A machine learning-based survival prediction model of high grade glioma by integration of clinical and dose-volume histogram parameters

Abstract

Purpose: Glioma is the most common type of primary brain tumor in adults, and it causes significant morbidity and mortality, especially in high-grade glioma (HGG) patients. The accurate prognostic prediction of HGG is vital and helpful for clinicians when developing therapeutic strategies. Therefore, we propose a machine learning-based survival prediction model by analyzing clinical and dose-volume histogram (DVH) parameters, to improve the performance of the risk model in HGG patients.

Methods: Eight clinical variables and 39 DVH parameters were extracted for each patient, who received radiotherapy for HGG with active follow-up. Ninety-five patients were randomly divided into training and testing cohorts, and we employed random survival forest (RSF), support vector machine (SVM), and Cox proportional hazards (CPHs) models to predict survival. Calibration plots, concordance indexes, and decision curve analyses were used to evaluate the calibration, discrimination, and clinical utility of these three models.

Results: The RSF model showed the best performance among the three models, with concordance indexes of 0.824 and 0.847 in the training and testing sets, respectively, followed by the SVM (0.792/0.823) and CPH (0.821/0.811) models. Specifically, in the RSF model, we identified age, gross tumor volume (GTV), grade, Karnofsky performance status (KPS), isocitrate dehydrogenase (IDH), and D99 as important variables associated with survival. The AUCs of the testing set were 92.4%, 87.7%, and 84.0% for 1-, 2-, and 3-year survival, respectively. According to this model, HGG patients can be divided into high- and low-risk groups.

Conclusion: The machine learning-based RSF model integrating both clinical and DVH variables is an improved and useful tool for predicting the survival of HGG patients.

Keywords: DVH features; high-grade glioma; machine learning; random survival forest; survival prediction.

FIGURE 1

Flow chart of patient inclusion

FIGURE 2

An example of treatment planning in a patient with HGG

FIGURE 3

(A and B) Out‐of‐bag (OOB) error rate to assess the quality of the survival prediction for HGG in the RSF model. The OOB error rate was 21.11% with mtry = 1, nodesize = 7 (A), and ntree value = 2000 (B). (C and D) Minimal depth to predict the survival of HGG patients in the RSF model. We identified six significant factors associated with survival, whose minimal depth was lower than the threshold of 2.247. Age was of the highest importance, followed by GTV, grade, KPS, IDH, and D99

FIGURE 4

(A) The ROC curves of the training set in the RSF model. The AUCs were 85.6% (95% CI [75.63%, 95.65%]), 85.4% (95% CI [74.81%, 96.08%]), and 91.4% (95% CI [81.92%, 99.99%]) for 1‐, 2‐, and 3‐year survival, respectively. (B) The ROC curves showed that the AUCs were 92.4% (95% CI [83.63%, 99.99%]), 87.7% (95% CI [75.65%, 99.78%]), and 84.0% (95% CI [68.82%, 99.14%]) for 1‐, 2‐, and 3‐year survival in the testing set, respectively. (C and D) HGG patients were divided into high‐ and low‐risk groups according to the RSF model. The low‐risk group had a longer OS time than the high‐risk group in both the training (HR = 9.075, 95% CI [3.603, 22.86], p < 0.0001) (C) and testing sets (HR = 17.4394, 95% CI [3.738, 81.37], p < 0.0001) (D)

FIGURE 5

(A) The prediction accuracy with a certain number of features in the SVM model. When the number of features was four, it achieved the optimal accuracy of 77%. (B) The top four features included in the SVM model were age, grade, GTV and CTV1. (C and D) In the SVM model, the AUCs for 1‐, 2‐, and 3‐year survival were 82.6% (95% CI [71.95%, 93.21%]), 83.3% (95% CI [71.88%, 94.73%]), and 88.5% (95% CI [77.51%, 99.53%]) in the training set (C) and 89.5% (95% CI [76.93%, 99.99%]), 87.1% (95% CI [76.25%, 97.98%]), and 82.3% (95% CI [67.38%, 97.39%]) in the testing set, respectively (D)

FIGURE 6

In the CPH model, the AUCs for 1‐, 2‐, and 3‐year survival were 83.5% (95% CI [72.36%, 94.66%]), 87.1% (95% CI [76.96%, 97.24%]), and 91.6% (95% CI [82.05%, 99.99%]), respectively, in the training set (A) and 87.9% (95% CI [74.01%, 99.99%]), 87.3% (95% CI [76.04%, 98.68%]), and 84.9% (95% CI [68.93%, 99.99%]) in the testing set, respectively (B)

FIGURE 7

DCA curves for the clinical benefit and the corresponding scope of application of the three models in the training (A) and testing sets (B). The RSF model had greater net benefits than the SVM and CPH models in the testing set