The prognostic value of lymph node yield in the earliest stage of colorectal cancer: a multicenter cohort study

This is a multicenter retrospective cohort study. Patients diagnosed with T1 CRC in 13 participating hospitals (1 academic and 12 non-academic hospitals) between 1 January 2000 and 31 December 2014 were selected from the Netherlands Cancer Registry. The electronic medical records of all patients were reviewed. Only cases in which the local pathologist clearly confirmed the diagnosis T1 CRC in the pathology report were selected, which was defined as tumors with invasion through the muscularis mucosa and into, but not beyond, the submucosa [22]. Patients were included if they were treated with surgical resection of T1 CRC. Transanal endoscopic microsurgery was considered an endoscopic treatment, as no lymphadenectomy is performed. Exclusion criteria were hereditary predisposition for CRC, inflammatory bowel disease, synchronous CRC (defined as CRC in the previous 5 years before detection of T1 CRC, or elsewhere in the colorectum at the time of detection of T1 CRC), non-CRC-related death within 1 year, non-adenocarcinoma, neo-adjuvant radiotherapy, unknown number of retrieved LNs, and AJCC stage IV T1 CRC at diagnosis.

This study was approved by the Medical Ethics Review Committee of the University Medical Center Utrecht (reference number 15-487/C) and was carried out in accordance with the Helsinki Declaration. The study conforms to the STROBE guideline for cohort studies [23].

In each participating center, study variables were collected from the electronic medical records, and the corresponding endoscopy, surgery, pathology, and radiology reports. Primary endpoint was incidence of recurrent cancer, either local or distant. Local recurrence was defined as malignant tissue at the site of the anastomosis. Distant recurrence was defined as metastasis to extra-colonic organs, bone or peritoneum confirmed with imaging or histology. A new primary CRC elsewhere in the colon or rectum was defined as a metachronous lesion, not as recurrence. Secondary endpoint was prevalence of LNM at time of surgery, defined as positive LNs in the resection specimen as reported in the pathology report.

The determinant of interest was number of retrieved LNs, as reported in the pathology reports. We dichotomized LN yield using 10 retrieved LNs as cut-off, as this is the lowest recommended minimum for AJCC stage II CRC in current (inter)national guidelines [9, 11, 12].

Established risk factors (based on previous literature) for recurrent cancer and LNM were considered potential confounding factors [3, 4, 7, 24]. Data on potential clinical confounders were collected from the medical records and endoscopy reports, and included age, sex, tumor location (right colon vs. left colon vs. rectum), tumor size and tumor morphology (pedunculated vs. non-pedunculated) [3, 7, 24]. Right colon was defined as caecum, ascending and transverse colon including the splenic curve. Left colon was defined as the descending and sigmoid colon. Tumor morphology was defined as pedunculated if the presence of a stalk or Paris 0-Ip was reported in the endoscopy report [25]. Flat and sessile tumors were defined as non-pedunculated T1 CRCs [26]. Potential histological confounders were collected from the pathology reports, and included lymphovascular invasion (absent vs. present), depth of submucosal invasion (superficial vs. deep invasion), and grade of differentiation (good vs. moderate vs. poor) [4]. Deep submucosal invasion was defined as invasion depth ≥ 1 mm or sm2/3 for non-pedunculated T1 CRC, and Haggitt 4 for pedunculated T1 CRC [4]. For the endpoint recurrent cancer, the list of potential confounders was extended with LNM as a confounding factor [7].

Additional patient characteristics (not causally related to recurrent cancer or LNM and therefore not considered potential confounders) and follow-up characteristics were collected from medical records. Patient characteristics were body mass index (calculated as weight in kilograms divided by height in meters squared) and comorbidity according to the American Society of Anesthesiologists Physical Status classification [27]. Follow-up was performed according to routine clinical care. Follow-up started at the date of diagnosis and ended at the date of detection of recurrence, death, or last follow-up.

Categorical data were expressed as frequencies and percentages; continuous variables as means with standard deviation (SD) or medians with interquartile range (IQR).

Primary aim of this study was to evaluate the association between number of retrieved LNs and recurrent cancer. Despite our large multicenter cohort of patients spanning an inclusion period of many years, the absolute number of recurrences was expected to be small, since recurrent cancer after surgical resection of T1 CRC is a rare event. We therefore used propensity scores allowing for the adjustment for more confounders than would have been feasible using standard multivariable adjustment approaches [28‐30]. Inverse probability weighting (IPW) was used to account for baseline differences in predictive characteristics between patients with LN yield < 10 vs. ≥ 10. A propensity score was derived by fitting a logistic regression model with the dichotomized number of retrieved LNs as the dependent variable and potential confounders as predictors. We adjusted for confounders in a two-step approach. For the primary analysis, we adjusted for clinical confounding factors (i.e., age, sex, tumor location, tumor size, tumor morphology, and presence of LNM). Secondary supporting analyses were performed to additionally correct for histological confounding factors (i.e., lymphovascular invasion, depth of submucosal invasion, grade of differentiation), which were not included in the primary analysis as they were missing in a considerable number of cases (Table 1). Continuous variables (i.e., age and tumor size) were fitted in the propensity models by restricted cubic spline functions, dummies were used for categorical variables (no interaction terms). Then, the inverse of the (propensity score-derived) predicted probability for actual LN yield was used to weigh patients in a Cox regression model relating the dichotomized LN yield as the sole determinant of recurrent cancer, yielding hazard ratios (HR) for a high (≥10) vs. low (< 10) retrieval of LNs. Patients who did not develop recurrent cancer were censored at the last follow-up moment or date of death. No violation of the proportionality of the hazard assumption was observed following inspection of the scaled Schoenfeld residuals. Bootstrapping was performed to obtain two-sided P values and 95% confidence intervals (CI) for IPW-derived estimates.

Assuming that a low LN yield risks understaging and these understaged patients would fail to receive chemotherapy that could prevent recurrence [31], we additionally repeated our main analysis restricted to patients in whom no LNM were observed [32]. Analyses were performed in the same manner with IPW analysis as described above, adjusting for the same potential confounders.

Furthermore, to assess the sensitivity of our results for the way we dichotomized LN yield in our analyses, the number of retrieved LNs was also analyzed continuously in univariable regression analyses. Moreover, sensitivity analyses were performed using 12 retrieved LNs as cut-off (using similar analysis as used for 10 retrieved LNs as cut-off), as this cut-off is often used as a quality measure for adequate staging in AJCC stage II CRC [9].

Several checks were performed to evaluate the appropriateness of the IPW analysis, namely (1) baseline characteristics before and after IPW adjustment for confounders were compared; (2) the c-index of the propensity models was assessed using the propensity model’s own IPWs (a c-index of 0.5 indicates a successfully obtained balance of confounders after adjusting; a c-index of 1.0 indicates extreme remaining imbalance); and (3) the maximum weight of a single patient used in the IPW-adjusted analysis to obtain balance in potential confounders was assessed as a quality instrument to assess whether a single or a few cases influence the risk estimate excessively [33, 34].

The secondary aim was to explore the association between LN yield and detection of LNM. Like recurrence, a low number of LNM were observed. Therefore, analyses were performed in the same manner with IPW analysis as described above with the presence of LNM as the endpoint, correcting for the same confounders except for the presence of LNM itself. Inverse probability weighted logistic regression analysis was used yielding odds ratios (OR) for a high (≥ 10) vs. low (< 10) retrieval of LNs on the outcome LNM.

Information on LN yield and the presence of LNM was available for all patients; however, several clinicopathological confounding variables had missing data (Table 1). As simply excluding patients with missing data is inefficient and increases the risk of selection bias, we used multiple imputation before data analysis [35]. Missing data was assumed to be missing at random. Multivariate imputation by chained equations (10 imputation datasets, 25 iterations, healthy convergence) was performed [36]. Rubin’s rules were used to pool results across imputation datasets [37]. Percentage of complete cases (i.e., percentage of patients with no imputed values) were reported for each analysis, together with the percentage of observed data points. For the primary outcome, a sensitivity analysis excluding patients with missing values was performed (complete-case analysis) to determine whether this agreed with the imputed results (Additional file 1: Table S1).

A total of 2253 patients with T1 CRC diagnosed between 2000 and 2014 were identified in the participating hospitals. A total of 1017 patients treated for pT1 CRC with surgical resection remained eligible for analysis (Fig. 1). Mean age of the cohort was 69.1 years (SD 9.6) and 54.6% of patients were male. A total of 405 patients (39.8%) had a LN yield ≥ 10. An overview of the baseline characteristics of patients with a LN yield < 10 vs. ≥ 10 before and after IPW adjustment for the confounders considered for the primary analysis (adjustment for clinical factors) and the secondary supporting analysis (adjustment for clinical and histological factors) is provided in Table 1. Before adjustment by IPW, patients with a LN yield ≥ 10 had T1 CRCs that were more often located in the right colon (37.5% vs. 15.2%, P < 0.001), had more often a non-pedunculated morphology (75.9% vs. 60.7%, P < 0.001), had a larger tumor size (2.9 vs. 2.6 cm, P = 0.003), more often had LNM (11.6% vs. 6.0%, P = 0.002), and less often showed lymphovascular invasion (20.2% vs. 29.9%, P = 0.04). Following adjustment, baseline characteristics between groups were comparable.

Median LN yield was 7 (IQR 3–12; range 0–52). Median LN yield increased over time from 4 (IQR 2–8) from 2000 to 2009 (N = 574) to 11 (IQR 7–15) from 2010 to 2014 (N = 443) (P < 0.001). If a minimum of 10 LNs retrieved was considered an adequate resection, this threshold value was achieved in 19.5% (112/574) of patients treated before 2010 versus 66.1% (293/443) of patients treated from 2010 onwards (P < 0.001). A scatterplot of LN retrieval over the years is presented in Fig. 2.

During the 4581 person-years of follow-up (median 49.0 months; IQR 19.6–81.5), 41 patients (4.0%) were diagnosed with recurrent cancer, corresponding to 9.0 events (95% CI 6.5–12.0) per 1000 person years of follow-up (Table 2). Recurrences occurred most frequently in distant organs (N = 23), followed by local (N = 11) or both local and distant recurrences (N = 7). Distant metastases were most often located in lung (N = 13) and liver (N = 16). Recurrences were detected after a median duration of 26.9 months (IQR 13.9–44.4).

Among 612 patients in whom < 10 LNs were retrieved, 37 patients (6.0%) developed recurrence. Among 405 patients in whom ≥ 10 LNs were retrieved, 4 patients (1.0%) developed recurrence (Table 2). In univariable analysis, LN yield ≥ 10 was associated with decreased risk for recurrence (unadjusted HR 0.27; 95% CI 0.10–0.76; P = 0.01) (Table 3). After adjustment for clinical confounders (i.e., age, sex, tumor location, tumor size, tumor morphology, and presence of LNM), LN yield ≥ 10 remained associated with a decreased risk for recurrence (adjusted HR 0.19; 95% CI 0.06–0.60; P = 0.005). Findings persisted after further adjusting for histological factors (adjusted HR 0.20; 95% CI 0.06–0.67; P = 0.009). Sensitivity analysis excluding patients with imputed values did not alter the outcomes (Additional file 1: Table S1).

In patients without LNM (N = 933 patients; N = 35 recurrences), LN yield ≥ 10 remained associated with decreased risk for recurrence, also after adjusting for clinicopathological factors (adjusted HR 0.23; 95% CI 0.06–0.81; P = 0.02). Kaplan–Meier curves of recurrence-free patients stratified for LN yield (< 10 vs. ≥ 10) and LNM (presence vs. absence) are presented in Fig. 3.

To assess the sensitivity of our results for the way we dichotomized LN yield, sensitivity analyses were performed with the number of retrieved LNs without thresholds and using 12 LNs as cut-off. Without thresholds, LN yield remained associated with a decreased risk for recurrence (per retrieved LN: HR 0.90; 95% CI 0.84–0.97; P = 0.007). Of the 749 patients in whom < 12 LNs were retrieved, 37 patients (4.9%) developed recurrences. Of the 268 patients in whom ≥ 12 LNs were retrieved, 4 patients (1.5%) developed recurrences. Similar to a LN yield ≥ 10, the HR of a LN yield ≥ 12 pointed towards a decreased risk for recurrence, albeit non-significant (adjusted HR 0.41; 95% CI 0.12–1.34; P = 0.14) (Additional file 1: Table S2).

LNM were detected in 84 patients, corresponding to a prevalence of 8.3% (84/1017) (Table 2); 1 positive LN was detected in 54 patients, 2 in 18 patients, 3 in 4 patients, and ≥ 4 in 8 patients, with a maximum of 8 positive LNs per patient. In 6 patients with LNM, recurrence occurred during follow-up. In univariable analysis, the number of retrieved LNs (without thresholds) was associated with increased detection of LNM (per retrieved LN: OR 1.04; 95% CI 1.01–1.07; P = 0.006).

LNM were detected in 37 patients (6.0%) with a LN yield < 10 and in 47 patients (11.6%) with a LN yield ≥ 10 (Table 2). In univariable analysis, LN yield ≥ 10 was positively associated with increased detection of LNM (unadjusted OR 2.04; 95% CI 1.30–3.20; P = 0.002) (Table 4). After adjusting for clinical confounding factors, LN yield ≥ 10 remained associated with increased detection of LNM (adjusted OR 2.09; 95% CI 1.28–3.42). Further adjusting for histological confounding factors did not alter the outcomes (adjusted OR 2.27; 95% CI 1.39–3.69; P = 0.001).