Random survival forests-based survival prediction for spinal chordomas

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Medicine

Background:

• Spinal chordomas are rare, malignant tumors with a high risk of recurrence and metastasis.
• Accurate survival prediction for spinal chordomas is challenging due to complex clinical and histological factors.

Purpose of the Study:

• To develop and validate a Random Survival Forest (RSF) model for predicting survival in spinal chordoma patients.
• To compare the performance of the RSF model against the traditional Cox proportional hazards model.

Main Methods:

• Retrospective analysis of 679 spinal chordoma patients from the SEER database (2000-2018).
• Construction of predictive models using RSF and Cox regression with clinical/demographic variables.
• Assessment of model performance via calibration (iBS), discrimination (iAUC, C-index), and clinical utility (DCA).

Main Results:

• The RSF model showed superior predictive performance (C-index: 0.790, iAUC: 0.731, iBS: 0.130) versus the Cox model (C-index: 0.770, iAUC: 0.679, iBS: 0.148).
• RSF demonstrated improved time-dependent predictions and clinical utility, with age and surgical intervention as key prognostic factors.
• RSF effectively stratified patients into distinct risk groups, validated by Kaplan-Meier analysis.

Conclusions:

• The RSF model significantly advances survival prediction for spinal chordomas, outperforming traditional methods in accuracy and clinical utility.
• Integrating the RSF model into practice may enhance personalized treatment strategies and improve patient outcomes.
• External validation is recommended to confirm the generalizability of the RSF model.