In the present study, the accuracy of clinical parameters along (CP), blood parameters along (BP), and combining both clinical and blood parameters (CBP), and published model (Tan) for noninvasive, individualized prediction of pCR in patients with LARC undergoing nCRT was compared. The proposed CBP model performs best compared with other models and thereby provides an effective tool for clinical decision making. The easy-to-use dynamic nomogram facilitated noninvasive estimation of pCR.
“Watch and Wait” was recommended as an objective strategy for LARC patients. How long to be waited between completion of radiation and achievement of cCR, it varied from different research institute. Habr-Gama et al. proposed that intervals between completion of radiation and achievement of cCR may beyond 16 weeks in most cases(23). Similarly, a randomized trial of 252 LARC patients reported that the pCR rate of the 12-week interval group (23.6%) was significantly higher than the 8-week interval group (12%)(24). The multivariate analysis outcomes of interval (OR 1.305, P = 0.01) in this study was consistent with these previous studies, indicated that despite frequent early responses and reduction in tumor burden, a complete response of tumor may take longer than we think.
There is growing evidence that several clinical parameters play a critical role in the treatment response of LARC(25). Despite “watch and wait” strategy have been observed benefit in favor of patients with LARC from organ preservation, surgery-related morbidity, and quality of life(6, 7). It is difficult to diagnose the primary tumor to be a cCR status based on the common methods in clinical practice, and further to predict pCR(9–11). A recent research showed that tumor differentiation grade presents the same predictive role for pCR (26). However, Ono et al. did not observe such associations(27). We found that tumor differentiation grade (P = 0.02) was an independent risk factor for pCR in our study. This discrepancy could be due to a small sample size or the different nCRT regimens. Previous studies showed that compared with the pre-nCRT CEA level, both the post-nCRT CEA level and the change pattern of CEA during nCRT, are closely related to nCRT response of LARC patients(28, 29). We found post-nCRT CEA level (OR 0.743, P = 0.046) was an another potentially important predictor of pCR. This might due to the latter values reflect the degree of nCRT response. This suggested that the performance of a predictive model would be enhanced using data obtained during or after CRT.
Recently, various blood parameters, which reflecting systematic inflammatory response and nutrition status, seem to affect the tumor response to nCRT in patients with malignancy(14, 18, 19). Of note, numerous studies have demonstrated that the systematic inflammatory response and nutrition status could destroy immune systems and increase tumor resistance to nCRT(30, 31). In multivariate analysis, other than FARI (OR 0.718, P = 0.036) and SGR (OR 2.555, P = 0.002), both NLR, LMR, PLR, SII, and PNI are failed to predict pCR. A potential reason for this inconsistent result is that except from FARI and SGR, these blood parameters were all leukocyte-based markers. However, in our study cohorts, all patients received nCRT, which would cause the bone marrow suppression. This may lead to these leukocyte-based markers fail to reflect the real systemic inflammatory response of patients, lowering the accuracy of these markers in predicting pCR. This is in line with the findings of Wang et al.(32) and our previous study(18).
Since the poor performance of singe-predictor to predicting pCR(9–11), predicting models integrating multiple parameters are receiving more and more attention in the field of rectal cancer as increasing evidence is gained about their promising performance in predicting pCR. Recent studies from Ren et al.(12) and Zhang et al.(33) found that the pCR predicting models comprised with clinical parameters and nCRT regimens performed exceedingly well, and the C-index were 0.793 and 0.802, respectively. However, the patients in this study were underwent the same nCRT regimen so it was impossible to explore its effect on pCR. In addition to the clinical based models, predicting models combined with blood parameters were also showed good performance in pCR prediction. A pCR predicting model based on NLR, LMR, and neutrophil-monocyte to lymphocyte ratio (NMLR) achieving an AUC of 0.75(34). Though promising, these published models have been limited by the lack of an independent validation cohort. Besides, more detail information of these models, such as sensitivity, specificity, and accuracy have not clarified. Considering different tumor behaviors are determined by many interactions among multiple factors. Hence, in the current study, a CBP model combined with clinical and blood parameters was built. Among the four models, only CBP model (AUC = 0.816 in training cohort and AUC = 0.752 in validation cohort) achieved an AUC > 0.75, indicating acceptable prediction. The result of calibration and DCA further confirmed our previous conjectures. Moreover, when comparing CBP model with the CP, BP, and Tan model through NRI and IDI, the improvement in prediction accuracy was all significant in training cohort (all P < 0.05, NRI test, IDI test), and CBP model showed significant improvement compared with CP (P < 0.05, NRI test, IDI test) and Tan model in validation cohort (P < 0.05, IDI test).
Finally, the developed dynamic nomogram allows the CBP model to be used conveniently in clinical practice.
The present study has limitations. First, our study is limited by its retrospective design, selection bias cannot be ruled out. Second, because of the small sample size, randomization resulted in imbalance of DTAV and PNI between training and validation cohort. Additionally, possibly for the same reason, some NRI and IDI results in validation cohort showed a consistent tend with the training cohort, but the difference did not reach statistical significance. Finally, although we categorized the patients into independent training and validation cohorts according to their TRG status, the CBP model may therefore perform less well in other situations because of the lack of an external validation. A multicenter approach would have given more external validity to our estimates.