Machine- and Human-learned Helynx Knowledge Graph for Clinical Trials

Author: Chess Stetson

An automated solution has long been sought for the problem of assigning cancer patients to clinical trials for which they qualify. Unstructured data from the patient's electronic medical record and the clinical trial documentation could be mined for features that would indicate a match between patient and trial. NLP/ML based systems of this kind have been reported. Yet such a system has yet to find widespread adoption in the clinical community, possibly because a recommender system requires a very high degree of accuracy in order to gain trust in the clinical community. We report a machine- and human-based method for parsing clinical inclusion/exclusion criteria from unstructured text with very high accuracy (ROC>.95) developed with Helynx software. We embed the clinical data into a human- and machine-readable Helynx data structure, and then enable a process of recursive machine-learning and human feature-selection, while maintaining a held-out test set to prevent snooping. A system like this could be employed to recommend more accurate matches between patients and clinical trials than have been previously possible.

Co Author/Co-Investigator Names/Professional Title: Chess Stetson, Ph.D., Boris Revechkis, PhD., Ted Giardello, Jae Kim, MD

Funding Acknowledgement (If Applicable): Bonnie J. Addario Lung Cancer Foundation