Thyroid cancer postoperative bleeding risk prediction system

Abstract

The invention relates to the technical field of medical data processing, in particular to a thyroid cancer postoperative bleeding risk prediction system. The system comprises a prediction probability obtaining module used for obtaining feature data of a patient, forming a training set and a verification set, and obtaining the bleeding prediction probability of each sample in the verification set; the clustering analysis module is used for clustering the bleeding prediction probability of each sample in the verification set to obtain clusters, and analyzing each cluster to obtain a standard deviation distance; the credible analysis module is used for acquiring a predicted deviation index; obtaining a clinical credibility score of the prediction model according to the standard deviation distance, the prediction deviation index and an AUC value of the prediction model; the prediction model adjusting module is used for determining whether the prediction model needs to be optimized or not and obtaining a final prediction model; and predicting whether the patient bleeding after operation by using the final prediction model. According to the method, the postoperative bleeding risk of the patient can be accurately predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a system for predicting postoperative bleeding risk of thyroid cancer. Background Art

[0002] The prediction of bleeding risk after thyroid cancer surgery is mainly to help medical staff assess the risk of postoperative bleeding complications, so as to take preventive measures, optimize patient treatment plans and nursing measures, thereby promoting postoperative recovery, improving patient survival rates, and reducing medical costs;

[0003] Predicting the risk of postoperative bleeding after thyroid cancer surgery is typically a binary classification problem: predicting whether a patient will experience postoperative bleeding (bleeding / no bleeding). This is typically done using a logistic regression model to help doctors make more accurate clinical decisions while explaining the relative contribution of each variable to bleeding risk. This is crucial for clinicians because it provides intuitive and easy-to-understand analysis results. To ensure the model's effectiveness and accuracy in the prediction task, the regression model needs to be evaluated and validated to ensure it can provide reliable decision-making support for doctors. How to effectively evaluate the model so that it can more accurately predict the risk of postoperative bleeding after thyroid cancer surgery in clinical practice is a pressing issue. Summary of the Invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a system for predicting the risk of bleeding after thyroid cancer surgery. The technical solutions adopted are as follows:

[0005] One embodiment of the present invention provides a system for predicting the risk of bleeding after thyroid cancer surgery, the system comprising:

[0006] The prediction probability acquisition module is used to obtain the patient's characteristic data to form a training set and a validation set; the training set is used to train the prediction model and obtain the bleeding prediction probability of each sample in the validation set;

[0007] The cluster analysis module is used to cluster the bleeding prediction probability of each sample in the validation set to obtain clusters; construct a line graph based on the average bleeding prediction probability and actual bleeding rate of the samples in each cluster; and obtain the standard deviation distance based on the distance from each sample to the straight line on the line graph;

[0008] The credibility analysis module is used to obtain the prediction deviation index based on the difference between the predicted probability of bleeding and the true label of each sample in the validation set; and to obtain the clinical credibility score of the prediction model based on the standard deviation distance, the prediction deviation index and the AUC value of the prediction model;

[0009] The prediction model adjustment module is used to determine whether the prediction model needs to be optimized based on the AUC value and clinical credibility score, and to obtain the final prediction model; the final prediction model is used to predict whether the patient will bleed after surgery.

[0010] Preferably, the patient's characteristic data is obtained to form a training set and a validation set, including:

[0011] The characteristic data of a patient is a sample, and the samples corresponding to all patients constitute the training set and validation set.

[0012] Preferably, a line graph is constructed based on the average predicted probability of bleeding and the actual bleeding rate of samples in each cluster, including:

[0013] The average predicted probability of bleeding of all samples in a cluster was calculated to obtain the average predicted probability of bleeding. The ratio of the number of samples with actual bleeding in a cluster to the number of all samples in the cluster was calculated as the actual bleeding rate of the cluster. A line graph was constructed using the average predicted probability of bleeding of each cluster as the horizontal axis and the actual bleeding rate of each cluster as the vertical axis.

[0014] Preferably, obtaining the standard deviation distance based on the distance between each sample and the straight line on the line graph includes:

[0015] Obtain the distance between the point corresponding to each cluster on the line graph and the X=Y straight line respectively, and record it as the first distance; obtain the difference between the ordinate of the point corresponding to a cluster on the line graph and the ordinate corresponding to the abscissa of the point on the X=Y straight line, and record it as the ordinate difference of the point corresponding to the cluster; if the ordinate difference of the point corresponding to a cluster is greater than zero, then use the sum of the ordinate difference and the first preset value as the exponent of the first distance corresponding to the point corresponding to the cluster to obtain the first distance mapping value of the point corresponding to the cluster; if the ordinate difference of the point corresponding to a cluster is less than or equal to zero, then the first distance mapping value of the point corresponding to the cluster is the first distance corresponding to the point corresponding to the cluster; sum the first distance mapping values ​​of the points corresponding to all clusters on the line graph to obtain the standard deviation distance.

[0016] Preferably, the prediction deviation index is obtained based on the difference between the predicted probability of bleeding and the true label of each sample in the validation set, including:

[0017] Calculate the difference between the predicted probability of bleeding and the true label of each sample, and record it as the predicted difference feature of the sample; use the predicted probability of bleeding of each sample to perform weighted summation on the absolute value of the predicted difference feature of each sample and normalize it to obtain the cumulative predicted difference feature; obtain the sum of the absolute values ​​of the difference between the predicted difference feature of each sample and the mean of the predicted difference features of all samples and normalize it to obtain the fluctuation feature; add the cumulative predicted difference feature and the fluctuation feature to obtain the prediction deviation index.

[0018] Preferably, the clinical credibility score of the prediction model is obtained according to the standard deviation distance, the prediction deviation index and the AUC value of the prediction model, comprising:

[0019] The corresponding value is obtained by taking the prediction deviation as an index of the standard deviation distance, and the error index is obtained by normalization; the difference between the first preset value and the error index is added to the AUC value of the prediction model and averaged to obtain the clinical credibility score of the prediction model.

[0020] Preferably, whether the prediction model needs to be optimized is determined according to the AUC value and the clinical credibility score, and the final prediction model is obtained, comprising:

[0021] If the AUC value of the prediction model is greater than the AUC threshold value and the clinical credibility score is greater than the clinical credibility score threshold value, the prediction model does not need to be optimized, and the prediction model is the final prediction model; if the AUC value of the prediction model is less than or equal to the AUC threshold value or the clinical credibility score is less than or equal to the clinical credibility score threshold value, the prediction model is optimized to obtain the final prediction model.

[0022] The embodiment of the present application has at least the following beneficial effects: the present application obtains the feature data of the patient, constructs the training set and the validation set, and then trains the prediction model (Logistic regression model), obtains the bleeding prediction probability of each sample in the validation set, and then performs clustering analysis to obtain the standard deviation distance; then the prediction deviation index is obtained according to the difference between the bleeding prediction probability of each sample in the validation set and the real label; the clinical credibility score of the prediction model is obtained according to the standard deviation distance, the prediction deviation index and the AUC value of the prediction model, and finally whether the prediction model needs to be optimized is determined according to the AUC value and the clinical credibility score, and the final prediction model is obtained, so that the final prediction model can output more reliable bleeding prediction probability on the basis of ensuring the sorting accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, and the advantages thereof, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0024] Figure 1 A system block diagram of a thyroid cancer postoperative bleeding risk prediction system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive objectives, the following describes in detail the specific implementation, structure, features and effects of a thyroid cancer postoperative bleeding risk prediction system according to the present application, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0027] The specific scheme of the thyroid cancer postoperative bleeding risk prediction system provided by the present application is described in detail below in combination with the accompanying drawings.

[0028] Embodiment:

[0029] The main application scenario of the present application is: for binary classification problems, ROC-AUC is a common evaluation index, especially in the case of thyroid cancer postoperative bleeding, which is extremely unbalanced in class (the actual bleeding sample accounts for a very low proportion of the whole), which can effectively evaluate the overall performance of the model; in the traditional way, by selecting the prediction results at different thresholds, comparing the real sample labels, obtaining the true positive rate (TPR) and false positive rate (FPR) corresponding to the threshold, connecting all TPR-FPR into a curve to obtain the ROC curve, and taking the integral of the curve on the domain as the AUC value as the index to judge the model; but the size of the traditional AUC value represents whether the model has the ability to put "true bleeding patients" in front of "non-bleeding patients" (for example: randomly select one patient from all bleeding patients, and also select one from non-bleeding patients, the larger the AUC value, the higher the prediction probability given to the bleeding patient), which means that the traditional AUC does not consider the fitting degree of the model's prediction to the actual situation, and it is difficult to reveal the deviation problem of the model output probability and the actual bleeding risk. Therefore, the evaluation index needs to be improved in order to improve the accuracy of the model and thus improve the prediction accuracy of postoperative bleeding.

[0030] Please refer to Figure 1 which shows a system block diagram of a thyroid cancer postoperative bleeding risk prediction system provided by an embodiment of the present application, the system comprising the following modules:

[0031] A prediction probability acquisition module is configured to acquire feature data of a patient to form a training set and a validation set; and train a prediction model using the training set and acquire a bleeding prediction probability of each sample in the validation set.

[0032] First, a large amount of data of patients who have undergone thyroid cancer surgery in the past needs to be obtained, which is referred to as patient feature data, mainly including but not limited to: patient basic information: height, weight, gender, age, etc.; coagulation function related: prothrombin time, activated partial thromboplastin time (APTT), fibrinogen (FIB), etc.; platelet indicators: platelet count (PLT), mean platelet volume (MPV), platelet distribution width (PDW); intraoperative and postoperative information: operation time (minutes), operation bleeding volume (mL), postoperative drainage volume (mL / hour), operation type, etc.

[0033] After obtaining the feature data of each patient, the feature data needs to be preprocessed, mainly including data cleaning and partial classification variable coding. Specifically:

[0034] 1) Missing value processing.

[0035] Missing values in clinical data caused by not measuring, input errors or patient refusal to check, etc. can be filled by median, mean, etc.

[0036] 2) Abnormal value identification.

[0037] Caused by equipment failure or input error, keep the possible pull model, check with experts or directly set medical critical value filtering rules, set upper and lower boundaries, replace extreme values with upper and lower boundary values.

[0038] 3) Categorical variable coding.

[0039] Logistic regression cannot directly process string or nominal categorical variables, so for limited and unordered categorical variables (such as operation type: type A, type B, type C), etc. or ordered categorical variables, they can be converted into data form using one-hot encoding, numerical mapping, etc.

[0040] 4) Standardization of continuous variables.

[0041] Continuous variables such as intraoperative bleeding volume, age, and PT value have large differences in numerical distribution, which will affect the training efficiency of the model. Standardization processing is performed on them, but this operation is not performed on binary classification variables or one-hot encoded features. This operation is a prior art and will not be described in detail here.

[0042] Wherein, the characteristic data of a patient is a sample, and the sample corresponding to all patients forms a training set and a validation set.Wherein the model for predicting the probability of bleeding after surgery of a patient in this application is a Logistic regression model (prediction model).With "whether bleeding after surgery" as the target, clinical variables (the characteristic data of the patient) are used as input features, and the probability of bleeding after surgery of the patient is predicted.Thus, the training set is utilized to train the prediction model, and then the sample in the validation set is brought into verification, and the bleeding prediction probability of each sample in the validation set is obtained, which is used for subsequent evaluation of the prediction model.

[0043] The cluster analysis module is used to cluster the predicted bleeding probability of each sample in the validation set to obtain clusters; construct a line graph based on the average predicted bleeding probability and actual bleeding rate of samples in each cluster; and obtain the standard deviation distance based on the distance from each sample to the straight line on the line graph.

[0044] The above obtained the predicted probability of bleeding for each sample in the validation set. Furthermore, the prediction results were combined to traverse all thresholds to obtain TPR-FPR and generate the ROC curve.

[0045] Because it's impossible to directly verify the credibility of the model's output probabilities, analyzing the actual bleeding rates corresponding to different predicted probability intervals can indirectly reflect the consistency between the model's output probabilities and the actual risk, thereby assessing whether the model's predictive ability is close to the ideal state. First, the predicted probabilities of all samples are clustered, and the average predicted probability in each cluster is used to reflect the average level of prediction within that cluster. Then, a line graph is constructed that plots the average predicted probability of each cluster against the actual bleeding rate. Theoretically, a perfect model output probability and the actual bleeding rate will reach a standard 1:1 ratio. Therefore, by adding a straight line X = Y to the line graph and calculating the distance between the actual line and each point on the X = Y line, the deviation of the predicted probability output by the current model from the ideal state can be determined.

[0046] In reality, there are only two possible answers, "yes" or "no", for whether a patient bleeds after surgery, and the model outputs a probability value between 0 and 1. Therefore, it is impossible to directly verify the credibility of the probability output by the model one by one. Ideally, the higher the probability output by the model, the higher the actual bleeding rate should be. Therefore, by analyzing the actual bleeding ratio corresponding to the model in different predicted probability intervals, the consistency between the model output probability and the actual risk can be indirectly reflected, thereby evaluating whether the model's predictive ability is close to the ideal state.

[0047] At the same time, patients' clinical characteristics often exhibit a distinct hierarchical structure, such as tumor size, invasion of key anatomical structures, intraoperative bleeding volume, and the presence of coagulation disorders or hypertension. These key variables, in combination, determine the patient's postoperative bleeding risk, causing the risk level within the population to be more likely to manifest as discrete stratifications such as "low risk," "medium risk," or "high risk," rather than continuous and smooth changes. This characteristic structure is preserved during model training and reflected in the predicted probabilities output by the model. As a result, the model's output probabilities are often highly concentrated within certain specific intervals, forming distinct probability clusters. Therefore, we first cluster the predicted probabilities of bleeding for each sample in the validation set using DBSCAN to obtain clusters. Due to the aforementioned reasons, these predicted probabilities of bleeding are usually clearly separated, allowing DBSCAN to effectively demarcate and perform adaptive classification.

[0048] Next, the average predicted probability of bleeding of all samples in a cluster was calculated to obtain the average predicted probability of bleeding, and the ratio of the number of samples that actually bled in a cluster to the number of all samples in the cluster was calculated as the actual bleeding rate of the cluster. A line graph was constructed using the average predicted probability of bleeding of each cluster as the horizontal axis and the actual bleeding rate of each cluster as the vertical axis.

[0049] A line graph of the average predicted probability of bleeding versus the actual bleeding rate was obtained. The graph shows the correspondence between different predicted probabilities of bleeding and the actual bleeding rate. The predicted probability of bleeding based on the output of the prediction model means that the model believes that the patient has a probability of bleeding. Compared with the actual bleeding rate, theoretically, the probability output by a perfect model and the actual bleeding rate will reach a standard 1:1. Therefore, a straight line X=Y is added to the line graph to express the relationship between the predicted probability output by a theoretically perfect model and the actual bleeding rate. Then, by calculating the distance between the actual broken line and each point on the X=Y straight line, the deviation of the predicted probability output by the current model from the ideal state can be known.

[0050] At the same time, in actual clinical treatment, it is known that the impact of missed detection and false positives on postoperative bleeding in thyroid cancer is extremely different. False positives only lead to a waste of medical resources, while missed detections directly threaten the patient's life safety. Therefore, in actual treatment, the impact of missed detections on the confidence of the model prediction results should be further increased. From the above analysis, it can be seen that in the broken line graph, when the broken line is above the X=Y line (it does not need to be completely above the line, and part of the broken line is also above it), it means that the model predicted probability is lower than the actual bleeding rate, indicating that the model underestimates the risk, that is, there is a tendency to miss diagnosis. Therefore, the distance between this part and the X=Y line should be increased to enhance the impact on the confidence of the model prediction probability.

[0051] In this way, the standard deviation distance can be obtained based on the distance from each sample to the straight line on the line graph. Specifically, the distance between the point corresponding to each cluster on the line graph and the X=Y straight line is obtained, which is recorded as the first distance; the difference between the vertical coordinate of the point corresponding to a cluster on the line graph and the horizontal coordinate of the point on the X=Y straight line is obtained, which is recorded as the vertical coordinate difference of the point corresponding to the cluster; if the vertical coordinate difference of the point corresponding to a cluster is greater than zero, the sum of the vertical coordinate difference and the first preset value is used as the index of the first distance corresponding to the point corresponding to the cluster, and the first distance mapping value of the point corresponding to the cluster is obtained; if the vertical coordinate difference of the point corresponding to a cluster is less than or equal to zero, the first distance mapping value of the point corresponding to the cluster is the first distance corresponding to the point corresponding to the cluster; the first distance mapping values ​​of the points corresponding to all clusters on the line graph are summed to obtain the standard deviation distance.

[0052] The specific calculation model of standard deviation distance is:

[0053]

[0054] Among them, U represents the standard deviation distance. The larger the result, the greater the deviation between the model's predicted probability and the actual bleeding rate of the sample, the less accurate the model's expression of risk, and the more likely there is a tendency to miss diagnosis; n represents the number of clusters, that is, the number of points on the line graph; d i (x i ,l X＝Y ) represents the i-th point x on the polyline i The distance from the X=Y line is the distance between the point corresponding to the ith cluster on the line and the X=Y line, which is the first distance between the point corresponding to the ith cluster on the line; F(α i ) represents the index of the first distance corresponding to the point on the polyline corresponding to the i-th cluster; Represents the first distance mapping value of the point corresponding to the i-th cluster on the polyline; Represents the ordinate of the i-th point on the polyline, Indicates the ordinate of the abscissa of the i-th point on the line corresponding to the X=Y line; α i It represents the difference in the ordinates of the points corresponding to the i-th cluster on the broken line. When the difference is greater than 0, it means that the point is above the broken line, that is, there is a tendency of missed diagnosis, and the distance between this part and the X=Y line should be increased. At this time, F(α i ) is 1+α i , when the difference is less than or equal to 0, F(α i ) is 1, and the first preset value is 1. Thus, the standard deviation distance can be obtained.

[0055] The credibility analysis module is used to obtain the prediction deviation index based on the difference between the predicted probability of bleeding and the true label of each sample in the validation set; and to obtain the clinical credibility score of the prediction model based on the standard deviation distance, the prediction deviation index and the AUC value of the prediction model.

[0056] In the above analysis, since the standard deviation distance U only reflects the average level of the model's overall credibility, in order to obtain a more accurate indicator of the confidence of the prediction results, it is necessary to further analyze the deviation issues under different bleeding prediction probabilities. This is because the bleeding prediction probability output by the prediction model reflects the model's confidence in whether the patient will bleed. Therefore, errors caused by high prediction values ​​should be more strictly penalized. Therefore, by calculating the difference between the bleeding prediction probability of a sample and its true label and assigning different weights based on the bleeding prediction probability, the prediction model's errors in the high prediction range are amplified, thereby more accurately reflecting the reliability of the prediction model in high-risk clinical judgment scenarios. At the same time, by calculating the fluctuation of the difference between the bleeding prediction probability of a sample and its true label across samples, the stability and consistency of the model's performance on different samples are reflected.

[0057] The prediction deviation index is calculated based on the difference between the predicted probability of bleeding and the true label for each sample in the validation set. Specifically, the difference between the predicted probability of bleeding and the true label for each sample is calculated and recorded as the sample's predicted difference feature. The absolute value of each sample's predicted difference feature is weighted and normalized using the predicted probability of bleeding to obtain the cumulative predicted difference feature. The absolute value of the difference between each sample's predicted difference feature and the mean of the predicted difference features of all samples is summed and normalized to obtain the fluctuation feature. The cumulative predicted difference feature and the fluctuation feature are added together to obtain the prediction deviation index. The true label value is 0 or 1, with 0 indicating no bleeding and 1 indicating bleeding.

[0058] The calculation model of the prediction deviation index is as follows:

[0059]

[0060] e i =p i -r i ,

[0061] Among them, V represents the prediction deviation index. The larger it is, the more likely the model will make mistakes under high predictions, indicating that the clinical usability is worse and the risk control ability is lower. At the same time, the performance of the model on different samples is unstable; Norm represents the normalization operation; N represents the number of samples in the validation set; p i represents the predicted probability of bleeding of the i-th sample, r i represents the true label of the i-th sample, e ia difference between the bleeding prediction probability of the i-th sample and the true label, i.e., a prediction difference feature, a cumulative prediction difference feature; is a fluctuation feature, reflecting the fluctuation of prediction errors among samples, i.e., the greater the output result, the more unstable the model performs on different samples, and the lower the prediction reliability. By calculating the difference between the bleeding prediction probability of the sample and the true label, and by giving different weights to the difference values according to the prediction probability, the errors of the model in the high prediction interval are highlighted to more truly reflect the reliability of the model in the clinical high-risk judgment scenario.

[0062] By the standard deviation distance and the prediction deviation index, a calculation formula about the prediction confidence of the model is comprehensively constructed to obtain an index that can reflect the confidence of the prediction result of the model, and the clinical credibility score is obtained by combining the AUC value obtained in the traditional way. The credibility score and the traditional AUC value are used to construct a model evaluation vector to comprehensively determine the effect of the prediction model. Because the standard deviation distance represents the overall deviation of the prediction model from the true risk, it represents the current overall state of the prediction model. However, in order to highlight the deviation of the prediction model at high prediction probability, which has a greater impact in practice, even if the overall deviation of the prediction model is not large, but the model has more deviation at high prediction probability, and the model behavior is unstable, the output result of the prediction model is not very reliable in actual clinical practice. Therefore, the standard deviation distance is used as the basis, the prediction deviation index is used as the index, and the clinical credibility score is comprehensively constructed.

[0063] Specifically, the prediction deviation is used as the index of the standard deviation distance to obtain a corresponding value and normalize to obtain an error index; the difference between the first preset value and the error index is added to the AUC value of the prediction model and averaged to obtain the clinical credibility score of the prediction model.

[0064] The specific calculation model of the clinical credibility score of the prediction model is:

[0065]

[0066] where ω represents the clinical credibility score of the prediction model, the higher the credibility score, the better the model performs in ranking and probability output, and the more consistent the prediction result of the model with the actual possible occurrence; AUC is the AUC value of the prediction model, and the AUC value of the prediction model is obtained by a known technique which will not be described in detail here; U represents the standard deviation distance, V represents the prediction deviation, and U V represents the corresponding value obtained by using the prediction deviation as the index of the standard deviation distance, the first preset value is 1, Norm[U V ] represents the error index, and Norm represents the normalization operation.

[0067] The standard deviation distance measures the overall deviation between all model predictions and their corresponding true labels, reflecting the model's stability and fit quality at the overall sample level. It can be considered a fundamental characterization of the model's current overall performance. Therefore, when constructing clinical credibility scoring metrics, it should serve as the "base" term of the scoring system, expressing the average deviation of the current prediction model from the ideal model.

[0068] However, in actual clinical applications, model deviations in high prediction probability intervals often pose more serious risks. For example, if misjudgments still occur when the prediction probability is close to 1.0, serious misdiagnosis or missed diagnosis may result. Therefore, relying solely on the overall mean error may not effectively identify potential model instability in critical probability intervals.

[0069] To this end, a forecast deviation index (FDI) was introduced as an index term to dynamically adjust the standard deviation distance. The FDI comprehensively evaluates the model's deviation performance and forecast volatility in high-probability samples, representing the model's reliability in "high-confidence decision-making scenarios." By amplifying its impact exponentially, the model's credibility score is significantly lowered in areas of large deviation or instability, reflecting the practical application orientation of "prioritizing key points."

[0070] In summary, the standard deviation distance provides a basic assessment of the model's overall performance, while the prediction deviation index, through exponential scaling, guides the scoring to focus on the stability of high-risk clinical areas. The combination of the two forms a credibility scoring system for practical decision-making scenarios. This yields a clinical credibility score for the prediction model.

[0071] The prediction model adjustment module is used to determine whether the prediction model needs to be optimized based on the AUC value and clinical credibility score, and to obtain the final prediction model; the final prediction model is used to predict whether the patient will bleed after surgery.

[0072] After obtaining the clinical credibility score and the AUC value of the traditional method, the prediction model can be evaluated based on these two indicators. When evaluating the prediction model, the AUC of the traditional method should first be greater than the initial threshold to ensure the basic ranking ability of the prediction model. Then, based on this foundation, the higher the clinical credibility score ω, the better the model prediction effect.

[0073] The AUC threshold and clinical credibility score threshold are set based on the scenario and actual needs. The AUC value of the prediction model must first be greater than the AUC threshold. Then, if the AUC value is greater than the AUC threshold, the clinical credibility score must also be greater than the clinical credibility score threshold. At this point, no further optimization of the prediction model is required and the prediction model at this point is considered the final prediction model. If the AUC value of the prediction model is less than or equal to the AUC threshold or the clinical credibility score is less than or equal to the clinical credibility score threshold, the prediction model is optimized to obtain the final prediction model. The final prediction model is then used to predict whether the patient will bleed after surgery.

[0074] Optimizing the prediction model requires refining the model structure, reconstructing features, or calibrating probabilities until it meets the requirements. Optimizing the prediction model is a well-known technique and will not be elaborated on here. This ensures that the final prediction model outputs more reliable prediction probabilities while ensuring accurate ranking, more consistent with clinical logic, and effectively ensuring patient safety.

[0075] In summary, this application clusters the bleeding prediction probabilities of the samples in the validation set through DBSCAN, calculates the average prediction probability in each category, and the proportion of actual bleeding patients in this category, and obtains a line graph of the average prediction probability of the category-actual bleeding rate. The standard deviation distance of the model is calculated according to the line graph, and the difference between the predicted probability of the sample and its true label is used to obtain the prediction deviation index. The clinical credibility score and evaluation vector of the model are constructed by combining the AUC value obtained by the traditional method. The model is evaluated by the evaluation vector, so that the final prediction model can output a more reliable prediction probability while ensuring accurate sorting.

[0076] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0077] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A system for predicting the risk of bleeding after thyroid cancer surgery, characterized in that: The system includes: The prediction probability acquisition module is used to obtain the patient's characteristic data to form a training set and a validation set; the training set is used to train the prediction model and obtain the bleeding prediction probability of each sample in the validation set; The cluster analysis module is used to cluster the bleeding prediction probability of each sample in the validation set to obtain clusters; construct a line graph based on the average bleeding prediction probability and actual bleeding rate of the samples in each cluster; and obtain the standard deviation distance based on the distance from each sample to the straight line on the line graph; The credibility analysis module is used to obtain the prediction deviation index based on the difference between the predicted probability of bleeding and the true label of each sample in the validation set; and to obtain the clinical credibility score of the prediction model based on the standard deviation distance, the prediction deviation index and the AUC value of the prediction model; The prediction model adjustment module is used to determine whether the prediction model needs to be optimized based on the AUC value and clinical credibility score, and to obtain the final prediction model; the final prediction model is used to predict whether the patient will bleed after surgery.

2. The system for predicting postoperative bleeding risk of thyroid cancer according to claim 1, characterized in that: The patient's characteristic data is obtained to form a training set and a validation set, including: The characteristic data of a patient is a sample, and the samples corresponding to all patients constitute the training set and validation set.

3. The system for predicting postoperative bleeding risk of thyroid cancer according to claim 1, characterized in that: The line graph is constructed based on the average predicted bleeding probability and actual bleeding rate of samples in each cluster, including: The average predicted probability of bleeding of all samples in a cluster was calculated to obtain the average predicted probability of bleeding. The ratio of the number of samples with actual bleeding in a cluster to the number of all samples in the cluster was calculated as the actual bleeding rate of the cluster. A line graph was constructed using the average predicted probability of bleeding of each cluster as the horizontal axis and the actual bleeding rate of each cluster as the vertical axis.

4. The system for predicting postoperative bleeding risk of thyroid cancer according to claim 1, characterized in that: The step of obtaining the standard deviation distance based on the distance between each sample and the straight line on the line graph includes: Obtain the distance between the point corresponding to each cluster on the line graph and the X=Y straight line respectively, and record it as the first distance; obtain the difference between the ordinate of the point corresponding to a cluster on the line graph and the ordinate corresponding to the abscissa of the point on the X=Y straight line, and record it as the ordinate difference of the point corresponding to the cluster; if the ordinate difference of the point corresponding to a cluster is greater than zero, then use the sum of the ordinate difference and the first preset value as the exponent of the first distance corresponding to the point corresponding to the cluster to obtain the first distance mapping value of the point corresponding to the cluster; if the ordinate difference of the point corresponding to a cluster is less than or equal to zero, then the first distance mapping value of the point corresponding to the cluster is the first distance corresponding to the point corresponding to the cluster; sum the first distance mapping values ​​of the points corresponding to all clusters on the line graph to obtain the standard deviation distance.

5. The system for predicting postoperative bleeding risk of thyroid cancer according to claim 1, characterized in that: The prediction deviation index is obtained based on the difference between the predicted probability of bleeding and the true label of each sample in the validation set, including: Calculate the difference between the predicted probability of bleeding and the true label of each sample, and record it as the predicted difference feature of the sample; use the predicted probability of bleeding of each sample to perform weighted summation on the absolute value of the predicted difference feature of each sample and normalize it to obtain the cumulative predicted difference feature; obtain the sum of the absolute values ​​of the difference between the predicted difference feature of each sample and the mean of the predicted difference features of all samples and normalize it to obtain the fluctuation feature; add the cumulative predicted difference feature and the fluctuation feature to obtain the prediction deviation index.

6. The system for predicting postoperative bleeding risk of thyroid cancer according to claim 1, characterized in that: Obtaining the clinical credibility score of the prediction model according to the standard deviation distance, the prediction deviation index, and the AUC value of the prediction model includes: The predicted deviation is used as the index of the standard deviation distance to obtain the corresponding value and normalize it to obtain the error index; the difference between the first preset value and the error index is added to the AUC value of the prediction model and the average is calculated to obtain the clinical credibility score of the prediction model.

7. The system for predicting postoperative bleeding risk of thyroid cancer according to claim 1, characterized in that: Determining whether the prediction model needs to be optimized based on the AUC value and the clinical credibility score, and obtaining the final prediction model, includes: If the AUC value of the prediction model is greater than the AUC threshold and the clinical credibility score is greater than the clinical credibility score threshold, the prediction model does not need to be optimized and the prediction model is the final prediction model; if the AUC value of the prediction model is less than or equal to the AUC threshold or the clinical credibility score is less than or equal to the clinical credibility score threshold, the prediction model is optimized to obtain the final prediction model.

Images