A Model to Predict Treatment Failure in Patients Undergoing Upfront Surgery for Resectable Colorectal Liver Metastases

Abstract

Introduction.

Patients who recur in the first year after resection of colorectal liver metastases (CRLM) do poorly. The aim of our study was to predict treatment failure in patients undergoing upfront resection with a nomogram.

Methods.

Data from patients resected between 1991 and 2019 were randomly split (70:30) into two cohorts. Treatment failure was defined as any recurrence or death within 12 months. A nomogram was constructed using multivariable logistic regression on the training cohort and validated using the testing cohort.

Results.

Overall, 783 patients were included. Primary tumor characteristics included 50% left-sided: 75.2% T3/4 and 56.5% node-positive. The median disease-free interval was 10 months, median number of metastases was 1 (1–50), and with a median size of 3.6 cm (0.2–22); 222 (28.3%) patients recurred within 1 year. Recurrence was mostly extrahepatic with or without liver involvement (150/222, 67.6%). Curative-intent treatment was possible in 37.8% of these patients. Primary location, T-stage and node status, disease-free interval, and number and size of metastases were associated with treatment failure. The area under the curve from the validation of the model was 0.6 (95% confidence interval 0.52–0.68). Patients with a high-risk of treatment failure (≥40%) had a worse survival from the landmark time of 12 months from surgery compared with those with low-risk (2-years: 82% vs. 70%; p = 0.0002).

Conclusions.

Primary location, T stage, node status, disease-free interval, and number and size of metastases are associated with treatment failure. The survival of patients with a probability of treatment failure ≥40% is unfavorable. Future trials investigating the role of neoadjuvant therapy in these high-risk patients are warranted.

Colorectal cancer is the third most common cause of cancer-related deaths worldwide.¹ Metastases from the primary tumor are frequent and affect the liver as the only site in approximately 30% of patients.² When feasible, complete resection is the optimal treatment modality for patients with colorectal liver metastases (CRLM) and is associated with 5-year survival rates ranging from 35% to 60% and the potential for cure.^3,4 Despite these overall favorable results, most patients experience recurrence of disease, most commonly within 2 years of surgery.⁵

Perioperative chemotherapy (including neoadjuvant treatment) combined with surgery has become a commonly recommended treatment strategy for patients with CRLM. While the benefit of systemic treatment in patients with unresectable metastases is well established, allowing for conversion to resectable disease and prolonging survival, there is minimal evidence to support neoadjuvant chemotherapy instead of upfront surgery in resectable CRLM.^6,7 To date, multicenter randomized trials have shown a small improvement in progression-free survival but have failed to demonstrate any survival benefit when a neoadjuvant approach was utilized.^8,9 As a result, both the National Comprehensive Cancer Network (NCCN) and the European Society of Medical Oncology (ESMO) guidelines recommend either treatment strategy.^10–12

The theoretical benefits of neoadjuvant systemic treatment in the setting of resectable CRLM includes an assessment of response to chemotherapy, guiding the use of adjuvant treatment, and sparing the risks of surgery in patients with rapidly progressing disease.^6,13 Furthermore, response may allow complete resection with less extensive surgery. However, some patients might progress during neoadjuvant treatment or develop major chemotherapy-related liver damage precluding resection.^13–17 Submitting all patients with resectable CRLM to neoadjuvant chemotherapy is an inefficient strategy. Only 5% of patients will progress during treatment, thereby limiting the selection of patients with poor prognosis.⁸ Furthermore, responsiveness, assessed by volumetric response, to chemotherapy is not associated with survival.¹⁸ However, among patients undergoing upfront surgery, there is a subset of patients who rapidly progress and who may have benefited from neoadjuvant chemotherapy or other selection strategies.

This study aimed to investigate the outcomes of a large cohort of patients undergoing upfront surgery for resectable CRLM and identify those experiencing early recurrence as a surrogate for a group who may be more likely to benefit from neoadjuvant chemotherapy or novel selection strategies. Indeed, patients who recur early after surgery have a poor outcome and could be considered as a failure of treatment as this likely represents occult microscopic disease at the time of liver resection.¹⁹ We sought to predict the probability of early treatment failure using clinicopathological characteristics.

MATERIALS AND METHODS

All consecutive patients who underwent upfront liver resection for resectable CRLM at Memorial Sloan Kettering Cancer Center (MSKCC) from 1991 through 2019 were retrospectively analyzed from a prospectively maintained database containing demographic, clinical, operative, pathologic, and follow-up data. A preoperative extent-of-disease evaluation included computed tomography and/or magnetic resonance imaging. ¹⁸F-fluorodeoxyglucose-positron emission tomography/computed tomography (¹⁸FDG PET/CT) was used selectively. Lesions were considered resectable if complete tumor removal with a negative margin was thought to be feasible based on preoperative imaging, leaving enough healthy future liver remnant with appropriate vascular inflow, outflow, and biliary drainage. Two-stage resections were excluded. The primary tumor was removed previously or at the time of liver resection. Patients who received systemic chemotherapy as neoadjuvant treatment were excluded, as were those who were treated exclusively with thermal ablation. Patients who did not receive adjuvant systemic treatment after liver resection were also excluded.

The preoperative carcinoembryonic antigen (CEA) level analyzed was the highest value assessed preoperatively. The number and size of lesions were based on pathologic examination. A clinical risk score (CRS) was calculated for each patient using nodal status of the primary tumor, the disease-free interval from the primary tumor to liver metastases, number of metastases, preoperative CEA level, and size of the largest tumor.²⁰ Patients with scores of 0, 1, or 2 were classified as low CRS and patients with scores of 3, 4, or 5 were classified as high CRS. In the case of adjuvant hepatic artery infusion (HAI) chemotherapy, this was administered concurrently with systemic therapy. Follow-up time was calculated from the date of liver resection to date of the last follow-up or date of death. Overall survival (OS) was calculated based on the survivorship status (deceased or alive) at the last follow-up.

Statistical Analysis

The total cohort of treatment-naïve patients was randomly split into a training dataset group (two-thirds) and a validation dataset (one-third), stratified by year of surgery to ensure appropriate temporal representation. The primary outcome was treatment failure, defined as any recurrence or death within 12 months from surgery. This endpoint was chosen based on previous literature showing poor prognosis in patients with recurrence within 1 year from surgery.^19,21,22 Patients without the 12-month follow-up were excluded. A nomogram for predicting the risk of treatment failure was built on data from the training set and validated on data from the test set. Baseline covariates to be considered in the model included age at resection, sex, primary tumor location, T and N stage, disease-free interval from the primary tumor to metastases diagnosis, and number and size of the largest CRLM. The final nomogram included variables that were significantly associated with the risk of treatment failure from the univariate logistic regression model, as well as primary tumor location, which has been shown to be a risk factor for treatment failure.²³ The number of CRLM was truncated at 10, the largest size of the tumor was capped at 20 cm, and the disease-free interval was truncated at 60 months, as there were very little data outside of these values. To permit a nonlinear relationship, continuous variables in the nomogram were modeled with restricted cubic splines with three knots.²⁴ The discriminatory performance of the nomogram was measured on the validation dataset, using the concordance statistic (c-statistic), i.e. the probability that a randomly selected subject who experienced the outcome will have a higher predicted probability of having the outcome occur compared with a randomly selected subject who did not experience the event.^24,25 The c-statistic is equivalent to the area under the receiver operating characteristics (ROC) curve, with 0.5 corresponding to random chance and 1.0 representing correct predictions for all patients.^24,25 In addition to discriminatory performance, the nomogram was further evaluated with the calibration curve by plotting the predicted 1-year treatment failure against the observed outcome for the validation dataset. If the points fall on or near the 45-degree line, the model is said to have good calibration and the predicted outcome matches the observed outcomes. If the points are above the 45-degree line, the model is said to underestimate outcome probabilities, and if the points fall below the 45-degree line, the model is said to overestimate outcome probabilities.

Finally, we considered patients in the validation cohort with a prediction of treatment failure of ≥0.4 to be in the high-risk group and those with a prediction of treatment failure of <0.4 to be in the low-risk group. This cut-off was applied after a literature search identified overall rates of recurrence within 1 year following hepatectomy for CRLM being around 40% in most series, including our previous paper on a similar population.¹⁹ OS from the landmark time of 12 months was calculated and compared between these two groups using a log-rank test.

All statistical analyses were performed using R 3.6.0 software (R Foundation for Statistical Computing, Vienna, Austria). All p-values were two-sided and p-values <0.05 were considered indicative of a significant difference.

RESULTS

During the study period, 3085 liver resections for CRLM were performed. Overall, 1008 (32.7%) patients proceeded to surgery without neoadjuvant chemotherapy, of whom 225 (22.3%) did not undergo adjuvant systemic treatment and were excluded. Finally, our study population included 783 treatment-naïve patients undergoing liver resection for CRLM and adjuvant chemotherapy (Table 1). The primary colorectal tumor was mostly located in the left colon (50%), followed by the right colon (30%) and the rectum (20%); 589 patients (75.2%) had a primary tumor that was T3 or more and 443 (56.5%) were lymph node-positive.

TABLE 1.

Demographics and clinicopathological characteristics of the overall population and training and validation sets

	Overall population (n = 783)	Training set (n = 535)	Validation set (n = 248)
Year of surgery
1991–2000	386 (49.0%)	266 (50.0%)	120 (48.0%)
2001–2010	220 (28.0%)	149 (28.0%)	71 (29.0%)
2011–2019	177 (23.0%)	120 (22.0%)	57 (23.0%)
Age, years	62 (23–89)	62 (23–88)	63 (26–89)
Male sex	343 (44.0%)	247 (46.0%)	101 (41.0%)
Body mass index, kg/m2	26.8 (15.1–61.7)	27.0 (16.7–61.7)	26.6 (15.1–51.6)
Primary tumor
Rectum	155 (20.0%)	111 (21.0%)	44 (18.0%)
Left colon	385 (50.0%)	262 (50.0%)	123 (50.0%)
Right colon	230 (30.0%)	151 (29.0%)	79 (32.0%)
Primary tumor T stage
T1	37 (4.7%)	22 (4.1%)	15 (6.0%)
T2	102 (13.0%)	70 (13.0%)	32 (12.9%)
T3	511 (65.2%)	354 (66.1%)	157 (63.3%)
T4	78 (9.9%)	56 (10.4%)	22 (8.8%)
Primary tumor N stage
N0	333 (42.5%)	225 (42.0%)	108 (43.5%)
N1	278 (35.5%)	189 (35.3%)	89 (35.8%)
N2	165 (21.0%)	116 (21.6%)	49 (19.7%)
Primary tumor adjuvant treatment	137 (49.0%)	98 (50.0%)	39 (46.0%)
Synchronous liver metastases	333 (43.0%)	228 (43.0%)	105 (42.0%)
Disease-free interval, months	10 (0–180)	10 (0–90)	11 (0–180)
Number of metastases	1 (1–50)	1 (1–50)	1 (1–21)
Metastases size, cm	3.6 (0.2–22)	3.7 (0.2–22)	3.5 (0.4–20)
Preoperative CEA, ng/mL	12 (0–12,325)	12 (0–6870)	13 (0–12,325)
Clinical risk score, low/high	491/240	334/169	157/61
Major hepatectomy	391 (52.0%)	270 (52.0%)	121 (51.0%)
R0 resection	729 (93.0%)	499 (93.0%)	230 (93.0%)
Overall postoperative morbidity	282 (45.9%)	193 (45.9%)	89 (45.8%)
90-day postoperative mortality	5 (0.7%)	3 (0.6%)	2 (0.9%)
Adjuvant hepatic artery pump chemotherapy	222 (28.0%)	146 (27.0%)	76 (31.0%)

CEA carcinoembryonic antigen, ASA American Society of Anesthesiologists

Missing values: BMI, n = 227; ASA, n = 194; primary tumor location, n = 13; primary tumor T stage, n = 55; primary tumor N stage, n = 7; disease-free interval, n = 1; number of metastases, n = 9; metastases size, n = 8; preoperative CEA, n = 112; clinical risk score, n = 52; major hepatectomy, n = 33; postoperative morbidity, n = 169; postoperative mortality, n = 70

After a median of 10 months (range 0–180) from the primary tumor resection, patients presented with a median of 1 (range 1–50) CRLM. The median size of the largest liver metastasis was 3.6 cm (range 0.2–22). Two hundred and twenty-two patients (28.3%) had a recurrence within 1 year of liver resection: 72 (32.4%) of these patients recurred in the liver as the only site, and 150 (67.6%) had a recurrence in distant organs with or without liver involvement. A salvage curative-intent treatment in the form of resection or ablation was utilized in the minority of patients with recurrence within 1 year (84 patients, 37.8%), while the majority were treated with palliative chemotherapy (138 patients, 62.2%). The study population was randomly divided into training (n = 535) and validation sets (n = 248). Demographics and clinicopathological characteristics were similar between the groups (Table 1).

At univariate analysis, primary tumor T stage and lymph node status, the time interval between primary tumor and metastases, and number and size of metastases were significantly associated with treatment failure (Table 2). The final multivariate prediction model is presented in Fig. 1 as a nomogram. The ROC of the nomogram had an area under the curve of 0.68 (95% confidence interval [CI] 0.63–0.73) in the training set (Fig. 1). The nomogram was then applied to the validation set, where the ROC had an area under the curve of 0.60 (95% CI 0.52–0.68).

TABLE 2.

Univariate analysis of predictive factors of treatment failure within 1 year

	Univariate analysis
	HR (95% CI)	p-Value
Age, years	0.99 (0.98–1.01)	0.35
Male sex	0.82 (0.56–1.19)	0.29
Primary tumor
Rectum	–	–
Left colon	1.02 (0.63–1.70)	0.93
Right colon	1.50 (0.88–2.59)	0.14
Primary tumor T stage
T1–T2	–	–
T3–T4	1.91 (1.12–3.41)	0.023
Primary tumor N stage
N0	–	–
N?	2.29 (1.54–3.44)	< 0.001
Disease-free interval, months	0.98 (0.97–0.99)	< 0.001
Number of metastases	1.17 (1.07–1.28)	< 0.001
Metastases size, cm	1.08 (1.02–1.14	0.007
Preoperative CEA, ng/mL
≤200	–	–
>200	1.43 (0.70–2.80)	0.31

Statistically significant values given in bold

CEA carcinoembryonic antigen, HR hazard ratio, CI confidence interval

FIG. 1.

(A) Training set nomogram to estimate the probability of treatment failure in patients undergoing upfront surgery for colorectal liver metastases. (B) Receiver operating characteristic curve in the training set. (C) Calibration plot of the model. CRLM colorectal liver metastases

In the validation cohort, 59 (27%) patients were considered as high-risk based on our nomogram (probability of treatment failure C40%), while 161 (73%) were considered as low risk. High-risk patients had a significantly worse OS from the landmark time of 12 months from surgery (2 years: 82% vs. 70%, low-risk and high-risk, respectively; p = 0.0002) (Fig. 2).

FIG. 2.

Overall survival curve from the landmark time of 12 months from surgery for patients with a high versus low probability of treatment failure

DISCUSSION

In this study, we have shown that 28% of patients with resectable CRLM who undergo upfront resection and adjuvant chemotherapy will develop early recurrence within 12 months from surgery, mostly presenting as extrahepatic disease with a low chance of potentially curative options and a relatively poor long-term outcome. Location, T stage, and lymph node status of the primary tumor, the time interval from the primary and the metastatic disease, and number and size of CRLM were associated with treatment failure.

Despite the belief that the window of ‘curability’ might be missed by delaying surgery with neoadjuvant treatment in resectable CRLM, some argue that due to the significant risk of early systemic dissemination, all patients should undergo chemotherapy first, regardless of the initial resectability of their metastases.^6,26 Indeed, early and rapid progression on neoadjuvant chemotherapy likely identifies a group of patients who would not benefit from surgery. Unfortunately, modern chemotherapy induces hepatic parenchymal damage that significantly affects the postoperative outcomes of surgery.^15,16,27,28 Furthermore, only 5% of patients will progress during neoadjuvant treatment, making neoadjuvant chemotherapy an inefficient method of patient selection. Therefore, defining a group of patients with resectable disease who fail early and do poorly helps to identify a group that requires better preoperative selection mechanisms.

In 2008, Nordlinger et al. reported on a trial that randomized patients with resectable CRLM to either six cycles of perioperative FOLFOX or surgery alone.^8,9 No difference in OS was found.⁹ The reasons why chemotherapy did not show a benefit are unclear but multiple hypotheses could be made. First, chemotherapy likely only delayed inevitable recurrences, thereby having no impact on OS. Second, the resectability rate was the same in both arms and only a very small group of patients (5%) did not undergo surgery due to tumor progression, supporting the idea that neoadjuvant chemotherapy does not select patients with poor prognosis well. Failure to fully account for the heterogeneity of patients presenting with CRLM is likely a major contributing factor. As the inclusion criteria were strict, the disease burden included in the EORTC trial was limited. Indeed, only patients with one to four liver metastases were included; more than 50% had only one have a poor prognosis with a high rate of significant undetectable viable micrometastasis at the time of resection.³² Indeed, most of the early recurrences were extrahepatic in our study, with few patients amenable to potentially curative salvage treatments. Furthermore, it has been shown that factors associated with early recurrences are related to the biology of the tumor (primary T and N stage, disease-free interval, and burden of liver disease), while factors associated with later recurrences are also associated with the surgical procedure itself (R1 resection rate, use of ablation).^36,37 For the above reasons, we believe that a recurrence within 1 year from upfront resection of CRLM can be considered a failure of treatment from an oncological point of view, such that a different approach or more effective mechanism of patient selection could have resulted in a better outcome.

In this study, we have shown that if a patient has an estimated probability of treatment failure of 40% or more (i.e., right-sided, T3N+ colon tumor, with four synchronous metastases, the largest measuring 6 cm), they are more likely to have a worse prognosis and should be considered as high-risk. Retrospective analyses have suggested that neoadjuvant chemotherapy for patients with resectable metastases and high-risk profiles is associated tumor and more than 25% had only two. Since the concept of resectability has dramatically evolved over the last 2 decades, what is currently considered resectable can also be more often biologically aggressive. Although these patients all have stage IV disease by definition, combinations of prognostic factors translate into different tumor biology, risk of recurrence, and survival after resection.^29–31 Patients with high-risk features such as multiple and large metastases may be the ones who might benefit from neoadjuvant chemotherapy and who were underrepresented in the trial. Therefore, risk stratification of patients presenting with resectable CRLM could help to design randomized studies to definitively answer the question. In the current study, we provide a nomogram to estimate the probability of treatment failure in patients who are considered for upfront resection, identifying those with high rates of early recurrence and poor outcomes. Early recurrence has been previously defined by different time cut-offs; some authors considered a short 6-month interval from resections while others lengthened the threshold to 24 months.^32–36 In a previous study, we have shown that in patients who were recurrence-free at 1 year, the probability of surviving to 10 years was 64%, compared with 24% for those who had recurred within the first year (p < 0.0001).¹⁹ Therefore, we consider patients recurring within 1 year to with improved outcomes.³ The CHARISMA trial is an ongoing, multicenter, randomized study comparing surgery with or without neoadjuvant chemotherapy in patients with resectable CRLM and a CRS of 3–5, therefore bearing poor prognosis and being considered as high-risk. The primary outcome was OS, and results are expected with great interest.³⁸ However, the sample size was calculated on a very large, estimated difference in survival, which is unlikely to be confirmed. As learned from previous trials, if a difference in survival exists, it is probably small and will require a large-scale multicenter trial to be demonstrated.

The risk stratification we are proposing in the current study derives from a treatment-naïve population that was selected for surgery only based on the characteristics of the disease, without considering the response to systemic treatment. Previously proposed prediction models for early recurrence were created on a heterogeneous population with both resectable and unresectable CRLM who both underwent neoadjuvant treatment or upfront resection.^20,32–36 Furthermore, intraoperative and postoperative predictors were included, to guide the use of adjuvant chemotherapy or salvage treatments (i.e., repeated resections). To our knowledge, this is the first time a nomogram has been provided from a large sample of a treatment-naïve population. This represents a homogenous group of patients who were all defined technically resectable upfront. Despite this, some recurred early and had a subsequent worse prognosis. Considering that we only used preoperative predictors, the clinician would know upfront the estimated risk of treatment failure for the specific patient. Given the nature of the study, we cannot speculate that high-risk patients would have benefited from neoadjuvant chemotherapy, but we suggest that a high-risk profile according to our nomogram, could serve as a selection for randomized trials investigating the role of such treatment.

This study has some limitations, including its retrospective nature and single-center population. Furthermore, the nomogram can only be applied to patients who are considered for upfront resection and adjuvant chemotherapy. Despite the fact that this represents the most common clinical scenario of resectable CRLM, patients presenting after downstaging chemotherapy, or who are not considered for chemotherapy at all, could not be fitted into our model. It is unclear why 225 patients (excluded from our study) did not undergo chemotherapy either before or after resection, but comorbidities, surgical complications, loss of follow-up, and treatment in the early period of the study could have played a role. Genetics and mutation profiles have been previously linked to prognosis in patients with colorectal cancer and liver metastasis.³⁹ Furthermore, they have been shown to improve prediction of prognosis for patients with CRLM.⁴⁰ Unfortunately, the availability of genomic data was limited to the most recent population in our database and therefore could not be included in our prediction model. In addition, circulating tumor DNA dynamics have recently shown good association with survival in CRLM, identifying patients with high recurrence risk.⁴¹ Unfortunately, data were not available for our patients due to this being a relatively new technology. These are certainly steps forward towards a better understanding of the complex heterogeneity in the outcomes of patients with CRLM. The inclusion of further information about the tumor in models to predict oncological outcomes would improve accuracy and reliability. Tumor genomics and circulating DNA should therefore be implemented in clinical practice to prompt more studies investigating the association with survival and the ability to predict prognosis. Finally, in this study, 222 patients underwent adjuvant hepatic arterial infusion chemotherapy. We acknowledge that this represents a potential source of bias as it could impact the results of this group of patients given the decreased risk of liver and overall recurrence.⁴²

CONCLUSION

Twenty-eight percent of patients who undergo upfront surgery for resectable CRLM recur within 1 year, most often systemically, and therefore fail treatment. Location, T stage and lymph node status of the primary tumor, disease-free interval, and number and size of metastases are associated with treatment failure. Despite the use of adjuvant therapy, survival for patients with a predicted probability of treatment failure of ≥40% is unfavorable. Future randomized trials investigating the role of neoadjuvant therapy or other novel selection strategies in these high-risk patients are warranted