Modeling Medicine: Developing Modern Machine Models for Healthcare

“If you had a specialist who was seeing [similar cases] all day, what would they say?” This is the question that a Yale research group led by Loren Laine, Interim Chief of Digestive Diseases at Yale, seeks to answer in predicting whether or not a patient showing symptoms of upper gastrointestinal bleeding can be discharged from the emergency room.

To make this prediction, the researchers utilized a machine learning model—created by lead author Dennis Shung—based on data from four medical centers as well as a validation dataset from two additional medical centers around the world. Shung’s machine learning model outperformed traditional risk assessment systems, which are based on static datasets, by two-fold. In other words, this new model was able to identify approximately twice as many patients who had no immediate medical need and could be discharged from the hospital. “It is like having another very specialized and experienced assistant in the room,” Shung said.

The team believes that the incorporation of machine learning technologies into medicine possesses great potential. “The beautiful thing about machine learning is that it picks up on a lot of patterns that people would not be able to pick up,” Shung said. The technology works best by matching the comprehensive, lateral thinking of physicians with the extraordinary analytical capacity of the machine models. However, this is not to say that physicians will be completely replaced anytime soon. “You still need a lot of human oversight,” Shung qualified.

With this early success, the researchers hope to expand the scope of their work by accessing electronic medical records in order to boost the data that feeds into the machine model. “Imagine taking it from twenty-four to six thousand variables,” Shung explained.