With machine learning on the rise, DCRI’s Eric Peterson, MD, MPH, takes a retrospective look at prior efforts to implement risk prediction models and what lessons can be drawn from the past for current and future care.
Machine learning holds great potential for improving patient risk prediction; however, clinicians and health systems still will have challenges getting widespread adoption of these predictive models into clinical care, wrote DCRI’s Eric Peterson, MD, MPH, in a recent JAMA Viewpoint.
The use of automated machine learning algorithms for risk prediction could help physicians and patients make more informed care decisions. However, Peterson notes, machine learning does have limitations; in some instances, machine learning models do not outperform traditional regression models, and clinicians are sometimes frustrated that they cannot identify the factors that influence machine learning models’ predictions.
In order to realize machine learning’s full potential in health care, Peterson writes, implementation is critical. He points to some of the first computer generation models that were developed at Duke in 1968. While the Duke prediction tools were proven to perform as well or better than skilled clinicians, the models were never fully implemented in clinical care. One challenge was that physicians were not trained in statistics or probabilistic thinking and so did not see the new models’ value. “This failure to adopt predictive analytics into practice would be relived again and again in medicine,” Peterson notes.
The availability of electronic health records (EHRs) could be helpful during the implementation of machine learning models, but challenges still exist in this space, as well, such as developing data standards that can be used across health systems and formatting the data to ensure it is usable.
Multiple initiatives need to be put into place in order to leverage the power of EHRs combined with machine learning to improve patient care. Initiatives should target the entire health care delivery, from practitioners to patients; better quantitative training is needed to empower clinicians to use novel predictive models, while improved communication tools could better educate patients on their risk.
To read more from Peterson on the implementation of machine learning models, read the Viewpoint piece in JAMA.