Dynamic Prediction of Rectal Cancer Relapse and Mortality Using a Landmarking-Based Machine Learning Model: A...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Rectal cancer (RC) recurrence impacts nearly 30% of patients post-treatment.
• Early detection through surveillance is crucial for managing RC relapse.
• Traditional survival prediction methods lack dynamic adaptability.

Purpose of the Study:

• To develop a machine learning algorithm for predicting rectal cancer prognosis and relapse risk.
• To profile patient risk and timing of relapse after curative resection.
• To compare a novel landmarking approach with classical survival analysis methods.

Main Methods:

• Analysis of 2450 rectal cancer patients using landmark analysis.
• Comparison of a Cox model (Model A) with a landmarking-based Random Survival Forest (RSF) competing risk algorithm (Model B).
• Bootstrapped validation and systematic hyperparameter tuning for model robustness.

Main Results:

• Model B (RSF) demonstrated superior predictive accuracy (C-index 0.95) compared to Model A (Cox, 0.78).
• The landmarking approach enabled dynamic, individualized predictions of RC relapse.
• Identified key clinical factors influencing survival outcomes over time.

Conclusions:

• The landmark approach enhances survival analysis for rectal cancer by incorporating time-dependent variables.
• This method provides a more precise and dynamic tool for patient prognosis.
• Supports improved clinical decision-making for rectal cancer management.