Preoperative prediction of lymph node status in patients with colorectal cancer. Developing a predictive model using machine learning

PURPOSE: Develop a prediction model to determine the probability of no lymph node metastasis (pN0) in patients with colorectal cancer.

METHODS: We used data from four Danish health databases on patients with colorectal cancer diagnosed between 2001 and 2019. The registries were harmonized into one common data model (CDM). Patients with clinical T4 tumors, undergoing palliative or acute surgery, and patients undergoing neoadjuvant therapy were excluded. Preoperative data was used to train the model. A postoperative model including tumor-specific variables potentially available after local tumor resection was also developed. Additionally, both models were compared with a model based on age, sex, and clinical N stage to resemble current standards. A Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression analysis for prediction was used.

RESULTS: In total, 35,812 patients with 16,802 variables were identified in the CDM, and 194 variables affected the probability of pN0 preoperative. The area under the receiver operating characteristic curve (AUROC) was 0.64 (95% CI 0.63-0.66), and the area under the precision-recall curve (AUPRC) was 0.75 (95% CI 0.74-0.76). The mean predicted risk was 0.649, observed risk was 0.650, and calibration-in-large was 0.998. Adding histopathological data from the tumor improved the model slightly by increasing AUROC to 0.69. In comparison, the AUROC of the current standard clinical staging model was 0.57.

CONCLUSION: Using Danish National Patient Registry data in a machine learning-based predictive model showed acceptable results and outperforms current tools for clinical staging in predicting pN0 status in patients scheduled for CRC surgery.