The long-term survival characteristics of a cohort of colorectal cancer patients and baseline variables associated with survival outcomes with or without time-varying effects

We assessed and ruled out the potential correlation between the categorical variables (Table 1) based on the pair-wise Pearson’s correlation coefficient value (Additional file 1: Table S1).

A set of widely investigated survival outcomes were examined in order to conduct a comprehensive investigation. The endpoints of overall survival (OS), disease-specific survival (DSS), recurrence-free survival (RFS), metastasis-free survival (MFS), recurrence or metastasis-free survival (RMFS), and event-free survival (EFS) were death from any cause, death from colorectal cancer, diagnosis of local recurrence, diagnosis of distant metastasis, diagnosis of recurrence or metastasis, and diagnosis of recurrence, metastasis, or death from colorectal cancer, respectively. Survival times were calculated starting at the date of diagnosis till the date of the first occurrence/observation of the endpoint (or the date of last contact); in multi-event outcomes, the latter date was the date of the first event. In each survival outcome, data were censored at the date of last contact for patients who have not experienced the events of interest during their follow-up. Data on the survival outcomes are summarized in Table 2.

IBM SPSS Statistics (version 25) was used for Kaplan-Meier method that generated the survival curves. Univariate Cox models were fitted for variables for each of the survival outcomes. PH assumption test [37] was performed using cox.zph function [38] in R (ver. 3.5.0) [39] using the default “km” function to obtain the transformed survival times. Multivariate Cox models were constructed using backward selection method, and when the PH assumption was violated, Cox model with time-varying effects (assuming piece-wise constant hazard ratios—this model is also called change-point Cox model [29, 30]) was used. Full multivariate models with all baseline variables were first checked for the PH assumption. For the variables that violated the PH assumption, cutoff time points before and after which the PH assumption was satisfied were obtained, starting from the variable with the lowest p value of the PH assumption test. The proper cutoff time points were selected based on the approach described in Pavelitz et al. [27] and Klein and Moeschberger [29]. In this study, a set of cutoff time points ranging from 0.5 to 18.5 years and with increments of 0.5 years was considered. The proper cutoff time point is ideally the one which makes the model (1) having the largest maximized log partial likelihood and (2) with the PH assumptions being satisfied before and after the time point. If the model with the largest maximized log partial likelihood did not satisfy the proportional hazards at both time intervals separated by the tested time point, then the one with the second largest maximized log partial likelihood value was tested. This step was repeated until a model was obtained that satisfied the criteria. The corresponding cutoff time point was then deemed to be the proper cutoff time point. In cases when the cutoff time points made the model having an infinite upper limit of the 95% confidence interval (CI) for hazard ratio (HR) of a variable, the cutoff point with the next largest maximized log partial likelihood was considered. This is because infinite limits suggest that a valid effect estimation cannot be made. In addition, rarely a proper cutoff time point for a variable was not identifiable. For example, for stage III in the OS analysis, a single time point that satisfies the PH assumption in both time periods (before and after the time point) was not identified. We then introduced additional time points in one of the time periods where the PH assumption was violated. However, this step did not identify any proper time points in this region. In this case, we analyzed this variable with the next one in line (i.e., the next variable with the smallest p value of the PH assumption test) and tested all the possible combinations of the cutoff time points to identify the proper cutoff points of both variables at the same time. Once the proper cutoff time points were identified and included in the model, variables with p values ≥ 0.05 in the model were removed one by one, starting with the one with the largest p value. During this process, if any of the remaining variables violated the PH assumption, the cutoff time point was identified/re-identified for this variable based on the method described above, followed by re-fitting of the model. The variables in the final model for each outcome measure reported in this manuscript have a p value < 0.05 either over the follow-up time (i.e., variables with no time-varying effects) or in at least one time period defined by the cutoff time points (i.e., variables with time-varying effects).

Age at diagnosis was examined as a continuous variable in this study. A p value < 0.05 was considered significant. All statistical analyses were performed by R (ver. 3.5.0) [39] or IBM SPSS Statistics (version 25).

Univariate associations between clinico-demographic and molecular variables and survival outcomes are summarized in Fig. 2. All variables investigated, except the familial risk status, were associated with at least one survival outcome. Several variables also violated the PH assumption of the Cox regression models. For these variables, Kaplan-Meier curves showing the survival probabilities over time are depicted in Additional file 1: Figure S1-S3. Those variables that violated the PH assumption and were significantly associated with the outcomes (univariate Cox regression p value < 0.05) tended to have separated curves with no visible crossing of the curves (type A variables; Additional file 1: Figure S1 and Figure S3). In contrast, it was clearly observable that those variables that violated the PH assumption but were not significantly associated with the outcomes (i.e., univariate Cox regression p value ≥ 0.05) tended to have their curves crossed (type B variables; Additional file 1: Figure S2).

Multivariable models are shown in Additional file 1: Tables S2-S7, and the main findings are summarized in Fig. 2. Age at diagnosis was associated with overall survival as well as metastasis-related outcomes (Additional file 1: Tables S2, S5-S6). The effect size of this variable on the risk of death from any cause became slightly larger after 10.5 years post-diagnosis (Additional file 1: Table S2). In contrast, increasing age at diagnosis was associated with decreased risks of MFS (HR 0.98; Additional file 1: Table S5) and RMFS (HR 0.98; Additional file 1: Table S6). Other demographic variables (sex and familial risk) as well as tumor histology and grade were not associated with any of the survival outcomes.

Tumor location and BRAF Val600Glu mutation status were associated with all outcomes. The effects of these two variables either remained constant or varied with time on different disease outcomes (Fig. 2; Additional file 1: Tables S2-S7). For example, the results of RFS, MFS, and EFS analyses showed that rectal cancer patients compared to colon cancer patients had shorter times to events throughout the follow-up time with no detectable time-varying effects (Additional file 1: Tables S4-S5, and S7). In the OS analysis, no significant difference between the rectal and colon cancer patients was detected prior to 2 years following diagnosis. However, after this time point, the risk for rectal cancer patients became significantly higher (HR 1.68; Additional file 1: Table S2). Also, while in the early years RMFS and DSS times did not significantly differ between the rectal and colon cancer patients, after 3 years in RMFS and after 6.5 years in DSS, the event risk significantly increased for the rectal cancer patients (HR 3.91 for RMFS and 5.97 for DSS; Additional file 1: Tables S3 and S6). The presence of BRAF Val600Glu mutation was associated with shorter overall and disease-specific survival times within the first 2.5 years post-diagnosis (HR 2.18 for OS and 3.05 for DSS), but not after that (Additional file 1: Tables S2-S3). This mutation was also significantly associated with an increased risk of recurrence after 4 years following diagnosis (HR 7.10; Additional file 1: Table S4). Last, patients with this tumor mutation had shorter MFS, RMFS, and EFS times without any time-varying effects (Additional file 1: Tables S5-S7).

Stage was associated with all outcomes except the risk of recurrence (Fig. 2). Patients with advanced stages had generally increased risks of outcome events, stage IV disease was a strong predictor of death, and stage III disease was a predictor of metastasis (Additional file 1: Tables S2-S3, S5-S7). For this variable, time-varying effects were found on death-related outcomes (i.e., OS, DSS, and EFS). Specifically, compared to stage I patients, stage III and stage IV patients had a much higher risk of death from any cause within the first year following diagnosis than later (Additional file 1: Table S2). Similar to this, for stage IV patients, the risk of death from colorectal cancer was much higher during the first year post-diagnosis (Additional file 1: Table S3). Additionally, the risk of having at least one disease-related event for stage III patients (EFS) was much higher within the first 1.5 years following diagnosis, which then decreased in magnitude (HR 6.02 within the first 1.5 years post-diagnosis versus 2.99 after that; Additional file 1: Table S7). The effects of disease stage on other outcome measures did not change over time.

Tumor MSI phenotype was associated with only metastasis-related outcome measures (MFS, RMFS, and EFS; Fig. 2). MSI-H tumor phenotype had a protective effect (Additional file 1: Tables S5-S7). Unlike other variables, MSI status had no time-varying effects.

Last, the two treatment-related variables, adjuvant chemotherapy and adjuvant radiotherapy treatment statuses, showed different association patterns in this observational cohort. While the associations of the adjuvant chemotherapy treatment were detected in death-related outcomes (i.e., OS, DSS, and EFS), adjuvant radiotherapy treatment was only associated with MFS (Fig. 2). These effects were non-proportional during the follow-up (i.e., varied over time). Specifically, within the first year following diagnosis, adjuvant chemotherapy had strong and significant protective effects on OS, DSS, and EFS, after which this effect was not detectable in the EFS analysis but was still significant in the OS and DSS analyses, albeit with a decreased effect size (HR 0.05 for OS, 0.15 for DSS, and 0.40 for EFS within the first year post-diagnosis and 0.56 for OS and 0.50 for DSS after this time point) (Additional file 1: Tables S2-S3 and S7). Whereas, compared to patients who did not receive adjuvant radiotherapy, patients who received adjuvant radiotherapy had an increased risk of metastasis (HR 6.00) after 5.5 years following diagnosis (Additional file 1: Table S5).

Multivariable models that considered the time-varying effects yielded a number of interesting findings (Additional file 1: Tables S2-S7). Regarding the demographic features, increasing age at diagnosis was associated with a small but significant increased risk of mortality throughout the follow-up time. This risk became slightly larger after the initial 10.5 years (OS; Additional file 1: Table S2). It is not surprising that younger patients had a lower risk of death, as they normally would have fewer comorbidities, lower chances of dying from other causes, and are likely to receive aggressive and intense treatments [42] that may contribute to their longer survival times. The slight increase in the risk of death after a decade can be explained by the aging of the patients in the cohort. On the other hand, small but long-term effects were detected for age at diagnosis on metastasis-related outcomes (MFS and RMFS) where decreased age was associated with worse MFS/RMFS times (Additional file 1: Tables S5-S6). It is reported by other studies that younger colorectal cancer patients present with advanced diseases [42, 43]. In our cohort 32.8% and 27.7% of the younger patients (age at diagnosis < 65) and older patients, respectively, were diagnosed with a stage III disease—this may explain the increased metastasis risk in the young patients. In contrast to age, another demographic variable, sex, was not associated with any of the survival outcomes examined in this study. The role of patient sex in prognosis is controversial: some studies support that female patients have a better prognosis compared to male patients [44‐46] while others do not find such a sex-based difference [21, 47]. We observed a better survival for female patients in the univariate analysis, but this association was not retained in the multivariable models. Additionally, consistent with other studies [10, 48], familial risk status, while it is a risk factor for the development of colorectal cancer [11, 49], had no significant relation to any survival outcomes investigated. Therefore, in this cohort, age at diagnosis has emerged as the only demographic factor with a predictive role.

In our analysis, MSI status was predictive of only metastasis-related outcomes (MFS, RMFS, and EFS; Additional file 1: Tables S5-S7), and its effects remained stable during the entire follow-up. MSI-H is a known marker with protective effects on patient survival [7, 50], possibly due to its biological effect on metastasis through its association with increased immune cell infiltration [51]. Thus, our results are consistent with these previous findings but additionally emphasize that the MSI status predicts the risk of metastasis even long after the diagnosis. As the most important prognostic marker, stage was a predictor of the majority of the outcome measures investigated (Fig. 2). As expected, increased disease stage was generally associated with increased risk of events, but in some cases, the hazard ratios significantly differed before and after a time point. Interestingly, such effects were detected in death-related outcomes. Specifically, fluctuating HRs were detected for stage III patients in the OS and EFS analyses and for stage IV patients in the OS and DSS analyses. In these cases, the risk of event was much higher for the patients immediately after the diagnosis (i.e., within 1–1.5 years) compared to later. This time relationship may be attributed to the advanced disease at diagnosis and/or the post-surgical complications that are known to lead to early death [52‐54]. We note that in a previous study on OS, similar findings (i.e., non-proportional effects of stage III and stage IV disease) were reported [25].

Two variables were associated with all outcomes examined in this study, and tumor location was one of them. Tumor location is one of the most widely examined clinical variables in colorectal cancer and is used in the clinic for prognostic estimations as well as surveillance and treatment-related decisions. In our study, rectal tumors compared to colon tumors were associated with worse RFS, MFS, and EFS times throughout the follow-up with no time-varying effect (Additional file 1: Tables S4-S5, S7). It is known that rectal cancer patients have a higher risk of recurrence and metastasis [55, 56], which is also shown by our results. However, our results additionally showed that the rectal tumors had sustained these constant/continuous effects over a long time after the diagnosis. In contrast, in the OS, DSS, and RMFS analyses, we observed time-varying effects of tumor location. The DSS and RMFS model data were particularly interesting. In both models, rectal cancer patients tended to have worse outcomes compared to colon cancer patients, but this difference reached significance only after certain time points. In the RMFS model, the risk for increased recurrence/metastasis became significantly higher for rectal cancer patients only after the initial 3 years. Since recurrence and metastasis indicate disease progression, RMFS data may be particularly relevant for clinical surveillance purposes and may suggest that the rectal cancer patients who survived the first 3 years without disease progression may need to be carefully surveilled after this time period. Additionally, a similar and a later effect was observed in the DSS model, where the risk of death from colorectal cancer significantly increased for the rectal cancer patients after 6.5 years. The non-proportional effect of tumor location on DSS has been observed by others as well [21]. In our study, the increased risk of disease progression for rectal cancer patients after 3 years (Additional file 1: Table S6) may explain their increased risk of disease-specific death after 6.5 years (Additional file 1: Table S3).

Like tumor location, BRAF Val600Glu mutation status was associated with all outcome measures (Additional file 1: Tables S2-S7). This mutation is one of the most studied tumor mutations in colorectal cancer as well as other cancer sites, such as ovarian cancer [57], thyroid cancer [58], lung cancer [59], and melanoma [9]. In our study, patients with this tumor mutation had increased risk of two metastasis-related outcomes throughout the follow-up (the highest risk being associated with metastasis-free survival; HR 3.46). Such a relationship between mutant BRAF and metastasis was previously reported in other cohorts [60, 61]. This mutation was also associated with a shorter time to recurrence after 4 years. It is not immediately clear how this mutation may influence the recurrence risk in colorectal cancer, but the association of this mutation with tumor recurrence has been reported in papillary thyroid cancer as well [58]. In addition to these, previously, BRAF Val600Glu mutation has been associated with the increased risk of mortality in colorectal cancer [13, 61‐64]. In our study, in two death-related outcomes (OS and DSS), this mutation emerged as a predictor of death early after diagnosis (within the first 2.5 years). Interestingly, this group of patients also tended to have better DSS times if they survived the initial 2.5 years following diagnosis, but this did not reach significance levels (HR 0.14, p = 0.0505; Additional file 1: Table S3). BRAF Val600Glu mutation status is the only variable identified in this study that was both an early-event (OS and DSS) and late-event (RFS) marker. The reason why this mutation has such effects remains unknown and warrants more investigations. Overall, the wide-spectrum of associations detected for this mutation in this study further strengthen this gene’s importance in colorectal cancer.

Adjuvant therapy is given based on the clinical and disease characteristics to help control the disease (e.g., to reduce/eliminate the recurrence and/or metastasis risk) and to improve the survival outcomes of patients. In our patient cohort, patients who received adjuvant chemotherapy had better survival outcomes (OS, DSS, and EFS) than those patients who did not receive it. These effects were especially stronger within the first year following diagnosis (OS, DSS, and EFS models; Additional file 1: Tables S2-S3, S7). The changing effects (from strong to weaker protective effects) may reflect the slightly diminishing effects of therapy after the treatment duration, which is usually no more than a year [65]. Time-varying effects for chemotherapy treatment were detected in other cancers as well, such as breast cancer [66‐68]. On the other hand, adjuvant radiotherapy status was associated with only MFS (Additional file 1: Table S5). Initially, MFS times did not differ significantly for the patients who did or did not receive this treatment. However, after 5.5 years following diagnosis, those patients who received radiotherapy had increased risk of developing their first metastases compared to patients who did not receive this treatment. The exact mechanisms through which adjuvant radiotherapy can have a late effect on MFS of patients is not clear, but it is known that in some cases, radiation treatment increases the risk of metastasis [69‐71]. As these authors discussed [69‐71], a variety of potential mechanisms can explain this effect, such as the appearance or development of radiation-resistant tumor cells, changes in the tumor microenvironment or immune system response over time, or suppression of the tumor progression by radiation treatment that initially delays the tumor metastasis. These previous and our findings emphasize the need for new research revenues and potentially prolonged surveillance for late-onset metastatic lesions in colorectal cancer patients who are treated with adjuvant radiotherapy.