Accommodating time-varying heterogeneity in risk estimation under the Cox model: a transfer learning approach

Area of Science:

• Biostatistics
• Machine Learning
• Oncology

Background:

• Cancer registries offer large datasets for clinical research.
• Inflammatory breast cancer (IBC) risk estimation requires precise individualized predictions.
• Existing methods lack robust strategies for transferring knowledge between disparate cancer data cohorts.

Purpose of the Study:

• To develop a transfer learning approach for improving individual risk estimation in IBC patients.
• To address the challenge of time-varying heterogeneity between source (cancer registries) and target (single cancer center) cohorts.
• To enhance the precision of death risk prediction using external data sources.

Main Methods:

• A transfer learning framework under the Cox proportional hazards model.
• Lasso penalties applied to regression coefficients and baseline hazards for adaptive information borrowing.
• Jointly solving discrepancies between source and target cohorts for robust risk estimation.

Main Results:

• The proposed method significantly improves individualized risk estimation precision compared to using the target cohort alone.
• The approach demonstrates robustness against cohort differences, outperforming direct data combination.
• A more accurate risk model was developed for the MD Anderson IBC cohort by leveraging the National Cancer Database.

Conclusions:

• Transfer learning offers a powerful strategy for enhancing cancer risk prediction by integrating diverse data sources.
• The developed method effectively handles time-varying cohort heterogeneity for improved prognostic accuracy.
• This approach provides a valuable tool for precise individualized risk assessment in inflammatory breast cancer.