Building Flexible, Scalable, and Machine Learning-Ready Multimodal Oncology Datasets.

Area of Science:

• Biomedical Informatics
• Computational Oncology
• Data Science

Background:

• Rapid growth of heterogeneous medical data necessitates integration for comprehensive disease understanding and treatment optimization.
• Complex diseases like cancer require integrated data for precision medicine and personalized therapies.
• Existing biomedical data silos limit research capabilities and hinder the development of advanced analytical models.

Purpose of the Study:

• To propose the Multimodal Integration of Oncology Data System (MINDS), a flexible, scalable, and cost-effective metadata framework.
• To efficiently fuse disparate data from public sources into an interconnected, patient-centric framework.
• To overcome limitations of current data silos and advance oncology data integration for research and clinical applications.

Main Methods:

• Developed MINDS, a cloud-native metadata framework for consolidating multimodal oncology data.
• Integrated over 41,000 cases from public repositories like the Cancer Research Data Commons (CRDC).
• Implemented features for data provenance tracking, auto-scaling, access controls, and dynamic data access.

Main Results:

• Consolidated over 41,000 cases, achieving high compression ratios from 3.78 PB of source data.
• Achieved sub-5-second query response times for interactive data exploration.
• Provided an interface for exploring cross-data type relationships and building cohorts for machine learning models.

Conclusions:

• MINDS offers a scalable, cost-effective solution for integrating heterogeneous oncology data, overcoming existing silos.
• The framework empowers researchers with enhanced analytical capabilities for uncovering diagnostic and prognostic insights.
• MINDS represents a significant advancement toward evidence-based personalized cancer care and future data integration strategies.