A machine learning prediction model for cancer risk in patients with type 2 diabetes based on clinical tests

1.Introduction

Over the past two decades, the incidence of type 2 diabetes and its complications have rapidly increased worldwide, accounting for 90% of all diabetes cases [1]. As this trend continues, the incidence of cancer among patients with type 2 diabetes has increased significantly [2]. A number of large-scale epidemiological studies and meta-analyses have demonstrated that type 2 diabetes is associated with the development of various cancers, including pancreatic, colorectal, thyroid, esophageal, and gynecological cancers [3, 4, 5, 6, 7]. Moreover, research indicates that this association may be attributed to hyperglycemia and hyperinsulinemia [8]. Since different cancers have different manifestations and characteristics, they are difficult to identify in patients with diabetes through routine screening. If one can comprehensively analyze and fully utilize the relevant laboratory indicators in patients with type 2 diabetes, the risk of cancer in patients with type 2 diabetes can be predicted.

Machine learning, an important field of artificial intelligence, is a general term for a class of algorithms that can learn from numerous datasets to predict the characteristics of new samples and perform required tasks [9]. A previous study by Choudhury [10] demonstrated the effectiveness of artificial intelligence algorithms in evaluating the early diagnosis and prognosis of tumors. Furthermore, artificial intelligence has been shown to assist greatly in cancer screening even when using the clinical data of patients alone [11]. Compared to traditional statistical methods, machine learning performs more objectively in classification and prediction, with better classification results. Currently, there are various machine learning algorithms based on different weak learners, such as logistic regression, naive Bayes, support vector machines, and decision tree classifiers. A previous study found that a sub-discipline of machine learning requiring less user input but more data and processing power has provided great promise in assisting physicians in achieving accurate diagnoses [12]. Furthermore, classifiers such as the multilayer perceptron (MLP), random forest (RF), and decision tree (DT) can accurately predict the survival of patients with cancer [13]. More and more machine learning methods have been applied in cancer diagnosis and prognosis and have shown great potential [14, 15, 16].

This study aimed to establish a machine-learning prediction model by selecting important features and predicting the risk of cancer in patients with type 2 diabetes based on tumor markers and other routine laboratory tests.

2.Materials and methods

2.1Sample information

The data for this study were collected from June 2013 to September 2022 from all hospitalized patients diagnosed with type 2 diabetes at the Shaanxi Provincial People’s Hospital. This study was approved by the Ethics Committee of the Shaanxi Provincial People’s Hospital. All participants were fully informed about the study and provided written informed consent.

Of the 100120 cases, 9352 (9.34%) were patients with cancer. According to statistics, the cancer types mainly included intestinal (20.4%), lung (16.2%), prostate (9.9%), breast (8.4%), liver (8.1%), gastric (6.9%), bladder (5.2%), pancreatic (4.9%), bone (4.6%), lymphoma (4.0%), kidney (2.5%), ovarian (1.8%), esophageal (1.7%), thyroid (1.4%), cervical (0.9%), leukemia (0.8%), glioma (0.7%), thymoma (0.7%), nasopharyngeal (0.3%), and pituitary (0.2%) cancer, comprising 20 different types of cancer (Fig. 1).

Figure 1.

Percentages of 20 different types of cancer in this study.

2.3Feature selection

Feature selection methods are used to remove redundant features from all available features to improve the efficiency of machine-learning models and reduce overfitting. Important feature combinations were effectively identified and validated by combining significance tests (P< 0.05), importance ranking of variables, and the minimum description length (MDL) algorithm [21]. This exploration of the importance of each feature in extracting data variables helps identify the optimal feature combination for model discrimination and improves model accuracy.

The MDL algorithm considers each relevant variable as a simple predictive model and compares and scores these individual models using their respective MDL measures to determine the relevant variables. The formula used was: Si (Modeli, D) = S (Modeli) + S (Ci), where Si (Modeli, D) is the total size obtained after building a simple predictive model using the ith attribute, S (Modeli) is the size of the simple predictive model built using the ith attribute, and S (Ci) is the total size of all the prediction errors after building a simple predictive model using the ith attribute.

Ultimately, the following criteria were used to select important features (laboratory indicators) for modeling: 1) features discovered by the feature selection algorithm are preferred features; 2) selected features should cover as many different aspects of routine checks as possible, such as blood routine, liver and kidney function, and electrolytes; 3) selected features should be as independent of each other as possible, i.e., they should reduce multicollinearity and multivariate correlation as much as possible; 4) the number of features should be proportional to the amount of data; and 5) medical expertise and practical experience should be considered.

Figure 2.

Construction of the machine-learning predictive model.

3.Results

3.3Model performance

By continuously adding feature variables and testing classifier performance, we found that the model classification performance peaked when 107 feature variables were used. The mean AUC values of the top 10 classifiers was 0.770 (0.691–0.852), as shown in Fig. 3.

Figure 3.

Change of ROC AUC curve along with gradually added features.

Table 2

Parameters of the five best classifiers

Classifier	AUC	Accuracy	Sensitivity	Specificity
CatBoost	0.852	0.796	0.761	0.832
Light gradient boosting	0.836	0.771	0.750	0.793
XGBoost	0.830	0.754	0.739	0.770
Gradient boosting	0.829	0.767	0.731	0.805
Random forest	0.833	0.779	0.746	0.813

Based on the principle of decision trees, the five best classifiers were obtained: random forest, CatBoost, light gradient boosting, gradient boosting, and XGBoost models, as listed in Table 2. The AUC values of the five classifiers were all > 0.80 and ranged from 0.829 (gradient boosting) to 0.852 (CatBoost). The precision, sensitivity, and specificity ranges were 0.754–0.796, 0.731–0.761, and 0.770–0.832, respectively (Fig. 4). The CatBoost classifier model achieved the best performance (AUC: 0.852, precision: 0.796, sensitivity: 0.761, and specificity: 0.832) and was considered the optimal model for predicting cancer risk in patients with type 2 diabetes.

Figure 4.

ROC curves of the five superior classifiers.

4.Discussion

Many studies have been conducted on cancer complications in patients with type 2 diabetes. The main mechanisms by which diabetes increases the risk of cancer are as follows: 1) Diabetes and cancer share common risk factors, such as family history (genetics), obesity, high-calorie diet, smoking, and alcohol consumption. 2) High blood glucose in patients with diabetes leads to a decrease in insulin sensitivity, compensatory hyperinsulinemia, and elevated levels of insulin-like growth factor (IGF). High concentrations of growth factors activate insulin receptors, thereby activating the insulin/IGF axis signaling pathway, which, in turn, activates signaling pathways to promote the proliferation and metastasis of cancer cells. 3) Long-term hyperglycemia provides sufficient energy for cancer cell proliferation, which can cause oxidative stress and promote cancer cell proliferation [17, 18, 19, 20]. Many studies have shown that diabetes increases the risk of liver, breast, throat, endometrial, bladder, and kidney cancers. A large study in the United States demonstrated that the risk and type of cancer in patients with type 2 diabetes varies by sex and organ system [19]. These studies do not completely agree on the results of cancer in different organ systems, and the disparities may be due to differences in patient populations, regions, course of disease, and selected analysis factors.

Currently, early screening for cancer is mainly based on a comprehensive evaluation of medical history, family history, imaging data, and blood tumor markers. Imaging diagnosis still requires a subjective assessment by a doctor, but omissions or misjudgments can occur. Blood tumor markers are one of the most commonly used screening methods for early-stage cancer. However, high levels are only present in some patients, and sensitivity is lacking in patients with low concentrations [21]. The diagnosis of cancer using a single test method has limitations, while comprehensive judgment based on clinical data extraction has great advantages. A previous study by Sharma et al. indicated that including data such as age, blood glucose level, and weight change in the model could predict the cancer risk of patients with diabetes and achieved good results with a sensitivity of 78%, specificity of 80%, and AUC of 0.87 [22]. In addition, a British study showed that models including age, BMI change, smoking, diabetes medication, proton pump inhibitors, hemoglobin, total cholesterol, and other indicators could predict the occurrence of pancreatic cancer in high-risk groups, with a sensitivity of 11%, specificity of 99.7%, and AUC of 0.82 [23]. Moreover, Ben et al. demonstrated that glucose level alone could be used as an indicator of pancreatic cancer risk [24]. We therefore considered that artificial intelligence could be used to extract disease-related data from patients to establish a simple, efficient, and population-adaptable pre-screening model, and subsequently established a prediction model using multiple classifiers and data from patients’ routine clinical tests. This can significantly increase the population participation rate and improve the screening efficiency for the disease, thereby improving patient prognosis.

The present study is similar to Choudhury’s study [10] in that multiple classifiers were used to construct models based on clinical data for early cancer diagnosis. Classifiers include MLP, voted perceptron, Clojure classifier, kernel logistic regression, stochastic gradient descent, AdaBoost, Hoeffding tree, and a primal estimated sub-gradient solver for support vector machines (s-Pegasos) [10]. In Choudhury’s study, the classification accuracy of AdaBoost was 71.29%, which was better than that of the other algorithms. However, the diagnostic efficiency obtained in our study differed because the number and category of indicators entered into the model were different. This study used 261 clinical characteristic variables and 16 machine learning algorithms based on the optimal combination of 107 characteristics to establish a prediction model. Our study used blood tumor markers as the main input, combined with routine blood, liver and kidney function, and other routine laboratory tests to establish a more sensitive and specific prediction model. In our study, the AUC values were all > 0.80, indicating that routine test indicators related to type 2 diabetes and cancer were identified through data mining, and that efficient prediction models could be established using machine learning algorithms.

The 16 machine-learning algorithms used in this study are commonly used in many research projects, and different algorithms have different advantages and disadvantages owing to their different principles. The random forest model performed well in this study because of its randomness, which makes it less prone to overfitting and able to handle high-dimensional input samples without dimensionality reduction, thereby achieving an AUC of 0.833. In addition, boosting algorithms are mainly divided into AdaBoost and gradient-boosting decision trees, with the CatBoost and XGBoost algorithms optimized based on a gradient boosting decision tree and known for their parallel computing and ability to simulate nonlinear effects, achieving high efficiency and accuracy in processing large datasets. These optimized algorithms can be applied in future studies to further improve model accuracy and stability. The specificity indicators of the three models (CatBoost, gradient boosting, and random forest) exceeded 80%, whereas the sensitivity indicators of the top five models did not reach 80%, possibly because of the large number of cancer types and relatively insufficient sample size, which made the model less effective for extracting the characteristic features of these patients. This can be improved in the future by increasing the sample size, thereby enabling the model to extract more typical features for modeling different types of cancer. In summary, using the 107 routine examination indicators selected in this study as variables for modeling enabled us to predict a high-risk population of cancer in patients with type 2 diabetes based on routine examination indicators. This can help in early intervention and prevention of cancer in clinical practice.

5.Conclusion

This study used routine laboratory data from patients with type 2 diabetes to construct a predictive model for cancer risk using 16 machine learning classifiers. The catBoost classifier model displayed good sensitivity and specificity, and could promote early risk screening and cancer prevention, thereby improving patient outcomes. Due to the large number of cancer types, the current number of atients is relatively insufficient to achieve high-efficiency prediction.

Funding

This work was supported by the Key Research and Development Program of Shaanxi Province (2017ZDXM-SF-064), Natural Science Basic Research Project of Shaanxi Province (No.2022JQ-925), and Shaanxi Provincial People’s Hospital Incubation Fund (2021YJY-03).

Supplementary data

The supplementary files are available to download from http://dx.doi.org/10.3233/THC-230385.