XGBoost Classifier Based on Computed Tomography Radiomics for Prediction of Tumor-Infiltrating CD8+ T-Cells in Patients With Pancreatic Ductal Adenocarcinoma

Abstract

Objectives: This study constructed and validated a machine learning model to predict CD8⁺ tumor-infiltrating lymphocyte expression levels in patients with pancreatic ductal adenocarcinoma (PDAC) using computed tomography (CT) radiomic features.

Materials and methods: In this retrospective study, 184 PDAC patients were randomly assigned to a training dataset (n =137) and validation dataset (n =47). All patients were divided into CD8⁺ T-high and -low groups using X-tile plots. A total of 1409 radiomics features were extracted from the segmentation of regions of interest, based on preoperative CT images of each patient. The LASSO algorithm was applied to reduce the dimensionality of the data and select features. The extreme gradient boosting classifier (XGBoost) was developed using a training set consisting of 137 consecutive patients admitted between January 2017 and December 2017. The model was validated in 47 consecutive patients admitted between January 2018 and April 2018. The performance of the XGBoost classifier was determined by its discriminative ability, calibration, and clinical usefulness.

Results: The cut-off value of the CD8⁺ T-cell level was 18.69%, as determined by the X-tile program. A Kaplan-Meier analysis indicated a correlation between higher CD8⁺ T-cell levels and better overall survival (p = 0.001). The XGBoost classifier showed good discrimination in the training set (area under curve [AUC], 0.75; 95% confidence interval [CI]: 0.67-0.83) and validation set (AUC, 0.67; 95% CI: 0.51-0.83). Moreover, it showed a good calibration. The sensitivity, specificity, accuracy, positive and negative predictive values were 80.65%, 60.00%, 0.69, 0.63, and 0.79, respectively, for the training set, and 80.95%, 57.69%, 0.68, 0.61, and 0.79, respectively, for the validation set.

Conclusions: We developed a CT-based XGBoost classifier to extrapolate the infiltration levels of CD8⁺ T-cells in patients with PDAC. This method could be useful in identifying potential patients who can benefit from immunotherapies.

Keywords: CD8 positive T lymphocytes; contrast-enhanced computed tomography images; pancreatic ductal adenocarcinoma; prognosis; radiomics.

Figure 1

Flow chart visualizing the patient selection process.

Figure 2

Radiomics workflow.

Figure 3

X-tile analysis of survival data in patients with pancreatic ductal adenocarcinoma (A, B) The optimal cut-off CD8⁺ T-cell level of 18.69%, determined by X-tile, is used to define the CD8⁺ T-high and CD8⁺ T-low groups. (C) CD8⁺ T in the CD8⁺ T-low group and the CD8⁺ T-high group. The chart includes a box plot, density plot, and dot plot. The 25th and 75th percentiles are shown as connecting lines between groups. (D) The Kaplan-Meier curve and log-rank test suggest that patients in the CD8⁺ T-high group survive significantly longer than those in the CD8⁺ T-low group.

Figure 4

Comparison between patients with low and high CD8⁺ T-cell infiltration (A–C) Patient 1: A 65-year-old man with PDAC in the CD8⁺ T-high group. (A) CD8⁺ T-cell infiltration is high (×20). (B) The axial portal-phase CT image shows an infiltrative, low-attenuation mass (arrows) located at the pancreatic head. (C) The prediction probability of low CD8⁺ T infiltration was 80.58% by XGBoost classifier. (D–F) Patient 2: A case of a 49-year-old man with PDAC in the CD8⁺ T-low group. (D) CD8⁺ T-cell infiltration is low (×20). (E) The axial portal-phase CT image shows an infiltrative, low-attenuation mass (arrows) located at the pancreatic body and tail. (F) The prediction probability of low CD8⁺ T-cell infiltration is 70.07% by XGBoost classifier.

Figure 5

Receiver operating characteristic (ROC) curves and calibration curves of the extreme gradient boosting (XGBoost) classifier (A) ROC curves of the XGBoost classifier in the training and validation set. (B) Calibration curves of the XGBoost classifier in the training and validation set.

Figure 6

Decision curve analysis (DCA) for the extreme gradient boosting (XGBoost) classifier. The y-axis represents the net benefit. The gray line represents the hypothesis that all patients had high CD8⁺ T-cell infiltration. The black line shows the hypothesis that all patients had low CD8⁺ T-cell infiltration. The x-axis shows the threshold probability, which is where the expected benefit of treatment is equal to the expected benefit of avoiding treatment. The decision curves show that with a threshold probability greater than 0.16, using the prediction model to predict CD8⁺ T-cell infiltration adds more benefit than the treat-all-patients as high CD8⁺ T-cell infiltration scheme or the treat-none as low CD8⁺ T-cell infiltration scheme in the training set.