To the best of our knowledge, this is the first study to apply competing risk analysis to TLI research in patients with NPC. Among T4 NPC patients, death may occur prior to the TLI occurrence. Since such competing risk events exclude the event of interest, competing risk events should be taken into account in cumulative incidence calculation or prognostic models. We prefer the competing risk methodology over the Kaplan-Meier method or standard Cox regression model. The frequently used Kaplan-Meier method, in contrast to the CIF, censors patients who die prior to TLI. This leads to overestimated probabilities of TLI because dead subjects (and thus censored) are considered to experience TLI in the future. Similarly, in the Fine and Gray model, the competing events are not censored but are treated as different (competing) events.
Incidence of TLI after IMRT in patients with T4 NPC
The crude incidence of TLI in patients with T4 NPC has been reported to range from 7.5 to 28% [
12,
17,
24‐
27]. In this study, 63 patients developed TLI with an incidence of 12.5%, which was consistent with previous reports. The differences in the follow-up duration might cause a deviation in the incidence. TLI is a late complication with a long latency; the longer the follow-up period, the greater the possibility of TLI occurrence. Zhou et al. [
11] reported that the 5-year cumulative incidence of TLI in patients with NPC after two-dimensional radiotherapy (2D-RT) was 34.9%, which was much higher than the finding in our study (13.2%). In a previous study involving 973 NPC patients after 2D-RT at our center, the 4-year cumulative incidence of TLI was 15.3% (data not published). This finding was higher than that reported in this study (9.6%). On the one hand, these data suggest that IMRT could better protect the TL than 2D-RT in T4 NPC. On the other hand, compared with early-stage NPC patients, the T4 patients had a shorter survival period, and some patients died before TLI occurrence. We calculated the CIF by appropriate accounting. However, the analysis in the paper mentioned above was performed using the Kaplan-Meier method, which censors death as it occurs and over-estimates the CIF in the presence of competing risks. In addition, when comparing the incidence of TLI with older studies, it is worth noting whether an MRI was used as a routine examination to detect TLI during the follow-up and whether the endpoints were the same. Moreover, due to patient loss to follow-up and a relatively short follow-up period, the incidence of TLI might be underestimated, and long-term outcomes should be updated in the future.
Clinical and dosimetric factors associated with TLI occurrence
Several clinical factors, such as the prescribed total dose, dose per fraction, OTT, chemotherapy, age and T category, have been reported to be associated with TLI occurrence [
9,
28‐
30]. Since our study was a case-control study in which each case was matched to a control by age and gender and all patients had T4 NPC, we could not analyze the differences in age, gender and T category between the two groups. Moreover, due to the standardized treatment modalities for T4 NPC at our center, this study failed to draw such conclusions. The study conducted by Zhou et al. [
16], which involved 1883 NPC patients, suggested that the use of concurrent cetuximab is a possible independent risk factor for TLI on a multivariate analysis. However, given the very small numbers of patients using targeted biologic agents, we found that concurrent cetuximab or nimotuzumab was not significantly correlated with TLI in our study.
Emami et al. [
31] first reported that a dose of 60 Gy to 1/3 of the brain would cause a 5% risk of radiation-induced brain necrosis within 5 years. However, this study was based on literature reviews before 1991 (predating 3-dimensional conformal radiotherapy and IMRT), and this estimation is likely out-of-date in the era of IMRT. In 2010, Lawrence et al. [
32] predicted incidences of 5 and 10% radiation brain necrosis to occur at doses of 72 Gy and 90 Gy in 2-Gy fractions, suggesting that a small volume of the brain could tolerate a higher dose.
This study presents a large data set used to assess the relationship among various treatment factors and TLI development in T4 NPC patients after IMRT. The advantages of our study include the use of standardized chemo-radiotherapy regimens, homogenous patient cohorts, and consistency in the TL contouring with one oncologist. These advantages help us minimize surveillance bias and focus on the dosimetric parameters. In multivariate competing risk regression model, TL V20 and D1cc were found to be predictive factors of TLI after IMRT and D1cc was considered as the most predictive. It was generally believed that TLI occurrence was related to a small volume of high-dose radiation, although the tolerance dose was not exactly the same [
12,
17‐
20]. D1cc acted as a predictor of TLI, which is consistent with the conclusions of most studies mentioned above [
12,
17,
18]. Among these reports, D1cc < 58 Gy, D1cc < 62.83 Gy and D1cc < 62.8 ± 2.2 Gy were found to be the tolerance doses of TL. However, these data seem to be conservative to ensure adequate tumor coverage for T4 NPC. In the previously published literature, patients with T1-T4 NPC were included, and for the T4 patients, the differentiation might not be good. The dose limit of the TL mentioned above is mainly calculated for TD 5/5 (radiation dose that could cause a 5% risk of complication in 5 years after irradiation) and the risk of TLI was analyzed without using death as a competing risk event, which could lead to an overestimation of the risk of developing TLI. According to clinical experience and actual research, the dose tolerance of the TL is higher than expected. After considering death a competing risk event among the 506 T4 NPC patients, under the premise of without comproming tumor coverage, the 5-year cumulative incidence of TLI was 13.2%, which was within the acceptable limits. In this nested case-control study, the 5-year cumulative incidence of TLI in the group with D1cc ≤71.14 Gy was 13.2%, which was much lower than the 5-year cumulative incidence of TLI in the group with D1cc > 71.14 Gy (62.2%,
P < 0.001). Considering the balance of tumor coverage and TLI occurrence, we proposed a dose limit of D1cc ≤71.14Gy. A study involving 1887 NPC patients [
16] demonstrated that the risk of TLI dramatically increased at a dose ≥70 Gy and that a dose of no more than 70 Gy was relatively safe. In a retrospective study involving 749 NPC patients, D0.5 cc < 73.66 Gy was found to be helpful in reducing the incidence of TLI [
19]. Considering the treatment outcome, Ng et al. [
33] proposed that Dmax ≤72 Gy is a new dose constraint for the TL. These results are similar to our results that D1cc ≤71.14 Gy could be a useful dose constraint.
Previous studies have shown that the primary prevention strategy for TLI is to avoid high dose delivery to the TL during IMRT. However, in this study, V20 was found to be another independent predictor of TLI occurrence, which might be regarded as a strange dose constraint for TLI. Su et al. [
13] reported in their study involving 259 NPC patients treated with IMRT that V40 is highly predictive of TLI. As the dosimetric parameters V20 and V40 were highly correlated with each other (Spearman rank correlation coefficient of 0.832,
P < 0.001), V20 acted as a predictor of TLI, which is consistent with the results reported by Su et al. [
13]. In addition to high-dose radiation, low-dose radiation also plays important roles in the occurrence of TLI. Animal experiments [
34,
35] suggested that exposing large volumes of adjacent normal tissues to lower-dose radiation could decrease the tolerance to toxicity when a small volume was exposed to a higher dose. When the brain is irradiated, oligodendrocytes are damaged, and when exposed to a large volume of lower-dose radiation, oligodendrocyte progenitor cells lose their proliferative ability. The damaged oligodendrocytes cannot be promptly renewed and replaced, affecting the repair of normal brain tissues that are exposed to high doses and ultimately leading to brain injury [
36]. Potentially, a large volume of surrounding brain tissue exposed to low-dose radiation may have effects on the cellular environment and microvasculature that could increase the likelihood of brain injury. Although IMRT optimizes the dose distribution in the tumors, high dose to TLs overlapped with tumors seems to be inevitable in T4 NPC. Furthermore, the TLs receive a large volume of low- to moderate-dose irradiation, thus leading to the occurrence of TLI.
Target under-dosage due to OAR constraints appears to be correlated with an increased risk of local failure. The volume of under-dosage (< 66.5 Gy), even with a very small GTV volume (< 3.4 cm
3), was highly detrimental to local control [
33]. So, the importance of an adequate tumor coverage should be kept in mind. For T4 NPC, under the premise of tumor coverage, patients with D1cc ≤71.14 Gy and V20 ≤ 42.22 cc had significantly lower risks and cumulative incidences of TLI than those with D1cc > 71.14 Gy and V20 > 42.22 cc. These results suggested that D1cc ≤71.14 Gy and V20 ≤ 42.22 cc could be useful dose constraints of the TL without obviously interfering with the tumor coverage. In this study, the maximum point dose was not found to be the best predictor of TLI. Only the previously determined maximum dose constraint was inadequate for preventing TLI without considering the shape of the DVH. Efforts should also be made to limit a small volume of high-dose radiation and a large volume of low-dose radiation.
As a case-control study, it is possible that these results were affected by patient selection bias. Due to the small number of events available, the statistical ability was limited. In addition, various TL contouring methods could result in different dosimetric parameters [
23]. A contouring guideline is needed to ensure that dosimetric parameters are comparable and reduce inter-institutional differences. Notably, different TPS may have effects on the dosimetric parameters. Our results were based on single-center data, and further validation through multicenter cooperation and consistency in the TL delineation are needed.