A machine learning based prediction model for short term efficacy of nasopharyngeal carcinoma

Abstract

The radiological dosimetric parameters and clinical features were screened by machine learning to construct a prediction model for the short-term efficacy of locally advanced Nasopharyngeal Carcinoma (LANPC). Patients diagnosed with Nasopharyngeal Carcinoma were retrospectively collected in the study. Twenty-four clinical features and twelve radiological dosimetric features were included. Three machine learning algorithms were used to construct predictive models for the short-term efficacy of LANPC. Kaplan-Meier log-rank method was used to compare the prognosis of patients with different efficacies in the model. The reliability of the model was evaluated using the calibration curve and the area under the curve (AUC). There were 194 patients who met the inclusion criteria. Among the three models being constructed, Random forest (RSF) model showed the best predictive ability, with AUC values of 1.000 in the training group and 0.944 in the test group, followed by XGBoost decision tree (GBDT) model (0.866/0.849) and decision tree (DT) model (0.848/0.783). In RSF model, the 3-year and 5-year overall survival rates of patients in complete remission (CR) group were 98.9% (95% CI 0.9688-1.0000) and 89.7% (95% CI 0.8256-0.9752), respectively.While for patients in non-CR group, the 3-year and 5-year overall survival (OS) rate was 100% (95%CI 1.000~1.000) and 98.8% (95% CI 0.9652-1.0000), respectively. There has statistically significant difference between the two groups (P = 0.0037). RSF model constructed by machine-learning algorithm based on radiological dosimetric parameters and clinical characteristics can better predict the short-term efficacy of LANPC, and is an effective tool to evaluate the short-term efficacy of different LANPC patients during treatment.

Keywords: Locoregionally advanced nasopharyngeal carcinoma; Machine-learning algorithm; Radiological dosimetric parameters; Short-term efficacy.

Fig. 1

A The important variables of DT model. As shown in A, the most important feature was PGTVnx D95, followed by PGTVnd_R Daverage, PGTVnx Volume, PGTVnd_R Dmax and Lateral pterygoid muscle. B ROC curve for DT model. The AUC value In the training set was 0.848. While in test cohort, the AUC value was 0.783.

Fig. 2

A The important variables for GBDT. B The top 5 important variables of GBDT model. From A and B the most important variable was PGTVnx D95, followed by PGTVnd-L volume, optic nerve-R Dmax, PGTVnd-L Dmin, and PGTVnx volume. C ROC curve for GBDT model. The AUC value of GBDT model was 0.866 in training cohort and 0.849 in the test cohort.

Fig. 3

A The OOB error rate to assess the quality of the short-term effects prediction for LANPC in the RSF model. The lowest OOB error rate of 24.82% with an mtry value of 3 and the ntree value of 500. B The important variables of RSF model. The top five important variable factors were PGTVnx D95, PGTVnx Daverage, PGTVnx volume, age, and PGTVnd_R volume by using randomForest method. C ROC curve for RSF model. The AUC value of RSF model is 1.000 in training set, and of 0.944 in test set.

Fig. 4

A Decision curves analysis of three models in training cohort, B Decision curves analysis of three models in test cohort. RSF model had superior predictive abilities for clinical decision-making than the DT and GBDT models both in the training and test groups.

Fig. 5

The survival curves for CR group and non-CR group. the 3-year and 5-year overall survival rates of non-CR group were 100% (95%CI: 1.000 ~ 1.000) and 98.8% (95%CI: 0.9652–1.0000), respectively. The 3-year and 5-year overall survival rates of patients in CR group were 98.9% (95%CI: 0.9688–1.0000) and 89.7% (95%CI: 0.8256–0.9752), respectively, there was significant difference in two groups (P = 0.0037).

Fig. 6

The flow chart of the study.