CN116246752B - A method to generate and use a prediction model for nausea and vomiting after general anesthesia - Google Patents

Abstract

The present application provides a method of generating and using a predictive model for nausea and vomiting (PONV) after general anesthesia, comprising: s1, converting patient information in a PONV sample set into a feature vector; s2, randomly generating t training subsets from the PONV sample set, and generating a random forest model by utilizing each training subset; s3, randomly selecting q eigenvectors from the eigenvectors, and splitting and growing at each node of the decision tree; s4, calculating the weight of each decision tree; s5, calculating the PONV incidence rate, wherein the PONV incidence rate is the weighted sum of all the decision tree prediction results; and S6, traversing the parameters t and q, repeatedly executing the steps S2-S5, comparing the output PONV incidence rate with a true value, and taking the random forest with the weight with the best prediction result as the optimal PONV prediction model. The invention can optimize the PONV prevention and treatment strategy, assist individuation and multimode prevention and treatment of PONV, improve the postoperative satisfaction of patients, reduce the physiological and psychological damage caused by the PONV to the patients, and the model is helpful for optimizing and homogenizing the medical quality of the PONV prevention and treatment work of medical institutions at all levels and improving the PONV prevention and treatment level of the whole medical system.

Description

Generation and use of a prediction model for nausea and vomiting after general anesthesia

Technical field

The present invention relates to the fields of medicine and artificial intelligence, and specifically to a method for generating and using a prediction model for nausea and vomiting after general anesthesia.

Background technique

Existing statistics show that the probability of postoperative nausea in patients under general anesthesia is about 30%, and the probability of vomiting is about 50%. Although there are currently a series of treatments and measures to intervene in postoperative nausea and vomiting (hereinafter referred to as PONV) under general anesthesia, the incidence of PONV has not significantly decreased in the more than 100 years of the development of anesthesiology. In some groups with high-risk factors, the incidence of PONV is as high as 80%. PONV may also cause serious complications such as dehydration, electrolyte imbalance, wound dehiscence, and bleeding, leading to delayed discharge from the hospital after surgery, increased medical expenses, and poor surgical experience.

Risk factors currently known to affect the incidence of PONV are mainly divided into three categories: 1. Patient-related factors; 2. Anesthesia-related factors; 3. Surgery-related factors. According to the study by Apfel et al., among many factors, there are four main factors that are closely related to the risk of PONV: female gender, history of motion sickness (MS) or PONV, non-smoking, and postoperative opioid use. Therefore, the Apfel simplified risk score is currently commonly used in clinical practice as a predictor of PONV. When 0-4 of the above risk factors are present, the incidence rates of PONV are 10%, 21%, 39%, 61% and 79%. The degree of PONV is assessed using the visual analogue scale (VAS): a 10cm ruler is used as the ruler. One end is 0, indicating no nausea and vomiting, and the other end is 10, indicating the most severe nausea and vomiting that is unbearable (1 to 4 is mild, 5 to 6 is moderate, 7 to 10 is severe).

Additionally, the latest guidance recommends putting the accuracy of risk scores into perspective. Even when the score is applied rigorously, the risk of PONV is often underestimated. The aforementioned classic risk factors may have varying degrees of impact on the actual incidence of PONV. Taking into account individual differences, some patients who are predicted to be at low risk for PONV before surgery may still experience severe nausea and vomiting after surgery.

Therefore, current guidelines recommend that patients under general anesthesia should be routinely given prophylactic antiemetic drugs during surgery. However, there is still some controversy regarding the efficacy and safety of routine use of preventive drugs. In addition to the possible side effects of antiemetics, currently commonly used preventive antiemetics such as 5-HT receptor antagonists such as ondansetron do not achieve the desired antiemetic effect. Studies have shown that more than 35% of Patients receiving ondansetron still developed PONV, which is refractory PONV.

The current problem that plagues clinical practice is that the factors affecting PONV are numerous and complex, and individual differences are large, making it difficult to predict patients and making treatment measures ineffective.

Contents of the invention

In order to overcome the above-mentioned shortcomings of existing technologies, optimize the precise prevention and treatment strategy of PONV, facilitate individualized and multi-modal prevention and treatment of PONV, better identify high-risk patients with PONV, improve patient satisfaction, improve patient postoperative experience, and reduce the physiological consequences of PONV on patients. and psychological damage, to promote rapid perioperative recovery. The present invention proposes a method for generating a prediction model for nausea and vomiting after general anesthesia. It combines patient information with artificial intelligence algorithms to generate a prediction model, and uses epidemiological big data to coordinate and Feedback optimizes the prediction model, which includes statistical analysis of risk factors for nausea and vomiting after general anesthesia and the effect of existing antiemetic measures, thereby accurately predicting the incidence of PONV and medication information.

The method for generating a prediction model for nausea and vomiting after general anesthesia proposed by the present invention includes:

S1. Convert the patient information in the PONV sample set into feature vectors;

S2. Randomly generate t training subsets from the PONV sample set, and use each training subset to generate a random forest;

S3. Randomly select q feature vectors from the feature vectors, split and grow at each node of the decision tree of the random forest;

S4. Calculate the weight of each decision tree;

S5. Calculate the PONV incidence rate, which is the weighted sum of all decision tree prediction results;

S6. Traverse parameters t and q, repeat steps S2-S5, and compare the output PONV incidence rate with the true PONV value. The weighted random forest corresponding to t and q with the best prediction results is the optimal PONV prediction. Model.

Further, in step S1, the patient information includes basic personal information, physiological information, past medical history information, surgical information, and anesthesia information.

Further, in step S1, the correlation coefficient of the feature is calculated through the regression algorithm, and then the Wald value of the feature is calculated, and the feature whose Wald value is greater than the predetermined threshold is selected as the feature to be used.

Further, in step S4, it includes:

S41. Calculate the correlation coefficient between each feature vector of the sample and the incidence of PONV;

S42. Calculate the weight for each decision tree. The calculation formula is: weight = the sum of correlation coefficients corresponding to the features included in the decision tree.

Further, the weight of each decision tree is normalized.

Further, the correlation coefficient is Pearson correlation coefficient.

Further, if the correlation coefficient is a negative number, the absolute value is taken as the correlation coefficient.

Further, in step S6, when comparing the output PONV incidence rate with the actual PONV value, the determination coefficient R ² is calculated. The calculation formula is:

In the formula: p represents the number of samples; _yi represents the true PONV value of the i-th sample; Represents the PONV incidence rate of the i-th sample; Represents the average of the true PONV values of p samples.

Further, the PONV prediction model also includes a medication prediction model, which is trained through the PONV sample set and used to output the PONV medication preferred by the patient.

According to another aspect of the present invention, a method for using a PONV prediction model is proposed, including:

Collect patient information from patients;

Input the patient information into the PONV prediction model to obtain the patient's PONV incidence rate, medication recommendations and analysis reports.

The beneficial effects of the present invention are:

(1) The model is generated through big data samples and integrates the clinical experience of medical experts and patient feedback, which can optimize PONV prevention and treatment strategies, facilitate individualized and multi-modal prevention and treatment of PONV, and improve patient satisfaction;

(2) It can predict and provide personalized treatment plans, predict the probability of postoperative nausea and vomiting, classify according to this probability, and further provide targeted treatment plans.

(3) The generated model integrates clinical experience feedback from medical experts and patients, and the patient information is input into the PONV prediction model to obtain the patient's PONV incidence rate and medication information. The PONV prediction model can also generate reports for doctors and technicians to review. The content of the report includes: the number of sample sets, patient information used, coefficient of determination, number and weight of decision trees, etc.

(4) A small program questionnaire or an intelligent voice question and answer can be designed according to the present invention to obtain the patient information required by the present invention, and then through the model generated by the present invention, the probability of nausea and vomiting and recommended medication measures for the patient are finally given. . Making PONV prevention and treatment more simple, accurate and effective will help optimize and homogenize the medical quality of PONV prevention and treatment work in medical institutions at all levels, and improve the overall PONV prevention and treatment level of the medical system.

Description of the drawings

Figure 1 is a schematic flowchart of a method for generating a PONV prediction model according to an embodiment of the present invention;

Figure 2 is a schematic flowchart of a method of using a PONV prediction model according to an embodiment of the present invention.

As shown in the figures, in order to clearly implement the structures or methods of the embodiments of the present invention, specific symbols are marked in the figures, but this is only for illustration and is not intended to limit the present invention to this specific equipment and environment. According to specific needs, those of ordinary skill in the art can adjust and modify these components, labels, and environments, and the adjustments and modifications made are still included in the scope of the appended claims.

Detailed ways

The speed estimation capability testing system and usage method provided by the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

In the following description, various aspects of the present invention will be described; however, it will be apparent to one of ordinary skill in the art that the present invention may be implemented using only some or all of its structures. Specific numbers, arrangements and sequences are set forth for clarity of explanation, but it will be apparent that the invention may be practiced without these specific details. In other cases, well-known features will not be described in detail in order not to obscure the invention.

As shown in Figure 1, the method for generating a prediction model for nausea and vomiting after general anesthesia proposed by the present invention includes:

S1. Convert the patient information in the PONV sample set into feature vectors;

S2. Randomly generate t training subsets from the PONV sample set, and use each training subset to generate a random forest;

S3. Randomly select q feature vectors from the feature vectors, split and grow at each node of the decision tree of the random forest;

S4. Calculate the weight of each decision tree;

S5. Calculate the PONV incidence rate, which is the weighted sum of all decision tree prediction results;

S6. Traverse parameters t and q, repeat steps S2-S5, compare the output PONV incidence rate with the true value of PONV, and the weighted random forest corresponding to t and q with the best prediction results is the optimal PONV. Predictive model.

In step S1, patient information includes basic personal information, such as height, weight, hospital, department, etc.; physiological information, including heart rate, blood pressure, blood oxygen saturation, hemodynamic changes, etc.; past medical history information, such as motion sickness. (MS) or PONV history, morning sickness, non-smoking, etc.; surgical information, such as type of surgery, length of surgery, surgical position, laparoscopic surgery, pneumoperitoneum, etc.; anesthesia information, such as anesthesia method, dose of inhaled anesthetic, obesity, surgical Post-opioid use, postoperative restlessness, patient satisfaction, etc.; PONV real value includes actual occurrence and severity score. There may be missing information in the above information, such as pneumoperitoneum in the surgical information. If there is no such information, it will be 0. PONV true values include actual occurrences and severity scores, which are used for comparison with predicted values for the sample.

Then the patient information of the sample is encoded as a feature, that is, the text or graphics are converted into feature vectors. The encoding can use existing methods, such as word2vec model, one-hot encoding, etc. Patient information in the PONV sample set can be extracted from the medical records of each hospital.

In the patient information, there are some features that have no or very little impact on whether the patient will end up with nausea and vomiting, such as the hospital, hospitalization number, etc. in the personal basic information, while some features such as the length of surgery, type of surgery, and some special surgical positions have a significant impact PONV has a greater impact after general anesthesia. For example, blood pressure in physiological information, intraoperative blood pressure reduction can lead to changes in the function of the vestibular organ and blood circulation in the digestive tract and vomiting center, leading to the occurrence of PONV; type of surgery in surgical information, such as gastrointestinal surgery, cavity The increase in intra-abdominal pressure caused by pneumoperitoneum during endoscopic surgery, the mechanical stimulation of the gastrointestinal tract during intraoperative operations, and the changes in anatomical structure after general anesthesia will all affect the occurrence of PONV.

Therefore, in one embodiment, the correlation between each feature and PONV is evaluated through a regression algorithm, and feature selection is performed. The logistic regression model is as follows:

Among them, X is the feature vector, β is the correlation coefficient, which can reflect the impact of features on PONV, and P is patient satisfaction,

According to the logistic regression model, the Wald value of each feature is calculated. The larger the value, the greater the influence. Select the feature with a Wald value greater than the predetermined threshold as the feature to be used ultimately.

In steps S2-S5, the random forest model consists of multiple decision trees. For the sample vector x, each decision tree h _i (x) (i indicates which decision tree) predicts the result of the sample vector relatively independently. After the random forest model obtains the prediction results of all decision trees, for regression problems, the random forest model calculates the mean of the prediction results given by all decision trees as the final prediction result. However, such a model cannot reflect the specific relationship between features and final results. Therefore, the present invention gives different weights through correlation coefficients for decision trees with different predictive capabilities, overcoming the problem of the same weight of decision trees in the traditional random forest model, and thereby improving The prediction accuracy of the PONV prediction model.

In step S2, bootstrap sampling technology can be used to randomly generate t training subsets from the PONV sample set. Generating decision trees through random forests is a commonly used neural network algorithm, which will not be described here.

In step S3, q eigenvectors are randomly selected from the eigenvectors, and the optimal attributes are selected for splitting according to predetermined rules. Each decision tree grows to the maximum extent and is completely split without pruning during the process. Splitting and growing decision trees are existing methods and will not be described again here.

In step S4, the method of calculating the weight of the decision tree includes:

S41. Calculate the correlation between each feature vector and the incidence of PONV. For example, you can use the Pearson correlation coefficient to calculate the set of Pearson correlation coefficients, recorded as R={R ₁ , R ₂ ,...R _m }, m is 1 to q. Other correlation coefficient calculation methods can also be used to obtain the correlation between each feature vector and the PONV incidence rate, such as the mutual information method.

S42. Calculate the weight for each decision tree h _k (x). The calculation formula is: weight = the sum of correlation coefficients corresponding to the features included in the decision tree. Preferably, the weights are normalized.

Because the Pearson correlation coefficient ranges from -1 to 1, when the correlation coefficient is less than 0, it means that the two show a negative correlation, and the influence between the two still exists. Therefore, the correlation coefficient of the feature is used to calculate the feature weight. When calculating the absolute value of the correlation coefficient, the absolute value of the correlation coefficient is used to calculate the corresponding feature weight.

For each decision tree, the training subset used is different, and the features used by each decision tree are not necessarily the same, so the weight of each decision tree is different. The PONV prediction model obtained by this method can better and more accurately reflect the relationship between analgesia-related feature vectors and the incidence of PONV, improving the accuracy of prediction.

In step S5, the decision tree h _k (x) (k is 1 to t) generated through the previous steps is a random forest model, in which the weight corresponding to each decision tree is w ₁ , w ₂ ,...w _t , so A random forest with weights is formed. For the feature vectors in the sample set generated in step S1, the model prediction results are expressed as the prediction results of each decision tree multiplied by the corresponding weights and then added together to obtain the PONV incidence rate.

In step S6, the output PONV incidence rate is compared with the true value of the PONV incidence rate corresponding to the sample, and the determination coefficient R ² is calculated. The calculation formula is:

In the formula: p represents the number of samples; _yi represents the true PONV value of the i-th sample; Represents the PONV incidence rate of the i-th sample; Represents the average of the true PONV values of p samples. The larger the coefficient of determination ^R2 , the better the prediction result.

When generating a PONV prediction model, traverse t and q, calculate the coefficient of determination for each traversal, select t and q with the largest coefficients of determination as the parameters of the optimal random forest, and use the optimal random forest as the PONV prediction model (the Random forest has the weight of each decision tree calculated by S4), t is 1~the number of samples in the PONV sample set, q is 1~the number of sample feature vectors.

Furthermore, as the number of sample sets increases, the method of the present invention can be used for multiple calculations to obtain a real-time PONV prediction model.

In one embodiment, the patient information also includes PONV degree, and this information is also input into the random forest as a feature to further improve the accuracy of the PONV prediction model.

In one embodiment, the PONV prediction model also includes a medication prediction model. The medication prediction model is also a neural network, such as a random forest model, used to output the medications used by the patient. Specifically, the random forest model can handle classification problems, so a conventional random forest model is used, the patient's feature vector is input into the random forest model, and the drugs used by the patient are output as predicted values, and the PONV sample set is trained after general anesthesia. It can be trained. The random forest generation and training here are commonly used methods and will not be described again. The trained medication prediction model can output medication information, such as drug name and dosage, based on the input patient information.

In one embodiment, the medication prediction model can also obtain drug information that can be used by sub-optimal and sub-sub-optimal patients, so as to provide doctors with multiple choices.

As shown in Figure 2, the usage method of the PONV prediction model proposed by the present invention includes:

Collect patient information;

The patient information is input into the PONV prediction model to obtain the patient's PONV incidence and medication information. In addition, the PONV prediction model can also generate reports for doctors and technicians to review. The content of the report includes: the number of sample sets, patient information used, coefficient of determination, number and weight of decision trees, etc.

A small program questionnaire or an intelligent voice question and answer can be designed according to the present invention to obtain the patient information required by the present invention, and then through the model generated by the present invention, the probability of nausea and vomiting and recommended medication measures for the patient are finally given.

Finally, it should be noted that the above embodiments are only used to describe the technical solutions of the present invention and do not limit the technical methods. The present invention can be extended to other modifications, changes, applications and embodiments in application, and therefore it is considered that all such Modifications, changes, applications, and embodiments are within the spirit and scope of the present invention.

Claims (7)

1. A method for generating a prediction model for nausea and vomiting after general anesthesia, characterized in that the generating method includes: S1. Convert the patient information in the PONV sample set into feature vectors; S2. Randomly generate t training subsets from the PONV sample set, and use each training subset to generate a random forest; S3. Randomly select q feature vectors from the feature vectors, split and grow at each node of the decision tree of the random forest; S4. Calculate the weight of each decision tree; S5. Calculate the PONV incidence rate, which is the weighted sum of all decision tree prediction results; S6. Traverse parameters t and q, repeat steps S2-S5, and compare the output PONV incidence rate with the true PONV value. The weighted random forest corresponding to t and q with the best prediction results is the optimal PONV prediction. Model; In step S1, the correlation coefficient of the feature is calculated through the regression algorithm, and then the Wald value of the feature is calculated, and the feature with a Wald value greater than the predetermined threshold is selected as the feature to be used. The logistic regression model is as follows: Among them, X is the feature vector, β is the correlation coefficient, which can reflect the impact of features on PONV, and P is patient satisfaction; In step S4, it includes: S41. Calculate the correlation coefficient between each feature vector of the sample and the incidence of PONV; S42. Calculate the weight for each decision tree. The calculation formula is: weight = the sum of correlation coefficients corresponding to the features included in the decision tree; In step S6, when comparing the output PONV incidence rate with the true PONV value, the determination coefficient R ² is calculated. The calculation formula is: In the formula: p represents the number of samples; _yi represents the true PONV value of the i-th sample; Indicates the PONV incidence rate of the i-th sample;/> Represents the average of the true PONV values of p samples. 2. The generation method according to claim 1, characterized in that in step S1, patient information includes basic personal information, physiological information, past medical history information, surgical information, and anesthesia information. 3. The generation method according to claim 1, characterized in that the weight of each decision tree is normalized. 4. The generation method according to claim 1, characterized in that the correlation coefficient is a Pearson correlation coefficient. 5. The generation method according to claim 4, characterized in that if the correlation coefficient is a negative number, the absolute value is taken as the correlation coefficient. 6. The generation method according to any one of claims 1 to 5, characterized in that the PONV prediction model also includes a medication prediction model, and the medication prediction model is trained through a PONV sample set to output the PONV preferred by the patient. Medication. 7. A method for using a prediction model for nausea and vomiting after general anesthesia, which is characterized by including: Collect patient information; The patient information is input into the PONV prediction model described in any one of claims 1 to 6, and the patient's PONV incidence rate, medication recommendations and analysis reports are obtained.