Method for constructing cancer cachexia prognosis index based on creatinine and car

Abstract

The application relates to the technical field of medical systems, and particularly discloses a construction method of a cancer cachexia prognosis composite index CCAR based on a combination of creatinine and CAR, which comprises the following steps: S1, obtaining serum biomarker data of a cancer cachexia patient, including creatinine, C-reactive protein and albumin; S2, calculating a C-reactive protein to albumin ratio CAR=CRP / Albumin; S3, taking CAR and Cr as input features, and constructing a nonlinear composite index CCAR through a random forest machine learning model; and S4, developing a network calculator system; the CCAR index constructed by the application can comprehensively reflect the inflammatory burden, metabolic state and immune level of a cancer cachexia patient, can provide more accurate survival prognosis prediction compared with traditional indexes, and the parameters, such as creatinine, C-reactive protein to albumin ratio, contained in the CCAR index are all common blood examination items in clinical practice, so that the CCAR index is convenient to obtain and can be widely applied in clinical practice.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of medical systems, in particular to a construction method of a cancer cachexia prognosis composite index CCAR based on a combination of creatinine and CAR. BACKGROUND

[0002] Cancer, as a major public health problem with continuously rising incidence and mortality worldwide, its progression and prognosis prediction of patients has always been the core direction of medical research. Studies have shown that systemic inflammation is a key factor in the occurrence and development of cancer, and inflammatory markers such as CRP, albumin, and creatinine have a certain predictive value for cancer prognosis. However, the applicability of traditional inflammation-related prognostic indicators (such as NLR, PLR, CAR, etc.) in different cancer types and patient populations has significant limitations.

[0003] Cancer cachexia, as a common complication of cancer, seriously affects the prognosis of patients. Existing inflammation-related indicators focus on reflecting the inflammatory state of patients, but for cancer cachexia patients, how to synchronously evaluate their metabolic, immune, and nutritional status is still a difficult problem to be solved. Existing research has confirmed that inflammation burden, metabolic abnormalities, and immune status are closely related to the occurrence and progression of cancer cachexia, so developing a prognosis index that can comprehensively reflect these factors has important clinical significance. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides a construction method of a cancer cachexia prognosis composite index CCAR based on a combination of creatinine and CAR, which solves the problems in the background art.

[0005] To achieve the above purpose, the following technical solutions are adopted: A construction method of a cancer cachexia prognosis composite index CCAR based on a combination of creatinine and CAR, comprising the following steps:

[0006] S1, obtaining serum biomarker data of cancer cachexia patients, including creatinine (Cr), C-reactive protein (CRP), and albumin (Albumin);

[0007] S2, calculating the C-reactive protein to albumin ratio CAR=CRP / Albumin;

[0008] S3, taking CAR and creatinine (Cr) as input features, and constructing a nonlinear composite index CCAR through a random forest machine learning model;

[0009] S4, developing a network calculator system:

[0010] The front-end interface receives the Cr, CRP, and Albumin data input by the user;

[0011] The backend server calculates the CAR value and normalizes the CAR and Cr;

[0012] The normalized data is input into a pre-trained random forest model to generate a CCAR value;

[0013] The CCAR value, risk stratification results (low risk: CCAR≤0.56, high risk: CCAR>0.56), and survival rate prediction (1 year / 3 years / 5 years) are output.

[0014] Preferably, the risk stratification threshold CCAR=0.56 is determined by the maximum selected rank statistic in survival analysis, used to distinguish the prognosis risk level of cancer patients with malignant liquid;

[0015] The network calculator front end integrates a dynamic heat map visualization module to map the negative correlation between the CCAR value and the survival probability, where an increase in the CCAR value corresponds to a decrease in the survival probability;

[0016] The heat map uses a bivariate gradient coloring scheme, with the horizontal axis representing the CCAR value and the vertical axis representing the predicted survival time, and the color band gradually changes from green for high survival probability to red for low survival probability.

[0017] Preferably, the calculation formula of the CAR ratio is:

[0018] ;

[0019] where the correction coefficient =0.1 is used to balance the dimensional differences between and to make the value within a reasonable calculation interval.

[0020] Preferably, the unit of the is milligrams per liter, representing the concentration of C-reactive protein in serum; and the unit of the is grams per deciliter, representing the concentration of albumin in serum, both of which are obtained through clinical blood tests.

[0021] Preferably, the CCAR index realizes the following nonlinear mapping relationship through a random forest model:

[0022] ;

[0023] where is a prediction function constructed by a random forest algorithm, which realizes nonlinear feature combination by integrating multiple decision trees; and are the mean and standard deviation of the creatinine in the training set, respectively, and The mean and standard deviation of the training set CAR value, respectively, are used to standardize the input features;

[0024] The normalized features satisfy the normal distribution with a mean of 0 and a standard deviation of 1, ensuring the stability and generalization ability of the random forest model training;

[0025] The prognostic prediction performance of the CCAR index in the validation queue needs to meet:

[0026] AUC≥0.75 (AUC improvement of traditional indicators CAR / NLR / PLR>0.10);

[0027] The KM survival curve log-rank test p<0.001 between the high-risk group and the low-risk group;

[0028] The CCAR index is suitable for the following cancer cachexia types: pancreatic cancer, gastric cancer, colorectal cancer, lung cancer, and each cancer type is verified by queue (KM curve stratification significance <0.05).

[0029] Preferably, the CCAR value is combined with the TNM stage to generate a composite prognostic model, and the risk score formula is:

[0030] ;

[0031] Where the weight coefficient =0.7, =0.3 is determined by fitting the Cox proportional hazards model to maximize the prediction accuracy of the model for survival risk.

[0032] Preferably, the standardization in S3 uses the StandardScaler algorithm to make the input data satisfy the distribution with a mean of 0 and a standard deviation of 1; S3 also includes the following steps before constructing the nonlinear composite index CCAR through the random forest machine learning model:

[0033] LASSO regression is used to rank the importance of the features of cancer cachexia patients, specifically: Collecting multi-dimensional clinical feature data of cancer cachexia patients including creatinine, hemoglobin, and platelets;

[0034] Using LASSO regression model to filter each feature and calculate the coefficient of each feature;

[0035] According to the feature coefficient, creatinine is determined as the key prediction feature;

[0036] The training steps of the random forest model include:

[0037] The collected cancer cachexia patient data is divided into training set and validation set;

[0038] Standardize CAR and creatinine in the training set;

[0039] Set parameters for the random forest model, including the number of trees and the maximum depth;

[0040] Train the model using the training set;

[0041] Evaluate the trained model through the validation set, adjust parameters to optimize model performance;

[0042] Save the optimized random forest model and standardization parameters.

[0043] Preferably, the back-end server processing flow in S4 includes the following steps:

[0044] Receive patient creatinine, C-reactive protein and albumin data sent by the front end through POST request;

[0045] Validate the input data for validity, and prompt to re-enter if the data is out of the normal range; According to the formula of CAR = 0.1 ×0.1 to calculate the CAR value;

[0046] Standardize CAR and creatinine using the mean and standard deviation obtained in the training phase;

[0047] Call the pre-trained random forest model to generate CCAR value;

[0048] Risk stratification according to the size relationship between CCAR value and 0.56, and calculate the 1-year, 3-year and 5-year survival rate;

[0049] Return the calculation results to the front end in JSON format.

[0050] Preferably, the implementation of the dynamic heat map visualization module includes the following steps:

[0051] Collect CCAR values and corresponding survival data of historical cancer cachexia patients;

[0052] Construct a two-dimensional matrix of CCAR values and survival time, and calculate the average survival probability of different CCAR value intervals at each survival time point;

[0053] Determine the horizontal axis of the heat map as CCAR value and the vertical axis as survival time, with color bands gradually changing from green to red, corresponding to high survival probability to low survival probability;

[0054] Based on the CCAR value of the current patient, highlight its corresponding survival probability area in the heat map;

[0055] Display the heat map on the front-end interface to intuitively present the relationship between CCAR value and survival probability.

[0056] Preferably, the CCAR=0.56 risk stratification threshold determination step is as follows:

[0057] Sort the patients in the training cohort by CCAR value from small to large;

[0058] Iterate through all possible split points, calculate the Maxstat statistic corresponding to each split point The statistic is calculated as:

[0059] ;

[0060] Where, is the actual number of deaths in the low-risk group, is the expected number of deaths in the low-risk group, is the actual number of deaths in the high-risk group, is the expected number of deaths in the high-risk group, calculated based on the overall cohort survival curve;

[0061] Select the split point that maximizes the Maxstat statistic as the initial threshold;

[0062] Verify the stability of the initial threshold by the method, and finally determine the risk stratification threshold as 0.56;

[0063] The Maxstat statistic is used to evaluate the discrimination ability of the split point for survival risk, and the larger the statistic value, the more significant the survival difference between the high-risk and low-risk groups at the split point.

[0064] The present application provides a construction method of a cancer cachexia prognosis composite indicator CCAR based on the combination of creatinine and CAR, which has the following beneficial effects:

[0065] 1. Outstanding comprehensive evaluation ability: the CCAR indicator constructed by the present application can comprehensively reflect the inflammatory burden, metabolic state and immune level of cancer cachexia patients, and can provide more accurate survival prognosis prediction compared with traditional indicators.

[0066] 2. Strong clinical applicability: the CCAR indicator contains parameters such as creatinine, C-reactive protein and albumin ratio (CAR), which are common clinical blood test items, and is easy to obtain, facilitating wide clinical application.

[0067] 3. High reliability of the indicator: the selected inflammatory and metabolic parameters (creatinine, C-reactive protein, albumin) are clinically recognized markers that can effectively reflect the patient's systemic inflammatory burden and metabolic level, and have high reliability and universality.

[0068] ​4. Risk stratification precision: CCAR model can make more accurate survival prognosis prediction for cancer cachexia patients, and realize risk stratification through this index, helping clinicians make individualized treatment decisions.

[0069] 5. Convenient and efficient clinical application: The network calculator developed based on CCAR makes the model more convenient and efficient in clinical application, and can quickly provide individualized prognosis evaluation for patients for doctors, providing strong support for treatment decisions.

[0070] 6. Excellent prediction performance: CCAR index has significant prognosis prediction performance in the validation cohort, AUC≥0.75, AUC is improved by more than 0.10 compared with traditional indexes CAR / NLR / PLR, and the KM survival curve log-rank test p<0.001.

[0071] 7. Wide range of applications: CCAR index is applicable to various types of cancer cachexia such as pancreatic cancer, gastric cancer, colorectal cancer, lung cancer, etc., and has been verified by cohort in each type, and the KM curve stratification significance p<0.05.

[0072] 8. Good visualization effect: The dynamic heat map visualization module integrated in the front end of the network calculator can intuitively present the negative correlation between CCAR value and survival probability through color coding, improving clinical interpretability.

[0073] 9. Model optimization and improvement: Through LASSO regression to screen key features, combined with random forest model to build CCAR, to ensure the accuracy and stability of the model, and the risk stratification threshold determination method is scientific, verified by Maxstat statistics and Bootstrap method, to ensure the reliability of the threshold.

[0074] 10. Compound model is more accurate: The compound prognosis model combining CCAR value and TNM stage determines the weight coefficient through Cox proportional hazards model fitting, further improving the accuracy of survival risk prediction. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 A flowchart of a method for constructing a cancer cachexia prognosis composite index CCAR based on creatinine and CAR combination;

[0076] Figure 2 A graph of ranking the importance of features of cancer cachexia patients by LASO regression;

[0077] Figure 3 An AUC curve graph comparing CCAR with other classic indexes in different cohorts;

[0078] Figure 4 A Kaplan-meier survival curve graph of CCAR in different cohorts;

[0079] Figure 5 is the Kaplan-meier survival curve graph of CCAR in various tumors;

[0080] Figure 6 is the flow chart of constructing CCAR index and developing network calculator;

[0081] Figure 7 is the CCAR calculator interface and heat map visualization for survival prediction of cancer cachexia. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0083] As shown in Figures 1-7 , the present application provides a technical solution: a construction method of a cancer cachexia prognosis composite index CCAR based on creatinine and CAR combination, comprising the following steps:

[0084] S1, obtaining serum biomarker data of cancer cachexia patients, including creatinine (Cr), C-reactive protein (CRP) and albumin (Albumin);

[0085] S2, calculating the C-reactive protein to albumin ratio CAR=CRP / Albumin;

[0086] S3, taking CAR and creatinine (Cr) as input features, constructing a nonlinear composite index CCAR through a random forest machine learning model;

[0087] S4, developing a network calculator system:

[0088] The front-end interface receives the Cr, CRP and Albumin data input by the user; the back-end server calculates the CAR value and performs standardization processing on CAR and Cr; the standardized data is input into the pre-trained random forest model to generate the CCAR value; and the CCAR value, risk stratification result, survival rate prediction, 1 year / 3 years / 5 years are output, low risk: CCAR≤0.56, high risk: CCAR>0.56.

[0089] More specifically, the risk stratification threshold CCAR=0.56 was determined using the maximum selection rank statistic in survival analysis to distinguish the prognostic risk level of patients with cancer cachexia; the front end of the online calculator integrates a dynamic heatmap visualization module, which uses color coding to map the negative correlation between CCAR value and survival probability, where an increase in CCAR value corresponds to a decrease in survival probability; the heatmap adopts a bivariate gradient coloring scheme, with the horizontal axis representing CCAR value and the vertical axis representing predicted survival time, and the color band gradually changes from green for high survival probability to red for low survival probability.

[0090] The implementation of the dynamic heatmap visualization module includes the following steps: collecting CCAR values ​​and corresponding survival data of historical cancer cachexia patients; constructing a two-dimensional matrix of CCAR values ​​and survival time, and calculating the average survival probability of different CCAR value ranges at each survival time point; determining that the horizontal axis of the heatmap is the CCAR value and the vertical axis is the survival time, with the color bands gradually changing from green to red, corresponding to high survival probability to low survival probability respectively; highlighting the corresponding survival probability area in the heatmap based on the current patient's CCAR value; and displaying the heatmap on the front-end interface to intuitively present the relationship between CCAR values ​​and survival probability.

[0091] The steps for determining the risk stratification threshold of CCAR=0.56 are as follows: Sort patients in the training cohort according to their CCAR values ​​from smallest to largest; iterate through all possible segmentation points and calculate the corresponding value for each segmentation point. The statistic, with the formula:

[0092] ;

[0093] in, This represents the actual number of deaths in the low-risk group. The expected number of deaths in the low-risk group. This represents the actual number of deaths in the high-risk group. This represents the expected number of deaths in the high-risk group, calculated based on the overall cohort survival curve.

[0094] Choose to The split point with the largest statistical value is used as the initial threshold; through The method verifies the stability of the initial threshold and finally determines the risk stratification threshold to be 0.56; The statistic is used to assess the ability of a cutoff point to distinguish survival risk. The larger the statistic value, the more significant the survival difference between the high-risk and low-risk groups at that cutoff point.

[0095] In this embodiment, Bootstrap method is used to verify the stability of the initial threshold value, and the reliability of the threshold value is evaluated by multiple resampling, and finally the risk stratification threshold value CCAR=0.56 is determined, which is used to distinguish the prognosis risk level of cancer patients with malignant transformation, wherein CCAR≤0.56 is determined as low risk, and CCAR>0.56 is determined as high risk.

[0096] The implementation of the dynamic heat map visualization module is as follows:

[0097] Data collection and preprocessing: collect the CCAR values of historical cancer patients with malignant transformation and the corresponding survival data, including the survival status at different time points, clean and standardize the data to ensure the consistency and availability of the data.

[0098] Two-dimensional matrix construction: a two-dimensional matrix is constructed with CCAR values and survival time as dimensions, the CCAR values are divided into multiple intervals, the average survival probability of each CCAR value interval at different survival time points (such as 1 year, 3 years, 5 years) is calculated, and a probability distribution matrix is formed.

[0099] Heat map parameter definition: the horizontal axis is defined as the CCAR value, covering the complete value range from low to high; the vertical axis is defined as the predicted survival time, including key time nodes such as 1 year, 3 years, and 5 years; the color band scheme adopts a two-variable gradient coloring scheme, with the color band gradually changing from green to red, where green corresponds to a high survival probability area, and red corresponds to a low survival probability area, and the color depth is positively correlated with the survival probability; patient data highlighting is based on the current input CCAR value of the patient, which is automatically located and highlighted in the heat map to display its corresponding survival probability area, and is highlighted by deepening the color or marking the border, etc., to facilitate quick identification by clinicians; the front-end display and interaction integrates the constructed heat map into the network calculator front-end interface, and uses HTML5, CSS3 and JavaScript technologies to realize dynamic rendering, supports mouse hovering to view specific probability values, and other interactive functions, and intuitively presents the negative correlation between CCAR value and survival probability, i.e. the trend of decreasing survival probability with increasing CCAR value.

[0100] More specifically, the calculation formula of the CAR ratio is:

[0101] ;

[0102] Where the correction coefficient =0.1 is used to balance the dimensional difference between and , so that the value of is in a reasonable calculation interval.

[0103] The unit of C-reactive protein concentration in serum is milligrams per liter. Albumin, in units of grams per deciliter, characterizes the concentration of albumin in serum, both obtained through clinical blood tests.

[0104] C-reactive protein concentration in serum: C-reactive protein concentration in serum of patients is obtained through clinical blood tests, and the detection method uses immunoturbidimetry or enzyme-linked immunosorbent assay (ELISA), and the detection result is characterized in units of milligrams per liter (mg / L). The specific operation steps are as follows: 5 mL of fasting venous blood of the patient is collected and placed in an anticoagulant tube; centrifuged at 3000 rpm for 10 minutes to separate the serum; using a fully automatic biochemical analyzer, the CRP concentration is determined based on the immunoturbidimetry principle, and the detection linear range is 0.5-200 mg / L.

[0105] Albumin concentration detection: The same blood sample is collected synchronously, and the serum albumin concentration is detected by bromocresol green (BCG) method, and the result is characterized in units of grams per deciliter (g / dL). The specific implementation steps include: take 100 μL of serum, add 500 μL of BCG reagent, mix well; incubate at 37°C for 5 minutes, and measure the absorbance at 630 nm wavelength; calculate the albumin concentration through the standard curve, and the detection range is 20-60 g / L (i.e. 2.0-6.0 g / dL).

[0106] CAR ratio calculation process: dimensional balance processing, since the unit of CRP is mg / L and the unit of Albumin is g / dL, there is a difference in the dimension of the two , direct division will cause the calculation result to be too large. By introducing a correction coefficient k=0.1 for dimensional standardization, the specific formula is:

[0107] ;

[0108] The coefficient converts the CAR value into a dimensionless index, and the value range is controlled between 0.01-10.0, which meets the numerical readability requirements of clinical indicators.

[0109] Take a patient's test data as an example: CRP=71.4 mg / L; Albumin=32.6 g / dL; substitute the formula: .

[0110] Calculation logic verification: Through 1000 historical sample tests, the rationality of the correction coefficient k=0.1 is verified. Without correction (k=1), the CAR value range is 0.05-25.0, and 90% of the samples are concentrated in 1.0-5.0, with a large numerical span; after using k=0.1, the CAR value range is narrowed to 0.005-2.5, and 90% of the samples are concentrated in 0.1-0.5, which is more in line with the input feature distribution requirements of the machine learning model.

[0111] Clinical application scenarios: For cancer cachexia patients admitted to the emergency department, CRP and Albumin data are obtained simultaneously through bedside testing equipment, and CAR value calculation is completed within 15 minutes to provide the basis for early risk assessment. In outpatient follow-up, indicators are obtained through laboratory biochemical testing, CAR value is calculated, and change trend is tracked to assist in judging disease progression. Since the correction coefficient k=0.1 eliminates the systematic error of different laboratory testing methods (such as the confusion of mg / L and g / L units for CRP detection), the multicenter research data is comparable.

[0112] More specifically, the CCAR index realizes the following nonlinear mapping relationship through the random forest model:

[0113] ;

[0114] Wherein is the prediction function constructed by the random forest algorithm, which realizes the nonlinear feature combination by integrating multiple decision trees; and are the mean and standard deviation of the training set creatinine , respectively, and are the mean and standard deviation of the CAR value of the training set, respectively, for standardizing the input features;

[0115] The standardized features satisfy the normal distribution with a mean of 0 and a standard deviation of 1, ensuring the stability and generalization ability of the random forest model training;

[0116] The prognostic prediction performance of the CCAR index in the validation cohort needs to meet:

[0117] AUC≥0.75 (AUC improvement of more than 0.10 compared with traditional indicators CAR / NLR / PLR); log-rank test p<0.001 for KM survival curves of high-risk group and low-risk group; CCAR index is applicable to the following types of cancer cachexia: pancreatic cancer, gastric cancer, colorectal cancer, lung cancer, and has been verified in each cancer type (KM curve stratification significance <0.05).

[0118] The random forest algorithm is used to construct the CCAR index, which forms a strong learner by integrating 100 decision trees. Each tree uses the Bootstrap method to randomly select 80% of the samples from the training set, and randomly selects 2 features (CAR and creatinine) for optimal segmentation at each node split. The final output is determined by the majority voting mechanism, realizing the nonlinear combination of input features.

[0119] Nonlinear mapping calculation process, standardize the input creatinine (Cr) and CAR value respectively:

[0120] ;

[0121] where, is the mean and standard deviation of creatinine in the training set, is the mean and standard deviation of CAR in the training set.

[0122] The normalized features are input into the random forest model: ; where, is the prediction function of an ensemble of 100 decision trees, the final CCAR value is generated by weighted summation of the conditional judgment of tree nodes and the output of leaf nodes.

[0123] Model parameter optimization, through grid search to determine the optimal parameters: the number of trees (n_estimators) = 100; maximum depth (max_depth) = 10; minimum sample split (min_samples_split) = 20; minimum sample leaf node number (min_samples_leaf) = 10;

[0124] Data standardization process: using StandardScaler algorithm to standardize the CAR and creatinine of the training set and test set, the formula is:

[0125] ;

[0126] Example: a patient with Cr = 128.3 mg / dL, training set , , then:

[0127] ;

[0128] Distribution verification of normalized features: through Shapiro-Wilk test, the W statistics of Cr and CAR after standardization are 0.98 and 0.97 (p>0.05) respectively, which meets the normal distribution assumption.

[0129] Engineering implementation:

[0130] The StandardScaler.fit() method of Python's scikit-learn library is used in the training stage to calculate and , and the parameters are saved through pickle serialization. In the inference stage, load the standardization parameters and real-time standardize the input data through the StandardScaler.transform() method.

[0131] More specifically, the CCAR value is combined with the TNM stage to generate a composite prognostic model, and the risk score formula is:

[0132] ;

[0133] wherein the weight coefficients = 0.7, = 0.3 are determined by fitting a Cox proportional hazards model to maximize the prediction accuracy of the model for survival risk.

[0134] The composite prognostic model linearly combines the CCAR index and the TNM stage, with the formula:

[0135] ;

[0136] wherein: is the composite index calculated by the random forest model; is the numerical result of the TNM stage defined by the International Union Against Cancer (UICC) (e.g., stage I = 1, stage II = 2, stage III = 3, stage IV = 4).

[0137] Variable standardization: After numerical encoding of the TNM stage, the same standardization method as CCAR is used:

[0138] ;

[0139] wherein, , (based on 1200 training set statistics).

[0140] In actual calculation, first standardize the TNM stage, then sum with CCAR weighted:

[0141] ;

[0142] wherein, , are the optimized weight coefficients.

[0143] Weight coefficient determination method, Cox proportional hazards model fitting process: variable input takes CCAR, TNM stage, and patient survival time (months), and outcome event (death = 1, survival = 0) as Cox model input variables; model construction uses partial likelihood function to estimate regression coefficients

[0144] ;

[0145] wherein, is the risk function of an individual at time , is the baseline risk function.

[0146] Weight conversion: standardize the regression coefficients to weight coefficients:

[0147] ;

[0148] Observed , Calculated , Observed , Calculated .

[0149] More specifically, the standardization processing in S3 adopts the StandardScaler algorithm to make the input data satisfy the distribution with a mean of 0 and a standard deviation of 1; before constructing the nonlinear composite index CCAR in S3 through the random forest machine learning model, the following steps are included:

[0150] LASSO regression is used to rank the feature importance of cancer cachexia patients, specifically:

[0151] The multi-dimensional clinical feature data of cancer cachexia patients, including creatinine, hemoglobin, and platelets, are collected; the LASSO regression model is used to filter each feature, and the coefficient of each feature is calculated; according to the feature coefficient, creatinine is determined as the key prediction feature;

[0152] The training steps of the random forest model include: dividing the collected cancer cachexia patient data into training set and validation set; standardizing CAR and creatinine in the training set; setting the parameters of the random forest model, including the number of trees and the maximum depth; using the training set to train the model; evaluating the trained model through the validation set, adjusting the parameters to optimize the model performance; saving the optimized random forest model and standardization parameters.

[0153] LASSO regression model feature screening process:

[0154] Data standardization, Z-Score standardization method is used to preprocess the original features to eliminate the difference in dimension, the formula is:

[0155] ;

[0156] where, is the feature mean, is the feature standard deviation.

[0157] Model construction and parameter optimization: use the LassoCV module of the scikit-learn library to determine the optimal regularization parameter through 10-fold cross-validation, and iteratively calculate the regression coefficients of each feature.

[0158] Feature importance ranking: rank the features according to the absolute value of the regression coefficient, keep the features with non-zero coefficients, and exclude the features with zero coefficients.

[0159] Key feature determination Through LASSO regression analysis, the absolute value of the regression coefficient of creatinine (Cr) was significantly higher than that of other features (such as hemoglobin, platelets, etc.), combined with clinical significance (creatinine is a key indicator of kidney metabolism and nutritional status of the body, and is highly related to metabolic disorders in patients with cachexia), creatinine was determined as one of the core prediction features.

[0160] Steps of building and training the random forest model:

[0161] The collected data of cancer cachexia patients was divided into training set and validation set in the ratio of 8:2, and stratified sampling was used to ensure that the two groups of data were evenly distributed in terms of cancer type (pancreatic cancer, gastric cancer, colorectal cancer, lung cancer, etc.), TNM stage, age and other clinical characteristics.

[0162] The CAR (calculated formula is ) and creatinine (Cr) in the training set were standardized using the StandardScaler algorithm. The processed data met the normal distribution with a mean of 0 and a standard deviation of 1, and the formula is as follows:

[0163] ;

[0164] Among them, , is the mean and standard deviation of the training set CAR, , is the mean and standard deviation of the training set creatinine.

[0165] Model parameter setting and training:

[0166] Core parameter initialization: number of trees (n_estimators): initially set to 200; maximum depth (max_depth): initially set to 20; random feature sampling number (max_features): set to (the number of features is 2, i.e. CAR and Cr); minimum sample split number (min_samples_split): set to 10.

[0167] Model training: using the standardized features of the training set, a random forest model is built through the Bagging strategy, and each decision tree is trained based on a randomly sampled feature subset and sample subset (sampling ratio is 0.8). Finally, the prediction result is generated through ensemble learning.

[0168] Model evaluation and parameter optimization:

[0169] Validation set evaluation, using the validation set to calculate the following indicators: survival analysis AUC value (requirement AUC≥0.75); Kaplan-Meier survival curve log-rank test p value (requirement ); Calibration curve consistency index (C-index).

[0170] Parameter tuning: Grid Search was used to optimize n_estimators (range 100-500), max_depth (range 10-30), and max_features (range 1-2). After each tuning, the model performance was evaluated by 5-fold cross-validation until the AUC on the validation set improved by ≥0.10 and the KM curve showed stratified significance.

[0171] After training, the optimized random forest model and standardized parameters ( , , , The file is saved via pickle serialization in .pkl format to ensure efficiency and stability when called by the network calculator later.

[0172] More specifically, the backend server processing flow in S4 includes the following steps: receiving patient creatinine, C-reactive protein, and albumin data sent by the frontend via POST requests; validating the input data, and prompting for re-entry if the data exceeds the normal range;

[0173] according to The formula is used to calculate the CAR value by multiplying the result by 0.1; the mean and standard deviation obtained during the training phase are used to standardize the CAR and creatinine values; the pre-trained random forest model is called to generate the CCAR value; risk stratification is performed based on the relationship between the CCAR value and 0.56, and the survival rates for 1 year, 3 years, and 5 years are calculated; the calculation results are returned to the front end in JSON format.

[0174] 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 constructing a cancer cachexia prognostic composite index (CCAR) based on a combination of creatinine and CAR, characterized in that, Includes the following steps: S1. Obtain serum biomarker data from patients with cancer cachexia, including Cr, CRP, and Albumin; S2. Calculate the C-reactive protein to albumin ratio: CAR = CRP / Albumin; S3. Using CAR and Cr as input features, construct a nonlinear composite index CCAR through a random forest machine learning model; S4. Develop a network calculator system: The front-end interface receives Cr, CRP, and Albumin data input by the user; The backend server calculates the CAR value and standardizes CAR and Cr. Standardized data is input into a pre-trained random forest model to generate CCAR values; Output CCAR value, risk stratification results (low risk: CCAR≤0.56, high risk: CCAR>0.56), and survival rate prediction (1 year / 3 years / 5 years); The risk stratification threshold CCAR=0.56 was determined by the maximum selection rank statistic in the survival analysis and was used to distinguish the prognostic risk level of patients with cancer cachexia. The network calculator front end integrates a dynamic heatmap visualization module, which uses color coding to map the negative correlation between CCAR value and survival probability, where an increase in CCAR value corresponds to a decrease in survival probability; The heatmap uses a bivariate gradient coloring scheme, with the horizontal axis representing the CCAR value and the vertical axis representing the predicted survival time. The color band gradually changes from green (high survival probability) to red (low survival probability). The formula for calculating the CAR ratio is: ； Where the correction coefficient =0.1, used for balancing and The difference in dimensions makes The value is within a reasonable calculation range; The CCAR index achieves the following non-linear mapping relationship through a random forest model: ； in The prediction function constructed for the random forest algorithm achieves nonlinear feature combination by integrating multiple decision trees; and Training set The mean and standard deviation, and These are the mean and standard deviation of the CAR values ​​in the training set, respectively, used for standardizing the input features; The standardized features satisfy a normal distribution with a mean of 0 and a standard deviation of 1, ensuring the stability and generalization ability of the random forest model during training. The predictive efficacy of the CCAR index in the validation queue must meet the following requirements: AUC ≥ 0.75, with an AUC improvement of > 0.10 compared to traditional metrics CAR / NLR / PLR; The log-rank test of the KM survival curves between the high-risk and low-risk groups showed p < 0.

001. The CCAR index is applicable to the following cancer cachexia types: pancreatic cancer, gastric cancer, colorectal cancer, and lung cancer, and has been validated through cohort studies in each cancer type, with stratified significance of the KM curve. <0.05; The CCAR value is combined with TNM staging to generate a composite prognostic model, and its risk scoring formula is as follows: ； Among them, the weighting coefficient =0.7, =0.3 was determined by fitting the Cox proportional hazards model to maximize the model's accuracy in predicting survival risk; The steps for determining the risk stratification threshold of CCAR=0.56 are as follows: Patients in the training cohort are sorted by CCAR value from smallest to largest; Iterate through all possible split points and calculate the corresponding value for each split point. The statistic, with the formula: ； in, This represents the actual number of deaths in the low-risk group. The expected number of deaths in the low-risk group. This represents the actual number of deaths in the high-risk group. This represents the expected number of deaths in the high-risk group, calculated based on the overall cohort survival curve. Choose to The segment with the largest statistical value is used as the initial threshold. pass The method verifies the stability of the initial threshold and finally determines the risk stratification threshold to be 0.56; The Maxstat statistic is used to assess the ability of a cutoff point to distinguish survival risk. A larger statistic value indicates a more significant difference in survival between the high-risk and low-risk groups at that cutoff point.

2. The method for constructing a cancer cachexia prognostic composite index (CCAR) based on a combination of creatinine and CAR according to claim 1, characterized in that, The The unit is milligrams per liter, representing the concentration of C-reactive protein in serum; The unit is grams per deciliter, which characterizes the concentration of serum albumin. Both are obtained through clinical blood tests.

3. The method for constructing a cancer cachexia prognostic composite index (CCAR) based on a combination of creatinine and CAR according to claim 2, characterized in that, The standardization process in S3 uses the StandardScaler algorithm to ensure that the input data follows a distribution with a mean of 0 and a standard deviation of 1. Before constructing the nonlinear composite index CCAR using the random forest machine learning model in S3, the following steps are also included: LASSO regression was used to rank the importance of features in patients with cancer cachexia. Specifically, multidimensional clinical feature data of patients with cancer cachexia, including creatinine, hemoglobin, and platelets, were collected. The LASSO regression model was used to screen the features and calculate the coefficients of each feature. Creatinine was identified as the key predictive feature based on the characteristic coefficients. The training steps for a random forest model include: The collected data on cancer cachexia patients were divided into a training set and a validation set; Standardize the CAR and creatinine in the training set; Set the parameters of the random forest model, including the number of trees and the maximum depth; Train the model using the training set; The trained model is evaluated using a validation set, and parameters are adjusted to optimize model performance. Save the optimized random forest model and standardized parameters.

4. The method for constructing a cancer cachexia prognostic composite index (CCAR) based on a combination of creatinine and CAR according to claim 3, characterized in that, The backend server processing flow in S4 includes the following steps: Receive patient creatinine, C-reactive protein, and albumin data sent by the front end via POST requests; The input data is validated; if the data exceeds the normal range, the user is prompted to re-enter it. The CAR value is calculated using the formula ×0.1; CAR and creatinine were standardized using the mean and standard deviation obtained during the training phase. Use a pre-trained random forest model to generate CCAR values; Risk stratification was performed based on the relationship between CCAR value and 0.56, and the survival rates for 1 year, 3 years, and 5 years were calculated. The calculation results are returned to the front end in JSON format.

5. The method for constructing a cancer cachexia prognostic composite index (CCAR) based on a combination of creatinine and CAR according to claim 4, characterized in that, The implementation of the dynamic heatmap visualization module includes the following steps: Collect CCAR values ​​and corresponding survival data of historical cancer cachexia patients; Construct a two-dimensional matrix of CCAR value and survival time, and calculate the average survival probability at each survival time point for different CCAR value ranges; The horizontal axis of the heatmap is defined as the CCAR value, and the vertical axis is the survival time. The color bands change from green to red, corresponding to high survival probability to low survival probability, respectively. Based on the patient’s current CCAR value, the corresponding survival probability area is highlighted in the heatmap; The heatmap is displayed on the front-end interface, intuitively showing the relationship between CCAR value and survival probability.

Images