Double-blind evaluation and benchmarking of survival models in a multi-centre study

Area of Science:

• Biostatistics
• Medical Informatics
• Computational Biology

Background:

• Accurate time-to-event data modeling is vital for cancer research and clinical decisions.
• Predictive models for cancer mortality require rigorous validation.

Purpose of the Study:

• To evaluate the out-of-sample prediction accuracy of artificial neural networks against traditional survival models.
• To benchmark non-linear models against partial logistic spline, log-normal, and COX regression for cancer mortality prediction.

Main Methods:

• A double-blind evaluation using a dataset of 2880 samples shared via the GEOCONDA secure web environment.
• Comparison of predicted survival estimates with Kaplan-Meier empirical estimates across TNM staging groups.
• Quantification of survival prediction accuracy over time using the time-dependent C-index (C(td)) and calibration plots.

Main Results:

• All evaluated models demonstrated satisfactory performance, with time-dependent C-index values around 0.7.
• No systematic over or underestimation of survival was observed at 3 and 5 years follow-up.
• At 10 years, most models underestimated survival, with COX regression showing an overestimate.

Conclusions:

• Advanced flexible modeling algorithms offer comparable predictive performance to established methods for cancer survival data.
• The study provides a robust benchmarking methodology for evaluating survival prediction models.
• Findings support the use of modern computational approaches in cancer prognosis.