Our summary
The rise of wide-spread electronic health record (EHR) implementation coincided with healthcare advances that rapidly increased the volume and complexity of information clinicians can access about individual patients. Unlocking the promise of electronically-stored healthcare data to improve healthcare across a population of patients requires better development and application of tools that collect and synthesize digitally stored data.
The overall objective of this paper is to propose and demonstrate the value of a framework to evaluate and deploy a machine learning (ML)-based clinical tool that supports a clinician’s independent assessment of patient risk for an adverse event. The resulting ML-based tool identifies relevant medical information in documented EHR-derived real-world data (RWD), displays the data in a clinician-friendly format in context of its impact on adverse event risks and calculates an overall predicted risk level for an adverse event. A key differentiator of this framework is inclusion of a prospective evaluation step. The authors depict an example use case in the oncology setting in order to demonstrate the functionality and utility of the development and assessment framework.
Why this matters
This manuscript published in Nature Digital Medicine is impactful because it is one of the first frameworks for developing and evaluating a fit-for-purpose ML-based clinical predictive tool with retrospective and prospective research. The use of RWD from sources such as EHRs, registries, and claims data for the development of machine learning models that can improve patient outcomes and care delivery is a promising area of research. However, to date the vast majority of such research has been retrospective in nature, which has limited the practical benefits of this technology for patients and providers.
The proposed generalized framework provides guidance on crossing the chasm between research and practice, by positing six key steps that can enable the use of ML-based tools at the point of care effectively, responsibly and practically. The feasibility and value of this framework is demonstrated through a specific application in oncology risk-prediction. The use of ML-based tools to aid the preemptive identification of patients who are at risk for an adverse clinical event could improve overall patient care and safety through more efficient healthcare delivery and the prompting of an early intervention to mitigate severity.