Defining and developing machine learning models for early recurrence after curative surgery in stage I to III colorectal cancer: a multicenter cohort study in China

Background: The definition of early recurrence (ER) after curative surgery for colorectal cancer (CRC) is currently unclear, and studies of machine learning architecture in predicting CRC ER are lacking. Objective: To define the ER cutoff in stage I-III CRC patients, develop a machine learning model to predict ER and test the reliability of individual treatment recommendations. Methods: This multicenter cohort study included patients who underwent primary surgery between January 1, 2009, and December 31, 2019. Patients were followed up until recurrence, death, or July 14, 2023. The area under the curve (AUC) value was used to evaluate the optimal cut-off value of ER. Least absolute shrinkage and selection operator (LASSO) regression selected ER related factors. Nine deep learning models were developed and validated to predict ER. Findings: Of 6800 patients included in the study, 1378 had recurrence after the primary surgery. The optimal cutoff value for ER was 30 months (AUC 0.920; specificity 0.950; sensitivity 0.760). The overall survival (OS) of patients with ER was poor compared to that of patients without ER (37.6% vs. 72.7%, P < 0.001). Postoperative adjuvant therapy improved the OS for both stage II/III colon cancer and rectal cancer, but not for stage II rectal cancer without ER. Among nine deep learning models, the Random forest model has the best prediction performance, with promising results for ER prediction in both the training cohort (accuracy=0.999, sensitivity=0.999, specificity=1, F1 score=0.999), test cohort (accuracy=0.776, sensitivity=0.927, specificity=0.799, F1 score=0.868) and independent external validation cohort (accuracy=0.839, sensitivity=0.921, specificity=0.785, F1 score=0.91). The model can predict new datasets through the attachment program. Interpretation: The optimal cutoff for defining ER in stage I-III CRC is 30 months. Postoperative adjuvant therapy is not recommended in stage II rectal cancer without ER. Our machine learning model shows potential benefits in ER evaluation and treatment recommendation.