Comparing predictive models of glioblastoma multiforme built using multi-institutional and local data sources

Area of Science:

• Oncology
• Bioinformatics
• Data Science

Background:

• Increasing electronic health data necessitates national repositories for multi-institutional patient data sharing.
• These repositories aim to enhance data mining and predictive modeling for improved patient outcomes.
• Challenges exist in ensuring multi-site data utility and model applicability for non-contributing institutions.

Purpose of the Study:

• To examine the challenges of using National Cancer Institute (NCI) datasets for glioblastoma multiforme (GBM) prognostic modeling.
• To compare prognostic models built with NCI data against those built with single-institution data.

Main Methods:

• Development of multiple prognostic models using both NCI multi-institutional datasets and single-institution data.
• Comparative analysis of model performance and variable selection between the different data sources.

Main Results:

• Overall prognostic model performance was comparable between NCI datasets and single-institution data.
• Significant differences were observed in the specific variables selected for model generation across the data sources.
• This suggests complexities in harmonizing and mapping data resources between different institutional datasets.

Conclusions:

• Utilizing national cancer datasets for predictive modeling requires careful consideration of data harmonization and variable mapping.
• While multi-institutional data can yield comparable model performance, differences in variable selection highlight potential challenges for broader applicability.
• Further research is needed to develop robust methods for data integration and model generalizability across diverse datasets.