Radiomics studies in EC most focused on the prediction of lymph node metastasis, radiation-induced diseases in the earlier[18–20]. What’s more, there were still few researches in the prediction of both PFS and OS[16, 21–23]. The establishment of accurate prediction models of PFS and OS were conducive to clinical decision-making and are expected to improve the survival rate of advanced ESCC patients. Therefore, in the present study, we constructed and validated machine learning models to predict PFS and OS of non-surgical ESCC patients, which incorporated the clinical variables and CECT images. The C-index and AUC showed that combined models had a better performance than radiomics or clinic models alone. The results demonstrated that incorporated the clinical variables enhanced the combined models’ predictive efficacy both in PFS and OS. In other word, the pre-treatment CECT images would not provide enough information to predict the treatment outcomes and the clinical data are essential for patient’s survival prediction.
As for the prediction of PFS and OS, the combined models performed well with the prognostic accuracy over 70% based on clinical and radiomics features. The C-index of PFS prediction radiomics model, clinical model, combined model in the validation cohort was 0.64, 0.78, 0.79 and AUC was 0.676,0.823,0.833, respectively. The C-index of OS prediction radiomics model, clinical model, combined model in the validation cohort was 0.65, 0.64, 0.71 and the AUC was 0.646, 0.695, 0.768, respectively. Bohanes et al[24] found that gender and age had significant influence for the treatment outcomes of EC patients. But, in our study, age and sex were not involved in the model’s development as they were no statistically significant by univariate Cox regression. TNM stage was the most commonly prognosis prediction method in clinical. MES et al[25] and Zhao et al[26] combined TNM stage and other clinical factors to improve the prognostic predictive. In the PFS and OS prediction, N stage was filtered to develop clinical model in our study. But for the prediction of OS, M stage was also selected to develop models. Li et al[27] used deep learning to predict the treatment response to CCRT for ESCC patients which also included the M stage in the progress of model development.
Jayaprakasam et al[21] established radiomics model to predict PFS based on 72 ESCC patient’s PET/CT images and the AUC was 0.73 in the validation cohort. They first included the PET responders into survey. But, PET/CT examination was highly expensive than CT or MRI and the sample was small in the study. Luo et al[22] also developed a nomogram model for predicting local PFS based on CT images and C-index was 0.723 in validation cohort. This study included clinical response to develop model and obtained a fine result. Liu et al[28] have found that clinical complete response after neoadjuvant CRT was significantly correlated with survival of patients with ESCC. So, we also selected the clinical response, ORR and DCR into model’s building, which significantly enhanced the model’s prediction efficacy of PFS. After selecting, the tumor differentiation was also chosen to develop the PFS and OS prediction models. Barbetta et al[29] found that poor tumor differentiation was an independent risk factor for recurrence in EC patients. Qiu et al[30]incorporated radiomics and clinical features (including tumor differentiation) to predict postoperative recurrence risk of ESCC patients .And the C-index of validation cohort was 0.724 in their combined model. In previous study, researchers had proved that ECOG PS had significant prognostic effects on clinical response and survival[31, 32].And RE was regarded as one of the factors affecting patient’s prognosis[33]. ECOG PS was used to evaluated the patient’s physical performance before treatment and RE was the radiation-induced esophageal disease after treatment. In our study, ECOG PS and RE were also selected to predict PFS.
Except the tumor differentiation and N stage described above, M stage was also selected to develop the OS prediction models. Shi et al[34] concluded that metastatic lesions were closely related to the prognosis of patients. Li et al[27] used deep learning to predict the treatment response to CCRT for ESCC patients and the M stage was also involved in the progression of model development. Due to the heterogeneity of tumor, the treatment outcomes of patients might be different even with the same clinical features[11, 35, 36].
Radiomics is defined as the high-throughput extraction of image features from radiographic images[37, 38]. Radiomic features provide abundant additional information predictive of underlying tumor biology and behavior[39]. These signatures can be used alone or with other patient related data (e.g., pathological data, genomic data, clinical data) to predict tumor phenotyping, treatment response prediction and prognosis. Our study finial selected one shape texture, two GLSZM textures, two GLDM textures and one GLCM texture to develop PFS prediction model and one shape texture, one GLSZM texture, one GLDM texture to construct OS prediction model. Wavelet transformed features contain more information and are more difficult to explain than first-order and shape features, but also reflect more complex information about tumor heterogeneity[16]. Therefore, the prediction results based on Wavelet transformed features are consistent with the cognition of clinical outcomes.
Although our research developed the survival prediction model with good accuracy, there are still some limitations in our study. Due to it is a retrospective study, some clinical variables are not comprehensive enough. Furthermore, all of the recruited EC patients were confirmed ESCC by pathology examination, which were the advanced stage and lost the surgery opportunity. Therefore, the constructed model may be limited in esophageal adenocarcinoma (EAC) or surgery patients. Due to the treatment plan is so vital for patient’s survival, our study recruited the patients treated with CRT, which might limit the model’s therapy decision. Further study adding the genomics features in the model is willing to improve the accuracy of the treatment outcomes.