Open-source distributed learning validation for a larynx cancer survival model following radiotherapy

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Personalized treatment relies on validated prediction models.
• External validation is crucial for safe and effective implementation.
• Sharing patient data between institutions is hindered by privacy and legal barriers.

Purpose of the Study:

• To validate a larynx cancer survival model using distributed learning.
• To assess the feasibility of external model validation without data sharing.
• To ensure patient data privacy and institutional governance.

Main Methods:

• Developed open-source distributed learning software based on a stratified Cox proportional hazard model.
• Validated the Egelmeer et al. MAASTRO survival model across two international hospitals.
• Optimized a single scaling parameter and ensured no information leakage.

Main Results:

• Validated the larynx cancer survival model on 1095 patients (initially 1745).
• Achieved Harrell C-indices of 0.74 and 0.70 for the two centers.
• Identified a need for a scaling update and noted reduced precision for hypofractionation predictions.

Conclusions:

• Distributed learning successfully validated a survival model without central access to sensitive patient data.
• The original MAASTRO model performed well, comparable to its initial validation.
• Open-source software facilitates distributed learning and external model validation.