*Purpose: UNOS, SRTR, USRDS registries are rich in patient baseline data. However, all suffer from data attrition as patients move further out post-transplant or between different health providers. These data gaps can adversely affect prognostication models of graft and patient survival and so are often augmented by other data sources. Comparative effectiveness research (CER) platforms support dataflow from multiple data centers, enabling integration from a variety of sources into a single system. However, CERs have not yet been developed for transplantation. We outline why this platform provides application for machine-learning (ML) methodology and show how a system can enrich the predictive power and accuracy of transplant data models.

*Methods: Transplant data platform (TDP) utilizes cloud-computing protocols, accessing data stored in the cloud or in-house. Its triple-tier, Apache Spark-based data lakehouse architecture enables the storage of raw, cleaned and aggregated data. There is currently no standardization of patient data storage at different centers. Instead of enforcing a center-level data-standardization principle, TDP resolves the diversity of data-formats into a unified TIDY data matrix. This enables efficient cleaning and indexing of data in preparation for invocation by the integrated ML engine (TensorFlow). ML model-based insights are synthesized in FHIR format and future designs will support federation through a web-based portal.

*Results: TDP provides an augmented intelligence framework the transplant community can utilize. The compiled data can be leveraged by clinicians to support the development of federated data stores. It supports creation of anonymized patient cohorts in a clinically definable manner, as well as through automated feature selection as part of an AI workflow. It supports genotype to phenotype relationship analyses to provide precision medicine insights. Future evolution will enable researchers to share evolving ML models through the web portal to augment clinicians’ patient risk assessment.

*Conclusions: TDP hosts anonymized data from registries, transplant centers and other clinical sources. Its end-to-end architecture allows data ingestion, harmonization, mapping, analysis and dissemination. Being able to support both unstructured and structured data makes it practical to work with hospital EMRs and patient records. This is important as the field considers ways to support patient linkages and longitudinal analyses when building outcome orientation models ensuring secure provenance with strong governance.

« Back to 2021 American Transplant Congress