Machine Learning to Guide Clinical Decision-making

Using data to guide clinical decisions.

The modern clinician is faced with a deluge of digital information when he or she sits down to discuss clinical care with patients. While much of this information is redundant noise, some could be critical in making decisions. Further, since the information is in a digital format, computer algorithms can be developed to not only help sort signal from noise, but to use the information to find the best treatment approach for each patient. In this project, our goal is to use machine learning to make clinical predictions at the point of care, to help guide providers toward treatments that are more likely to produce a benefit for patients, as well as avoid unnecessary risks. Our team has developed several applications for point-of-care clinical decision support, with ongoing study into the efficacy and safety.

 Our applications include the following:

— A dose-guidance application for initiation and loading of the anti-arrhythmic medication dofetilide. This app uses simple rule-based guidance for selecting the initial dose, and then applies Markov chain simulation to examine the impact of changes in dose over the 3-day loading period.

— A rhythm-management application, called QRhythm, which examines clinical characteristics to determine the optimal rhythm-control strategy for patients to avoid future hospitalizations, stroke, death, and changes in treatment. This work is based on big data analysis of electronic health record data, in which we applied reinforcement learning to demonstrate feasibility of an application that learns in real-time how to find the best option for each patient. QRhythm has been pilot-tested among clinical providers, and we have plans for clinical trials in the coming months.

These projects are presently being conducted on the Anschutz Medical Campus, of the University of Colorado.