Predictive Value of the Total Bilirubin and CA50 Screened Based on Machine Learning for Recurrence of Bladder Cancer Patients

Introduction

Bladder cancer is the 10th most common cancer in the world, with approximately 573,000 new cases and 213,000 deaths every year. Therefore, bladder cancer is the 9th major cause of cancer deaths.¹ The incidence of bladder cancer is closely related to tobacco smoking, occupational exposure, bladder related diseases, diet and genetic factors. As a common malignant tumor in the urinary system, despite the optimal treatment for patients with bladder cancer, its high recurrence rate is still a nodus in clinical practice.² Although a good deal of research and progress have been made on new molecular markers used to monitor the occurrence and development of bladder cancer,^3,4 cystoscopy is still the preferred method to detect disease progress due to factors such as sensitivity and specificity. For patients undergoing total cystectomy, postoperative follow-up can only rely on imaging examinations,^5,6 which not only imposes an economic burden on the patients but also makes it difficult to detect some small lesions in the early stages. Therefore, there is an urgent need to develop a relatively objective and cost-effective high-sensitivity biomarker for detecting bladder tumor recurrence.

Clinical chemical testing and fluid analysis are the main screening methods in clinical laboratories.⁷ The change of each test index can be interpreted as a relationship with diseases, for instance, white blood cell, neutrophil, and C-reactive protein, which are considered to be closely related to inflammation.^8–10 Studies have shown that CA125, AFP, CEA, and NSE are tumor markers for ovarian cancer,¹¹ liver cancer,¹² gastric cancer,¹³ and lung cancer,¹⁴ respectively. In addition, the physiological significance of the alteration in ions, such as potassium ions, calcium ions, sodium ions, and chloride ions is complex.¹⁵ However, analyzing these indicators individually may lead to biases and limitations in the judgment of clinical diseases.¹⁶ Therefore, we urgently need to adopt various effective ways to process multidimensional datasets and effectively evaluate a large number of variables.

In the era of rapid computer development, the field of artificial intelligence (AI) has shifted from primarily theoretical research to real-world applications.¹⁷ The application of AI is not only reflected in scientific and research scenarios, but also closely related to our daily activities, such as social networks, smart devices and autopilot mode.^18,19 Machine learning (ML) belongs to AI, which can improve predictions by detecting high-dimensional and potential nonlinear effects of variables.^20,21 Tree based models, such as decision tree (DT), random forest (RF), adaptive boosting (AdaBoost), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost), have been widely applied in medical research.^22–24 However, the use of ML modalities has not been explored to a large extent. The few prior reports on ML technology in bladder cancer prediction focus on diagnosis and overall survival,^15,25,26 and there is no research on combining ML algorithms with clinical laboratory data to predict the recurrence of bladder cancer.

Herein, we collected clinical laboratory data of 345 bladder cancer patients from the Affiliated Hospital of Nantong University, mainly including biochemical markers. We explored the relationship between recurrence of bladder cancer and clinical variables through five different ML methods, including DT, RF, AdaBoost, GBM and XGBoost. In addition, the best two biomarkers that can predict the recurrence of bladder cancer were obtained by combining each algorithm, and the area under the area of receiving operating characteristic curve (AUC) was calculated to evaluate the performance of the model.

Methods

Study Population

This retrospective study enrolled the data of patients diagnosed with bladder cancer in the Affiliated Hospital of Nantong University from January 2012 to December 2017. According to some irresistible factors, the exclusion criteria for patients are as follows: 1. With incomplete data more than 20%; 2. With previous or coexisting cancers; 3. Without available follow-up data. Finally, we recruited a total of 345 eligible patients to participate in this study. In order to ensure a minimum follow-up of 5 years for patients, our final follow-up date was December 2023. This study followed the ethical principles in the Declaration of Helsinki and was approved by the Ethics Committee of the Affiliated Hospital of Nantong University. The specific workflow of this study is shown in Figure 1.

Figure 1 The schematic diagram of the overall workflow.

Data Collection

The data of this study are derived from 34 clinical parameters (Table 1) of bladder cancer patients at the first diagnosis: Age, White blood cell (WBC), Neutrophil count (NEU), Lymphocyte count (LYM), Mononuclear cell count (MONO), Eosinophil count (EO), Basophil cell count (BASO), Red blood cell (RBC), Hematocrit (HCT), Packed cell volume (PCV), Mean corpuscular volume (MCV), Mean corpuscular hemoglobin (MCH), Mean corpuscular hemoglobin concentration (MCHC), Red blood cell distribution width-SD (RDW-SD), Red blood cell distribution width-CV (RDW-CV), Platelet count (PLT), Platelet distribution width (PDW), Mean platelet volume (MPV), Thrombocytocrit (PCT), Total protein (TP), Albumin (ALB), Globulin (GLO), Total bilirubin (TBIL), Creatinine (CREA), Prealbumin (PAB), Kalium (K), Natrium (Na), Chlorine (CL), Cytokeratin 19 fragment antigen21-1 (CYFRA21-1), Carbohydrate antigen 50 (CA50), Carcinoembryonic antigen (CEA), Prothrombin time (PT), Fibrinogen (FIB), D-dimer (DD). The missing value rate of all variables is below 20%. According to the complete queue, use the simple mean interpolation method to interpolate missing values.

Table 1 Characteristics of Included Patients (345)

Machine Learning Models

In this study, we applied five supervised ensemble based ML algorithms: DT, RF, AdaBoost, GBM, and XGBoost. The effectiveness of all ML algorithms was evaluated using the Python programming language. Using Python software for programming, using Anaconda software to install libraries and packages. Basic data processing was performed through Python libraries such as Pandas and Numpy. The four ML methods DT, RF, AdaBoost, and GBM were implemented using the Sklearn package, while XGBoost used its own dedicated software package. According to the Python function train_ Test_ Split randomly divides the samples into training and testing sets, and each ML method matches the testing group based on the results of the training group, and then compares the accuracy of each ML algorithm. Finally, determine the importance of each variable in each ML algorithm through “feature_importance_”. The relevant parameters are as follows. For DT, the initial value of tree depth was set to 1–10 with a step size of 1. The kernel of this model was set to entropy or gini. For RF, the initial value of the tree number was set to 100 and increased by 100 until 500. The kernel of the model was set to entropy or gini. For AdaBoost, the step size was 0.01. For GBM, the step size was 0.005. For XGBoost, the initial value of eta was 0.01 to 0.2 with a step size of 0.05. The parameters of ML were adjusted through the training and validation groups of the dataset.

Statistical Analysis

We used IBM SPSS statistical software (Version 25.0) to draw the receiver operating characteristic (ROC) curves and evaluate the predictive effects of various clinical variables using the area under the receiver operating characteristic curve (AUC). The independent-sample t-test was employed for descriptive statistical analysis of continuous variables between the two groups. The value of p<0.05 was considered statistically significant.

Results

The correlation between each variable and bladder cancer recurrence can be determined by heat map (Figure 2). Among them, LYM, EO, CYFRA21-1 and CA50 were positively correlated with the recurrence of bladder cancer, while HCT, RDW-SD, PDW and TBIL were negatively correlated with the recurrence of bladder cancer.

Figure 2 Correlation between variables. The gradient from green to red represents a gradient from positive to negative correlation.

By comparing the results of five ML algorithms: in the DT model, we found that CA50, TP, TBIL, and EO have relatively high weights (Figure 3A); in RF algorithm, CYFRA21-1, TP, ALB and MONO were in the forefront in predicting the recurrence of bladder cancer (Figure 3B); in contrast, FIB, EO, MONO, and Age were the most prominent in the AdaBoost model (Figure 3C); in addition, the results of the GBM algorithm show that the top four important variables were Age, EO, TP, and D-dimer, respectively (Figure 3D); the risk factors of recurrence of bladder cancer predicted by the final XGBoost model were CA50, FIB, LYM, TBIL and HCT (Figure 3E). Univariate and multivariate analyses were performed to identify the prominently independent prognostic factors in recurrence of bladder cancer (Table 2).

Table 2 Univariate and Multivariate Analyses for Factors Predicting Recurrence in Bladder Cancer

Figure 3 Variable importance of features included in five ML algorithms for predicting the recurrence of bladder cancer. (A) DT. (B) RF. (C)AdaBoost. (D) GBM. (E) XGBoost.

Based on the above results, we can easily found that CA50 and TBIL were significant risk factors for predicting the recurrence of bladder cancer. In order to verify this result, we evaluated the prediction performance of CA50 and TBIL by drawing ROC curves. As shown in the figure, the AUC of CA50 is 0.602, p=0.038 (Figure 4A), while the AUC of TBIL is 0.585, p=0.014 (Figure 4B). We further speculated whether combining these two vital indicators would make predictive performance more meaningful than individual indicator. It was gratifying that the AUC value of the combination of the two variables reached 0.623, p=0.013 (Figure 4C), which was better than the ability of any individual index to predict the recurrence of bladder cancer.

Figure 4 ROC curve analysis. (A) CA50 predicts the AUC value of recurrence. (B) TBIL predicts the AUC value of recurrence. (C) CA50 and TBIL predicts the AUC value of recurrence.

Discussion

Bladder cancer is a highly heterogeneous tumor with at least 40 histological subgroups.²⁷ Compared with other solid tumors, bladder cancer has a higher mutation rate, with heterogeneity of epithelial biology, interstitial fibroblasts, matrix reaction and immune infiltration.²⁸ Therefore, recurrence is a common phenomenon in the clinical treatment of bladder cancer, which has aroused extensive attention. At present, ML algorithms based on AI have made significant progress in the past few years, especially in the medical field, where ML models have been widely used for disease diagnosis and prognosis assessment.^29,30 The ML models can identify high-dimensional nonlinear relationships between detection variables, effectively improving the performance of traditional nomograms.²⁶ As far as we know, our research is the first time to employ clinical laboratory data coupled with ML algorithms to evaluate the recurrence of bladder cancer and screen out the most valuable variables.

Bladder cancer is a fairly common and costly malignant tumor, as it requires lifelong routine monitoring and treatment after diagnosis.³¹ Tumor recurrence is the most common clinical phenomenon after treatment, and once cancer recurrence occurs, it may develop into high grade and bring huge economic burden to patients.³² In this study, we applied ML models to process multidimensional datasets and evaluate a large number of variables to predict bladder cancer recurrence. We explored five ML algorithms based on 34 clinical variables to predict recurrence in patients of bladder cancer. The collected 34 biological indicators cover various clinical parameters in the patient’s age and peripheral blood to achieve more accurate disease prediction. By integrating the results of five ML algorithms, we identified the key factors of the recurrence of bladder cancer, CA50 and TBIL. Our research not only improves prognosis diagnosis, but also greatly reduces medical costs for patients and provides valuable patient information for clinicians.

Bile acids are the final products of cholesterol catabolite, appearing in three forms: TBIL, DBIL, and IDBIL in hepatocytes and peripheral blood.³³ For a long time, serum bilirubin has been used as a marker for liver, gallbladder, and hematological disorders.³⁴ Interestingly, several studies in recent years have shown that high levels of serum bilirubin are beneficial for cancer patients.³⁵ A slightly higher TBIL level indicates a better prognosis for pancreatic cancer,³⁶ breast cancer³⁷ and lung cancer.³⁸ Therefore, the prognostic significance of TBIL is inconsistent in different types of tumors, and further research is needed to determine whether pretreatment level of TBIL is a protective or harmful prognostic factor for bladder cancer. In our study, a higher level of TBIL indicates that patients with bladder cancer are less prone to recurrence and have a better prognosis. To our best knowledge, this is the first study to show the beneficial prognostic value of TBIL in bladder cancer patients.

Although individual serum biomarkers are beneficial prognostic factors in cancer patient studies, single biomarkers may not be sufficient to predict disease progression in clinical settings. Combining other biomarkers together can improve their predictive ability.³⁹ In this study, we included CA50 to improve the predictive ability of recurrence of bladder cancer patients. CA50 is a glycolipid antigen that was first discovered in human colorectal cancer.⁴⁰ With continuous in-depth research, CA50 plays a vital role in cell growth and differentiation, indicating that tumor cells expressing glycolipid antigens may have increased proliferative activity.⁴¹ Previous studies have shown that CA50 is highly expressed in many types of cancers, such as pancreatic cancer, liver cancer, gastric cancer, colorectal cancer, and cervical cancer.⁴² The elevated levels of CA50 in patients with intrahepatic cholangiocarcinoma suggest that the expression of CA50 may help distinguish between benign and malignant bile duct diseases.⁴³ It has been reported that the level of serum tumor marker CA50 can serve as an auxiliary indicator for diagnosing early pancreatic ductal adenocarcinoma, which is of great significance for improving the diagnostic sensitivity and accuracy of early pancreatic ductal adenocarcinoma.⁴⁴ The high levels of serum CA50 detected in patients with colorectal cancer can be used to assist in monitoring disease recurrence and metastasis.⁴⁵ Consistently, our results showed that a high level of CA50 indicates a worse prognosis in patients with bladder cancer. In addition, if TBIL combined with CA50 to predict the recurrence of bladder cancer patients, the results revealed that the combined prediction effect is better than any individual prediction effect.

This study has important clinical application value. Preoperative laboratory examination is a low-cost and minimally invasive blood routine examination. If applied in clinical practice, it can effectively minimize medical costs. Due to objective limitations, there are some shortcomings in this study. Firstly, this is a single center study that includes a small number of cases, so it is necessary to further expand the sample size or conduct multicenter studies in subsequent studies. Moreover, due to the retrospective design was used in this study, there may be confounding factors that may affect the results. Therefore, more detailed prospective studies are needed to further elucidate these relationships. Finally, the data processing in this study only focuses on ML algorithms, which may lead to overfitting of the data. Therefore, traditional statistical analysis such as LASSO regression can be considered in combination to screen variables.

Conclusions

Accurate prognostic prediction tools are helpful for clinical decision-making of bladder cancer. In this retrospective study of 345 bladder cancer patients, ML methods revealed the correlation between clinical parameters and bladder cancer recurrence. Our research showed that the combination of TBIL and CA50 may improve the prognosis prediction of bladder cancer recurrence.