Helping Doctors Doctor: Using AI to Automate Documentation and "De-Autonomate" Health Care - Johnson, Kevin DP1 Details
Project Name: Helping Doctors Doctor: Using AI to Automate Documentation and “De-Autonomate” Health
Care
Project Summary
Clinical encounter documentation is one of the most time-consuming tasks of the ambulatory encounter,
taking approximately two hours for every hour spent with patients. Clinical note lengths in the US are longer
than those in other countries due to requirements to justify billing and complete quality care metrics. Not
surprisingly, clinical encounter documentation has become a major source of clinician burnout over the past
two decades. Although companies are selling technologies to transcribe what was discussed during the
encounter, these costly solutions reproduce what is already suboptimal about how we create and use EHR
information. There is a desperate need to reimagine the process of documenting as well as the content of
documenting a clinical encounter. In this application I propose to develop a new generation of automated
documentation algorithms—algorithms that can listen to the dialog between a patient and clinician, collect
quantitative data about these observations, combine those with existing electronic health record (EHR) data
and create relevant encounter summary information. These documentation algorithms will leverage the
remarkable progress we have made in computer vision, natural language processing, machine learning to
support image labeling, and other advances using EHR data. These novel computational approaches have yet
to be explored as alternative approaches to summarizing medical data collected in real time. As a pediatrician
and biomedical informatician who has acquired considerable expertise in real-world systems design,
implementation and medical data analytics, this project leverages many of my skills, though it is a departure
from my previous human-computer interface work. Rather, the goal of this project is to remove the burden of
documentation from clinicians to the extent possible. To achieve this goal, I will work with a multidisciplinary
group of collaborators, including computer scientists, technology engineers, and clinical domain experts.
Specifically, I will: (1) collect and analyze exam room video and annotations of the encounters to identify
salient characteristics of patients and their interaction with the clinician that led to specific diagnoses; (2)
apply natural language processing, deep learning, and computer vision methods to learn and characterize
patterns from vast streams of data using supervised and unsupervised learning methods. Combining these
techniques to directly impact what is documented and how it is generated is a new area of investigation for
me and an approach that promises to support innumerable other projects including identifying implicit bias in
clinical encounters, enabling a new class of real-time decision making and improving the usefulness of
encounter summaries.
