Patient and tumor characteristics

In total 110 mccRCC patients were finally included into this study. Training set has 77 patients (mean age ± standard deviation, 55.9 years ± 9.6; 55 men); test set has 33 patients (mean age ± standard deviation, 53.0 years ± 13.0; 27 men). In total, 33 (30.0%) patients presented VEGFR-TKI early resistance, 22 (28.6%) in training set and 11 (33.3%) in test set, with no statistical significance. Among 110 patients, tumor progression occurred in 99 patients, the 21 patients remained PR or SD. 45 patient deaths occurred in this cohort. The clinical characteristics of patients and tumors are summarized in Table 1. Male presented dominance in both training and test sets. Most of the patients had synchronous metastasis (89/110, 80.9%). Most patients were in the IMDC intermediate (78/110, 70.9%) and poor (25/110, 22.7%) group. Age, gender, body mass index(BMI), tumor largest dimension, tumor T staging, N staging, venous thrombus, metastatic status (synchronous/metachronous), WHO/ISUP grading, IMDC score and median PFS had no differences among two patient sets.

Table 1 Patient and tumor characteristicsRadiomic analysis and nomogram construction in predicting first-line VEGFR-TKI early resistance

In the final feature selection with the LASSO method, 12 features were included in the radiomic models (see Table S1). The radiomic signature was constructed with a Radscore calculated using the following formula:

$$\text = -0.24\times\text\,1 + 0.799\times\text\,2 + 0.268\times\text\,3 + 0.263\times\text\,4 + 0.298\times\text\,5 + (-0.111\times\text\,6) + 0.002\times\text\,7 + 0.229\times\text\,8+ (-0.095\times\text\,9) + (-0.099\times\text\,10) + (-0.81\times\text\,11) + 0.15\times\text\,12 + 0.511$$

The Rad-scores were significantly higher in the early resistant group than in the clinical beneficial group in the training and test sets (p<0.001 and p = 0.02, respectively; Wilcoxon’s rank-sum test). In the training set, the AUC (95% CI) were 0.81 (0.72, 0.90). Accuracy was 0.701 (95%CI: 0.586, 0.800). In the test set, the AUC (95% CI) were 0.79 (0.62, 0.96). Accuracy was 0.727 (95%CI: 0.545, 0.867). After univariate and multivariate logistic regression in training cohort, several clinical factors (T staging, N staging, IMDC score and WHO/ISUP grading) were confirmed to correlate with VEGFR-TKI resistance. Since WHO/ISUP grading need to be evaluated through pathological examination and sometimes cannot be evaluated by biopsy samples. For wider application of the predicting nomogram, we did not include WHO/ISUP grading in nomogram construction. Finally, a novel radiomic-based nomogram was generated by incorporating the three clinical factors and radiomic signature in the training set (Fig. 3). In the training set, the nomogram had the AUC (95% CI) of 0.83 (0.74, 0.92) and accuracy (95% CI) of 0.792 (0.685, 0.876). In the test set, the nomogram had the AUC (95%CI) of 0.88 (0.77, 1.00) and accuracy (95% CI) of 0.818 (0.645, 0.930) (Table 2; Fig. 4). The nomogram had the positive prediction value (PPV) of 0.75 in training set and 0.77 in test set; and the negative prediction value (NPV) of 0.91 in training set and 0.91 in test set. Cut-off value of the nomogram score is 1.18. The nomogram performed better than clinical model in training set (p = 0.02), and in test set (p = 0.04).

Fig. 3

Nomogram for predicting first-line VEGFR-TKI early resistance in metastatic clear cell renal cell carcinoma

Table 2 Model performances in predicting first-line VEGFR-TKI early resistance Fig. 4

ROC curves in training and test set for predicting first-line VEGFR-TKI early resistance in metastatic clear cell renal cell carcinoma

Correlation with progression-free survival and overall survival

Univariate cox regression demonstrated that venous thrombus, WHO/ISUP grading, sarcomatoid differentiation and nomogram score ≥ 1.18 correlated with PFS. Multivariate cox regression indicated that only nomogram score ≥ 1.18 was independent predictive factor of PFS. Nomogram score ≥ 1.18 had hazard ratio(HR) (95% CI) of 0.34(0.18, 0.65) with p<0.001 (Table 3). We classified patients with nomogram score ≥ 1.18 as low-risk group and patients with nomogram score<1.18 as high-risk group. In the training set, median PFS (95% CI) in low-risk group (n = 43) was 19.4 (9.8, 28.9) months, median PFS (95%CI) in high-risk group (n = 34) was 4.0 (2.7, 5.2) months (log rank p<0.001). In the test set, median PFS (95%CI) in low-risk group (n = 19) was 13.4 (6.9, 20.0) months, median PFS (95%CI) in high-risk group (n = 14) was 3.8 (2.4, 5.3) months (log rank p<0.001) (Table 4; Fig. 5).

Fig. 5

Progression-free survival of different risk groups of patients in training and test set

Cox regression analysis indicated that nomogram score ≥ 1.18 was the only prognostic factor of OS (HR 0.38 (0.20, 0.71), p = 0.002) in the training set. In the training set, median OS (95%CI) in low-risk group (n = 43) was 90.4 (60.7, 126.9) months, which was significantly longer than that in high-risk group (n = 34, OS (95%CI) = 60.4 (60.0, 61.9) months), with log rank p = 0.003. In the test set, median OS (95%CI) in low-risk group (n = 19) was 99.2 (49.8, 99.2) months, median OS (95%CI) in high-risk group (n = 14) was 32.1 (6.0, 55.2) months (log rank p = 0.009) (Table 4; Fig. 6). Figure 7 demonstrated two examples of low-risk and high-risk patients with the same clinical factors, different rad-scores and nomogram scores, who presented different responses to first-line VEGFR-TKI therapy and different short and long-time prognosis.

Table 3 Univariate and multivariate cox regression of factors correlated with PFS in training set Table 4 Progression-free survival and overall survival of different risk groups Fig. 6

Overall survival of different risk groups of patients in training and test set

Fig. 7

Examples-risk stratification of patients with different rad-scores and nomogram scores