Diabetes patient colonoscopy hypoglycemia risk prediction model and method

Abstract

This invention relates to the field of risk prediction technology, specifically to a model and method for predicting the risk of hypoglycemia during colonoscopy in diabetic patients. The model includes a feature extraction module, a weight analysis module, a model construction module, a model calibration module, and a risk prediction module. In this invention, core clinical features such as patient age, bowel preparation score, fasting time, and specific medication usage are extracted. A hypoglycemia risk nomogram prediction model is constructed using weight analysis. The model is then dynamically calibrated by calculating the prediction fitting deviation. These techniques enable precise quantification of the probability of hypoglycemia during colonoscopy in diabetic patients and provide early warning of high-risk conditions. This helps medical personnel identify potential risks in advance, optimize personalized fasting and medication care plans, effectively avoid serious adverse cardiovascular and neurological events, and ultimately ensure that diabetic patients can undergo colonoscopy normally.

Description

Technical Field

[0001] This invention relates to the field of risk prediction technology, and in particular to a model and method for predicting the risk of hypoglycemia during colonoscopy in diabetic patients. Background Technology

[0002] The field of risk prediction technology primarily involves collecting and analyzing large amounts of historical data and real-time information to estimate the probability of specific future events. Its main purpose is to help decision-makers identify potential risks in advance and formulate corresponding intervention strategies. It is widely used in numerous industries, including healthcare, finance, risk control, industrial safety, and transportation. Among these, the hypoglycemia risk prediction model for colonoscopy in diabetic patients is a specialized assessment tool developed for clinical settings. This tool addresses the risk of hypoglycemia in diabetic patients during preparation for or undergoing colonoscopy due to factors such as fasting and bowel preparation. It quantifies and warns of the probability of hypoglycemia in an individual patient before the procedure.

[0003] Colonoscopy is widely considered the reference standard for colorectal cancer screening, but the procedures requiring fasting and the use of laxatives can lead to complications such as hypoglycemia. Diabetic patients, due to their complex metabolic disorders, are at even higher risk of hypoglycemia during colonoscopy, posing not only a challenge to short-term glycemic management but also potentially leading to serious adverse events such as cardiovascular and neurological damage. Currently, existing guidelines lack specific recommendations on how to effectively prevent and manage hypoglycemia, particularly regarding personalized care strategies for diabetic patients (such as medication adjustments and optimization of fasting times), which still require further refinement. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a model and method for predicting the risk of hypoglycemia during colonoscopy in diabetic patients.

[0005] To achieve the above objectives, the present invention employs the following technical solution: a hypoglycemia risk prediction model for colonoscopy in diabetic patients includes: The feature extraction module acquires clinical medical records and blood glucose monitoring records of hospitalized diabetic patients, and extracts age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema status before colonoscopy, meglitinide use status, and sedative use status before colonoscopy as feature parameters, and combines them to obtain a set of independent clinical risk features. The weight analysis module analyzes the clinical independent risk feature set, obtains the risk association weights corresponding to each feature parameter in the clinical independent risk feature set, filters significant independent risk features and their corresponding risk association weights, and constructs a hypoglycemia risk feature mapping weight set. The model building module processes the hypoglycemia risk feature mapping weight set, determines the graphical scale projection of the hypoglycemia risk feature mapping weight set under the preset hypoglycemia risk score benchmark scale, and constructs an initial hypoglycemia risk nomogram prediction model. The model calibration module analyzes the initial hypoglycemia risk nomogram prediction model, obtains the area under the corresponding model prediction curve and the prediction fitting deviation. If the prediction fitting deviation is greater than the preset prediction tolerance threshold, the risk scale distribution state of the initial hypoglycemia risk nomogram prediction model is adjusted according to the area under the model prediction curve, and the calibrated hypoglycemia risk nomogram prediction model is output. The risk prediction module acquires the current clinical data of the diabetic patient to be tested, inputs it into the calibrated hypoglycemia risk nomogram prediction model for processing, obtains the corresponding hypoglycemia occurrence probability and compares it with the preset risk warning threshold, determines the risk status of the diabetic patient to be tested, and generates the hypoglycemia risk prediction result for the colonoscopy examination of the diabetic patient.

[0006] As a further aspect of the present invention, in the process of screening significant independent hazard features, if the risk association weight is greater than the preset significant association threshold, the corresponding feature parameter is extracted as a significant independent hazard feature; if the risk association weight is less than or equal to the preset significant association threshold, the corresponding feature parameter is removed. In the process of determining the risk status of a diabetic patient to be tested, if the probability of hypoglycemia is greater than the preset risk warning threshold, the diabetic patient to be tested is determined to have a high risk of developing hypoglycemia; if the probability of hypoglycemia is less than or equal to the preset risk warning threshold, the diabetic patient to be tested is determined to have a normal risk of developing hypoglycemia.

[0007] As a further aspect of the present invention, the clinical independent risk feature set includes patient vital signs indicators, preoperative preparation assessment items, and clinical medication tracking items; the hypoglycemia risk feature mapping weight set includes feature importance values, risk impact multipliers, and independent variable coefficients; the initial hypoglycemia risk nomogram prediction model includes variable scoring scales, comprehensive cumulative scoring items, and risk mapping transformation axes; the calibrated hypoglycemia risk nomogram prediction model includes assessment calibration charts, scale correction amounts, and optimized probability reference lines; and the hypoglycemia risk prediction results of colonoscopy in diabetic patients include patient risk level and abnormal blood glucose records.

[0008] As a further aspect of the present invention, the feature extraction module includes: The clinical medical record acquisition submodule acquires clinical medical record data of diabetic inpatients before colonoscopy and during colonoscopy bowel preparation. It extracts the values ​​of various basic physiological indicators from the clinical medical record data, compares the values ​​of various basic physiological indicators with the preset normal indicator baseline, calculates the absolute difference between the values ​​of various basic physiological indicators and the preset normal indicator baseline, and summarizes all absolute differences to obtain the basic physiological difference set. The monitoring record extraction submodule collects blood glucose monitoring records of diabetic inpatients within a specified time period from the start of bowel preparation to the end of colonoscopy, based on the basic physiological difference set. It extracts the blood glucose change values ​​of each time series in the blood glucose monitoring records, integrates the basic physiological difference set with the blood glucose change values ​​of each time series, calculates the comprehensive physiological change amplitude at each time point, and generates the time series physiological fluctuation amplitude. The risk feature combination submodule integrates the values ​​of various states with the temporal physiological fluctuation amplitude to form a feature mapping matrix, and establishes a clinically independent risk feature set based on the extraction age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema operation status before colonoscopy, meglitinide use status, and sedative use status before colonoscopy.

[0009] As a further aspect of the present invention, the weight analysis module includes: The risk weight analysis submodule, based on the clinical independent risk feature set, decomposes the independent feature parameter items of each dimension, calculates the corresponding partial regression coefficient value and standard error value of each independent feature parameter item under the multi-factor logistic regression rule, and merges the corresponding partial regression coefficient value and standard error value of each independent feature parameter item to obtain the risk association weight. The significant feature screening submodule compares the numerical difference between the risk association weight and the preset significant association threshold. When the risk association weight is greater than the preset significant association threshold, the corresponding feature parameter is extracted into the candidate region. When the risk association weight is not greater than the preset significant association threshold, the corresponding feature parameter is removed to obtain significant independent risk features. The mapping weight construction submodule extracts the feature weight values ​​corresponding to the significant independent risk features, processes the feature weight values ​​according to the preset weight allocation constant, calculates the proportion coefficient of each feature weight value in the overall feature space, combines the significant independent risk features and their corresponding proportion coefficients, and establishes a hypoglycemia risk feature mapping weight set.

[0010] As a further aspect of the present invention, the model building module includes: The scale weight normalization submodule obtains the preset hypoglycemia risk score benchmark scale, processes the weight values ​​of each item in the hypoglycemia risk feature mapping weight set according to the standard normalization function, calculates the relative mapping coordinate values ​​of each weight value within the preset interval, integrates the relative mapping coordinate values ​​of each item to construct a standardized one-dimensional numerical vector sequence, and obtains a unified scale weight set. The scale projection comparison submodule compares the values ​​of each item in the unified scale weight set with the coordinate coefficient values ​​of each node under the preset hypoglycemia risk scoring benchmark scale, calculates the offset distance vector corresponding to the values ​​of the unified scale weight set in the preset hypoglycemia risk scoring benchmark scale, and converts the offset distance vector to generate a graphical scale projection. The initial model construction submodule extracts the coordinate values ​​of the node positions in the graphical scale projection, draws the score scale lines of each risk feature factor in combination with the node position distribution coordinate values, arranges the score scale lines of each risk feature factor with the total score prediction axis to form a multivariate interactive interface architecture view, and establishes an initial hypoglycemia risk nomogram prediction model based on the multivariate interactive interface architecture view.

[0011] As a further aspect of the present invention, the model calibration module includes: The positive distribution analysis submodule extracts the prediction output classification results of the initial hypoglycemia risk nomogram prediction model according to the subject operating characteristic curve analysis, calculates the corresponding true positive rate and false positive rate at each prediction probability cutoff point, statistically analyzes the spatial distribution of the true positive rate and false positive rate, and establishes the true and false positive distribution index. The fitting deviation test submodule summarizes the coordinate value sequences of each item in the true and false positive distribution index to generate the area under the model prediction curve. It compares the actual observation frequency value and the expected occurrence frequency value corresponding to the area under the model prediction curve according to the goodness of fit test function, calculates the sum of the absolute discrete differences between the actual observation frequency value and the expected occurrence frequency value, and obtains the prediction fitting deviation. The model calibration output submodule compares the difference between the prediction fitting deviation and the preset prediction tolerance threshold. When the prediction fitting deviation is greater than the preset prediction tolerance threshold, the original risk scale distribution distance value is adjusted according to the area under the model prediction curve. When the prediction fitting deviation is not greater than the preset prediction tolerance threshold, the original structural parameters are maintained, and a calibrated hypoglycemia risk nomogram prediction model is generated.

[0012] As a further aspect of the present invention, the risk prediction module includes: The clinical data acquisition submodule acquires current clinical data during the preparation stage before colonoscopy for diabetic patients, extracts the real-time physiological index values ​​of various diabetic patients in the current clinical data, compares the real-time physiological index values ​​with the baseline healthy physiological index thresholds, calculates the deviation percentage coefficients corresponding to the real-time physiological index values, and integrates all deviation percentage coefficients to obtain the current clinical risk characteristic data. The probability assessment submodule inputs the percentage coefficients of each deviation in the current clinical risk feature data into the feature scale nodes of the calibrated hypoglycemia risk nomogram prediction model, calculates the individual risk score value corresponding to each feature scale node, and summarizes all individual risk score values ​​to output the probability of hypoglycemia occurrence. The risk warning determination submodule compares the probability of hypoglycemia with a preset risk warning threshold. When the probability of hypoglycemia exceeds the preset risk warning threshold, it outputs the high-risk status of the patient under test. When the probability of hypoglycemia does not exceed the preset risk warning threshold, it outputs the normal risk status of the patient under test, generating a hypoglycemia risk prediction result for colonoscopy of diabetic patients.

[0013] Methods for predicting the risk of hypoglycemia during colonoscopy in diabetic patients include the following steps: S1: Obtain clinical medical records and blood glucose monitoring records of hospitalized diabetic patients, and extract age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema status before colonoscopy, meglitinide use status and sedative use status before colonoscopy as feature parameters, and combine them to obtain a clinical independent risk feature set. S2: Analyze the clinical independent risk feature set, obtain the risk association weights corresponding to each feature parameter in the clinical independent risk feature set, screen significant independent risk features and their corresponding risk association weights, and construct a hypoglycemia risk feature mapping weight set; S3: Process the hypoglycemia risk feature mapping weight set, determine the graphical scale projection of the hypoglycemia risk feature mapping weight set under the preset hypoglycemia risk scoring benchmark scale, and construct an initial hypoglycemia risk nomogram prediction model. S4: Analyze the initial hypoglycemia risk nomogram prediction model, obtain the area under the corresponding model prediction curve and the prediction fitting deviation. If the prediction fitting deviation is greater than the preset prediction tolerance threshold, adjust the risk scale distribution state of the initial hypoglycemia risk nomogram prediction model according to the area under the model prediction curve, and output the calibrated hypoglycemia risk nomogram prediction model. S5: Obtain the current clinical data of the diabetic patient to be tested, input it into the calibrated hypoglycemia risk nomogram prediction model for processing, obtain the corresponding hypoglycemia occurrence probability and compare it with the preset risk warning threshold, determine the risk status of the diabetic patient to be tested, and generate the hypoglycemia risk prediction result of the colonoscopy examination of the diabetic patient.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by extracting core clinical characteristics such as patient age, bowel preparation score, fasting time, and specific medication use status, and combining them with weight analysis to construct a hypoglycemia risk nomogram prediction model, and by calculating the prediction fitting deviation for dynamic model calibration, this invention achieves accurate quantification of the probability of hypoglycemia during colonoscopy in diabetic patients and provides early warning of high-risk conditions. This helps medical personnel identify potential risks in advance, optimize personalized fasting and medication care plans, effectively avoid serious adverse cardiovascular and neurological events, and thus fully ensure that diabetic patients can undergo colonoscopy normally. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0016] Figure 1 This is a flowchart of the model of the present invention; Figure 2 This is a model framework diagram of the present invention; Figure 3 This is a flowchart of the feature extraction module of the present invention; Figure 4 This is a flowchart of the weight analysis module of the present invention; Figure 5 This is a flowchart of the model construction module of the present invention; Figure 6 This is a flowchart of the model calibration module of the present invention; Figure 7 This is a flowchart of the risk prediction module of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] Please see Figure 1 and Figure 2 Predictive models for the risk of hypoglycemia during colonoscopy in diabetic patients include: The feature extraction module acquires clinical medical records and blood glucose monitoring records of diabetic inpatients within a specified time period before and after colonoscopy. It extracts age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema status before colonoscopy, meglitinide use status, and sedative use status before colonoscopy as feature parameters and combines them to obtain a set of independent clinical risk features. The weight analysis module uses a multi-factor logistic regression model to analyze the clinical independent risk feature set, obtains the risk association weights corresponding to each feature parameter in the clinical independent risk feature set, and compares them with a preset significant association threshold. If the risk association weight is greater than the preset significant association threshold, the corresponding feature parameter is extracted as a significant independent risk feature. If the risk association weight is less than or equal to the preset significant association threshold, the corresponding feature parameter is removed. By screening significant independent risk features and their corresponding risk association weights, a hypoglycemia risk feature mapping weight set is constructed. The model building module obtains the preset hypoglycemia risk score benchmark scale, processes the hypoglycemia risk feature mapping weight set through the standard normalization function and mapping transformation rules, determines the graphical scale projection of the hypoglycemia risk feature mapping weight set under the preset hypoglycemia risk score benchmark scale, and constructs the initial hypoglycemia risk nomogram prediction model. The model calibration module analyzes the initial hypoglycemia risk nomogram prediction model through the receiver operating characteristic curve and the goodness-of-fit test function, obtains the area under the corresponding model prediction curve and the prediction fit deviation, and compares it with the preset prediction tolerance threshold. If the prediction fit deviation is greater than the preset prediction tolerance threshold, the risk scale distribution of the initial hypoglycemia risk nomogram prediction model is adjusted according to the area under the model prediction curve, and the calibrated hypoglycemia risk nomogram prediction model is output. The risk prediction module acquires the current clinical data of the diabetic patient to be tested, inputs it into the calibrated hypoglycemia risk nomogram prediction model for processing, obtains the corresponding probability of hypoglycemia occurrence, and compares it with the preset risk warning threshold. If the probability of hypoglycemia occurrence is greater than the preset risk warning threshold, the diabetic patient to be tested is determined to have a high risk of occurrence; if the probability of hypoglycemia occurrence is less than or equal to the preset risk warning threshold, the diabetic patient to be tested is determined to have a normal risk of occurrence, and generates the hypoglycemia risk prediction result for the colonoscopy examination of the diabetic patient.

[0019] The clinical independent risk feature set includes patient vital signs, preoperative preparation assessment, and clinical medication tracking; the hypoglycemia risk feature mapping weight set includes feature importance values, risk impact multipliers, and independent variable coefficients; the initial hypoglycemia risk nomogram prediction model includes variable scoring scales, comprehensive cumulative scoring items, and risk mapping transformation axes; the calibrated hypoglycemia risk nomogram prediction model includes assessment calibration charts, scale corrections, and optimized probability reference lines; the hypoglycemia risk prediction results for colonoscopy in diabetic patients include patient risk level and abnormal blood glucose records.

[0020] Please see Figure 2 and Figure 3 The feature extraction module includes: The clinical medical record acquisition submodule acquires clinical medical record data of diabetic inpatients before colonoscopy and during colonoscopy bowel preparation. It extracts the values ​​of various basic physiological indicators from the clinical medical record data, compares the values ​​of various basic physiological indicators with the preset normal indicator baseline, calculates the absolute difference between the values ​​of various basic physiological indicators and the preset normal indicator baseline, and summarizes all absolute differences to obtain the basic physiological difference set. The clinical medical record acquisition submodule directly extracts patients' clinical medical record data from the time of colonoscopy to bowel preparation from the hospital's internal medical information database. A data cleaning unit based on machine learning feature engineering removes invalid records lacking key indicators and extracts fasting blood glucose and glycated hemoglobin (HbA1c) values ​​as basic physiological indicators, laying a high-quality data foundation for subsequently building a hypoglycemia risk prediction model for colonoscopy in diabetic patients. The clinical medical record acquisition submodule calls preset baseline indicators, which are obtained by averaging the corresponding physiological indicators of 1000 healthy individuals. The baseline values ​​for fasting blood glucose are set at 5.5 mmol / L, and for HbA1c at 5%. The submodule then calculates the absolute difference between the extracted fasting blood glucose values ​​and the baseline values, and similarly calculates the absolute difference between the HbA1c values ​​and the baseline values, thus determining the absolute difference between each basic physiological indicator value and the preset baseline indicators. For example, when the fasting blood glucose level is 8.5 mmol / L and the glycated hemoglobin level is 7%, the absolute difference in fasting blood glucose is 3.0 mmol / L and the absolute difference in glycated hemoglobin is 2%. The clinical medical record collection submodule splices and organizes all the absolute differences calculated above, and summarizes all the absolute differences to obtain the basic physiological difference set.

[0021] The monitoring record extraction submodule collects blood glucose monitoring records of diabetic inpatients within a specified time period from the start of bowel preparation to the end of colonoscopy, based on the baseline physiological difference set. It extracts the blood glucose change values ​​of each time series in the blood glucose monitoring records, integrates the baseline physiological difference set and the blood glucose change values ​​of each time series, calculates the comprehensive physiological change amplitude at each time point, and generates the time series physiological fluctuation amplitude. The monitoring record extraction submodule acquires blood glucose monitoring records for diabetic inpatients from the start of bowel preparation to the end of colonoscopy within 24 hours using dynamic blood glucose monitoring equipment in the ward. It utilizes a time-series sliding window to extract instantaneous blood glucose readings at each time point, extracting various time-series blood glucose variation values. The submodule then retrieves the previously generated baseline physiological difference set and performs a weighted summation operation on the absolute difference in fasting blood glucose in the baseline physiological difference set with the corresponding time-series blood glucose variation values ​​at each time point to fuse the data, demonstrating an early exploration of multimodal feature fusion in machine learning systems. The monitoring record extraction submodule sets the baseline physiological weight coefficient to 0.4 and the time-series variation weight coefficient to 0.6 to calculate the comprehensive physiological variation amplitude at each time point. For example, when the absolute difference in fasting blood glucose is 3.0 mmol / L and the time-series blood glucose variation extracted at a certain time point is 4.5 mmol / L, the monitoring and recording extraction submodule multiplies 3.0 by 0.4 to get 1.2, multiplies 4.5 by 0.6 to get 2.7, and adds the two together to obtain the comprehensive physiological variation amplitude at that time point as 3.9 mmol / L. The monitoring and recording extraction submodule traverses all time points along the time axis and performs the above calculation, combining all comprehensive physiological variation amplitudes in chronological order to generate the time-series physiological fluctuation amplitude.

[0022] The risk feature combination submodule integrates the values ​​of various states with the temporal physiological fluctuations, extraction age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema operation status before colonoscopy, meglitinide use status, and sedative use status before colonoscopy to form a feature mapping matrix and establish a clinically independent risk feature set. The risk feature combination submodule continuously extracts various status values ​​from the medical information database. It extracts patient age values ​​and quantifies them by interval, assigning a value of 2 to those over 65 years old and 1 to others. It extracts the Boston Score for Colonoscopy Bowel Preparation, directly using the actual score. It extracts the fasting time before colonoscopy, using the actual value in hours. It determines the status of the enema procedure before colonoscopy, assigning a value of 1 if performed and 0 if not. It determines the use of meglitinides, assigning a value of 1 if used and 0 if not. It determines the use of sedatives before colonoscopy, assigning a value of 1 if used and 0 if not. The risk feature combination submodule uses the quantified status values ​​and the previously generated temporal physiological fluctuation amplitude as a column vector, performing a matrix-level concatenation operation to form a feature mapping matrix. This provides a standardized underlying data structure for predicting the risk of hypoglycemia during colonoscopy in diabetic patients. For example, an age of 68 is assigned a value of 2, a Boston score of 7, a fasting period of 12 hours, an enema administered, no meglitinides used, and a sedative used are assigned a value of 1. These are then combined with the aforementioned time-series physiological fluctuation amplitude of 3.9 to obtain a single-row matrix vector consisting of 7 values. The risk feature combination submodule establishes a clinically independent risk feature set based on multidimensional matrix data to fully meet the configuration requirements of machine learning algorithms for high-dimensional input spaces.

[0023] Please see Figure 2 and Figure 4 The weighting analysis module includes: The risk weight analysis submodule, based on the clinical independent risk feature set, decomposes the independent feature parameter items of each dimension, calculates the corresponding partial regression coefficient and standard error value of each independent feature parameter item under the multi-factor logistic regression rule, and merges the corresponding partial regression coefficient and standard error value of each independent feature parameter item to obtain the risk association weight. The risk weighting analysis submodule, based on the aforementioned output set of clinically independent risk features, breaks down each dimension's independent feature parameter item. This submodule incorporates a multi-factor logistic regression analysis engine with typical machine learning characteristics. It uses the occurrence of hypoglycemia during colonoscopy as the target variable (1 for occurrence, 0 for absence) and inputs each dimension's independent feature parameter item as an independent variable into the regression engine. The aim is to reveal deep relationships between variables through machine learning-driven data fitting. The risk weighting analysis submodule uses maximum likelihood estimation to fit the correlation between the independent and target variables, calculating the partial regression coefficients and standard errors for each dimension's independent feature parameter item under the multi-factor logistic regression rule. For example, for the independent feature parameter item of patient age, the risk weighting analysis submodule, after logistic regression, obtains a partial regression coefficient of 1.25 and a standard error of 0.35. The submodule establishes a one-to-one mapping between the calculated partial regression coefficients and standard errors for each dimension, packages and matches them, and merges the partial regression coefficients and standard errors for each dimension's independent feature parameter item to obtain the risk association weights. The risk weight analysis submodule performs the same analysis operations on all the decomposed feature dimensions.

[0024] The significant feature screening submodule compares the numerical difference between the risk association weight and the preset significant association threshold. When the risk association weight is greater than the preset significant association threshold, the corresponding feature parameter is extracted into the candidate region. When the risk association weight is not greater than the preset significant association threshold, the corresponding feature parameter is removed to obtain significant independent risk features. The salient feature selection submodule obtains the risk association weights generated above and extracts the partial regression coefficients and standard errors of each feature. The salient feature selection submodule divides the partial regression coefficients by the standard errors to calculate the test statistic. The salient feature selection submodule calls a preset significance threshold, set at 1.96 corresponding to a 5% significance level, and compares the test statistic after risk association weight conversion with the preset significance threshold. For example, the partial regression coefficient for the age dimension is 1.25 and the standard error is 0.35; dividing them yields a test statistic of approximately 3.57. The salient feature selection submodule determines that 3.57 is greater than 1.96, and when the risk association weight is greater than the preset significance threshold, the corresponding feature parameter, age, is extracted into the candidate region. Conversely, if the test statistic for a feature dimension is 1.20, it determines that 1.20 is not greater than 1.96, and when the risk association weight is not greater than the preset significance threshold, the corresponding feature parameter is removed. The salient feature selection submodule traverses all dimensions and performs comparisons to obtain significant independent risk features. This process, as the core feature selection step in machine learning, significantly reduces the curse of dimensionality and optimizes the overall performance of predicting hypoglycemia risk during colonoscopy in diabetic patients.

[0025] The mapping weight construction submodule extracts the feature weight values ​​corresponding to the significant independent risk features, processes the feature weight values ​​according to the preset weight allocation constant, calculates the proportion coefficient of each feature weight value in the overall feature space, combines the significant independent risk features and their corresponding proportion coefficients, and establishes the hypoglycemia risk feature mapping weight set. The mapping weight construction submodule extracts the feature weight values ​​corresponding to the aforementioned significant independent risk features. These feature weight values ​​directly reference the extracted partial regression coefficient values. The mapping weight construction submodule introduces a preset weight allocation constant to process the feature weight values; this constant is calculated by accumulating all partial regression coefficient values ​​within the significant independent risk feature set. The mapping weight construction submodule divides the partial regression coefficient value of each feature by the sum of all partial regression coefficient values ​​to calculate the proportion coefficient of each feature weight value in the overall feature space. For example, if the significant independent risk feature set includes a partial regression coefficient of 1.25 for age and 0.75 for fasting time, and the sum of all partial regression coefficient values ​​is 2.0, the mapping weight construction submodule divides the partial regression coefficient of age (1.25) by 2.0 to obtain a proportion coefficient of 0.625, and divides the partial regression coefficient of fasting time (0.75) by 2.0 to obtain a proportion coefficient of 0.375. The mapping weight construction submodule combines the significant independent risk features and their corresponding proportion coefficients to establish a hypoglycemia risk feature mapping weight set, providing a precise parameter foundation for subsequently building a highly interpretable machine learning prediction model.

[0026] Please see Figure 2 and Figure 5 The model building module includes: The scale weight normalization submodule obtains the preset hypoglycemia risk score benchmark scale, processes the weight values ​​of each item in the hypoglycemia risk feature mapping weight set according to the standard normalization function, calculates the relative mapping coordinate values ​​of each weight value within the preset interval, integrates the relative mapping coordinate values ​​of each item to construct a standardized one-dimensional numerical vector sequence, and obtains a unified scale weight set. The scaling and weight normalization submodule obtains a preset baseline scale for hypoglycemia risk scoring. This baseline scale is set as a linear interval of 100 points, with values ​​ranging from 0 to 100. The scaling and weight normalization submodule processes the weight values ​​of each item in the aforementioned hypoglycemia risk feature mapping weight set according to a standard normalization function to ensure that all indicators can be used for unbiased scheduling by the machine learning algorithm within the same mathematical framework. The scaling and weight normalization submodule multiplies each percentage coefficient by 100 to calculate the relative mapping coordinate value of each weight value within a preset interval. For example, the previously calculated percentage coefficient for age is 0.625 and the percentage coefficient for fasting time is 0.375. The scaling and weight normalization submodule multiplies 0.625 by 100 to obtain a relative mapping coordinate value of 62.5 for the age feature and multiplies 0.375 by 100 to obtain a relative mapping coordinate value of 37.5 for the fasting time feature. The scaling and weight normalization submodule performs a one-dimensional array concatenation operation on the obtained relative mapping coordinate values ​​to construct a standardized one-dimensional numerical vector sequence. The scale weight normalization submodule outputs a sequence containing values ​​such as 62.5 and 37.5, ultimately yielding a unified scale weight set.

[0027] The scale projection comparison submodule compares the values ​​of each item in the unified scale weight set with the coordinate coefficient values ​​of each node under the preset hypoglycemia risk score benchmark scale, calculates the offset distance vector corresponding to the unified scale weight set values ​​in the preset hypoglycemia risk score benchmark scale, and converts the offset distance vector to generate a graphical scale projection. The scale projection comparison submodule receives the previously generated unified scale weight set. It compares the values ​​in the unified scale weight set with the coordinate coefficients of each node under the preset hypoglycemia risk scoring baseline scale. The preset hypoglycemia risk scoring baseline scale is divided into a 10-part node coordinate system, containing coordinate node values ​​of tens of integers from 0 to 100. The scale projection comparison submodule selects each relative mapping coordinate value minus the baseline starting coordinate value of 0, and calculates the corresponding offset distance vector of the unified scale weight set value in the preset hypoglycemia risk scoring baseline scale. For example, the offset distance vector length obtained by subtracting 0 from the relative mapping coordinate value of the extracted age feature (62.5) is 62.5, and the offset distance vector length obtained by subtracting 0 from the relative mapping coordinate value of the fasting time feature (37.5) is 37.5. The scale projection comparison submodule converts all calculated offset distance vectors into extended line segments originating from the base point on the visualization interface, generating a graphical scale projection, thus transforming the abstract underlying weights of machine learning into concrete, visual evaluation coordinates.

[0028] The initial model construction submodule extracts the coordinate values ​​of the node positions in the graphical scale projection, draws the score scale lines of each risk feature factor in combination with the node position distribution coordinate values, arranges the score scale lines of each risk feature factor with the total score prediction axis to form a multivariate interactive interface architecture view, and establishes an initial hypoglycemia risk nomogram prediction model based on the multivariate interactive interface architecture view. The initial model building submodule extracts the coordinate values ​​of the node positions in the previously generated graphical scale projection. These coordinate values ​​are the previously calculated values ​​such as 62.5 and 37.5. The initial model building submodule then uses these node position coordinate values ​​to draw score scale lines for each hazard characteristic factor on the virtual chart. Using age as an identifier, the initial model building submodule draws a parallel horizontal line on the chart with a length corresponding to the 62.5 scale, and below it, a parallel horizontal line with a length corresponding to the 37.5 scale, using fasting time as an identifier. The initial model building submodule adds a total score prediction axis with a scale range of 0 to 100 at the bottom of the chart. The initial model building submodule aligns the zero-scale starting points of each feature scale line and arranges the score scale lines for each hazard characteristic factor with the total score prediction axis to form a multivariate interactive interface architecture view. The initial model building submodule directly establishes an initial hypoglycemia risk nomogram prediction model with white-box machine learning characteristics based on the generated multivariate interactive interface architecture view, providing an intuitive reading interface for clinical risk probability queries. This model will play a core discriminator role in predicting the risk of hypoglycemia during colonoscopy in diabetic patients.

[0029] Please see Figure 2 and Figure 6 The model calibration module includes: The positive distribution analysis submodule extracts the classification results of the initial hypoglycemia risk nomogram prediction model based on the analysis of the subject operating characteristic curve, calculates the true positive rate and false positive rate at each prediction probability cutoff point, statistically analyzes the spatial distribution of the true positive rate and false positive rate, and establishes the true and false positive distribution index. The positive distribution analysis submodule extracts a dataset of 500 patients from the validation group and inputs it into the aforementioned initial hypoglycemia risk nomogram prediction model. Following classic machine learning evaluation criteria, namely receiver operating characteristic (ROC) curve analysis, the positive distribution analysis submodule extracts the classification results predicted by the initial hypoglycemia risk nomogram prediction model. The positive distribution analysis submodule sets prediction probability cutoff points from 0.01 to 0.99 with a step size of 0.01, comparing predicted positive samples with actual labels at each prediction probability cutoff point. The positive distribution analysis submodule calculates the true positive rate and false positive rate corresponding to each prediction probability cutoff point. The true positive rate is calculated as the number of correctly predicted positives divided by the total number of actual positives, and the false positive rate is calculated as the number of incorrectly predicted positives divided by the total number of actual negatives. For example, when the cutoff point is set to 0.50, the calculated true positive rate is 0.85 and the false positive rate is 0.15. The positive distribution analysis submodule maps the paired data corresponding to each cutoff point to coordinates, statistically analyzes the spatial distribution of the true positive rate and the false positive rate, and establishes a true and false positive distribution index to refine the ability of the hypoglycemia risk prediction mechanism in colonoscopy for diabetic patients.

[0030] Table 1 lists the partial positive distribution data of the validation group patients extracted by the positive distribution analysis submodule. See Table 1.

[0031] Table 1. Distribution data of partial positive results in the validation group.

[0032] The fitting deviation test submodule summarizes the coordinate value sequences of each item in the true and false positive distribution index to generate the area under the model prediction curve. It compares the actual observation frequency value and the expected occurrence frequency value corresponding to the area under the model prediction curve according to the goodness of fit test function, calculates the sum of the absolute discrete differences between the actual observation frequency value and the expected occurrence frequency value, and obtains the prediction fitting deviation. The fit bias testing submodule reads the previously generated set of true and false positive distribution indicators. It summarizes the coordinate value sequences of each indicator and calculates the area enclosed by the line connecting the points and the horizontal axis using the trapezoidal area integral rule, generating the area under the model's prediction curve. This area directly reflects the model's overall predictive performance from a machine learning perspective. For example, the calculated area under the prediction curve is 0.88. Following the goodness-of-fit test function calculation logic, the submodule divides the samples into 10 evaluation groups based on risk probability. It extracts the average predicted probability as the expected frequency and the actual occurrence rate as the actual observation frequency for each group. The submodule compares the actual observation frequency with the expected frequency corresponding to the area under the model's prediction curve, calculating the sum of the absolute discrete differences between the two. For example, in a certain group, the expected frequency is 15 while the actual observed frequency is 12; the absolute difference is 3. The fitting bias test submodule sums the absolute discrete differences of all 10 groups to obtain the prediction fitting bias, thereby exploring the potential bias risk in the prediction of hypoglycemia risk during colonoscopy in diabetic patients.

[0033] The model calibration output submodule compares the difference between the prediction fitting deviation and the preset prediction tolerance threshold. When the prediction fitting deviation is greater than the preset prediction tolerance threshold, the original risk scale distribution distance value is adjusted according to the area under the model prediction curve. When the prediction fitting deviation is not greater than the preset prediction tolerance threshold, the original structural parameters are maintained, and a calibrated hypoglycemia risk nomogram prediction model is generated. The model calibration output submodule extracts the aforementioned prediction fitting deviation values. The submodule sets a preset prediction tolerance threshold of 15 and compares the prediction fitting deviation with this threshold. This decision-making step is essentially a recalibration process for the machine learning model's output probability. When the calculated prediction fitting deviation is 18, the submodule determines that 18 is greater than 15. If the deviation exceeds the preset tolerance threshold, the submodule adjusts the original risk scale distribution distance based on the area under the model's prediction curve, multiplying the distance between each feature scale line by 0.95 for compression correction. When the calculated deviation is 12, the submodule determines that 12 is not greater than 15. If the deviation is not greater than the preset tolerance threshold, the original structural parameters remain unchanged. The submodule outputs the adjusted or unchanged model structure to generate a calibrated hypoglycemia risk nomogram prediction model, successfully locking in a highly accurate prediction of hypoglycemia risk during colonoscopy in diabetic patients.

[0034] Please see Figure 2 and Figure 7 The risk prediction module includes: The clinical data acquisition submodule acquires current clinical data during the preparation stage before colonoscopy for diabetic patients, extracts the real-time physiological index values ​​of various diabetic patients in the current clinical data, compares the real-time physiological index values ​​with the baseline healthy physiological index thresholds, calculates the deviation percentage coefficients corresponding to the real-time physiological index values, and integrates all deviation percentage coefficients to obtain the current clinical risk characteristic data. The clinical data acquisition submodule directly obtains current clinical data from the hospital's physiological monitoring equipment during the pre-colonoscopy preparation phase for diabetic patients. This submodule extracts real-time physiological indicators for each diabetic patient from the current clinical data, including real-time blood glucose levels. It retrieves baseline healthy physiological indicator thresholds from the database, fixing the healthy blood glucose threshold at 6.0 mmol / L. The submodule compares each real-time physiological indicator value with the baseline healthy physiological indicator threshold, subtracting the healthy blood glucose threshold from the real-time blood glucose value and then dividing by the healthy blood glucose threshold to calculate the deviation percentage coefficient for each real-time physiological indicator value. For example, if the received real-time blood glucose value for the patient is 7.5 mmol / L, subtracting 6.0 gives 1.5, dividing by 6.0, and multiplying by 100 to calculate a deviation percentage coefficient of 25%. The clinical data acquisition submodule performs the same processing on all acquired physiological data, integrating all deviation percentage coefficients to obtain the current clinical risk characteristic data, achieving real-time data injection for machine learning inference.

[0035] The probability assessment submodule inputs the percentage coefficients of various deviations in the current clinical risk feature data into the feature scale nodes of the calibrated hypoglycemia risk nomogram prediction model, calculates the individual risk score value corresponding to each feature scale node, and summarizes all individual risk score values ​​to convert and output the probability of hypoglycemia occurrence. The probability assessment submodule retrieves the integrated current clinical risk feature data. It inputs the percentage coefficients of each deviation from the current clinical risk feature data into the corresponding feature scale nodes of the calibrated hypoglycemia risk nomogram prediction model for matching. The submodule queries the corresponding scale number at the position of each deviation percentage coefficient in the model and calculates the individual risk score value corresponding to each feature scale node. For example, the calculated blood glucose deviation percentage coefficient of 25% corresponds to a single risk score of 15 points, while other physiological deviation percentages such as fasting yield a score of 10 points. The submodule uses an addition rule to summarize all individual risk scores to obtain a total score of 25 points. The submodule then combines the model's embedded total score and probability mapping rule, optimized by machine learning, to output the probability of hypoglycemia occurrence. For example, if the mapping rule sets 25 points to a 30% incidence rate, the submodule ultimately outputs this 30% value as the hypoglycemia occurrence probability result, thus effectively implementing a quantitative evaluation of hypoglycemia risk prediction for diabetic patients undergoing colonoscopy.

[0036] Table 2 lists the final scores and probability results of some patients obtained from the probability assessment submodule. See Table 2.

[0037] Table 2. Detailed Table of Patient Prediction Result Scores and Probabilities

[0038] The risk warning determination submodule compares the probability of hypoglycemia with the preset risk warning threshold. When the probability of hypoglycemia exceeds the preset risk warning threshold, it outputs the high-risk status of the patient under test. When the probability of hypoglycemia does not exceed the preset risk warning threshold, it outputs the normal risk status of the patient under test and generates the hypoglycemia risk prediction result for colonoscopy of diabetic patients. The risk warning determination submodule reads the hypoglycemia occurrence probability output by the probability assessment submodule. It also reads the system's preset risk warning threshold, set at 45%. The submodule executes numerical comparison logic, comparing the hypoglycemia occurrence probability with the preset risk warning threshold. For example, for the aforementioned patients with a hypoglycemia occurrence probability of 55%, the risk warning determination submodule determines that 55% is greater than 45%, and outputs a high-risk status message for the patient when the hypoglycemia occurrence probability exceeds the preset risk warning threshold. For another calculated hypoglycemia occurrence probability of 30%, the submodule determines that 30% is not greater than 45%, and outputs a routine risk status message for the patient when the hypoglycemia occurrence probability does not exceed the preset risk warning threshold. The risk warning determination submodule encapsulates the determined risk status and probability results into text to generate a hypoglycemia risk prediction result for colonoscopy in diabetic patients, profoundly demonstrating the key value of introducing machine learning technology in strengthening clinical intelligent early warning mechanisms.

[0039] Methods for predicting the risk of hypoglycemia during colonoscopy in diabetic patients include the following steps: S1: Obtain clinical medical records and blood glucose monitoring records of hospitalized diabetic patients, and extract age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema status before colonoscopy, meglitinide use status and sedative use status before colonoscopy as feature parameters, and combine them to obtain a clinical independent risk feature set. S2: Analyze the clinical independent risk feature set, obtain the risk association weights corresponding to each feature parameter in the clinical independent risk feature set, screen significant independent risk features and their corresponding risk association weights, and construct a hypoglycemia risk feature mapping weight set; S3: Process the hypoglycemia risk feature mapping weight set, determine the graphical scale projection of the hypoglycemia risk feature mapping weight set under the preset hypoglycemia risk score benchmark scale, and construct the initial hypoglycemia risk nomogram prediction model. S4: Analyze the initial hypoglycemia risk nomogram prediction model, obtain the area under the corresponding model prediction curve and the prediction fitting deviation. If the prediction fitting deviation is greater than the preset prediction tolerance threshold, adjust the risk scale distribution of the initial hypoglycemia risk nomogram prediction model according to the area under the model prediction curve, and output the calibrated hypoglycemia risk nomogram prediction model. S5: Obtain the current clinical data of the diabetic patient to be tested, input it into the calibrated hypoglycemia risk nomogram prediction model for processing, obtain the corresponding hypoglycemia occurrence probability and compare it with the preset risk warning threshold, determine the risk status of the diabetic patient to be tested, and generate the hypoglycemia risk prediction result of the colonoscopy examination of the diabetic patient.

[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A model for predicting the risk of hypoglycemia during colonoscopy in diabetic patients, characterized in that, include: The feature extraction module acquires clinical medical records and blood glucose monitoring records of hospitalized diabetic patients, and extracts age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema status before colonoscopy, meglitinide use status, and sedative use status before colonoscopy as feature parameters, and combines them to obtain a set of independent clinical risk features. The weight analysis module analyzes the clinical independent risk feature set, obtains the risk association weights corresponding to each feature parameter in the clinical independent risk feature set, filters significant independent risk features and their corresponding risk association weights, and constructs a hypoglycemia risk feature mapping weight set. The model building module processes the hypoglycemia risk feature mapping weight set, determines the graphical scale projection of the hypoglycemia risk feature mapping weight set under the preset hypoglycemia risk score benchmark scale, and constructs an initial hypoglycemia risk nomogram prediction model. The model calibration module analyzes the initial hypoglycemia risk nomogram prediction model, obtains the area under the corresponding model prediction curve and the prediction fitting deviation. If the prediction fitting deviation is greater than the preset prediction tolerance threshold, the risk scale distribution state of the initial hypoglycemia risk nomogram prediction model is adjusted according to the area under the model prediction curve, and the calibrated hypoglycemia risk nomogram prediction model is output. The risk prediction module acquires the current clinical data of the diabetic patient to be tested, inputs it into the calibrated hypoglycemia risk nomogram prediction model for processing, obtains the corresponding hypoglycemia occurrence probability and compares it with the preset risk warning threshold, determines the risk status of the diabetic patient to be tested, and generates the hypoglycemia risk prediction result for the colonoscopy examination of the diabetic patient.

2. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that: In the process of screening significant independent hazard features, if the risk association weight is greater than the preset significant association threshold, the corresponding feature parameter is extracted as a significant independent hazard feature; if the risk association weight is less than or equal to the preset significant association threshold, the corresponding feature parameter is removed. In the process of determining the risk status of a diabetic patient to be tested, if the probability of hypoglycemia is greater than the preset risk warning threshold, the diabetic patient to be tested is determined to have a high risk of developing hypoglycemia; if the probability of hypoglycemia is less than or equal to the preset risk warning threshold, the diabetic patient to be tested is determined to have a normal risk of developing hypoglycemia.

3. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that: The clinical independent risk feature set includes patient vital signs, preoperative preparation assessment, and clinical medication tracking; the hypoglycemia risk feature mapping weight set includes feature importance values, risk impact multipliers, and independent variable coefficients; the initial hypoglycemia risk nomogram prediction model includes variable scoring scales, comprehensive cumulative scoring items, and risk mapping transformation axes; the calibrated hypoglycemia risk nomogram prediction model includes assessment calibration charts, scale corrections, and optimized probability reference lines; the hypoglycemia risk prediction results of colonoscopy in diabetic patients include patient risk level and abnormal blood glucose records.

4. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that, The feature extraction module includes: The clinical medical record acquisition submodule acquires clinical medical record data of diabetic inpatients before colonoscopy and during colonoscopy bowel preparation. It extracts the values ​​of various basic physiological indicators from the clinical medical record data, compares the values ​​of various basic physiological indicators with the preset normal indicator baseline, calculates the absolute difference between the values ​​of various basic physiological indicators and the preset normal indicator baseline, and summarizes all absolute differences to obtain the basic physiological difference set. The monitoring record extraction submodule collects blood glucose monitoring records of diabetic inpatients within a specified time period from the start of bowel preparation to the end of colonoscopy, based on the basic physiological difference set. It extracts the blood glucose change values ​​of each time series in the blood glucose monitoring records, integrates the basic physiological difference set with the blood glucose change values ​​of each time series, calculates the comprehensive physiological change amplitude at each time point, and generates the time series physiological fluctuation amplitude. The risk feature combination submodule integrates the values ​​of various states with the temporal physiological fluctuation amplitude to form a feature mapping matrix, and establishes a clinically independent risk feature set based on the extraction age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema operation status before colonoscopy, meglitinide use status, and sedative use status before colonoscopy.

5. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that, The weight analysis module includes: The risk weight analysis submodule, based on the clinical independent risk feature set, decomposes the independent feature parameter items of each dimension, calculates the corresponding partial regression coefficient value and standard error value of each independent feature parameter item under the multi-factor logistic regression rule, and merges the corresponding partial regression coefficient value and standard error value of each independent feature parameter item to obtain the risk association weight. The significant feature screening submodule compares the numerical difference between the risk association weight and the preset significant association threshold. When the risk association weight is greater than the preset significant association threshold, the corresponding feature parameter is extracted into the candidate region. When the risk association weight is not greater than the preset significant association threshold, the corresponding feature parameter is removed to obtain significant independent risk features. The mapping weight construction submodule extracts the feature weight values ​​corresponding to the significant independent risk features, processes the feature weight values ​​according to the preset weight allocation constant, calculates the proportion coefficient of each feature weight value in the overall feature space, combines the significant independent risk features and their corresponding proportion coefficients, and establishes a hypoglycemia risk feature mapping weight set.

6. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that, The model building module includes: The scale weight normalization submodule obtains the preset hypoglycemia risk score benchmark scale, processes the weight values ​​of each item in the hypoglycemia risk feature mapping weight set according to the standard normalization function, calculates the relative mapping coordinate values ​​of each weight value within the preset interval, integrates the relative mapping coordinate values ​​of each item to construct a standardized one-dimensional numerical vector sequence, and obtains a unified scale weight set. The scale projection comparison submodule compares the values ​​of each item in the unified scale weight set with the coordinate coefficient values ​​of each node under the preset hypoglycemia risk scoring benchmark scale, calculates the offset distance vector corresponding to the values ​​of the unified scale weight set in the preset hypoglycemia risk scoring benchmark scale, and converts the offset distance vector to generate a graphical scale projection. The initial model construction submodule extracts the coordinate values ​​of the node positions in the graphical scale projection, draws the score scale lines of each risk feature factor in combination with the node position distribution coordinate values, arranges the score scale lines of each risk feature factor with the total score prediction axis to form a multivariate interactive interface architecture view, and establishes an initial hypoglycemia risk nomogram prediction model based on the multivariate interactive interface architecture view.

7. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that, The model calibration module includes: The positive distribution analysis submodule extracts the prediction output classification results of the initial hypoglycemia risk nomogram prediction model according to the subject operating characteristic curve analysis, calculates the corresponding true positive rate and false positive rate at each prediction probability cutoff point, statistically analyzes the spatial distribution of the true positive rate and false positive rate, and establishes the true and false positive distribution index. The fitting deviation test submodule summarizes the coordinate value sequences of each item in the true and false positive distribution index to generate the area under the model prediction curve. It compares the actual observation frequency value and the expected occurrence frequency value corresponding to the area under the model prediction curve according to the goodness of fit test function, calculates the sum of the absolute discrete differences between the actual observation frequency value and the expected occurrence frequency value, and obtains the prediction fitting deviation. The model calibration output submodule compares the difference between the prediction fitting deviation and the preset prediction tolerance threshold. When the prediction fitting deviation is greater than the preset prediction tolerance threshold, the original risk scale distribution distance value is adjusted according to the area under the model prediction curve. When the prediction fitting deviation is not greater than the preset prediction tolerance threshold, the original structural parameters are maintained, and a calibrated hypoglycemia risk nomogram prediction model is generated.

8. The hypoglycemia risk prediction model for colonoscopy in diabetic patients according to claim 1, characterized in that, The risk prediction module includes: The clinical data acquisition submodule acquires current clinical data during the preparation stage before colonoscopy for diabetic patients, extracts the real-time physiological index values ​​of various diabetic patients in the current clinical data, compares the real-time physiological index values ​​with the baseline healthy physiological index thresholds, calculates the deviation percentage coefficients corresponding to the real-time physiological index values, and integrates all deviation percentage coefficients to obtain the current clinical risk characteristic data. The probability assessment submodule inputs the percentage coefficients of each deviation in the current clinical risk feature data into the feature scale nodes of the calibrated hypoglycemia risk nomogram prediction model, calculates the individual risk score value corresponding to each feature scale node, and summarizes all individual risk score values ​​to output the probability of hypoglycemia occurrence. The risk warning determination submodule compares the probability of hypoglycemia with a preset risk warning threshold. When the probability of hypoglycemia exceeds the preset risk warning threshold, it outputs the high-risk status of the patient under test. When the probability of hypoglycemia does not exceed the preset risk warning threshold, it outputs the normal risk status of the patient under test, generating a hypoglycemia risk prediction result for colonoscopy of diabetic patients.

9. A method for predicting the risk of hypoglycemia during colonoscopy in diabetic patients, characterized in that, The method is used to implement the model according to any one of claims 1-8, and includes the following steps: S1: Obtain clinical medical records and blood glucose monitoring records of hospitalized diabetic patients, and extract age, Boston score for colonoscopy bowel preparation, fasting time before colonoscopy, enema status before colonoscopy, meglitinide use status and sedative use status before colonoscopy as feature parameters, and combine them to obtain a clinical independent risk feature set. S2: Analyze the clinical independent risk feature set, obtain the risk association weights corresponding to each feature parameter in the clinical independent risk feature set, screen significant independent risk features and their corresponding risk association weights, and construct a hypoglycemia risk feature mapping weight set; S3: Process the hypoglycemia risk feature mapping weight set, determine the graphical scale projection of the hypoglycemia risk feature mapping weight set under the preset hypoglycemia risk scoring benchmark scale, and construct an initial hypoglycemia risk nomogram prediction model. S4: Analyze the initial hypoglycemia risk nomogram prediction model, obtain the area under the corresponding model prediction curve and the prediction fitting deviation. If the prediction fitting deviation is greater than the preset prediction tolerance threshold, adjust the risk scale distribution state of the initial hypoglycemia risk nomogram prediction model according to the area under the model prediction curve, and output the calibrated hypoglycemia risk nomogram prediction model. S5: Obtain the current clinical data of the diabetic patient to be tested, input it into the calibrated hypoglycemia risk nomogram prediction model for processing, obtain the corresponding hypoglycemia occurrence probability and compare it with the preset risk warning threshold, determine the risk status of the diabetic patient to be tested, and generate the hypoglycemia risk prediction result of the colonoscopy examination of the diabetic patient.

Images