LASSO-derived nomogram prediction model for lymph node metastasis in colorectal cancer: a retrospective analysis

Patients

Data spanning from January 2021 to December 2023 were collected by researchers at Zhuzhou Hospital Affiliated to Xiangya School of Medicine, Central South University. Inclusion criteria encompassed patients with a first diagnosis of CRC during this period. Demographic variables such as age and gender were required, alongside hematological testing indicators including white blood cells, red blood cells, platelets, hemoglobin, neutrophils, lymphocytes, red blood cell distribution width, mismatch repair (MMR), and CRC tumor markers (MLH1, PMS2, MSH2, MSH6, KRAS, BRAF, NRAS, PIK3CA). Additionally, patients underwent colorectal tumor resection either during their first hospitalization or at a later stage of resection in cases with distant metastases. Detailed pathological data including size, gross staging, histologic type, tumour, node, metastasis (TNM) stage, vascular tumor thrombus, and nerve involvement were obtained. The confirmed diagnosis of CRC required validation through at least two imaging examinations or histopathologic diagnosis. Exclusion criteria comprised incomplete information, lack of essential clinicopathological factors, receipt of adjuvant treatments before obtaining pathological information, and presence of other malignant tumors. Ultimately, 869 patients were included for studying diagnostic risk factors in CRC patients with LNM. In the context of CRC, N0 Indicates no regional LNM, while stages N1 and N2 signify LNM. Randomization divided CRC patients into a training set (435 patients) and a validation set (434 patients) using a random number generation algorithm. Specifically, we used the set.seed(123) function in R to ensure reproducibility, and the sample() function was applied to randomly allocate 50% of the patients to the training set and the remaining 50% to the validation set. The randomization was performed without replacement, ensuring that no patient appeared in both sets. The training set was utilized for constructing the nomogram, while the test set was used for validation purposes. This study was approved by the Ethics Committee of Zhuzhou Hospital Affiliated to Xiangya School of Medicine, Central South University, and the ethical approval number was KY2024027-01. We received a waiver of the need for informed consent from participants of our study.

Statistical analysis

Frequency data were presented as counts (percentages), and differences between groups were assessed using the chi-square test. The normality of continuous data was evaluated using the Kolmogorov–Smirnov method. Data that met the criteria of normal distribution and homogeneity of variances were described as Mean ± SD, and independent samples t-test was used for comparisons between groups. Non-normally distributed data or those with uneven variances were described using the median with an interquartile range (M(Q1, Q3)). LASSO and binary logistic regression models were utilized to construct the clinical nomogram. All tests were two-sided with a significance level of α = 0.05, where differences were considered statistically significant when P < 0.05. Statistical analyses were performed using SPSS 26.0 and R (v 4.2.1; R Core Team, 2022) software.

Comparison of clinical factors between training and validation sets

Following the predefined inclusion and exclusion criteria, a total of 869 patients were enrolled in the study, with 435 assigned to the training set and 434 to the validation set using a randomized method. Examination of Table 1 indicated no significant disparities in clinical characteristics between the two sets, suggesting the robustness of the randomization process. The number of all CRC patients with LNM was 380 (43.73%), including 196 (45.06%) in the training set and 184 (42.40%) in the validation set.

Identification of risk factors for LNM in CRC patients

Comparative analysis comparing CRC patients with and without LNM revealed significant differences in various clinical factors (such as histological type, T-stage, M-stage, vascular tumor thrombus, nerve involvement, as well as genetic markers including PMS2, MSH2, KRAS, BRAF, PIK3CA, and counts of leukocytes and neutrophils) between the two groups (P < 0.05, as shown in Table 2). Conversely, gender, age, tumor site, tumor size, general classification, as well as markers such as MLH1, MSH6, MMR, NRAS, and hematological parameters including red blood cells, platelets, hemoglobin, lymphocytes, and red blood cell distribution width did not exhibit significant differences between the two patient groups (P > 0.05, as presented in Table 2). To further assess the relationships between the variables in the training set, Spearman’s rank correlation analysis was conducted. The results showed that most of the factors had correlation coefficients between −0.2 and 0.2, indicating a weak or negligible correlation between the majority of variables. A heatmap illustrating these correlations is presented in Fig. S1. The lack of strong correlations between variables suggests minimal multicollinearity, supporting the robustness of the model.

LASSO regression analysis was utilized to further refine the factors influencing LNM of CRC patients (Simon et al., 2011; Tibshirani, 1997). The “glnmnet” R package was used for variable selection and shrinkage (Friedman, Hastie & Tibshirani, 2010). The independent variables in LASSO regression are a matrix of clinically relevant candidate factors, while the response variable represents the occurrence of LNM in the training set. The penalty parameter (λ) of the model was determined by ten-fold cross-validation. The findings indicate that with 11 variables, the binomial deviance value of the regression model is minimized. The selected variables at this point include: T stage, M stage, vascular tumor thrombus, nerve involvement, PMS2, MSH2, KRAS, BRAF, PIK3CA, leukocyte count, and neutrophil count (Fig. 1).

Development of nomogram for predicting LNM in CRC patients

Multivariate analysis was conducted on the 11 single-factor variables identified through LASSO regression. A binary logistic regression model was then established to predict the incidence of LNM, using all 11 selected variables. The model is expressed as: ${\displaystyle \text{logit}\left(p\right)=-5.645+0.638\cdot \mathrm{T}\text{stage}+0.213\cdot \mathrm{M}\text{stage}+1.194\cdot \text{Vascular tumor}\mathrm{thrombus}+0.701\cdot \text{Nerve involvement}+1.250\cdot \mathrm{PMS}2+1.572\cdot \mathrm{MSH}2+0.533\cdot \text{KRAS}+1.687\cdot \text{BRAF}-1.649\cdot \mathrm{PIK}3\mathrm{CA}-0.145\cdot \text{Leukocyte}+2.093\cdot \text{Neutrophils}}$

The results revealed that T stage, vascular tumor thrombus, PMS2, MSH2, KRAS, BRAF, PIK3CA, leukocyte, and neutrophil were independent predictors for the incidence of LNM (all with P < 0.05, as shown in Table 3). The Hosmer-Lemeshow test chi-square value was 10.206, and the significance was 0.251, indicating that the model had good goodness of fit. For each 1-unit increase in T-stage, the likelihood of lymph node metastasis (LNM) is associated with a 1.892-fold increase. The presence of vascular tumor thrombus is associated with a 3.301-fold higher risk of LNM compared to its absence. Mutations in PMS2, MSH2, KRAS, and BRAF are associated with a 3.489, 4.815, 1.704, and 5.402-fold increased risk of LNM, respectively, compared to non-mutation. Each additional unit of neutrophil count is associated with an 8.113-fold increase in the likelihood of LNM. The presence of a PIK3CA mutation is associated with a 0.192-fold reduced risk of LNM compared to non-mutation. Lastly, for each 1-unit increase in leukocyte count, the likelihood of LNM is associated with a 0.865-fold decrease.

A nomogram was developed to clinically predict LNM using independent predictors (T stage, vascular tumor thrombus, PMS2, MSH2, KRAS, BRAF, PIK3CA, leukocyte count, and neutrophil count) identified from a multivariate logistic regression model (Fig. 2). Among these predictors, T stage (AUC: 0.605), vascular tumor thrombus (AUC: 0.605), leukocyte (AUC: 0.577), and neutrophil (AUC: 0.572) exhibited the highest AUC values. Notably, the predictive performance of the nomogram (AUC: 0.751) surpassed that of individual variables (Fig. 3A). The concordance index (C-index) was calculated as 0.751 (95% CI [0.728–0.774]), indicating good predictive ability. The calibration curve (Fig. 3B) closely matches the reference line, suggesting excellent concordance between predicted and observed values of the nomogram. Decision curve analysis (DCA) revealed that the nomogram provided greater net clinical benefit compared to a single clinicopathological feature, as evidenced by its deviation from the reference lines (Fig. 3C), indicating its superiority in clinical prediction for patients.

Validation of nomogram for predicting LNM in CRC patients

In the validation group, the nomogram exhibited robust predictive performance for LNM. The validation set ROC curve (Fig. 4A) showcased the model’s discriminative prowess in distinguishing patients with and without LNM. Notably, the nomogram achieved an AUC value of 0.710, surpassing individual clinical indicators, underscoring its good discriminatory ability. The calibration curve (Fig. 4B) further underscored the alignment between predicted probabilities and observed outcomes, indicating the model’s reliability. Additionally, the DCA curve (Fig. 4C) depicted the clinical utility of the model across various threshold probabilities, demonstrating enhanced net benefit compared to single clinicopathological factors. These findings collectively underscore the robust discriminative power, calibration, and significant clinical utility of the prediction model. They suggest its potential to improve accurate risk assessment and guide treatment decisions for CRC patients with LNM, thereby profoundly impacting clinical decision-making.