A method for establishing an elderly esophageal cancer treatment mode and long-term prognosis model

Abstract

The application relates to the technical field of esophageal cancer treatment, and discloses an establishment method of an old esophageal cancer treatment mode and a long-term prognosis model, which comprises the following steps: step 1, collecting data; step 2, data processing; step 3, establishing a random survival forest model; step 4, accuracy evaluation of the random survival forest model established in step 3; and step 5, applying the random survival forest model to clinical evaluation of the old esophageal cancer treatment mode and the survival prognosis model. The method can establish the old esophageal cancer treatment mode and the long-term prognosis model by using the random survival forest method, can give specific risk ratios and survival curves according to the number and range of lymph nodes cleaned by a surgical plan before surgery, can provide a clinical basis for the surgical treatment of old esophageal cancer patients, and solves the problems that an effective surgical plan cannot be made before surgery, the number of cleaned lymph nodes and the surgical outcome cannot be predicted and the like in the existing clinical treatment process.

Description

Technical Field

[0001] This invention relates to the field of esophageal cancer treatment technology, and discloses a treatment model for elderly esophageal cancer and a method for establishing a long-term prognostic model. Background Technology

[0002] Although there are increasingly more treatment options available for esophageal cancer, such as surgery, chemotherapy, radiotherapy, and immunotherapy, elderly esophageal cancer patients have a significantly higher risk of death after surgery compared to younger patients due to their multiple underlying diseases, poor organ function, and shorter life expectancy. Furthermore, previous literature reports that the perioperative mortality (POM) rate for elderly esophageal cancer patients after surgical treatment ranges from 5.6% to 18.0%. Therefore, any potential surgical or non-surgical complications could be fatal for elderly esophageal cancer patients, and the selection of treatment options for them should be approached with greater caution.

[0003] Esophageal cancer can metastasize early through the complex lymphatic network within the esophageal wall. Therefore, surgical dissection of regional lymph nodes is of significant clinical importance. Scientific and rational lymph node dissection is the most crucial step in radical esophagectomy. However, the optimal number and extent of lymph nodes to be dissected remain controversial. As early as the 1980s, Japanese scholars proposed a negative correlation between cervical and upper mediastinal lymph node metastasis and long-term survival after esophagectomy, thus suggesting extensive three-field lymph node dissection during radical esophagectomy. However, opposing three-field lymph node dissection persists. European and American scholars argue that cervical lymph node metastasis is distant metastasis rather than local lesion. Furthermore, three-field lymph node dissection carries a higher incidence of surgical complications, with significantly higher rates of anastomotic leakage and vocal cord paralysis compared to two-field lymph node dissection. A single-center, prospective, randomized controlled phase III clinical trial in my country also indicated that three-field lymph node dissection did not significantly improve long-term survival in patients with squamous cell carcinoma of the middle and lower thoracic esophagus. Furthermore, the number of metastatic lymph nodes removed during esophageal cancer surgery not only affects the accurate staging of (y)pN and the decision-making process for postoperative adjuvant therapy, but is also an important predictor of postoperative recurrence and metastasis. Therefore, determining a reasonable number of lymph nodes to remove is one of the key focuses of current esophageal surgery. The latest NCCN guidelines recommend that at least 15 lymph nodes be removed during esophageal cancer surgery. According to a retrospective study by the International Multicenter Esophageal Cancer Collaboration Group, among 3859 patients with esophageal or esophagogastric junction adenocarcinoma who underwent neoadjuvant therapy, the long-term prognostic benefit threshold for patients with ypN0 stage was 25 lymph nodes, while the threshold for ypN+ stage patients reached 30 lymph nodes. However, elderly esophageal cancer patients often have poor nutritional status and weak physiological reserves. Extensive lymph node removal can further increase surgical trauma and the risk of postoperative complications such as recurrent laryngeal nerve injury, esophagobronchial fistula, chylothorax, and bleeding, thereby reducing the long-term survival rate of patients. Paulus et al. retrospectively analyzed data from elderly esophageal cancer patients who underwent surgery at their center between 2000 and 2012. The incidence of complications involving the heart, lungs, and surgical site was 24%, 27%, and 6%, respectively. Anastomotic leakage occurred in 18% of elderly patients, and 3% required reoperation due to postoperative bleeding. A 2018 study by Bakhos also confirmed these findings, showing that the perioperative mortality risk at 30 days (5.6% vs. 3.4%, P < 0.001) and 90 days (11.4% vs. 7.3%, P < 0.001) was significantly higher in 16,388 elderly esophageal cancer patients compared to younger patients. Therefore, for elderly esophageal cancer patients, preoperative surgical planning, the extent and number of lymph node dissections, and prediction of surgical outcomes are urgent clinical issues that need to be addressed.

[0004] With the advancement of computer technology, machine learning has been widely applied in clinical practice and has played a significant role in the diagnosis and treatment of many diseases. Therefore, there is an urgent need to design a method that can effectively apply computer technology to the treatment of elderly esophageal cancer patients, such as how to formulate surgical plans before surgery, determine the scope and number of lymph node dissections, predict surgical outcomes, and establish a long-term prognostic model for elderly esophageal cancer, in order to solve the problems existing in the above-mentioned technologies. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method for establishing a treatment model and long-term prognosis model for elderly esophageal cancer. This method offers advantages such as providing specific risk ratios and survival curves based on the number of lymph nodes to be removed preoperatively, personalized analysis of risk factors affecting long-term survival in elderly patients and their specific weights, and providing clinical evidence for precise surgical treatment of elderly esophageal cancer patients. It solves problems existing in current clinical treatment processes, such as the inability to formulate effective surgical plans preoperatively, predict the extent and number of lymph node removals, and predict surgical outcomes.

[0007] (II) Technical Solution

[0008] To achieve the aforementioned objectives of developing effective surgical plans preoperatively, predicting the extent and number of lymph node dissections, and forecasting surgical outcomes, this invention provides the following technical solution:

[0009] A method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly, including...

[0010] Step 1: Collect data;

[0011] Step 2: Process the data collected in Step 1;

[0012] Step 3: Train the random survival forest model on the training set and validation set processed in Step 2;

[0013] Step 4: Evaluate the accuracy of the random survival forest model established in Step 3;

[0014] Step 5: Apply the random survival forest model validated in Step 4 to clinically evaluate treatment modalities and survival prognosis models for elderly patients with esophageal cancer.

[0015] Preferably, step 2 involves processing the data collected in step 1, including...

[0016] Step 2.1: Identify duplicate cases: Check and remove identical cases from both the training and validation sets;

[0017] Step 2.2: Remove irrelevant cases: Analyze esophageal cancer patient cases to identify cases that are irrelevant to the random survival forest model, and remove incomplete, damaged, or irrelevant cases.

[0018] Step 2.3: Categorical variable encoding: Use hot encoding to encode the categorical variables in the training and validation sets, and use them as input cases to train the random survival forest model.

[0019] Preferably, step 3 involves training the training and validation sets processed in step 2 using a random survival forest model. The process of establishing the random survival forest model includes...

[0020] Step 3.1: Construct a decision tree based on log-rank substatistics;

[0021] Step 3.2: Integrate and aggregate the multiple decision trees constructed in Step 3.1 to obtain the RSF model;

[0022] Step 3.3: Optimize the hyperparameters of the RSF model to obtain the most accurate model after cross-validation.

[0023] Preferably, step 3.1, the process of constructing a decision tree based on log-rank substatistics, includes step 3.1.1: Let there be n training samples in the training set, D = {(xi, ti, δi), i = 1, ..., n};

[0024] Where xi=(xi1,...,xip) is the p-dimensional feature vector of the i-th patient; ti is the survival time; δi∈{0,1} is the event indicator;

[0025] Step 3.1.2: In the training sample D, for the parent node `parent`, candidate split points `j∈{1, ..., p}`, and tangent point `s`, define the individual set and risk set of the two child nodes as follows:

[0026] Dparent(j,s)={i:xi∈D(parent)}

[0027] Dleft(j,s)={i:xi∈Dparent(j,s),xij≤s}

[0028] Dright(j,s)={i:xi∈Dparent(j,s),xij>s}

[0029] Rl(t)={i:xi∈Dleft(j,s),ti≥t}

[0030] Rr(t)={i:xi∈Dright(j,s),ti≥t}

[0031] Where Dparent(j, s) is the set of individuals of the parent node of the candidate split point j and the tangent point s in the training sample D; Dleft(j, s) is the set of individuals of the left child nodes of the candidate split point j and the tangent point s in the training sample D; Dright(j, s) is the set of individuals of the right child nodes of the candidate split point j and the tangent point s in the training sample D; Rl(t) is the left rotation, and Rr(t) is the right rotation;

[0032] Step 3.1.3: Let Λl(t) and Λr(t) be Nelson-Aalen cumulative hazard estimates of Rl(t) and Rr(t):

[0033] Λl(t)=∑ti≤tI(xi∈Rl(ti))δi / |Rl(ti)|

[0034] Λr(t)=∑ti≤tI(xi∈Rr(ti))δi / |Rr(ti)|

[0035] Furthermore, we define the U statistic and the V statistic:

[0036] Ul(t)=Nl(t)-∫0tYl(u)dΛ0(u)

[0037] Ur(t)=Nr(t)-∫0tYr(u)dΛ0(u)

[0038] Vl(t)=Nl(t)+∫0t dΛl(u) / Yl(u)

[0039] Vr(t)=Nr(t)+∫0tdΛr(u) / Yr(u)

[0040] Where Nl(t) = ∑ti≤tδiI(xi∈Rl(ti)) is the cumulative number of events in Rl(t), Yl(t) = |Rl(t)| is the size of the risk set; Nr(t) = ∑ti≤tδiI(xi∈Rr(ti)) is the cumulative number of events in Rr(t), Yr(t) = |Rr(t)| is the size of the risk set; Λ0(t) is the cumulative hazard at the Nelson-Aalen baseline for the entire data D;

[0041] Step 3.1.4: From Step 3.1.3, the log-rank fractional statistic is:

[0042] Wlog-rank(j,s)=Ul(∞) / [Vl(∞)]1 / 2+Ur(∞) / [Vr(∞)]1 / 2

[0043] Solving for (j*, s*) = argmax(j, s)Wlog-rank(j, s) gives the optimal splitting rule for the current node.

[0044] Step 3.1.5: Repeat steps 3.1.1-3.1.4 above until the stopping condition is met, and construct a single survival decision tree T.

[0045] Preferably, step 3.2, which integrates and aggregates the multiple decision trees constructed in step 3.1, includes step 3.2.1: finding the terminal node hx(Ti) of the new individual x = (x1,...,xp) from each decision tree Ti;

[0046] Step 3.2.2: Calculate the Nelson-Aalen cumulative hazard of individual terminal node hx(Ti):

[0047] ΛTi(t|x)=∑xj∈hx(Ti)δjI(tj≤t) / [∑xk∈hx(Ti)I(tk≥tj)]

[0048] Step 3.2.3: Integrate and average the cumulative danger ΛTi(t|x) of the N trees to obtain the Λ(t|x) estimate of the RSF model:

[0049] Λ(t|x)=(1 / N)∑i=1:NΛTi(t|x)

[0050] Based on the Λ(t|x) estimate of RSF, the survival function of patient x in the RSF model is estimated according to the semi-parametric proportional hazard model Λ(t|x,β)=Λ0(t)exp(x'β):

[0051] S(t|x,β)=exp[-Λ(t|x)exp(x'β)]

[0052] Where β is an unknown regression parameter vector.

[0053] Preferably, step 3.3, the process of tuning the hyperparameters of the RSF model, includes:

[0054] Five-fold cross-validation is employed, and the grid hyperparameter space is traversed to evaluate the consistency index and Brier score. The hyperparameter combination that performs optimally on the validation set is then selected. Specifically, this includes...

[0055] Step 3.3.1: Divide the D layer in the SEER data into 5 disjoint subsets {D1,...,D5};

[0056] Step 3.3.2: For each set of hyperparameters θ = {N, max_depth, mtry, ...}:

[0057] 1) Construct the RSFθ model by using the i-th fold Di as the validation set and the remaining D\Di as the training set;

[0058] 2) Calculate the consistency index C-indexθ, Brier score, and Brierθ evaluation index on the validation set Di;

[0059] Step 3.3.3: For each set of hyperparameters θ, calculate the mean and variance of C-indexθ and Brierθ at 5 folds;

[0060] Step 3.3.4: Select the optimal hyperparameter θ* that maximizes the mean of C-indexθ or minimizes the mean of Brierθ.

[0061] Preferably, step 3.3, the process of tuning the hyperparameters of the RSF model, includes:

[0062] The importance of each feature Xj is assessed using the Mean Perturbation Importance (MPI) method.

[0063] MPI(Xj)=(1 / B)∑b=1:B|C(Db',FSRFθ*)-C(Db'j,FSRFθ*)|

[0064] Where Db' is the out-of-bag test sample; Db'j is the version of feature Xj after random permutation on Db' in the b-th repetition; C(·) is the evaluation metric; FSRFθ* is the random survival forest model constructed based on θ*.

[0065] Preferably, step 4 uses a consistency index to evaluate the accuracy of the random survival forest model established in step 3, specifically including...

[0066] Step 4.1: Given a training dataset D = {(Xi, Ti, δi), i = 1, ..., n};

[0067] Where Xi = (Xi1, ..., Xip) is the p-dimensional clinical feature vector of the i-th patient, Ti is the survival time, δi∈{0,1} is the event indicator, and 1 = death, 0 = review;

[0068] Step 4.2: In the N distinct survival decision trees {T1, ..., TN} of the RSF model, for each decision tree Ti targeting individual x, estimate the Nelson-Aalen cumulative baseline hazard function as follows:

[0069] ΛTi(t|x)=Σj∈hx(Ti)δjI(Tj≤t) / [Σk∈hx(Ti)I(Tk≥Tj)]

[0070] Where hx(Ti) is the terminal node to which individual x belongs on Ti;

[0071] Step 4.3: Obtain the cumulative risk estimate of x from the RSF model through average ensemble:

[0072] ΛRSF(tx)=(1 / N)Σi=1:NΛTi(t|x)

[0073] From the semi-parametric proportional risk model:

[0074] Λ(t|x,β)=Λ0(t)exp(x'β)

[0075] The estimated survival function for individual x is:

[0076] S(t|x,βRSF)=exp[-ΛRSF(t|x)exp(x'βRSF)]

[0077] Where βRSF is the vector of regression parameters to be estimated;

[0078] Step 4.4: Adjust the hyperparameters θ = {N, max_depth, mtry, min_node_size, sample_fraction};

[0079] Where N is the number of survival decision trees, max_depth is the maximum tree depth, mtry is the size of the random feature subset when a node splits, min_node_size is the minimum sample size to terminate a split, and sample_fraction is the Bootstrap sampling ratio.

[0080] Preferably, step 4.4, adjusting the hyperparameter θ, includes:

[0081] Step 4.4.1: Using a K-fold cross-validation strategy, evaluate the consistency index C-index of the RSF model under different θ values ​​on the validation set, defined as:

[0082] C=Pr(S(Xi|βRSF(θ))>S(Xj|βRSF(θ))|Ti>Tj)

[0083] Where (Xi, Ti) and (Xj, Tj) are two independent test samples, and S(Xi|βRSF(θ)) and S(Xj|βRSF(θ)) are the predicted survival function values ​​of Xi and Xj based on the RSF model constructed according to parameter θ.

[0084] Step 4.4.2: In the K-fold CV cross-validation strategy, the training set D is randomly divided into K disjoint subsets {D1, ..., DK}. For each fold k = 1, ..., K, the hyperparameter θ is adjusted as follows:

[0085] 1) Let Dk be the validation set, and the rest D\Dk be the training set;

[0086] 2) Construct an RSFθ model on the training set D\Dk to obtain βRSF(θ,k);

[0087] 3) On the validation set Dk, calculate the consistency score for each pair of samples (i, j): ci,j = I(S(Ti|Xi, βRSF(θ,k)) > S(Tj|Xj, βRSF(θ,k))), where Ti > Tj;

[0088] 4) Calculate CVCk of this fold as CVCk = Σi<j ci,j / (nk(nk-1) / 2), where nk is the sample size of Dk. Finally, the C-index of the model under this parameter θ is:

[0089] C(θ) = (1 / K) Σk=1:K CVCk

[0090] Traverse the hyperparameter space θ∈Θ, select θ* that maximizes C(θ), and obtain the optimal RSF model.

[0091] Preferably, the process of adjusting the hyperparameter θ in step 4.4 includes

[0092] Quantify the relative importance contribution of each feature to the prediction through the Mean Perturbation Importance MPI:

[0093] MPI(Xj) = (1 / B) Σb=1:B |C(Db', FRSFθ*) - C(Db'j, FRSFθ*)|

[0094] Where Db' is the Out-of-bag test sample of the b-th repetition; Db'j is the version after randomly permuting the feature Xj on Db'; FRSFθ* is the RSF model constructed according to the optimal parameter θ*; C(·) is the evaluation indicator.

[0095] (III) Beneficial Effects

[0096] Compared with the prior art, the present invention provides a method for establishing a treatment mode and long-term prognosis model for elderly esophageal cancer, which has the following beneficial effects:

[0097] 1. The present invention uses the random survival forest method to establish a treatment mode and long-term prognosis model for elderly esophageal cancer, and constructs an artificial intelligence analysis system for the first time through interdisciplinary integration. This method focuses on key clinical issues such as the range and number of lymph node dissections in esophageal surgery, and provides clinical basis for precise surgical treatment of elderly esophageal cancer patients;

[0098] 2. This method uses the random survival forest model to establish a treatment mode and long-term prognosis model for elderly esophageal cancer. Compared with traditional models, this model can process high-dimensional datasets or missing data, can adaptively handle complex non-linear and interaction effects, and also has the advantages of high computational efficiency and no dependence on specific distribution assumptions, which greatly improves the accuracy and complexity of the artificial intelligence analysis system;

[0099] 3. This invention establishes a random survival forest model and applies it to the treatment model and long-term prognosis model of elderly esophageal cancer. It can not only predict the long-term survival outcome of elderly esophageal cancer patients in a simple way, but also provide specific risk ratios and survival curves based on the number and extent of lymph nodes to be removed in the surgical plan before surgery. At the same time, it can further personalize the analysis of risk factors affecting the long-term survival of elderly patients and their specific weights, opening the "black box" of survival prognosis and laying a solid foundation for its further widespread application in clinical practice. In addition, it introduces computer expertise into medical practice, providing new ideas and methods for solving clinical problems in the future. Attached Figure Description

[0100] Figure 1 This is a flowchart illustrating the establishment of the treatment model and long-term prognosis model for esophageal cancer in the elderly according to the present invention.

[0101] Figure 2 This is a graph showing the importance of overall patient characteristics in the SEER database according to Embodiment 2 of the present invention.

[0102] Figure 3 This is the entry diagram of the machine model analysis interface in Embodiment 2 of the present invention;

[0103] Figure 4 This is a curve graph illustrating the treatment model and long-term prognosis model for elderly esophageal cancer in Embodiment 2 of the present invention, which predicts surgical risk and clinical prognosis based on the characteristics of esophageal cancer patients and the number of lymph nodes dissected.

[0104] Figure 5 This is a diagram illustrating the impact of different clinicopathological features on the survival prognosis of elderly esophageal cancer patients, as shown in Embodiment 2 of the present invention, regarding the treatment model and long-term prognosis model.

[0105] Among them, Figure 1 In this context, SEER Database represents the SEER database; Random forest represents the random forest survival model; Cox proportional model represents the proportional hazards regression model; Hazard Rate represents the hazard rate; patient's clinical information represents patient clinical data; artificial intelligence assistant represents the treatment model and long-term prognosis model for elderly esophageal cancer; and Treatment decision represents the prediction of treatment outcomes. Detailed Implementation

[0106] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0107] Example 1: Please refer to Figure 1 The present invention provides a technical solution:

[0108] A method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly, including...

[0109] Step 1: Collect data

[0110] Case Collection: Esophageal cancer patient case data were downloaded from the SEER database as the training cohort; and esophageal cancer patients meeting the inclusion criteria at Shaanxi Provincial People's Hospital from January 2013 to June 2023 were included as the external validation set for the model. Inclusion criteria: 1) Patients pathologically confirmed with squamous cell esophageal cancer; 2) Patients aged >60 years; 3) Patients who underwent esophagectomy with more than one lymph node dissection during the operation; 4) Patients with a life expectancy of at least 30 days. Exclusion criteria: 1) Patients whose lymph nodes were not identified or did not exist during treatment. Individuals with positive results for regional lymph node dissection; 2) Patients with uncertain or missing tumor size data; 3) Patients with missing or incomplete postoperative survival data; 4) Patients with multifocal esophageal cancer; 5) Patients with a history of malignant tumors; Clinical data to be collected include demographic information (age and sex), pathological characteristics of esophageal cancer (TNM stage, histological type, tumor size, number of regional lymph nodes dissected, number of positive lymph nodes), preoperative imaging data (including tumor size, number of suspicious enlarged lymph nodes), surgical procedure, and postoperative survival time;

[0111] Follow-up methods: This study followed the NCCN guidelines for follow-up of postoperative esophageal cancer patients, utilizing a combination of outpatient and online methods such as telephone and WeChat. Follow-up was recommended every 3 months for the first 1-2 years, every 6 months for the 3rd-5th years, and annually thereafter. Follow-up included medical history taking and physical examination. Imaging examinations such as complete blood count, blood biochemistry (liver and kidney function, protein, tumor markers, etc.), and chest and abdominal CT scans were performed based on the clinical situation. If recurrence or metastasis was suspected, PET-CT, MRI, and bone scans were recommended, and timely radiotherapy and chemotherapy were initiated. The follow-up period was at least >24 months.

[0112] Step 2: Process the data collected in Step 1 (clean the case data):

[0113] Step 2.1: Identify duplicate cases:

[0114] Check if there are duplicate cases in the training set and validation set (in the dataset, training set: test set = 4:1), and remove duplicates to ensure that each case is unique in the entire dataset;

[0115] Step 2.2: Delete irrelevant cases:

[0116] Analyze esophageal cancer patient cases to identify records that are irrelevant to the random survival forest model; because the model itself is used to predict survival rates and to determine the optimal number of lymph node dissections for new patients based on this model, incomplete, damaged, or irrelevant cases will be removed from the case set.

[0117] Step 2.3: Categorical Variable Encoding: The patient case set used for training and validation in this embodiment contains many categorical variables, such as TNM staging (a commonly used clinical method for tumor staging, where T generally refers to the tumor status, N generally refers to the presence and number of lymph node metastases, and M generally refers to the presence or absence of distant metastases). Because the machine learning algorithm in this embodiment requires the input data to be numerical, and in order to clearly represent the relationship between categories and avoid the algorithm incorrectly assuming sequentiality, which could lead to implicit assumptions embedded in the model and cause prediction bias, this embodiment uses hot coding to classify and encode the categorical clinical features (training set and validation set) in this embodiment and use them as input cases to train the random survival forest model.

[0118] Step 3: Train the random survival forest model on the training and validation sets processed in Step 2 to establish the random survival forest model.

[0119] Each decision tree integrated in the random survival forest is constructed by recursively splitting the data according to a specific splitting criterion. This process involves selecting the optimal features and split points to optimally split the selected esophageal cancer cases from the public database according to the chosen splitting criterion. In this embodiment, the purpose of selecting the splitting criterion is to improve the separation of survival curves between different groups. We choose a Log-rank test to measure the significant difference in survival curves between different groups. Based on this, we integrate and aggregate multiple decision trees to reduce overfitting and provide a model that minimizes empirical error for prediction. Finally, we tune the hyperparameters of the random survival forest to obtain the most accurate model after cross-validation. The specific process includes...

[0120] Step 3.1: Construct a decision tree based on log-rank substatistics

[0121] Step 3.1.1: Let there be n training samples in the training set, D = {(xi, ti, δi), i = 1, ..., n}

[0122] Where xi = (xi1, ..., xip) is the p-dimensional feature vector of the i-th patient, including covariates such as age, gender, and stage; ti is the survival time; δi∈{0,1} is the event indicator (1 = death, 0 = review);

[0123] Step 3.1.2: In the training sample D, for the parent node `parent`, candidate split points `j∈{1, ..., p}`, and tangent point `s`, define the individual set and risk set of the two child nodes as follows:

[0124] Dparent(j,s)={i:xi∈D(parent)}

[0125] Dleft(j,s)={i:xi∈Dparent(j,s),xij≤s}

[0126] Dright(j,s)={i:xi∈Dparent(j,s),xij>s}

[0127] Rl(t)={i:xi∈Dleft(j,s),ti≥t}

[0128] Rr(t)={i:xi∈Dright(j,s),ti≥t}

[0129] Where Dparent(j, s) is the set of individuals of the parent node of candidate split point j and tangent point s in training sample D; Dleft(j, s) is the set of individuals of the left child nodes of candidate split point j and tangent point s in training sample D; Dright(j, s) is the set of individuals of the right child nodes of candidate split point j and tangent point s in training sample D; Rl(t) is the left rotation, and Rr(t) is the right rotation.

[0130] Step 3.1.3: Let Λl(t) and Λr(t) be Nelson-Aalen cumulative hazard estimates of Rl(t) and Rr(t):

[0131] Λl(t)=∑ti≤tI(xi∈Rl(ti))δi / |Rl(ti)|

[0132] Λr(t)=∑ti≤tI(xi∈Rr(ti))δi / |Rr(ti)|

[0133] Furthermore, we define the U statistic and the V statistic:

[0134] Ul(t)=Nl(t)-∫0tYl(u)dΛ0(u)

[0135] Ur(t)=Nr(t)-∫0tYr(u)dΛ0(u)

[0136] Vl(t)=Nl(t)+∫0t dΛl(u) / Yl(u)

[0137] Vr(t)=Nr(t)+∫0tdΛr(u) / Yr(u)

[0138] Where Nl(t) = ∑ti≤tδiI(xi∈Rl(ti)) is the cumulative number of events in Rl(t), Yl(t) = |Rl(t)| is the size of the risk set; Nr(t) = ∑ti≤tδiI(xi∈Rr(ti)) is the cumulative number of events in Rr(t), Yr(t) = |Rr(t)| is the size of the risk set; Λ0(t) is the cumulative hazard at the Nelson-Aalen baseline for the entire data D;

[0139] Step 3.1.4: From Step 3.1.3, the log-rank fractional statistic is obtained as follows:

[0140] Wlog-rank(j,s)=Ul(∞) / [Vl(∞)]1 / 2+Ur(∞) / [Vr(∞)]1 / 2

[0141] Solving for (j*, s*) = argmax(j, s)Wlog-rank(j, s) gives the optimal splitting rule for the current node.

[0142] Step 3.1.5: Repeat steps 3.1.1 to 3.1.4 above until the stopping condition (such as minimum number of samples or maximum depth) is met, and construct a single survival decision tree T;

[0143] Step 3.2: Integrate and aggregate the multiple decision trees constructed in Step 3.1.

[0144] To reduce variance, the RSF (Random Survival Forest) model needs to generate N distinct decision trees {T1, ..., TN}; for a new individual x = (x1, ..., xp):

[0145] Step 3.2.1: Find the terminal node hx(Ti) of individual x in each decision tree Ti;

[0146] Step 3.2.2: Calculate the Nelson-Aalen cumulative hazard of individual terminal node hx(Ti):

[0147] ΛTi(t|x)=∑xj∈hx(Ti)δjI(tj≤t) / [∑xk∈hx(Ti)I(tk≥tj)]

[0148] Step 3.2.3: Integrate and average the cumulative danger ΛTi(t|x) of the N trees to obtain the Λ(t|x) estimate of the RSF model:

[0149] Λ(t|x)=(1 / N)∑i=1:NΛTi(t|x)

[0150] Based on the Λ(t|x) estimate of RSF, the survival function of patient x in the RSF model is estimated according to the semi-parametric proportional hazard model Λ(t|x,β)=Λ0(t)exp(x'β):

[0151] S(t|x,β)=exp[-Λ(t|x)exp(x'β)]

[0152] Where β is an unknown regression parameter vector, which can be obtained by Cox partial likelihood estimation or other methods;

[0153] Step 3.3: Tune the hyperparameters of the RSF model to obtain the most accurate model after cross-validation.

[0154] The RSF model has several hyperparameters that need to be tuned:

[0155] —Number of trees N: Usually taken as 100≤N≤1000;

[0156] —Maximum tree depth max_depth;

[0157] —The size of the random feature subset mtry when a node splits (generally mtry = √p);

[0158] —Minimum number of nodes to terminate a split;

[0159] —Bootstrap self-sampling ratio sample_fraction (recommended 0.632≤fraction≤0.8);

[0160] Five-fold cross-validation (5-fold CV) is employed, and the grid hyperparameter space is traversed to evaluate metrics such as the consistency index (C-index) and Brier score. The hyperparameter combination that performs optimally on the validation set is selected. The specific steps are as follows:

[0161] Step 3.3.1: Divide the D layer in the SEER data into 5 disjoint subsets {D1,...,D5};

[0162] Step 3.3.2: For each set of hyperparameters θ = {N, max_depth, mtry, ...}:

[0163] 1) Construct the RSFθ model by using the i-th fold Di as the validation set and the remaining D\Di as the training set;

[0164] 2) Calculate evaluation metrics such as consistency index C-indexθ and Brier score Brierθ on the validation set Di;

[0165] Step 3.3.3: For each set of hyperparameters θ, calculate the mean and variance of C-indexθ and Brierθ at 5 folds;

[0166] Step 3.3.4: Select the optimal hyperparameter θ* that maximizes the mean of C-indexθ or minimizes the mean of Brierθ;

[0167] Furthermore, in step 3.3, the importance of each feature Xj can also be evaluated using the Mean Permutation Importance (MPI):

[0168] MPI(Xj)=(1 / B)∑b=1:B|C(Db',FSRFθ*)-C(Db'j,FSRFθ*)|

[0169] Where Db' is the out-of-bag test sample; Db'j is the version of feature Xj after random permutation on Db' in the b-th repetition; C(·) is the evaluation metric, such as C-index; FSRFθ* is the random survival forest model constructed based on θ*;

[0170] MPI(Xj) actually measures the average change in model prediction performance before and after a random perturbation feature Xj, and can be used to quantify the relative importance contribution of each feature to the prediction.

[0171] Step 4: Evaluate the accuracy of the random survival forest model established in Step 3.

[0172] In this embodiment, we use the consistency index (C-index) to evaluate the accuracy of the random survival forest model. The consistency index assesses the degree of consistency between the survival time order predicted by the survival model and the actual observed order; its value ranges from 0 to 1, where 1 represents complete consistency (the predicted and observed survival times are completely consistent), and 0.5 represents random ordering (no consistency exceeding randomness); the specific process is as follows:

[0173] Step 4.1: Given a training dataset D = {(Xi, Ti, δi), i = 1, ..., n}

[0174] Where Xi = (Xi1, ..., Xip) is the p-dimensional clinical feature vector of the i-th patient, Ti is the survival time, and δi∈{0,1} is the event indicator (1 = death, 0 = review);

[0175] Step 4.2: The RSF model (Random Survival Forest model) is based on the idea of ​​ensemble learning, growing N different survival decision trees {T1, ..., TN}. For each decision tree Ti, the Nelson-Aalen cumulative baseline risk function for individual x is estimated as follows:

[0176] ΛTi(t|x)=Σj∈hx(Ti)δjI(Tj≤t) / [Σk∈hx(Ti)I(Tk≥Tj)]

[0177] Where hx(Ti) is the terminal node to which individual x belongs on Ti;

[0178] Step 4.3: Obtain the cumulative risk estimate of x from the RSF model through average ensemble:

[0179] ΛRSF(tx)=(1 / N)Σi=1:NΛTi(t|x)

[0180] This leads to the semi-parametric proportional risk model:

[0181] Λ(t|x,β)=Λ0(t)exp(x'β)

[0182] We can estimate the survival function of individual x:

[0183] S(t|x,βRSF)=exp[-ΛRSF(t|x)exp(x'βRSF)]

[0184] Where βRSF is the vector of regression parameters to be estimated;

[0185] Step 4.4: To obtain the optimal RSF model, we need to fine-tune a series of hyperparameters θ = {N, max_depth, mtry, min_node_size, sample_fraction}, including the number of trees N, the maximum tree depth max_depth, the size of the random feature subset mtry when splitting a node, the minimum sample size min_node_size to terminate a split, and the bootstrap sampling ratio sample_fraction, etc. The specific process includes...

[0186] Step 4.4.1: We employ a K-fold cross-validation (CV) strategy to evaluate the consistency index C-index of the RSF model under different θ values ​​on the validation set, defined as:

[0187] C=Pr(S(Xi|βRSF(θ))>S(Xj|βRSF(θ))|Ti>Tj)

[0188] where (Xi, Ti) and (Xj, Tj) are two independent test samples, and S(Xi|βRSF(θ)) and S(Xj|βRSF(θ)) are the predicted survival function values for Xi and Xj from the RSF model constructed according to parameter θ;

[0189] Step 4.4.2: The C-index essentially measures the discriminative power of the model for ranking the survival time of pairs of test samples; in K-fold CV, we randomly divide the training set D into K disjoint subsets {D1, ..., DK}, and for each fold k=1, ..., K: the tuning process of the hyperparameter θ includes:

[0190] 1) Let Dk be the validation set, and the remaining D\Dk be the training set;

[0191] 2) Construct the RSFθ model on the training set D\Dk to obtain βRSF(θ, k);

[0192] 3) On the validation set Dk, calculate the concordance score for each pair of samples (i, j): ci,j = I(S(Ti|Xi, βRSF(θ, k))>S(Tj|Xj, βRSF(θ, k))), Ti>Tj;

[0193] 4) Calculate the CVCk of this fold as CVCk=Σi<j ci,j / (nk(nk-1) / 2), where nk is the sample size of Dk. Finally, the C-index of the model under this parameter θ is:

[0194] C(θ)=(1 / K)Σk=1:K CVCk

[0195] We traverse the hyperparameter space θ∈Θ, select θ* that maximizes C(θ), and obtain the optimal RSF model;

[0196] In step 4.4, in addition to the C-index, we can also adopt other evaluation metrics such as the Brier score, and combine with the characteristics of clinical data to comprehensively evaluate the performance of the RSF model through cross-validation or model analysis on the holdout set;

[0197] In addition, in step 4.44, the RSF can also quantify the relative importance contribution of each feature to the prediction through Mean Permutation Importance (MPI):

[0198] MPI(Xj)=(1 / B)Σb=1:B|C(Db',FSRFθ*)-C(Db'j,FSRFθ*)|

[0199] Where Db' is the out-of-bag test sample of the b-th repetition; Db'j is the version after randomly permuting feature Xj on Db'; FSRFθ* is the RSF model constructed based on the optimal parameters θ*; C(·) is the evaluation index, such as C-index;

[0200] MPI(Xj) actually reflects the average change of the model evaluation index before and after the random perturbation Xj, thus measuring the importance of each feature to the survival prediction task.

[0201] Step 5: Apply the random survival forest model validated in Step 4 to clinically evaluate treatment modalities and survival prognosis models for elderly patients with esophageal cancer.

[0202] Based on an externally validated random survival forest model, clinical characteristics (test set) of esophageal cancer patients who did not undergo lymph node dissection were input into the random survival forest. The number of lymph nodes dissected was varied to obtain corresponding 5-year survival curves. In addition, the system generates line graphs of risk coefficients based on different numbers of lymph node dissections, providing data guidance for experts on the extent and scope of lymph node dissection. Simultaneously, the artificial intelligence system analyzes the weights of prognostic factors in detail and provides specific risk ratios. The specific establishment process includes...

[0203] Step 5.1: User Interface Development Details

[0204] - Use the WPF framework for interface development and define interface elements using the XAML markup language;

[0205] - Lay out the main window in MainWindow.xaml, including the title bar, menu bar, toolbar, main area, etc.;

[0206] - Use Panel controls such as Grid / StackPanel to layout the input area and result display area;

[0207] - Place controls such as TextBox and ComboBox in the input area for inputting patient information;

[0208] - Use controls such as DataGrid / ListView to preview the entered patient information;

[0209] - Use TabControl in the results area, with each TabItem page displaying different results;

[0210] - Develop a custom UserControl to draw survival curve charts in TabItems;

[0211] - Write a code-behind file, XAML.cs, to handle control events and data binding;

[0212] - Use DataBinding to bind input controls and internal data models;

[0213] - Implement button click event handling using the ICommand interface;

[0214] Step 5.2: Implementation of the Survival Model Calculation Module

[0215] - Define core data structure classes such as Tree, Node, and Forest;

[0216] - Implement dataset splitting, node growth, and pruning functions for the ID3 algorithm;

[0217] - Calculate the Nelson-Aalen hazard function and Log-rank statistic for survival data;

[0218] - Implement the bootstrap method to construct the out-of-bag estimator;

[0219] - Generate a randomly generated survival forest;

[0220] - Integrate the predictions from individual trees to produce the final survival function estimate;

[0221] - Design cross-validation strategies and optimize hyperparameters such as the number of trees and maximum depth;

[0222] - Improve computational efficiency by using thread pools / task parallel libraries, etc.

[0223] Step 5.3: Data Processing Module Functions

[0224] - Parse the input patient data and map strings / numerical values ​​into internal feature vectors;

[0225] - Fill in missing data and encode categorical data;

[0226] - Construct a sample dataset (DataSet) and call the model to calculate the baseline survival curve;

[0227] - Iterate through the number of lymph nodes removed, construct a new DataSet, and calculate the predicted survival curve;

[0228] - Store the survival curve results as a key-value pair sequence of time-survival rates;

[0229] - Calculate the hazard ratio and its statistical significance level;

[0230] - Unify the encapsulation of result data as attributes of the Report class;

[0231] Step 5.4: Implementation of the report generation module

[0232] - Use the WPF graphics framework's System.Windows.Shapes and System.Windows.Media namespaces to draw graphical elements;

[0233] -Build a Plot control and encapsulate basic methods for drawing line charts, survival curves, etc.

[0234] - In the OnRender method of the Plot control, draw the corresponding graph based on the input data points;

[0235] - Use the Graphics class in GDI+ to draw text, legends, and other elements on a bitmap;

[0236] - Use third-party reporting tools such as ActiveReports to design formatted Word / PDF report templates;

[0237] -Inject the drawn images, risk ratio text, and other data into the report template to generate a report;

[0238] - Use the ReportPreviewControl to display the report;

[0239] - Offers the ability to export to Word / PDF files;

[0240] Step 5.5: Testing, Deployment, and Maintenance

[0241] - Use the NUnit testing framework to write unit test cases;

[0242] - The Mock component is used to test interface interactions and uses data test stubs to test UI interactions;

[0243] - Use interface testing tools to record and replay UI test cases;

[0244] - Configure continuous integration for the project to automatically run tests and generate reports;

[0245] - Use tools such as InstallShield to package and generate an installer;

[0246] - Conduct beta testing with QA and clinical experts, and get feedback on any issues;

[0247] - Collect user feedback, maintain and fix bugs, and release new version updates;

[0248] - Based on new clinical data and needs, retrain the survival model and update the parameters;

[0249] - Introducing the DevOps model to automate the build, test, and release processes.

[0250] Example 2: Please refer to Figure 2-5 To verify the method for establishing the treatment model and long-term prognosis model for elderly esophageal cancer as described in Example 1, and the effectiveness of the obtained treatment model and long-term prognosis model for elderly esophageal cancer, this example is used for verification.

[0251] Step 1: In data collection, this study first collected data from 647 elderly patients with esophageal squamous cell carcinoma who met the criteria in the SEER database. Squamous cell carcinoma accounted for the majority of histological types (87.01%); squamous cell carcinoma with keratosis was the second most common histological type, accounting for 9.58% of cases; 67.54% of esophageal cancer patients were in stage N0, 24.11% in stage N1, 6.64% in stage N2, and 1.70% in stage N3; in terms of T staging, 22.10% of patients were in stage T1, 19.31% in stage T2, 51.93% in stage T3, and 6.64% in stage T4; approximately 41.26% of patients underwent esophagectomy and partial gastrectomy, and 16.84% underwent esophagectomy and gastrectomy. Simultaneously, we collected data from 102 patients with esophageal squamous cell carcinoma who underwent surgery at the Department of Thoracic Surgery of Shaanxi Provincial People's Hospital in January 2013. Their basic information is shown in Table 1 below.

[0252] Table 1: Characteristics of esophageal cancer patients in the SEER database and the Department of Thoracic Surgery, Shaanxi Provincial People's Hospital

[0253]

[0254]

[0255]

[0256] Step Two: Simultaneously, we input the SEER data into the elderly esophageal cancer treatment model and long-term prognosis model; based on the model and patient data, the importance of the overall characteristics is as follows (see below). Figure 2 (as shown);

[0257] exist Figure 2In the Chinese: impact on model output indicates the effect on the model output; Regional nodes examined (1988+) indicates nodes in the examined region; CS tumor size indicates tumor size; Regional nodes positive (1988+) indicates positive regional nodes; Age recode_75-79 years indicates the age range of elderly patients covered; Grade_Moderately differentiated indicates tumor grade; Sex_Female indicates patient sex; SurgPrim Site indicates surgical method; Sum of 34 other features indicates other features.

[0258] (1) In Figure 2 In our study, we identified three key characteristics of overall survival, listed from top to bottom: number of lymph nodes dissected, N stage, and tumor size. We found that as the number of lymph nodes dissected increased, the risk score gradually decreased, while the overall survival of patients prolonged. For N stage, we observed that N0 stage patients had better clinical prognosis than other N stage patients. This suggests that the treatment model and long-term prognosis model for elderly esophageal cancer can effectively predict treatment outcomes for elderly esophageal cancer patients.

[0259] (2) We utilized the treatment model and long-term prognostic model for elderly esophageal cancer constructed in Example 1 (e.g. Figure 3 As shown in the operation interface), under the premise of verifying the accuracy of the treatment model and long-term prognosis model for elderly esophageal cancer, we use AI technology to analyze the surgical risk and prognostic survival of different patients under different lymph node dissection statuses based on their clinical characteristics (e.g., Figure 4 (as shown);

[0260] Among them, Figure 4 In the diagram, the horizontal axis "Survival Probability Under Different Examined Regional Nodes" represents the patient's survival probability under different regional nodes; the vertical axis "Risk Score" represents the risk score; "Patient 1" and "Patient 2" represent Patient 1 and Patient 2, respectively; and "examined regional nodes" represent the lymph nodes in the examined regions.

[0261] (3) Using the treatment model and long-term prognostic model for elderly esophageal cancer constructed in Example 1, we can analyze the risk contribution of different esophageal cancer patients' clinical characteristics to their long-term prognosis in a personalized manner (e.g., Figure 5As shown, Patient 1 and Patient 2 represent Patient 1 and Patient 2, respectively; this provides a concrete basis for clinicians to develop personalized treatment strategies, and also allows patients and their families to have a more intuitive understanding of the risks of surgical treatment.

[0262] As demonstrated by the above verification process, the treatment model and long-term prognostic model for elderly esophageal cancer established by the method described in Example 1 can assess the long-term survival outcomes and prognostic factors of elderly esophageal cancer patients. It also provides specific risk ratios and survival curves based on different lymph node dissection ranges and numbers, and offers personalized analysis of prognostic factors affecting long-term survival and their specific weights. This provides a solid clinical basis for surgeons to accurately formulate surgical plans, provide personalized perioperative management, and conduct long-term follow-up observations. Furthermore, in terms of doctor-patient communication, the artificial intelligence system can visualize and digitally present the surgical risks and survival status of elderly esophageal cancer patients, providing a practical tool for family members to participate in plan formulation and understand medical risks, playing a positive role in building harmonious medical relationships. Finally, this method introduces computer expertise into medical practice, providing new ideas and methods for solving future clinical problems.

[0263] The aforementioned treatment model and long-term prognostic model for elderly esophageal cancer can be used to evaluate the long-term survival outcomes of elderly esophageal cancer patients. Its successful implementation is expected to provide important evidence-based support for surgeons' preoperative decisions, thereby providing individualized treatment plans for each patient, reducing the occurrence of postoperative complications, improving clinical efficacy, and alleviating the burden on patients.

[0264] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0265] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly, characterized in that: include Step 1: Collect data; Step 2: Process the data collected in Step 1; Step 3: Train the random survival forest model on the training set and validation set processed in Step 2; Step 4: Evaluate the accuracy of the random survival forest model established in Step 3; Step 5: Apply the random survival forest model validated in Step 4 to clinically evaluate treatment modalities and survival prognosis models for elderly patients with esophageal cancer.

2. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 1, characterized in that: Step 2 involves processing the data collected in Step 1. Step 2.1: Identify duplicate cases: Check and remove identical cases from both the training and validation sets; Step 2.2: Remove irrelevant cases: Analyze esophageal cancer patient cases to identify cases that are irrelevant to the random survival forest model, and remove incomplete, damaged, or irrelevant cases. Step 2.3: Categorical variable encoding: Use hot encoding to encode the categorical variables in the training and validation sets, and use them as input cases to train the random survival forest model.

3. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 1, characterized in that: Step 3 involves training the random survival forest model on the training and validation sets processed in Step 2. The process of building the random survival forest model includes... Step 3.1: Construct a decision tree based on log-rank substatistics; Step 3.2: Integrate and aggregate the multiple decision trees constructed in Step 3.1 to obtain the RSF model; Step 3.3: Optimize the hyperparameters of the RSF model to obtain the most accurate model after cross-validation.

4. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 3, characterized in that: Step 3.1, the process of constructing a decision tree based on log-rank substatistics, includes: Step 3.1.1: Let there be n training samples in the training set, D = {(xi, ti, δi), i = 1, ..., n}; Where xi=(xi1,...,xip) is the p-dimensional feature vector of the i-th patient; ti is the survival time; δi∈{0,1} is the event indicator; Step 3.1.2: In the training sample D, for the parent node `parent`, candidate split points `j∈{1, ..., p}`, and tangent point `s`, define the individual set and risk set of the two child nodes as follows: Dparent(j,s)={i:xi∈D(parent)} Dleft(j,s)={i:xi∈Dparent(j,s),xij≤s} Dright(j,s)={i:xi∈Dparent(j,s),xij>s} Rl(t)={i:xi∈Dleft(j,s),ti≥t} Rr(t)={i:xi∈Dright(j,s),ti≥t} Where Dparent(j, s) is the set of individuals of the parent node of the candidate split point j and the tangent point s in the training sample D; Dleft(j, s) is the set of individuals of the left child nodes of the candidate split point j and the tangent point s in the training sample D; Dright(j, s) is the set of individuals of the right child nodes of the candidate split point j and the tangent point s in the training sample D; Rl(t) is the left rotation, and Rr(t) is the right rotation; Step 3.1.3: Let Λl(t) and Λr(t) be Nelson-Aalen cumulative hazard estimates of Rl(t) and Rr(t): Λl(t)=∑ti≤tI(xi∈Rl(ti))δi / |Rl(ti)| Λr(t)=∑ti≤tI(xi∈Rr(ti))δi / |Rr(ti)| Furthermore, we define the U statistic and the V statistic: Ul(t)=Nl(t)-∫0tYl(u)dΛ0(u) Ur(t)=Nr(t)-∫0tYr(u)dΛ0(u) Vl(t)=Nl(t)+∫0t dΛl(u) / Yl(u) Vr(t)=Nr(t)+∫0tdΛr(u) / Yr(u) Where Nl(t) = ∑ti≤tδiI(xi∈Rl(ti)) is the cumulative number of events in Rl(t), Yl(t) = |Rl(t)| is the size of the risk set; Nr(t) = ∑ti≤tδiI(xi∈Rr(ti)) is the cumulative number of events in Rr(t), Yr(t) = |Rr(t)| is the size of the risk set; Λ0(t) is the cumulative hazard at the Nelson-Aalen baseline for the entire data D; Step 3.1.4: From Step 3.1.3, the log-rank fractional statistic is: Wlog-rank(j,s)=Ul(∞) / [Vl(∞)]1 / 2+Ur(∞) / [Vr(∞)]1 / 2 Solving for (j*, s*) = argmax(j, s)Wlog-rank(j, s) gives the optimal splitting rule for the current node. Step 3.1.5: Repeat steps 3.1.1-3.1.4 above until the stopping condition is met, and construct a single survival decision tree T.

5. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 3, characterized in that: Step 3.2, the process of integrating and aggregating the multiple decision trees constructed in Step 3.1, includes... Step 3.2.1: Find the terminal node hx(Ti) of the new individual x = (x1,...,xp) from each decision tree Ti; Step 3.2.2: Calculate the Nelson-Aalen cumulative hazard of individual terminal node hx(Ti): ΛTi(t|x)=∑xj∈hx(Ti)δjI(tj≤t) / [∑xk∈hx(Ti)I(tk≥tj)] Step 3.2.3: Integrate and average the cumulative danger ΛTi(t|x) of the N trees to obtain the Λ(t|x) estimate of the RSF model: Λ(t|x)=(1 / N)∑i=1:NΛTi(t|x) Based on the Λ(t|x) estimate of RSF, the survival function of patient x in the RSF model is estimated according to the semi-parametric proportional hazard model Λ(t|x,β)=Λ0(t)exp(x'β): S(t|x,β)=exp[-Λ(t|x)exp(x'β)] Where β is an unknown regression parameter vector.

6. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 3, characterized in that: Step 3.3, the process of tuning the hyperparameters of the RSF model, includes... Five-fold cross-validation is employed, and the grid hyperparameter space is traversed to evaluate the consistency index and Brier score. The hyperparameter combination that performs optimally on the validation set is then selected. Specifically, this includes... Step 3.3.1: Divide the D layer in the SEER data into 5 disjoint subsets {D1,...,D5}; Step 3.3.2: For each set of hyperparameters θ = {N, max_depth, mtry, ...}: 1) Construct the RSFθ model by using the i-th fold Di as the validation set and the remaining D\Di as the training set; 2) Calculate the consistency index C-indexθ, Brier score, and Brierθ evaluation index on the validation set Di; Step 3.3.3: For each set of hyperparameters θ, calculate the mean and variance of C-indexθ and Brierθ at 5 folds; Step 3.3.4: Select the optimal hyperparameter θ* that maximizes the mean of C-indexθ or minimizes the mean of Brierθ.

7. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 3, characterized in that: Step 3.3, the process of tuning the hyperparameters of the RSF model, includes... The importance of each feature Xj is assessed using the Mean Perturbation Importance (MPI) method. MPI(Xj)=(1 / B)∑b=1:B|C(Db',FSRFθ*)-C(Db'j,FSRFθ*)| Where Db' is the out-of-bag test sample; Db'j is the version of feature Xj after random permutation on Db' in the b-th repetition; C(·) is the evaluation metric; FSRFθ* is the random survival forest model constructed based on θ*.

8. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 1, characterized in that: Step 4 uses a consistency index to evaluate the accuracy of the random survival forest model established in Step 3, specifically including... Step 4.1: Given a training dataset D = {(Xi, Ti, δi), i = 1, ..., n}; Where Xi = (Xi1, ..., Xip) is the p-dimensional clinical feature vector of the i-th patient, Ti is the survival time, δi∈{0,1} is the event indicator, and 1 = death, 0 = review; Step 4.2: In the N distinct survival decision trees {T1, ..., TN} of the RSF model, for each decision tree Ti targeting individual x, estimate the Nelson-Aalen cumulative baseline hazard function as follows: ΛTi(t|x)=Σj∈hx(Ti)δjI(Tj≤t) / [Σk∈hx(Ti)I(Tk≥Tj)] Where hx(Ti) is the terminal node to which individual x belongs on Ti; Step 4.3: Obtain the cumulative risk estimate of x from the RSF model through average ensemble: ΛRSF(tx)=(1 / N)Σi=1:NΛTi(t|x) From the semi-parametric proportional risk model: Λ(t|x,β)=Λ0(t)exp(x'β) The estimated survival function for individual x is: S(t|x,βRSF)=exp[-ΛRSF(t|x)exp(x'βRSF)] Where βRSF is the vector of regression parameters to be estimated; Step 4.4: Adjust the hyperparameters θ = {N, max_depth, mtry, min_node_size, sample_fraction}; where N is the number of survival decision trees, max_depth is the maximum tree depth, mtry is the size of the random feature subset at node splitting, min_node_size is the minimum sample size for terminating splitting, and sample_fraction is the Bootstrap sampling ratio.

9. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 8, characterized in that: The process of adjusting the hyperparameter θ in step 4.4 includes Step 4.4.1: Using the K-fold cross-validation strategy, the concordance index C-index of the RSF model under different θ is evaluated on the validation set, which is defined as: C=Pr(S(Xi|βRSF(θ))>S(Xj|βRSF(θ))|Ti>Tj) where (Xi, Ti) and (Xj, Tj) are two independent test samples, S(Xi|βRSF(θ)) and S(Xj|βRSF(θ)) are the predicted survival function values of Xi and Xj from the RSF model constructed according to the parameter θ; Step 4.4.2: In the K-fold CV cross-validation strategy, the training set D is randomly divided into K mutually disjoint subsets {D1, ..., DK}, and for each fold k=1, ..., K, the process of adjusting the hyperparameter θ includes: 1) Set Dk as the validation set, and the remaining D\Dk as the training set; 2) Construct the RSFθ model on the training set D\Dk to obtain βRSF(θ, k); 3) On the validation set Dk, calculate the concordance score for each pair of samples (i, j): ci,j=I(S(Ti|Xi, βRSF(θ, k))>S(Tj|Xj, βRSF(θ, k))), Ti>Tj; 4) Calculate CVCk of this fold as CVCk=Σi<j ci,j / (nk(nk-1) / 2), where nk is the sample size of Dk. Finally, the C-index of the model under this parameter θ is: C(θ)=(1 / K)Σk=1:K CVCk Traverse the hyperparameter space θ∈Θ, select θ* that maximizes C(θ), and obtain the optimal RSF model.

10. The method for establishing a treatment model and long-term prognostic model for esophageal cancer in the elderly as described in claim 8, characterized in that: The process of adjusting the hyperparameter θ in step 4.4 includes Quantify the relative importance contribution of each feature to the prediction through the Mean Perturbation Importance MPI: MPI(Xj)=(1 / B)Σb=1:B|C(Db',FSRFθ*)-C(Db'j,FSRFθ*)| where Db' is the Out-of-bag test sample of the b-th repetition; Db'j is the version after random permutation of feature Xj on Db'; FSRFθ* is the RSF model constructed according to the optimal parameter θ*; C(·) is the evaluation index.

Images