Prediction of 5-year postoperative survival and analysis of key prognostic factors in stage III colorectal cancer patients using novel machine learning algorithms

Abstract

Objective: This study explores the predictive value of clinical and socio-demographic characteristics for postoperative survival in stage III colorectal cancer (CRC) patients and develops a 5-year postoperative survival prediction model using machine learning algorithms.

Methods: Data from 13,855 stage III CRC patients who underwent surgery were extracted from the SEER database. Key variables, including marital status, gender, tumor location, histological type, T stage, chemotherapy status, age, tumor size, lymph node ratio, and others, were collected. Data were split into a 7:3 training-validation ratio. Optimal cutoff points for age, tumor diameter, and lymph node ratio were determined using X-tile software. Independent prognostic factors for postoperative survival in stage III colorectal cancer patients were identified through univariate and multivariate logistic regression as well as Lasso regression analyses. These factors were incorporated into machine learning models, including logistic regression, decision tree, LightGBM, and others. ROC curves, calibration curves, and decision curve analysis were used to assess model performance. External validation was performed using data from Shanxi Bethune Hospital.

Results: Optimal cutoff points were identified for age (65, 80 years), tumor size (29 mm, 74 mm), and lymph node ratio (0.11, 0.49). Both multivariate logistic regression and Lasso regression consistently identified marital status, tumor location, histological type, T stage, chemotherapy, radiotherapy, age, maximum tumor diameter, lymph node ratio, serum carcinoembryonic antigen (CEA) level, perineural invasion, and tumor differentiation as independent prognostic factors for 5-year postoperative survival in patients with stage III colorectal cancer (P < 0.05). The models showed excellent predictive performance with AUC values ranging from 0.766 to 0.791 in the validation cohort. Age, lymph node ratio, chemotherapy, and T stage were key factors. External validation confirmed model accuracy and clinical applicability.

Conclusion: This study developed and validated an interpretable machine learning model that predicts the 5-year postoperative survival of stage III CRC patients, offering potential for personalized treatment plans.

Keywords: SEER database; colorectal cancer; machine learning; prognostic model; survival prognosis.

Figure 1

Patient selection flowchart from the SEER database.

Figure 2

Optimal cutoff values for age (A), maximum tumor diameter (B), and lymph node ratio (C) determined by X-tile analysis.

Figure 3

Lasso regression path plot.

Figure 4

Nomogram.

Figure 5

Performance evaluation of different machine learning models. (A) ROC curves of each model in the training set; (B) ROC curves of each model in the validation set; (C) Calibration curves of each model in the training set; (D) Calibration curves of each model in the validation set; (E) DCA curves of each model in the training set; (F) DCA curves of each model in the validation set.

Figure 6

Feature importance ranking based on the LightGBM model.

Figure 7

SHAP summary plot for the LightGBM model.

Figure 8

Interactive SHAP plot for surviving patients.

Figure 9

Interactive SHAP plot for deceased patients.

Figure 10

Kaplan-Meier survival curves for different clinical subgroups. (A) Survival curves by age group; (B) Survival curves by LNR levels; (C) Survival curves by chemotherapy status; (D) Survival curves by T stage.

Figure 11

(A) ROC curve of Logistic model in external validation cohort. (B) ROC curve of LightGBM model in external validation cohort. (C) Calibration curves of the Logistic models in the external validation cohort. (D) Calibration curves of the LightGBM models in the external validation cohort. (E) DCA of the Logistic models in the external validation cohort. (F) DCA of the LightGBM models in the external validation cohort.