type, resection depth level, and resection site sensitivity level; the nonlinear mapping function is: in, The original weights of the i-th dimension in the multidimensional individualized fasting stress feature vector are: The weights are calibrated, and TLC represents the trauma grade coefficient. For the preset scaling factor, is the surgical procedure sensitivity coefficient corresponding to the i-th dimension.
4. The postoperative risk prediction and management system for digestive endoscopy based on artificial intelligence according to claim 3, characterized in that, The lightweight adaptation layer consists of an output bias vector and a scaling factor. The output of the general skeleton network is scaled by the scaling factor and then superimposed with the output bias vector to obtain the final risk prediction value. The parameters of the lightweight adaptation layer are updated using an exponential moving average based on local case data from the target hospital, while the parameters of the general backbone network remain frozen during the update process. The training process of the general backbone network adopts a multi-task learning framework, simultaneously predicting the postoperative food intolerance risk level, postoperative infection risk probability, and postoperative hospital stay regression value. Among these, the tasks are jointly optimized by sharing underlying feature representations. The parameter initialization of the lightweight adaptation layer adopts a meta-learning-based strategy, specifically including the following: a set of basic parameters is pre-trained based on multi-center historical data, and then the basic parameters are rapidly gradient-adapted based on the initial local case data from the target hospital to obtain initial parameters adapted to the target hospital.
5. The postoperative risk prediction and management system for digestive endoscopy based on artificial intelligence according to claim 4, characterized in that, In the protocol generation module, when the risk level is low, the individualized dietary guidance protocol includes a first postoperative fasting period and a first liquid diet; when the risk level is medium, the individualized dietary guidance protocol includes a second fasting period of the same length as the first, and a sugary electrolyte drink as the first food intake; when the risk level is high, the individualized dietary guidance protocol includes a third fasting period of the same length as the first, and recommendations for sugary drinks and preventative antiemetics; the individualized dietary guidance protocol output by the protocol generation module also includes a visual timeline diagram, which marks the recommended dietary content, expected eating time, and risk warning threshold for each stage.
6. The postoperative risk prediction and management system for digestive endoscopy based on artificial intelligence according to claim 5, characterized in that, It also includes a closed-loop update module, used to collect the patient's actual eating response data after implementing the individualized eating guidance plan, and calculate the gradient loss based on the actual eating response data; wherein, during backpropagation, the parameter gradient of the general backbone network is locked, and only the parameters of the lightweight adaptation layer are updated online incrementally; during the update process, the closed-loop update module uses a differential privacy mechanism to process the noise perturbation of the actual eating response data, and when the prediction accuracy of the updated lightweight adaptation layer on the validation set improves by more than a preset threshold, the original parameters are replaced.
7. The postoperative risk prediction and management system for digestive endoscopy based on artificial intelligence according to claim 6, characterized in that, The preoperative baseline construction module also includes a process deviation correction unit; the process deviation correction unit is used to obtain the steady-state drift value of the scaling factor in the lightweight adaptation layer within a preset statistical period, and determine the process deviation factor vector of the target hospital based on the steady-state drift value; the process deviation factor vector is used to perform affine transformation correction on the dynamic change slope and tolerance inflection point time to obtain the corrected dynamic change slope and the corrected tolerance inflection point time, and a multidimensional individualized fasting stress feature vector is constructed using the corrected dynamic change slope and the corrected tolerance inflection point time.
8. A method for predicting and managing postoperative risks after digestive endoscopy based on artificial intelligence, applied to the artificial intelligence-based postoperative risk prediction and management system for digestive endoscopy as described in claim 7, characterized in that, Includes the following steps: S1. Before surgery, blood glucose, hunger score and heart rate were collected from patients under normal eating and fasting conditions at time intervals. The slope of dynamic changes of each indicator and tolerance inflection point were calculated to construct a multidimensional individualized fasting stress feature vector. S2. Determine the process deviation factor based on the steady-state drift value of the scaling factor of the target hospital adaptation layer, and reconstruct the feature vector after correcting the slope and inflection point; S3. Based on the intraoperative surgical procedure type, resection depth and location, determine the trauma level coefficient, and redistribute the weights of each dimension of the feature vector through a nonlinear mapping function to obtain the calibrated stress characteristics. S4. Input the calibrated stress features into the prediction model, extract features through the frozen general backbone network, adjust the bias through the local lightweight adaptation layer, and output the postoperative food intolerance risk level. S5. Match and output individualized eating plans based on risk levels, with different levels corresponding to different fasting durations and initial food intake; S6. Collect actual patient eating response data to calculate gradient loss. During backpropagation, lock the general backbone network, perform incremental updates on the adaptation layer parameters under differential privacy noise perturbation, and replace parameters when the validation set accuracy exceeds the threshold.