Method and device for predicting muscle loss after liver transplantation of liver cancer

Abstract

The invention discloses a liver cancer post-transplantation muscle loss prediction method and device, and belongs to the technical field of biomedicine, and the method comprises the steps: constructing an unbalanced random forest model comprising a plurality of decision trees, independently predicting through the plurality of decision trees to obtain a plurality of high-risk or low-risk dichotomy prediction results, taking the proportion of the decision tree with the prediction result being high risk in all decision trees as the prediction probability of muscle loss after liver cancer liver transplantation finally output by the model; performing risk classification according to the prediction probability, and judging the postoperative muscle loss level of the patient as high risk, medium risk or low risk; constructing a decision tree model to layer prognosis of patients, and dividing the patients into a high-risk group, a medium-risk group and a low-risk group; and carrying out survival rate prediction in combination with liver transplantation standards and prognosis layering results of patients. According to the method, the muscle loss level after the liver transplantation can be accurately predicted, and a clinician can make a personalized treatment scheme and perform scientific management after the transplantation.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical technology, and in particular relates to a method and device for predicting muscle loss after liver transplantation for liver cancer. Background Art

[0002] Hepatocellular carcinoma (HCC) is a major cause of cancer-related death worldwide, and its severity cannot be ignored. Among the treatment methods, liver transplantation is a radical treatment method for some HCC patients. During the rehabilitation process after liver transplantation, muscle loss is a key factor that has not been fully valued but affects the patient's overall recovery and long-term survival outcomes.

[0003] Muscle loss, also known as sarcopenia in medicine, is a phenomenon of reduced muscle mass and decreased muscle function associated with multiple factors such as aging, disease progression, and surgical treatment. This phenomenon is particularly significant in liver transplant patients, which may not only lead to a decrease in physical strength and limited mobility, but may also further increase the risk of other complications, such as infection and thrombosis, thus seriously affecting the patient's postoperative recovery process and long-term quality of life.

[0004] Although previous studies have highlighted the prognostic significance of sarcopenia in patients undergoing liver transplantation, this area of ​​research remains understudied, especially for patients who do not suffer from sarcopenia before surgery, the dynamic changes in muscle loss have not been fully explored. This gap in understanding has hindered the development of effective prediction models, resulting in clinicians often lacking reliable tools to accurately determine which patients may experience significant muscle loss after surgery, and thus making it difficult to take effective intervention measures in a timely manner to prevent and reduce this risk, which not only affects the patient's treatment outcome and recovery process, but may also increase the waste of medical resources to a certain extent.

[0005] In recent years, with the rapid development and widespread application of machine learning technology, its application prospects in the field of health care have become increasingly broad, opening up new ways to predict complex clinical outcomes. Machine learning has also shown great potential and value in the prediction and intervention of muscle loss after liver transplantation. By collecting and analyzing multi-dimensional clinical data including basic patient information, preoperative and postoperative examination results, treatment records, etc., machine learning models can build more sophisticated and accurate prediction models, identify high-risk patient groups who may experience significant muscle loss after surgery, and thus provide clinicians with a more scientific basis for decision-making.

[0006] In summary, the severity of hepatocellular carcinoma and the impact of muscle loss after liver transplantation on patient recovery and long-term outcomes cannot be ignored, and the introduction of machine learning technology provides new ideas and means to solve this problem. In the future, with the continuous deepening of related research and the continuous improvement of technology, it is necessary to further develop a model that can predict muscle loss after liver transplantation in patients with hepatocellular carcinoma, play a more important clinical role in the prediction and intervention of muscle loss after liver transplantation, and stimulate further exploration in this field. Summary of the invention

[0007] In view of the above, the purpose of the present invention is to provide a method and device for predicting muscle loss after liver transplantation for liver cancer, which can achieve accurate prediction of the level of muscle loss after liver transplantation, and combine the patient prognosis stratification and survival rate prediction results to help clinical physicians formulate personalized treatment plans and scientific management after transplantation.

[0008] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention is as follows: An embodiment of the present invention provides a method for predicting muscle loss after liver transplantation for liver cancer, comprising the following steps: An unbalanced random forest model including several decision trees is constructed. Several high-risk or low-risk binary classification prediction results are obtained by independently predicting multiple patient clinical key indicators after input preprocessing through several decision trees. The proportion of decision trees with high-risk prediction results to all decision trees is used as the predicted probability of muscle loss after liver transplantation for liver cancer, which is the final output of the unbalanced random forest model. If the predicted probability is greater than or equal to the first risk threshold, the postoperative muscle loss level is determined to be high risk; if the predicted probability is less than the first risk threshold and greater than or equal to the second risk threshold, the postoperative muscle loss level is determined to be medium risk; if the predicted probability is less than the second risk threshold, the postoperative muscle loss level is determined to be low risk, wherein the first risk threshold is greater than the second risk threshold; A decision tree model was constructed to stratify the patient prognosis, in which patients with SMI less than or equal to the stratification threshold were directly classified into the high-risk group, patients with SMI greater than the stratification threshold and muscle loss level judged as high risk were classified into the medium-risk group, and patients with SMI greater than the stratification threshold and muscle loss level judged as medium risk or low risk were classified into the low-risk group; The survival rate was predicted by combining the liver transplantation standards and the patient prognostic stratification results. For patients who met the liver transplantation standards, if the patients were in the high-risk group, the 1-year survival rate after surgery was 71.5%, the 3-year survival rate after surgery was 53.2%, and the 5-year survival rate after surgery was 36.3%; if the patients were in the intermediate-risk group, the 1-year survival rate after surgery was 86.4%, the 3-year survival rate after surgery was 74.6%, and the 5-year survival rate after surgery was 63.9%; if the patients were in the low-risk group, the 1-year survival rate after surgery was 92.3%, the 3-year survival rate after surgery was 85.9%, and the 5-year survival rate after surgery was 75.7%.

[0009] Preferably, multiple key clinical indicators of patients are screened from a wide range of clinical indicators of patients by recursive feature elimination method, and multiple key clinical indicators of patients include α-L-fucosidase, glycocholate dehydrogenase, age, alpha-fetoprotein, international normalized ratio, alanine aminotransferase, eosinophil count, basophil count, alkaline phosphatase, adenosine deaminase, glutamyl transpeptidase and body mass index in the peripheral blood of patients before surgery.

[0010] Preferably, the preprocessing of the acquired multiple key clinical indicators of patients includes missing value processing, outlier processing, feature coding conversion and data standardization.

[0011] Preferably, in the unbalanced random forest model, the decision tree construction process includes: A plurality of key clinical indicators of patients after pretreatment collected from different patients are constructed into a data set, wherein the different patients include patients with high risk of postoperative muscle loss and patients with low risk of postoperative muscle loss. Each sample in the data set represents a plurality of key clinical indicators of patients after pretreatment of a patient, and each key clinical indicator of the patient is used as a feature; For each decision tree, a part of the features of each category contained in all samples in the data set is randomly selected as the candidate feature set. Starting from the root node, an optimal segmentation feature and its segmentation point are selected from the candidate feature set according to the Gini impurity as the judgment criterion to split the samples in the data set into two subsets and generate two new child nodes. After that, the above segmentation process is recursively repeated for each new child node until the stopping condition is met. The stopping condition is that the preset maximum depth is reached or the number of samples in the node is less than the preset value.

[0012] Preferably, the calculation formula of Gini impurity is: , Among them, the high-risk proportion and the low-risk proportion are the proportion of patients with high risk of postoperative muscle loss and the proportion of patients with low risk of postoperative muscle loss in each subset obtained after each node splits the data set. Each decision tree will try various possible split features and split points, and select the split scheme that can minimize the sum of the Gini impurities of each node.

[0013] Preferably, the first risk threshold is set to 0.7, and the second risk threshold is set to 0.3.

[0014] Preferably, the delamination threshold is set to 43.6 cm 2 / m 2 .

[0015] To achieve the above-mentioned purpose of the invention, the embodiment of the present invention further provides a device for predicting muscle loss after liver transplantation for liver cancer, which is implemented by the above-mentioned method for predicting muscle loss after liver transplantation for liver cancer, and includes: a model prediction module, a risk classification module, a prognosis stratification module and a survival prediction module; The model prediction module is used to construct an unbalanced random forest model including a plurality of decision trees, and independently predict a plurality of high-risk or low-risk binary classification prediction results based on a plurality of pre-processed clinical key indicators of the inputted decision trees, and the proportion of the decision trees with the prediction results of high risk to all the decision trees is used as the prediction probability of muscle loss after liver transplantation for liver cancer, which is finally output by the unbalanced random forest model; The risk classification module is used to determine the postoperative muscle loss level as high risk if the predicted probability is greater than or equal to a first risk threshold, determine the postoperative muscle loss level as medium risk if the predicted probability is less than the first risk threshold and greater than or equal to a second risk threshold, and determine the postoperative muscle loss level as low risk if the predicted probability is less than the second risk threshold, wherein the first risk threshold is greater than the second risk threshold; The prognostic stratification module is used to construct a decision tree model to stratify the patient prognosis, wherein patients whose SMI is less than or equal to the stratification threshold are directly classified as a high-risk group, patients whose SMI is greater than the stratification threshold and whose muscle loss level is determined to be high risk are classified as a medium-risk group, and patients whose SMI is greater than the stratification threshold and whose muscle loss level is determined to be medium risk or low risk are classified as a low-risk group; The survival prediction module is used to predict the survival rate by combining the liver transplantation criteria and the patient prognosis stratification results. For patients who meet the liver transplantation criteria, if the patient is in the high-risk group, the 1-year survival rate after surgery is 71.5%, the 3-year survival rate after surgery is 53.2%, and the 5-year survival rate after surgery is 36.3%; if the patient is in the medium-risk group, the 1-year survival rate after surgery is 86.4%, the 3-year survival rate after surgery is 74.6%, and the 5-year survival rate after surgery is 63.9%; if the patient is in the low-risk group, the 1-year survival rate after surgery is 92.3%, the 3-year survival rate after surgery is 85.9%, and the 5-year survival rate after surgery is 75.7%.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses a recursive feature elimination method to select 12 key clinical indicators from a wide range of patient clinical indicators and incorporate them into the final model input, which not only improves the prediction accuracy and interpretability of the model, but also reduces the computational complexity and promotes the optimization of clinical decision-making. After testing, the AUC of the model can reach 0.8503, effectively avoiding information redundancy and noise interference, thereby significantly improving the prediction accuracy of the unbalanced random forest model.

[0017] (2) The present invention predicts and obtains the predicted probability of muscle loss after liver transplantation through an unbalanced random forest model, determines the risk level based on the predicted probability, further stratifies the patient's prognosis based on the risk level, and predicts the survival rate based on the stratification results and liver transplantation standards. On the basis of being able to accurately predict the level of muscle loss after liver transplantation, combined with the patient's prognosis stratification and survival rate prediction results, it can provide doctors with more accurate and personalized clinical decision support. This helps doctors to more accurately assess the patient's condition and formulate a more reasonable treatment plan, thereby improving the patient's treatment effect and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 It is a flow chart of a method for predicting muscle loss after liver transplantation for liver cancer provided by an embodiment of the present invention.

[0020] Figure 2 : is a schematic diagram of the relationship between the muscle loss level and the prognosis of liver transplant recipients for liver cancer provided by an embodiment of the present invention; wherein the Kaplan-Meier survival curve shown in A shows the recurrence-free survival probability of the preoperative sarcopenia group and the non-sarcopenia group in the patient cohort based on the sarcopenia condition; the Kaplan-Meier survival curve shown in B shows the recurrence-free survival probability of the preoperative mild muscle loss group and the severe muscle loss group in the patient cohort based on the muscle loss condition; the Kaplan-Meier survival curve shown in C shows the recurrence-free survival probability of the preoperative mild muscle loss group and the severe muscle loss group in the patient cohort based on the muscle loss condition (after propensity matching); D represents the relative risk ratio between the recurrence results of severe muscle loss and mild muscle loss and the death results of severe muscle loss and mild muscle loss.

[0021] Figure 31 is a schematic diagram of the establishment and performance test of the unbalanced random forest model provided by an embodiment of the present invention; wherein A is an overview of the 46 features included in the model development, the inner circle represents the feature type division and its proportion of the 46 indicators, and the outer circle represents the data properties and its proportion corresponding to each feature type; B is a curve showing the average AUC changing with the number of features in the recursive feature elimination process; C is the model ROC curve of the preoperative non-sarcopenia group; D is a model waterfall chart, showing the contribution of each variable to the model prediction, for example, when AFU=16.9, the contribution to the model prediction is +0.155 (positive contribution), and when GPDA=44, the contribution to the model prediction is -0.0955 (negative contribution), f(x)=0.838 represents the predicted value of the model after integrating 12 indicators, and E[f(x)]=0.729 represents the expectation of f(x); E is a schematic diagram of the decision tree model; the Kaplan-Meier survival curve shown in F shows the recurrence-free survival rate of the high-risk group, the intermediate-risk group and the low-risk group based on the decision tree model.

[0022] Figure 4 It is a schematic diagram of the structure of a device for predicting muscle loss after liver transplantation for liver cancer provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0023] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0024] The inventive concept of the present invention is: in view of the problem of insufficient development of models for predicting muscle loss after liver transplantation in patients with hepatocellular carcinoma in the prior art, the embodiments of the present invention provide a method and device for predicting muscle loss after liver transplantation for liver cancer, by inputting 12 key clinical indicators of the patient into an unbalanced random forest model containing several decision trees for prediction and obtaining the predicted probability of muscle loss after liver transplantation for the patient, determining the risk level according to the predicted probability, further stratifying the patient's prognosis in combination with the risk level, and predicting the survival rate according to the stratification results and liver transplantation standards, thereby achieving accurate prediction of the level of muscle loss after liver transplantation, and helping clinical physicians to formulate personalized treatment plans and scientific management after transplantation in combination with the patient's prognosis stratification and survival rate prediction results.

[0025] Figure 1 FIG. 1 is a flow chart of a method for predicting muscle loss after liver transplantation for liver cancer provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides a method for predicting muscle loss after liver transplantation for liver cancer, comprising the following steps: S1. Construct an unbalanced random forest model containing several decision trees. Use several decision trees to independently predict several high-risk or low-risk binary classification prediction results based on multiple pre-processed patient clinical key indicators. The proportion of decision trees with high-risk prediction results to all decision trees is used as the predicted probability of muscle loss after liver transplantation for liver cancer, which is the final output of the unbalanced random forest model.

[0026] S1.1, Experimental methods.

[0027] (1) Study population and data collection.

[0028] Patients who underwent liver transplantation at the First Affiliated Hospital of Zhejiang University School of Medicine (Hangzhou, China) and Shulan Hospital (Hangzhou, China) between January 2015 and January 2021 and had postoperative pathologically confirmed hepatocellular carcinoma (HCC) were included in the study. Subjects were selected and 248 patients who underwent complete preoperative and postoperative CT scans were finally included. Peripheral blood indicators collected within 2 weeks before surgery, including complete blood count and liver function tests, were also included in the study. After liver transplantation, recipients were followed up until December 31, 2022. During the follow-up period, recipients were managed according to standard protocols. Alpha-fetoprotein (AFP) levels were measured every 1-2 months. Ultrasound, computed tomography (CT), or magnetic resonance imaging (MRI) were performed every 3-6 months after liver transplantation. Recurrence-free survival (RFS) was defined as the number of months from the date of surgery to the date of the first confirmed recurrence or death. The site of HCC metastasis was determined according to standard HCC guidelines based on radiological evidence. Clinical and anthropometric parameters were collected from the Chinese Liver Transplant Registry (CLTR) in strict accordance with the Human Organ Transplant Regulations and national legal requirements.

[0029] (2) Skeletal muscle assessment.

[0030] Skeletal muscle was assessed at the third lumbar vertebra (L3) on cross-sectional CT images using Sliceomatic software (version 5.0; Tomovision). The Hounsfield unit (HU) threshold for skeletal muscle was −29 to +150 HU. Muscle area was normalized to the square of the patient's height in meters (m). 2), the skeletal muscle index (SMI) was derived. The measurements of body composition were performed by two trained observers. Before the formal measurements, 50 CT images were randomly selected from the cohort and provided to the observers to assess the repeatability between observers. One week later, the two observers repeated the same task to assess the repeatability within the observer. The images of the above measurements were verified by an anatomical radiologist. Before the dynamic analysis, an accuracy test (repeated measurement of 30 CT images by two observers) was performed according to the previously reported method to avoid random errors caused by measurement. Once the accuracy error of the measurement is known, the amplitude of muscle change indicating true biological changes can be determined by the least significant change (LSC). The value with the larger error in the results of the two observers was used in the subsequent discussion.

[0031] (3) Model development, feature selection and model interpretation.

[0032] The data processing uses a hybrid sampling technique, including oversampling and undersampling, combined with 5-fold cross validation to avoid overfitting problems. A grid search benchmark analysis was performed among 50 machine learning models, and 10-fold cross validation was performed to obtain a stable model. AUC, accuracy, F1, LogLoss, precision, and recall were used to evaluate model performance.

[0033] The SHAP method was used to obtain the importance of features and the model was interpreted by the R package "mlr3". Feature selection was based on a strategy of adding features sequentially from 1 to all features (46 clinically available indicators) according to the importance level. The model with reduced features and comparable predictive ability (AUC) was considered the final model.

[0034] (4) Statistical analysis

[0035] Categorical variables are presented as numbers (percentages) and χ was used as appropriate. 2 or Fisher's exact test. Continuous variables are expressed as median (interquartile range) and compared using the Kruskal-Wallis test. Comparison of survival curves was examined by the Kaplan-Meier method and the log-rank test.

[0036] Propensity score matching (PSM) analysis was performed using the “MatchIt” R package (version 4.5.5) with a 1:2 matching ratio and the default nearest neighbor matching algorithm. A P value < 0.05 was considered significant for the analysis. All statistical analyses were performed by R language software (version 4.2.1).

[0037] S1.2, Experimental results.

[0038] (1) Baseline characteristics of patients.

[0039] A total of 248 patients who underwent liver transplantation met the inclusion criteria of the study. By examining all possible SMI values ​​in the cohort, patients with an SMI less than 43.6 were classified as having sarcopenia, resulting in 69 patients (27.9%) being diagnosed with sarcopenia. In the non-sarcopenic cohort, patients with a postoperative muscle loss of more than 16.8% were classified as having severe muscle loss, including 46 patients (25.7%).

[0040] Kaplan-Meier analysis showed that patients with sarcopenia had a poorer prognosis (e.g. Figure 2 In addition, the study found that patients with greater muscle loss had a worse prognosis (e.g. Figure 2 ), and this trend continued after 1:2 PSM (propensity matching) (as shown in Figure 2 Competing risk analysis suggested that the difference in RFS due to muscle loss was attributable to relapse rather than to non-relapse-related mortality (e.g. Figure 2 The following variables were then included in the univariate Cox analysis: sex, age, BMI (body mass index), HBV (hepatitis B virus) status, SMI change, SMI (skeletal muscle index), MELD (end-stage liver disease) score, AFP (alpha-fetoprotein) level (>400 μg / L), tumor differentiation, tumor number, microvascular invasion, and maximum tumor diameter. Multivariate Cox regression analysis was then performed, including the following factors: muscle loss, tumor differentiation, tumor number, microvascular invasion, maximum tumor diameter, and alpha-fetoprotein level. The results of the multivariate Cox regression are shown in Table 1 below, where HR represents the hazard ratio, 95% CI represents the 95% confidence interval, and P represents the probability value.

[0041] Table 1 Multivariate Cox regression results

[0042] The results of the study showed that severe muscle loss is an independent risk factor for recurrence in liver transplant recipients of hepatocellular carcinoma. However, there is still a lack of effective preoperative models to predict the extent of postoperative muscle loss in clinical practice. To solve this problem, it is necessary to develop an effective model for predicting muscle loss after liver transplantation for liver cancer.

[0043] (2) Model development.

[0044] In order to develop a machine learning model using clinical indicators to predict muscle loss in liver transplant patients, we first collected 46 clinical features of the patients (feature overview as shown in Figure 3The following are the main parameters of the clinical diagnosis and treatment: AGE (age), height, BMI (body mass index), HBV (hepatitis B virus) infection status, MELD (end-stage liver disease) score, AFP (alpha-fetoprotein), tumor differentiation, tumor number, maximum tumor diameter, total tumor size, microvascular invasion, albumin, albumin to globulin ratio, ALT (alanine aminotransferase), AST (aspartate aminotransferase), INR (international normalized ratio), monocyte count, neutrophil count, platelet count, SMI (skeletal muscle index), white blood cell count, lymphocyte count, E (eosinophil count), B (basophil count), hemoglobin concentration, red blood cell count, RDW (red blood cell volume distribution width), MHV (mean hemoglobin volume), MCHC (mean hemoglobin concentration), HGB (hemoglobin), PCT (platelet volume), GGT (glutamyl transpeptidase), ALP (alkaline phosphatase), indirect bilirubin, direct bilirubin, ChE (cholinesterase), TBA (total bile acid), GPDA (glycocholate dehydrogenase), AFU (α-L-fucosidase), ADA (adenosine deaminase), TG (triglyceride), TC (total cholesterol), HDL (high-density lipoprotein cholesterol), LDL (low-density lipoprotein cholesterol), VLDL (very low-density lipoprotein cholesterol) and fasting blood glucose. The performance of 50 different machine learning models was compared using 5-fold cross validation. Among them, the unbalanced random forest model used in the embodiment of the present invention showed the best performance for this clinical problem. The unbalanced random forest model is specially designed to process unbalanced result data and is very suitable for the clinical problems in this study.

[0045] To determine the optimal input features for the model, recursive feature elimination (RFE) was used to systematically reduce the number of features from 46 to 12. By comparison, using 12 features produced comparable performance to the model using all features (e.g. Figure 3 Therefore, in the present invention, these 12 key clinical indicators of patients (AFU, GPDA, AGE, AFP, INR, ALT, E, B, ALP, ADA, GGT and BMI) were included in the final model. The AUC of the unbalanced random forest model reached 0.8503 (as shown in Figure 1). Figure 3 as shown in C in the figure).

[0046] The SHAP method was used to interpret this unbalanced random forest model and provide insights for future research. Among the final 12 key clinical indicators of patients, the top 4 important features associated with muscle loss were AFU, GPDA, AGE, and AFP (e.g. Figure 3 as shown in D in the figure).

[0047] (3) Model training.

[0048] First, the 12 key clinical indicators collected from different patients after pretreatment were constructed into a data set after data preprocessing. These different patients included patients with high risk of postoperative muscle loss and patients with low risk of postoperative muscle loss. Each sample in the data set represented multiple key clinical indicators of a patient after pretreatment, and each key clinical indicator of the patient was used as a feature.

[0049] Data preprocessing includes: Missing value processing: For numerical features (such as ALT, AFP and other indicators), the median of the indicator is used for filling, and for categorical features (such as HBV infection status), the most common category is used for filling. Abnormal value processing: The 3 times standard deviation (3σ) principle is used to identify abnormal values, and values ​​that exceed the range of ±3 times the standard deviation of the mean are processed. Values ​​exceeding the upper limit are set to the upper limit value, and values ​​exceeding the lower limit are set to the lower limit value. Feature encoding conversion: Perform one-hot encoding conversion on categorical features, and convert non-numeric features into numerical features. Data standardization: Standardize all numerical features, using the standardization formula: (x-μ) / σ, where μ is the mean and σ is the standard deviation, to unify each feature value into the same scale range.

[0050] The model parameters are set as follows: the number of decision trees is set to 100, the maximum tree depth is automatically adjusted, the feature sampling ratio is set to the square root of the total number of features, and the sample sampling strategy uses balanced sampling. The internal structure of the model is actually a complex and sophisticated decision-making system. Imagine an expert group of 100 experienced doctors, each with his or her own unique diagnostic method, which is similar to the 100 decision trees in the model. Each decision tree is a unique set of decision rules learned through training data.

[0051] Next, the five-fold cross-validation method is used for training. The same parameter settings are used for each fold training, and the training performance indicators of each fold are recorded. For each decision tree, a part of the features of each category contained in all samples in each fold data set is randomly selected as the candidate feature set. Starting from the root node, an optimal segmentation feature and its segmentation point are selected from the candidate feature set according to the Gini impurity as the judgment criterion to split the samples in the data set into two subsets and generate two new child nodes. When patient data enters a decision tree, it goes through a series of "yes / no" judgments, just like a doctor's consultation. For example, the first node may ask: "Is the patient's AFP value greater than 400?" If yes, the data will flow to the right branch; if not, it will flow to the left branch. At each branch point (node), an optimal segmentation feature and segmentation point will be selected.

[0052] This optimal segmentation feature and segmentation point are constructed during training based on the Gini impurity indicator. The calculation formula for Gini impurity is: , Among them, the high-risk ratio and the low-risk ratio are the proportion of patients with high risk of postoperative muscle loss and the proportion of patients with low risk of postoperative muscle loss in each subset obtained after each node splits the data set. Assuming that 60% of the sample data of patients in the input data set corresponds to high risk and 40% to low risk, the Gini impurity = 1 - 0.6²- 0.4² = 0.48. Each decision tree will try various possible split features and split points, and select the split scheme that minimizes the sum of the Gini impurities of each node. Such a split will be better than a random split because it makes the patient risk level in each child node more "pure". After that, this split process will be repeated continuously, and the above split process will be recursively repeated for each new child node until the stopping condition is met. The stopping condition is to reach the preset maximum depth (such as 6 layers) or the number of samples in the node is less than the preset value (such as 5). This design can prevent the model from overfitting the training data.

[0053] After the unbalanced random forest model is trained, when predicting new patients, clinical indicators are first collected, and 12 key indicator data of patients are collected to ensure the integrity and accuracy of the data, and the data collection time and conditions are recorded. Then the collected data is preprocessed and input into the trained unbalanced random forest model. The 12 feature data will be input into 100 decision trees at the same time. Each tree will guide the patient to a leaf node according to its own decision rules and give a prediction result: high risk or low risk. Finally, the proportion of decision trees with high risk prediction results to all decision trees is used as the predicted probability of muscle loss after liver transplantation for liver cancer output by the unbalanced random forest model. For example, among 100 decision trees, 65 decision trees predict high risk and 35 decision trees predict low risk, then the final predicted probability of high risk of muscle loss is 0.65.

[0054] This integrated decision-making method has several important advantages: by allowing 100 decision trees to vote together, the impact of misjudgment that a single tree may have is reduced. Different trees may focus on different combinations of features, which improves the robustness of the model. The final probability output provides confidence information for the prediction, rather than a simple binary classification result. It is particularly noteworthy that this unbalanced random forest model also handles the common class imbalance problem in medical data through an unbalanced sampling strategy. When training each decision tree, it ensures that it sees enough minority class samples (high-risk cases) so that even if there are fewer high-risk cases in the actual data, the model can accurately identify them. This complex and sophisticated prediction mechanism ultimately achieved an AUC value of 0.8503, proving its reliability in clinical prediction. More importantly, its prediction results show a significant correlation with the actual prognosis of patients, which is the best proof of its clinical value.

[0055] S2, risk classification is performed based on the predicted probability, and the patient's postoperative muscle loss level is determined as high risk, medium risk, or low risk.

[0056] In the embodiment, the first risk threshold is set to 0.7, and the second risk threshold is set to 0.3.

[0057] If the predicted probability was ≥0.7, the level of postoperative muscle loss was judged as high risk; If 0.3≤prediction probability<0.7, the level of postoperative muscle loss was judged as medium risk; If the predicted probability was <0.3, the level of postoperative muscle loss was judged as low risk.

[0058] S3, construct a decision tree model to stratify patient prognosis and divide patients into high-risk group, medium-risk group and low-risk group.

[0059] In the embodiment, in order to stratify the prognosis of patients using simple preoperative blood tests and preoperative sarcopenia status, the predicted probability of the model was combined with the Hangzhou standard, and a simple decision tree model (such as Figure 3 For SMI≤43.6cm 2 / m 2 Patients with SMI>43.6cm were directly classified as high-risk group. 2 / m 2 Patients with a high risk of muscle loss were classified as intermediate risk. 2 / m 2 Patients with moderate or low risk of muscle loss were classified as low-risk group. This decision tree model is both easy to use and clinically effective.

[0060] S4. Survival rate prediction was performed by combining liver transplantation criteria and patient prognostic stratification results.

[0061] In the embodiment, the Hangzhou liver transplantation standard is included, which is mainly used to evaluate whether liver cancer patients are suitable for liver transplantation. The Hangzhou standard is more relaxed than the Milan standard and expands the scope of indications, but still maintains a good prognostic effect. According to the results of stratifying the prognosis of patients by the decision tree model, combined with the Hangzhou liver transplantation standard, the survival rate of patients who meet the Hangzhou liver transplantation standard is predicted (such as Figure 3 As shown in F, the blue curve is the high-risk group, the red curve is the medium-risk group, and the green curve is the low-risk group). P < 0.0001 indicates that there are significant differences in the possibility of survival among the three groups. If the patient is in the high-risk group, the 1-year survival rate after surgery is 71.5%, the 3-year survival rate after surgery is 53.2%, and the 5-year survival rate after surgery is 36.3%. If the patient is in the medium-risk group, the 1-year survival rate after surgery is 86.4%, the 3-year survival rate after surgery is 74.6%, and the 5-year survival rate after surgery is 63.9%. If the patient is in the low-risk group, the 1-year survival rate after surgery is 92.3%, the 3-year survival rate after surgery is 85.9%, and the 5-year survival rate after surgery is 75.7%. The above survival rates are calculated by analyzing the historical data of liver transplant patients in a large-scale retrospective study and calculating the survival rates of patients with different prognostic stratifications based on postoperative follow-up records. The survival rate is calculated by the statistical method of Kaplan-Meier survival analysis, which fully considers the various changes in the patient's survival status from surgery to the end of follow-up. Therefore, it is scientific and reliable and can provide an important reference for clinical decision-making.

[0062] In summary, the method for predicting muscle loss after liver transplantation for liver cancer provided by the embodiment of the present invention can accurately predict the level of muscle loss after liver transplantation for liver cancer, which helps clinical physicians to formulate personalized treatment plans and scientific management after transplantation, thereby providing patients with more personalized and efficient medical services, and has important clinical application value and scientific research significance.

[0063] Based on the same inventive concept, Figure 4 As shown, the embodiment of the present invention further provides a device 400 for predicting muscle loss after liver transplantation for liver cancer, comprising: a model prediction module 410, a risk classification module 420, a prognosis stratification module 430 and a survival prediction module 440.

[0064] The model prediction module 410 is used to construct an unbalanced random forest model including several decision trees, and independently predict several high-risk or low-risk binary classification prediction results based on multiple pre-processed patient clinical key indicators input by the several decision trees, and the proportion of decision trees with high-risk prediction results to all decision trees is used as the predicted probability of muscle loss after liver transplantation for liver cancer, which is the final output of the unbalanced random forest model.

[0065] The risk classification module 420 is used to judge the postoperative muscle loss level as high risk if the predicted probability is greater than or equal to a first risk threshold, to judge the postoperative muscle loss level as medium risk if the predicted probability is less than the first risk threshold and greater than or equal to a second risk threshold, and to judge the postoperative muscle loss level as low risk if the predicted probability is less than the second risk threshold, wherein the first risk threshold is greater than the second risk threshold.

[0066] The prognosis stratification module 430 is used to construct a decision tree model to stratify patient prognosis, wherein patients whose SMI is less than or equal to the stratification threshold are directly classified as a high-risk group, patients whose SMI is greater than the stratification threshold and whose muscle loss level is judged as high risk are classified as a medium-risk group, and patients whose SMI is greater than the stratification threshold and whose muscle loss level is judged as medium risk or low risk are classified as a low-risk group.

[0067] The survival prediction module 440 is used to predict the survival rate by combining the liver transplantation criteria and the patient prognosis stratification results. For patients who meet the liver transplantation criteria, if the patient is in the high-risk group, the 1-year survival rate after surgery is 71.5%, the 3-year survival rate after surgery is 53.2%, and the 5-year survival rate after surgery is 36.3%; if the patient is in the medium-risk group, the 1-year survival rate after surgery is 86.4%, the 3-year survival rate after surgery is 74.6%, and the 5-year survival rate after surgery is 63.9%; if the patient is in the low-risk group, the 1-year survival rate after surgery is 92.3%, the 3-year survival rate after surgery is 85.9%, and the 5-year survival rate after surgery is 75.7%.

[0068] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting muscle loss after liver transplantation for liver cancer, characterized in that: The following steps are involved: An unbalanced random forest model including several decision trees is constructed. Several high-risk or low-risk binary classification prediction results are obtained by independently predicting multiple patient clinical key indicators after input preprocessing through several decision trees. The proportion of decision trees with high-risk prediction results to all decision trees is used as the predicted probability of muscle loss after liver transplantation for liver cancer, which is the final output of the unbalanced random forest model. If the predicted probability is greater than or equal to the first risk threshold, the postoperative muscle loss level is determined to be high risk; if the predicted probability is less than the first risk threshold and greater than or equal to the second risk threshold, the postoperative muscle loss level is determined to be medium risk; if the predicted probability is less than the second risk threshold, the postoperative muscle loss level is determined to be low risk, wherein the first risk threshold is greater than the second risk threshold; A decision tree model was constructed to stratify the patient prognosis, in which patients with SMI less than or equal to the stratification threshold were directly classified into the high-risk group, patients with SMI greater than the stratification threshold and muscle loss level judged as high risk were classified into the medium-risk group, and patients with SMI greater than the stratification threshold and muscle loss level judged as medium risk or low risk were classified into the low-risk group; Survival rate was predicted by combining liver transplantation criteria and patient prognostic stratification results.

2. The method for predicting muscle loss after liver transplantation for liver cancer according to claim 1, characterized in that: A number of key clinical indicators for patients were screened from a wide range of clinical indicators for patients using the recursive feature elimination method. The key clinical indicators for patients included α-L-fucosidase, glycocholate dehydrogenase, age, alpha-fetoprotein, international normalized ratio, alanine aminotransferase, eosinophil count, basophil count, alkaline phosphatase, adenosine deaminase, glutamyl transpeptidase, and body mass index in the peripheral blood of patients before surgery.

3. The method for predicting muscle loss after liver transplantation for liver cancer according to claim 1, characterized in that: The preprocessing of multiple key clinical indicators of patients included missing value processing, outlier processing, feature coding conversion and data standardization.

4. The method for predicting muscle loss after liver transplantation for liver cancer according to claim 1, characterized in that: In the unbalanced random forest model, the decision tree construction process includes: A plurality of key clinical indicators of patients after pretreatment collected from different patients are constructed into a data set, wherein the different patients include patients with high risk of postoperative muscle loss and patients with low risk of postoperative muscle loss. Each sample in the data set represents a plurality of key clinical indicators of patients after pretreatment of a patient, and each key clinical indicator of the patient is used as a feature; For each decision tree, a part of the features of each category contained in all samples in the data set is randomly selected as the candidate feature set. Starting from the root node, an optimal segmentation feature and its segmentation point are selected from the candidate feature set according to the Gini impurity as the judgment criterion to split the samples in the data set into two subsets and generate two new child nodes. After that, the above segmentation process is recursively repeated for each new child node until the stopping condition is met. The stopping condition is that the preset maximum depth is reached or the number of samples in the node is less than the preset value.

5. The method for predicting muscle loss after liver transplantation for liver cancer according to claim 4, characterized in that: The calculation formula of Gini impurity is: ， Among them, the high-risk proportion and the low-risk proportion are the proportion of patients with high risk of postoperative muscle loss and the proportion of patients with low risk of postoperative muscle loss in each subset obtained after each node splits the data set. Each decision tree will try various possible split features and split points, and select the split scheme that can minimize the sum of the Gini impurities of each node.

6. The method for predicting muscle loss after liver transplantation for liver cancer according to claim 1, characterized in that: The first risk threshold is set to 0.7, and the second risk threshold is set to 0.

3.

7. The method for predicting muscle loss after liver transplantation for liver cancer according to claim 1, characterized in that: The stratification threshold is set to 43.6cm 2 / m 2 .

8. A device for predicting muscle loss after liver transplantation for liver cancer, implemented by the method for predicting muscle loss after liver transplantation for liver cancer according to any one of claims 1 to 7, characterized in that: include: Model prediction module, risk classification module, prognosis stratification module and survival prediction module; The model prediction module is used to construct an unbalanced random forest model including a plurality of decision trees, and independently predict a plurality of high-risk or low-risk binary classification prediction results based on a plurality of pre-processed clinical key indicators of the inputted decision trees, and the proportion of the decision trees with the prediction results of high risk to all the decision trees is used as the prediction probability of muscle loss after liver transplantation for liver cancer, which is finally output by the unbalanced random forest model; The risk classification module is used to determine the postoperative muscle loss level as high risk if the predicted probability is greater than or equal to a first risk threshold, determine the postoperative muscle loss level as medium risk if the predicted probability is less than the first risk threshold and greater than or equal to a second risk threshold, and determine the postoperative muscle loss level as low risk if the predicted probability is less than the second risk threshold, wherein the first risk threshold is greater than the second risk threshold; The prognostic stratification module is used to construct a decision tree model to stratify the patient prognosis, wherein patients whose SMI is less than or equal to the stratification threshold are directly classified as a high-risk group, patients whose SMI is greater than the stratification threshold and whose muscle loss level is determined to be high risk are classified as a medium-risk group, and patients whose SMI is greater than the stratification threshold and whose muscle loss level is determined to be medium risk or low risk are classified as a low-risk group; The survival prediction module is used to predict the survival rate by combining liver transplantation criteria and patient prognosis stratification results.

Images