Data-centric Artificial Intelligence and Cancer Research: Construction of a Real-World Head and Neck Treatment Data Repository

Background and Purpose

The performance and generalisability of machine learning (ML) models relies on high-quality data. Retrospective and prospective collection of high-quality data for research use whilst respecting data protection and patient privacy remains a challenge in the clinical environment. Currently, months of laborious extraction and clinical annotation are often necessary before data analysis can begin. We present a novel institutional federated data lake, utilising open-source software, to facilitate efficient production of ML models from Head and Neck Cancer (HNC) imaging and Radiotherapy (RT) data. This structured pipeline dramatically reduces the time associated with the production of ML models and real-world evidence generation. This paper describes our governance-compliant processes and provides a framework for establishing similar databases.

Materials and Methods

XNAT, is a powerful open-source imaging platform. Within our department, it forms a part of the local secure enclave for the purposes of federated learning in artificial intelligence projects and provides import, archiving, processing, search and secure distribution facilities for imaging and RT data.

Results

We have created a clinically annotated, carefully curated, data lake of 2,895 consenting HNC patients containing 22,170 relevant diagnostic, staging, treatment and monitoring imaging sets. Key recommendations for replication include infrastructure planning, robust patient and data selection criteria and prioritising patient consent and privacy.

Conclusions

This secure and extensible imaging and HNC RT cancer database set-up promises to be an exceedingly useful tool for research, revolutionising the time and cost associated with the production of ML models, making the process safer, faster and more efficient.

Highlights

• Real-world data is critical for building predictive clinical models.

• Curation of clinical data for machine learning can necessitate months of laborious extraction and annotation.

• Patients have the right to have their data handled with the highest standards of respect, security and governance.

• We describe a database infrastructure constructed under a transparent and safe system of control and stewardship.

• This secure and extensible structured database dramatically reduces data extraction time for AI-driven cancer research.