CDS Alumna Publishes A New Approach To Medical Diagnosis in Nature

CDS alumna Maya Rotmensch and former faculty member David Sontag apply NLP to healthcare

It always starts with a mysterious rash, or a sudden onset of fever and chills.

You type out your symptoms online — only to find website after website listing a gauntlet of diseases that may be responsible for your condition, each sounding as ghastly as the next. Which one is it? Will you live? And, how accurate are these diagnoses anyway?

Automated medical diagnosis tools are poised revolutionize healthcare, but only if we are able to improve their accuracy.

This is why CDS alumna and data scientist Maya Rotmensch, along with former CDS professor David Sontag, Courant graduate student Yoni Halpern, and researchers from Beth Insrael Deaconess Medical Center (BIDMC), are exploring a new approach for inventing better diagnostic tools, and recently published their findings in Nature.

Doctors currently use what are called Health Knowledge Graphs to diagnose their patients, like the one produced by Google.

These platforms aggregate data from medical textbooks, journals, and reliable online content, use statistical analysis and NLP to find relationships between symptoms and diseases, and then diagnose patients by comparing their symptoms to that data.

“However,” the researchers explain, “another potential source of data, currently underutilized, is the electronic medical record (EMR).”

EMRs are physician and nursing freeform notes about each patient’s medical condition, and their highly unstructured nature is why data scientists have, until now, avoided using them.

But, as the researchers point out, it is precisely their unstructured nature that makes these notes so valuable, for they have “the advantage of being closer to the actual practice of medicine than the idealized and curated information presented in textbooks and journals.”

The researchers collected over 200,000 EMRs from BIDMC to create a new health knowledge graph, compared their model against Google’s platform, and found that their graph achieved a strong precision rate, meaning that it is a brilliant candidate for real-life applications.

Read more about their work here.

By Cherrie Kwok

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.