Journal Article

Dynamic Prediction of Rectal Cancer Relapse and Mortality Using a Landmarking-Based Machine Learning Model: A Multicenter Retrospective Study from the Italian Society of Surgical Oncology—Colorectal Cancer Network Collaborative Group

Persiani, Roberto, Ferrero, Alessandro, Piccoli, Micaela, Reddavid, Rossella, Barberis, Andrea, Pedrazzani, Corrado, Pecchini, Francesca, Gili, Simona, Divizia, Andrea, Polastri, Roberto, Ortenzi, Monica, Foppa, Caterina, Cianflocca, Desiree, Sica, Giuseppe S., Pafundi, Donato P., Azzolina, Danila, Coco, Claudio, Anania, Gabriele, Bianco, Francesco, Biondi, Alberto, Turri, Giulia, Borghi, Felice, Manigrasso, Michele, Guerrieri, Mario, De Nardi, Paola, Sasia, Diego, Costanzo, Federico, Delrio, Paolo, Elmore, Ugo, Gallo, Gaetano, Milone, Marco, Moro, Jacopo, Solaini, Leonardo, Sammarco, Giuseppe, Mineccia, Michela, Spinelli, Antonino, Rega, Daniela, D’Ugo, Domenico, Rosati, Riccardo, Cavaliere, Davide, Maiello, Fabio

2025

Overview

Background: Almost 30% of patients with rectal cancer (RC) who submit to comprehensive treatment experience relapse. Surveillance plays a leading role in early detection. The landmark approach provides a more flexible and dynamic framework for survival prediction. Objective: This large retrospective study aims to develop a machine learning algorithm to profile the patient prognosis, especially the risk and the onset of RC relapse after curative resection. Methods: A cohort of 2450 RC patients were analyzed using landmark analysis. Model A applied a classical cause-specific Cox approach with a landmarking approach, while Model B implemented a landmarking-based RSF (random survival forest) competing risk algorithm. The two models were compared in terms of predictive and interpretative ability. A bootstrapped validation strategy was employed to validate the model’s performance and prevent overfitting. The best-performing hyperparameters were selected systematically, ensuring the model’s robustness within the landmark approach. The study assessed these factors’ importance and interactions using RSF and compared the predictive accuracy to that of the classical Cox model. Results: Model B outperformed Model A (mean C-index 0.95 vs. 0.78), capturing complex interactions and providing dynamic, individualized relapse predictions. Clinical factors influencing survival outcomes were identified across time with the landmark approach allowing for more accurate and timely predictions. Conclusions: The landmark approach offers an improvement over traditional methods in survival analysis. By accommodating time-dependent variables and the evolving nature of patient data, this approach provides a precise tool for profiling RC survival, thereby supporting more informed and dynamic clinical decision-making.

• Publisher: MDPI AG,MDPI
• Subject: Algorithms / Cancer / Cancer therapies / Care and treatment / Colorectal cancer / Colorectal carcinoma / Colorectal surgery / Datasets / Decision making / Development and progression / Health aspects / Laparoscopy / Learning algorithms / Machine learning / Medical prognosis / Mortality / Oncology / Patient outcomes / Patients / Predictions / Prognosis / Quality of life / Rectum / Relapse / Surgery / Surveillance / Survival / Survival analysis / Tumors