Operation treatment scheme rationality evaluation method and device based on machine learning

Abstract

The invention relates to a machine learning-based operation treatment scheme rationality evaluation method and device, and belongs to the technical field of medical data processing and analysis. According to the method, six different machine learning models including linear regression, K nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression or MLP regression models are adopted, and regression prediction is performed on hospitalization duration and hospitalization cost in a surgical special disease data set; and the reasonability of the surgical treatment path of the patient is evaluated from two aspects of hospitalization duration and hospitalization cost. According to the method, the hospitalization cost and the hospitalization duration in the thoracic surgery special disease data set are respectively subjected to regression prediction, and the thoracic surgery treatment scheme is ensured to reach the optimal state in the aspects of economic benefits and patient satisfaction as much as possible.

Description

Technical Field

[0001] The present invention relates to a method and device for evaluating the rationality of a surgical treatment plan based on machine learning, and belongs to the technical field of medical data processing and analysis. Background Art

[0002] Evaluating the rationality of surgical patient treatment plans can provide important data and decision-making support for improving the hospital's diagnosis and treatment standards and clinical research capabilities. The present invention comprehensively examines the treatment effect and cost-effectiveness from two key factors: hospital stay and hospitalization costs. Short-term hospitalization usually reflects the efficiency of the treatment plan, suggesting that the patient successfully recovers and is discharged within an appropriate time, indicating that the medical team has adopted an effective treatment strategy. However, if the length of hospital stay is abnormally extended, it may indicate that there are problems in the treatment process, such as slow progress in patient recovery or unnecessary medical procedures. On the other hand, the evaluation of hospitalization costs is also crucial. The ideal treatment plan should minimize the overall treatment cost and reduce the patient's medical burden while ensuring patient recovery. Low-cost treatment plans reflect the effective use of medical resources and avoid unnecessary expenses. On the contrary, high surgical costs may mean that there is a waste of resources in the treatment process, including excessive use of expensive medical technologies or drugs. Therefore, a reasonable treatment plan needs to achieve a balance between treatment effect, resource utilization and patient experience, ensuring the provision of high-quality medical services while minimizing the length of hospital stay and cost expenditure for patients. This comprehensive evaluation helps ensure that the treatment plan is optimal in terms of economic benefits and patient satisfaction. Summary of the invention

[0003] The present invention provides a method and device for evaluating the rationality of surgical treatment plans based on machine learning, which are used to perform regression prediction on the length of hospital stay and hospitalization costs in a thoracic surgery disease-specific data set, aiming to evaluate the rationality of the patient's surgical treatment path from the two perspectives of length of hospital stay and hospitalization costs.

[0004] The technical solution of the present invention is: the present invention provides a method for evaluating the rationality of a surgical treatment plan based on machine learning, using six different machine learning models including linear regression, K nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression or MLP regression model to perform regression prediction on the length of hospital stay and hospitalization costs in a surgical disease data set respectively; the rationality of the patient's surgical treatment path is evaluated from the two perspectives of length of hospital stay and hospitalization costs.

[0005] Furthermore, the surgical disease dataset is obtained in the following manner:

[0006] (1) Constructing the data set: Using data ETL tools, we extracted all the data including patient demographic characteristics, preoperative tests, surgical records, and medical history from the hospital's HIS system, LIS system, PACS system, surgical anesthesia system, and nursing system, and screened out specific data fields based on the CRF to form a thoracic surgery disease-specific data set;

[0007] (2) Cleaning data; data cleaning includes outlier processing and missing value processing;

[0008] The abnormal value processing includes unifying the format of abnormal data and eliminating special characters in the data;

[0009] The missing value processing includes filling in the numerical field with the mean and median; filling in the categorical field with the special value "-1"; and filling in the text field with the special value "None".

[0010] Furthermore, based on the prediction target, the random forest regression model and the XGBoost regression model were used in parallel to regress and predict the surgical treatment rationality index, and a surgical treatment plan rationality evaluation model with dual regression of time and cost was constructed; among them, the surgical treatment rationality index TRI was defined as:

[0011] TRI=w 1 ·normal(H cost )+w 2 ·normal(H stay )

[0012] Among them, H cost is the hospitalization cost, H stay is the length of hospital stay, normal(·) is the normalization function, w 1 , w 2 are the weights of hospitalization expenses and hospitalization duration respectively; hospitalization expenses and hospitalization duration are equivalent in evaluating the rationality of surgical treatment. 1 , w 2 Both weights are taken as 0.5.

[0013] The present invention also provides a surgical treatment plan rationality assessment device based on machine learning, including six different machine learning models, namely linear regression, K nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression or MLP regression model;

[0014] These six models are used to perform regression prediction on the length of hospital stay and hospitalization costs in the surgical disease dataset; the rationality of the patient's surgical treatment path is evaluated from the two perspectives of length of hospital stay and hospitalization costs.

[0015] Furthermore, the random forest regression model and the XGBoost regression model are further used to perform regression prediction on the surgical treatment rationality index in a parallel manner based on the prediction target, and a surgical treatment plan rationality evaluation model with dual regression of time and cost is constructed; wherein the surgical treatment rationality index TRI is defined as:

[0016] TRI=w 1 ·normal(H cost )+w 2 ·normal(H stay )

[0017] Among them, H cost is the hospitalization cost, H stay is the length of hospital stay, normal(·) is the normalization function, w 1 , w 2 are the weights of hospitalization expenses and hospitalization duration respectively; hospitalization expenses and hospitalization duration are equivalent in evaluating the rationality of surgical treatment. 1 , w 2 Both weights are taken as 0.5.

[0018] The beneficial effects of the present invention are:

[0019] 1. Extract medical data with relatively complete surgical information from the full data of thoracic surgery, construct a thoracic surgery disease-specific data set, and introduce it from multiple dimensions such as surgery type and type of thoracic surgery.

[0020] 2. Further clean the outliers and error values ​​in the dataset, and encode the numerical, categorical, text, and time fields in the dataset so that the dataset can be analyzed using machine learning algorithms.

[0021] 3. Use machine learning algorithms, such as linear regression, K-nearest neighbor, random forest, XGBoost, CatBoost and MLP, to perform regression prediction on the hospitalization costs and length of stay in the thoracic surgery disease dataset, and propose a dual regression model based on random forest and XGBoost based on the prediction results. Together with other machine learning algorithms, perform regression prediction on the TRI indicator that comprehensively considers costs and duration, to ensure that the thoracic surgery treatment plan achieves the best state in terms of economic benefits and patient satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of the rationality evaluation model of thoracic surgery treatment plan with double regression of time and cost of the present invention. DETAILED DESCRIPTION

[0023] Example 1: Figure 1As shown in the figure, a method for evaluating the rationality of surgical treatment plans based on machine learning adopts six different machine learning models including linear regression, K nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression or MLP regression model to perform regression prediction on the length of hospital stay and hospitalization costs in the surgical disease dataset respectively; the rationality of the patient's surgical treatment path is evaluated from the two perspectives of length of hospital stay and hospitalization costs.

[0024] Taking thoracic surgery as an example, the present invention proposes a thoracic surgery treatment plan rationality evaluation network based on the modeling of thoracic surgery data sets, and evaluates the comprehensive hospitalization time and hospitalization cost (Treatment Rationality Index, TRI) indicator, realizing machine learning modeling and analysis of the full amount of multi-source heterogeneous thoracic surgery data. The surgical treatment plan rationality evaluation network can also be applied to other departments.

[0025] Furthermore, the surgical disease dataset is obtained in the following manner:

[0026] (1) Constructing the data set: Using data ETL tools, we extracted all the data including patient demographic characteristics, preoperative tests, surgical records, and medical history from the hospital's HIS system, LIS system, PACS system, surgical anesthesia system, and nursing system, and screened out specific data fields based on the CRF to form a thoracic surgery disease-specific data set;

[0027] (2) Cleaning data; data cleaning includes outlier processing and missing value processing;

[0028] There are many outliers and missing values ​​in the thoracic surgery disease dataset. These erroneous data may affect the subsequent algorithm analysis, so they need to be processed and cleaned first to ensure the accuracy and reliability of the analysis results.

[0029] The abnormal value processing includes unifying the format of abnormal data and eliminating special characters in the data;

[0030] The appearance of heterogeneous values ​​in a data set may be caused by many factors. First, medical data often involves human physiology and disease status, and is therefore affected by individual differences, physiological variations, and disease specificity, resulting in data inhomogeneity. Second, due to the large number of clinical systems within the hospital, errors or inconsistencies may exist during data collection. For example, human input errors, equipment failures, or changes in experimental conditions may all lead to the generation of outliers.

[0031] Table 1 Dataset

[0032]

[0033] The presence of outliers may cause the model to misjudge the data, which will bring a series of negative effects on the later training and lead to a decrease in the prediction performance of the model. This is because outliers may deviate from the normal data distribution, affecting the model's learning of the overall data pattern. In addition, outliers may cause the model to overfit to extreme cases, causing the model to pay too much attention to outliers during the training process, affecting the normal applicability of the vast majority of patients' conditions in practical applications. As shown in the table above, in the thoracic surgery disease data set, the types of outliers that appear are roughly divided into two types: one is data format abnormality, such as the field that was originally a numerical type is mixed with Chinese or other special characters in the actual storage; the other is digital range abnormality, such as the general range of height should be "30-250", but in the actual data there will be abnormal data such as "1690.0". In response to the above two problems, the present invention unifies the format of abnormal data and eliminates special characters in the data. At the same time, for range abnormal data, it is processed according to its own information or other information, such as changing the height from "1690.0" to "169.0".

[0034] The missing value processing includes filling in the numerical field with the mean and median; filling in the categorical field with the special value "-1"; and filling in the text field with the special value "None".

[0035] The appearance of missing values ​​in the data set is related to the following factors. First, the collection of comprehensive information of in-hospital patients involves multiple medical systems. For example, the patient's physical examination data (such as height, weight, etc.) comes from the nursing system; past history data (such as chief complaint, past history, current medical history, family history, etc.) comes from the electronic pathology system. These medical systems often do not verify the integrity of the data when collecting data, resulting in incompleteness and limitations in medical data records. Secondly, due to individual differences among patients, differences in treatment plans, and the complexity of medical processes, some diagnosis and treatment data may not be fully recorded. For example, emergency surgery patients cannot undergo preoperative physical examinations. Finally, there are also missing data due to special circumstances, such as network failures, equipment offline, human omissions, etc., as well as patients' inability to provide certain information or refusal to participate in certain tests, which can also lead to missing data.

[0036] The existence of missing values ​​will affect the integrity of the sample and weaken the learning ability of the model, especially when the sample size is small. In addition, missing values ​​may cause the model to misjudge the real potential pattern, thereby affecting the prediction performance of the model, making the model overly dependent on existing data during training, thereby ignoring the information loss caused by missing data. In order to effectively deal with the impact of missing data, the missing values ​​in the data set are mainly processed from the two dimensions of "row" and "column". First, the feature missingness of each sample ("row") is counted, and samples with a feature missing ratio higher than 20% are deleted. In addition, for the remaining samples, each column ("column") of the sample is filled with missing values ​​in a targeted manner. Specifically, combined with the specific medical meaning of each field, the mean, median, etc. are used to fill the numerical field; the special value "-1" is used to fill the category field; the special value "None" is used to fill the text field, etc.

[0037] Furthermore, based on the prediction target, the random forest regression model and the XGBoost regression model were used in parallel to regress and predict the surgical treatment rationality index, and a surgical treatment plan rationality evaluation model with dual regression of time and cost was constructed; among them, the surgical treatment rationality index TRI was defined as:

[0038] TRI=w 1 ·normal(H cost )+w 2 ·normal(H stay )

[0039] Among them, H cost is the hospitalization cost, H stay is the length of hospital stay, normal(·) is the normalization function, w 1 , w 2 are the weights of hospitalization expenses and hospitalization duration respectively; hospitalization expenses and hospitalization duration are equivalent in evaluating the rationality of surgical treatment. 1 , w 2 Both weights are taken as 0.5.

[0040] Feature engineering plays an important role in machine learning algorithms. It provides more reasonable and recognizable inputs for the algorithms by processing and converting various information in the original data set. In medical data, feature engineering can extract valuable features for medical research from information such as age, gender, inspection, and examination, thereby providing support for subsequent data analysis and modeling. Therefore, the present invention encodes the numerical, category, text, and time fields in the thoracic surgery disease dataset so as to better apply them to machine learning models.

[0041] In order to select and construct an appropriate model, the present invention uses six machine learning algorithms to perform regression prediction on the hospitalization costs and hospitalization duration in the thoracic surgery disease dataset, and proposes a dual regression model based on the prediction results to perform regression prediction on the TRI index together with other machine learning algorithms. Among them, the TRI index is the linear weighted sum of hospitalization costs and hospitalization duration, which aims to evaluate the rationality of the treatment plan for thoracic surgery patients during hospitalization.

[0042] The present invention also provides a surgical treatment plan rationality assessment device based on machine learning, including six different machine learning models, namely linear regression, K nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression or MLP regression model;

[0043] These six models are used to perform regression prediction on the length of hospital stay and hospitalization costs in the surgical disease dataset; the rationality of the patient's surgical treatment path is evaluated from the two perspectives of length of hospital stay and hospitalization costs.

[0044] The present invention selects six machine learning models for modeling medical table data, and these models have different capabilities in modeling medical table data. Among them, linear regression is simple and highly explanatory, and is suitable for revealing the linear relationship between variables. In medical data analysis, it can provide intuitive relationship explanations for predicting data related to numerical variables such as patient health status and treatment effects; K nearest neighbor regression considers the nearest samples and is suitable for situations where there are local relationships in the data set. In the medical field, K nearest neighbor regression can be used to predict the treatment effect or disease development trend of neighboring patients in similar situations based on the characteristics of the patient. Random forest is an integrated model based on decision trees, which has high predictive performance and robustness. In the medical industry, random forest can be used to comprehensively consider multiple features to predict the patient's hospitalization duration, cost, etc., while reducing the risk of overfitting. XGBoost is a gradient boosting algorithm that improves the performance of the model by optimizing a collection of decision trees. In medical data analysis, XGBoost can be used to more accurately predict the patient's disease progression, drug response, etc., and performs well for high-dimensional and complex data sets; CatBoost is a gradient boosting algorithm designed specifically for categorical features, with strong robustness and high performance. In medical data, CatBoost can be used to process mixed data, such as patient information that contains both numerical and categorical features; MLP neural networks can capture complex relationships in data and are suitable for modeling large-scale high-dimensional data. In the medical field, neural network regression can be used to predict patients' physiological parameters, disease risks, etc., especially in deep learning applications.

[0045] In order to gain a deeper understanding of the hospitalization of patients undergoing thoracic surgery, the present invention firstly used the above six different machine learning models, including linear regression, K-nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression and MLP regression, to respectively perform regression prediction on the length of stay and hospitalization costs in the thoracic surgery disease dataset, aiming to evaluate the rationality of the patient's surgical treatment path from the two perspectives of length of stay and hospitalization costs. Based on the prediction target, a new rationality index for thoracic surgery treatment is proposed. The random forest regression model and the XGBoost regression model are further used to perform regression prediction on the rationality index of surgical treatment in parallel based on the prediction target, and a rationality evaluation model for surgical treatment plans with dual regression of time and cost is constructed; wherein, the surgical treatment rationality index TRI is defined as:

[0046] TRI=w 1 ·normal(H cost )+w 2 ·normal(H stay )

[0047] Among them, H cost is the hospitalization cost, H stay is the length of hospital stay, normal(·) is the normalization function, w 1 , w 2 are the weights of hospitalization expenses and hospitalization duration respectively; hospitalization expenses and hospitalization duration are equivalent in evaluating the rationality of surgical treatment. 1 , w 2 Both weights are taken as 0.5.

[0048] The present invention constructs a dual regression model integrating random forest algorithm and XGBoost algorithm to evaluate the rationality of thoracic surgery treatment plan based on two factors: hospitalization duration and hospitalization cost.

[0049] 1. Evaluation indicators:

[0050] In machine learning, the performance evaluation of regression tasks usually uses a series of indicators to measure the model's fit to the target variable and the prediction accuracy. In regression tasks, the mean absolute error (MAE) and the root mean squared error (RMSE) are

[78] They are evaluation indicators commonly used to measure the performance of prediction models. They are statistics that measure the difference between the model prediction value and the actual observation value.

[0051] MAE is the mean absolute value of the prediction error, which is calculated as the average of the absolute values ​​of the difference between the predicted value and the true value of each sample. The specific formula is:

[0052]

[0053] Where n is the number of samples, y i is the actual observed value, is the corresponding model prediction value.

[0054] RMSE is the root mean square of the prediction error. First, we need to calculate the square of the difference between the predicted value and the true value of each sample, then calculate the average of the squared difference, and finally calculate the square root of the mean. The specific formula is:

[0055]

[0056] MAE focuses on the absolute value of the prediction error and is insensitive to outliers. Whether it is a positive error or a negative error, it has the same impact on the model evaluation. RMSE pays more attention to the impact of large errors, because the square effect will amplify large errors, making RMSE more sensitive to outliers.

[0057] 2. Experimental effect analysis

[0058] Table 1

[0059]

[0060]

[0061] In order to select and construct an appropriate model, the present invention uses six machine learning algorithms to perform regression prediction on the hospitalization costs and length of stay in the thoracic surgery disease dataset, and proposes a dual regression model based on the prediction results, and performs regression prediction on the TRI indicator together with other machine learning algorithms. Among them, the TRI indicator is the linear weighted sum of hospitalization costs and length of stay, which aims to evaluate the rationality of the treatment plan for thoracic surgery patients during hospitalization. The experimental comparison results are shown in the table above. Under different regression objectives, the present invention mainly uses MAE and RMSE indicators to evaluate model performance, and the optimal test set performance is indicated in bold in the table above. This comprehensive method can effectively provide a comprehensive performance evaluation for the regression prediction task of medical data to support the selection of the most appropriate model.

[0062] In the above table, for the regression target of hospitalization expenses, the present invention compared different models, and the results showed that the random forest regression model and the CatBoost regression model achieved the best performance under the two evaluation indicators of MAE and RMSE, with a MAE value of 0.031 and a RMSE value of 0.041. For the regression target of hospitalization duration, it was observed that the random forest regression model reached the best under the MAE indicator, with a value of 0.012, while the XGBoost regression model achieved the best performance under the RMSE indicator, with a value of 0.022. In the regression target of TRI, the CatBoost regression model showed the best results under both MAE and RMSE indicators, with a specific MAE value of 0.019 and a RMSE value of 0.025. Finally, the dual regression model of time and cost proposed by the present invention based on the prediction results achieved better results than the CatBoost regression model in the regression target of TRI, with MAE and RMSE values ​​of 0.018 and 0.024, respectively.

[0063] In general, for these three regression targets, due to the robustness of the ensemble learning model to noisy data, its performance is usually significantly better than that of the simple linear model and K nearest neighbor regression. At the same time, due to the missing values ​​and unbalanced distribution in medical data, as well as the limitation of the number of experimental samples, it can be found that the performance of the model based on neural network regression (MLP) is slightly worse than that of the traditional ensemble learning algorithm or simple linear model. In addition, by comparing the performance of the model in various indicators, it can be observed that the gap between the training set error and the test set error is small, and the training set error slightly exceeds the test set error. This shows that the model successfully captures the potential correlation between indicators such as patient demographic characteristics, test examinations, and hospitalization expenses, length of stay, and treatment plan rationality indicators when learning the thoracic surgery disease dataset, and the degree of overfitting is relatively small.

[0064] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for evaluating the rationality of surgical treatment plans based on machine learning, characterized in that: Six different machine learning models, including linear regression, K-nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression, or MLP regression model, were used to perform regression prediction on the length of stay and hospitalization costs in the surgical disease dataset. The rationality of the patient's surgical treatment path is evaluated from two perspectives: length of hospitalization and hospitalization costs.

2. The method for evaluating the rationality of surgical treatment plans based on machine learning according to claim 1, characterized in that: The surgical disease dataset is obtained in the following way: (1) Constructing the data set: Using data ETL tools, we extracted all the data including patient demographic characteristics, preoperative tests, surgical records, and medical history from the hospital's HIS system, LIS system, PACS system, surgical anesthesia system, and nursing system, and screened out specific data fields based on the CRF to form a thoracic surgery disease-specific data set; (2) Cleaning data; data cleaning includes outlier processing and missing value processing; The abnormal value processing includes unifying the format of abnormal data and eliminating special characters in the data; The missing value processing includes filling in the numeric field with the mean and median; filling in the categorical field with the special value "-1"; and filling in the text field with the special value "None".

3. The method for evaluating the rationality of surgical treatment plans based on machine learning according to claim 1, characterized in that: Based on the prediction target, the random forest regression model and the XGBoost regression model were used in parallel to regress and predict the surgical treatment rationality index, and a surgical treatment plan rationality evaluation model with dual regression of time and cost was constructed; among them, the surgical treatment rationality index TRI was defined as: TRI=w1·normal(H cost )+w2·normal(H stay ) Among them, H cos t is the hospitalization cost, H stay is the length of hospital stay, normal(·) is the normalization function, w1 and w2 are the weights of hospitalization cost and hospitalization length, respectively. Hospitalization cost and hospitalization length are equivalent in evaluating the rationality of surgical treatment, and both weights w1 and w2 are taken as 0.

5.

4. A device for evaluating the rationality of surgical treatment plans based on machine learning, characterized in that: Includes six different machine learning models, including linear regression, K nearest neighbor regression, random forest regression, XGBoost regression, CatBoost regression or MLP regression model; These six models are used to perform regression prediction on the length of hospital stay and hospitalization costs in the surgical disease dataset; the rationality of the patient's surgical treatment path is evaluated from the two perspectives of length of hospital stay and hospitalization costs.

5. The device for evaluating the rationality of surgical treatment plans based on machine learning according to claim 4, characterized in that: The random forest regression model and the XGBoost regression model are further used to perform regression prediction on the surgical treatment rationality index in a parallel manner based on the prediction target, and a surgical treatment plan rationality evaluation model with dual regression of time and cost is constructed; wherein the surgical treatment rationality index TRI is defined as: TRI=w1·normal(H cost )+w2·normal(H stay ) Among them, H cost is the hospitalization cost, H stay is the length of hospital stay, normal(·) is the normalization function, w1 and w2 are the weights of hospitalization cost and hospitalization length, respectively. Hospitalization cost and hospitalization length are equivalent in evaluating the rationality of surgical treatment, and both weights w1 and w2 are taken as 0.5.

Images