Developing and Validating a Survival Prediction Model for NSCLC Patients Through Distributed Learning Across 3...

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Survival prediction tools for non-small cell lung cancer (NSCLC) patients undergoing chemoradiation or radiation therapy are currently limited in quality.
• Accurate survival prediction is crucial for treatment planning and patient management in NSCLC.

Purpose of the Study:

• To develop and validate a predictive model for 2-year survival in NSCLC patients treated with curative intent.
• To demonstrate the feasibility and effectiveness of a distributed learning approach for building predictive models from multi-institutional data.

Main Methods:

• A Bayesian network model was adapted for distributed learning using clinical data from 698 NSCLC patients across three institutions (Maastro Clinic, University of Michigan, The Christie).
• The model was trained on data from 559 patients and validated on 196 external patients, with 2-year post-treatment survival as the endpoint.
• Publicly available datasets and developed models are accessible for further research and application.

Main Results:

• The developed distributed learning model achieved an Area Under the Curve (AUC) of 0.66 on the external validation set.
• This performance was significantly better than a model based solely on T and N category (AUC 0.47, P<.001).
• Centralized and distributed learning approaches showed minimal differences in model performance, indicating the viability of distributed methods.

Conclusions:

• Distributed learning enables the development of predictive models using data from multiple institutions while mitigating data-sharing barriers.
• This approach represents a promising future direction for data sharing and collaborative research in healthcare.
• The developed model serves as a proof of concept for leveraging federated databases in clinical prediction.