Dr. Herzenstube is a practicing family physician who can make nothing of it.
AMIA is the professional society for health informatics. The AMIA annual symposium is the largest scientific informatics conference in the US. It brings together researchers, policymakers, industry leaders, and practitioners of health informatics from dozens of countries. I have been attending regularly since 2000 and it has been amazing to see the attendees and conference content grow in diversity as clinical information systems become more widespread.
AMIA always offers tutorial learning sessions before the official start of the conference and I have always tried to attend at least one. The chance to take a full day to participate in a structured, deep learning activity, taught by experts in the field, is a rare joy.
In line for coffee, I struck up a conversation with another tutorial attendee, a neonatologist at a major medical center. He also has a degree in informatics and spends “as much time as they’ll let me” on applied informatics projects in his institution, though there is no dedicated informatics department or team. Much of his time is spent working on their Epic system, into which he says they have “shoehorned” their neonatology workflows.
He is here to attend a CMIO workshop, hoping to learn ways to elevate his level of influence within his organization. It is heartening to see someone dedicated enough to the promise of informatics to push on against the headwind of an organization that doesn’t yet know how to effectively use him, but it is a shame that those headwinds are still so prevalent.
At the tutorial, I found myself sitting next to one of the luminaries of the informatics field, someone who has occupied most of the leadership positions at AMIA and is now a senior executive of a very large academic medical center. To my surprise, he explained that he, too, makes a point of attending at least one tutorial at every AMIA conference. There are few things as impressive to me as someone at the top of their field who still thinks they have something to learn.
At the morning break, I chatted with Jose, another tutorial attendee. Jose is an internist and part of the clinical informatics team at a large East Coast medical system. His interests include population health and chronic care management. One of the projects he’s working on is development of a homegrown application for their health coaches. Among the workflows this application will support is capture of PHQ-9 questionnaire results.
Jose recognized that there are LOINC codes for both PHQ-9 questions and answers and has been working with his development team to make sure that those codes are stored along with the questionnaire, increasing its ability to be re-used for reporting, decision support, interoperability, etc. Another great example of how informatics knowledge can make a difference in how health care organizations operate.
The NLP tutorial itself was certainly worthwhile, with instructors who were very knowledgeable and well-prepared. At the same time, it illustrated one of the challenges that faces AMIA and the field of informatics in general. Informatics is a “big tent” field whose adherent come from a wide variety of professional backgrounds and are working to solve a wide variety of problems. While this is a tremendous strength, it also creates challenges. In some cases, informaticists assume of each other familiarity with a particular set of knowledge or a shared set of priorities and interests.
This was evident in the NLP tutorial. The presenters spent much more time describing the steps in using a set of open-source tools to create NLP engines (including the mechanics of setting up the processing queue for new documents in a data repository) than they did describing the logic by which NLP engines work and how that can be optimized. It would have been a great introduction for a grad student considering building an NLP engine for their dissertation. The clinician attendees, hoping to learn how NLP could help manage clinical information and patient care at their organizations, seemed less well served. Still, without AMIA, the “in the trenches” folks and the “in the ivory tower” folks would rarely come into contact. I believe that both benefit from the interaction.
AMIA officially opened today with a plenary session with a keynote from Avi Rubin, an information security expert from Johns Hopkins, who gave a widely-viewed TED talk back in 2011 pointing out some serious security vulnerabilities of modern technology, including medical devices. His keynote today expanded on this landscape, which has only worsened. It was a very unsettling talk to hear and a cautionary tale to those who develop IT-enabled implantable devices or take care of people who have them.
After the keynote, the first set of conference sessions began. I attended a paper session on “Deep Phenotyping.” AMIA paper sessions fit four brief presentations into 90 minutes with a few minutes at the end for questions. If you’re not already very familiar with the topic and current research in the area, it’s tricky to keep up.
Phenotyping refers simply to solving the problem of identifying the phenotype of a person, i.e. classifying them according to some biological or health-related category, such as determining whether they’re diabetic or not diabetic. It’s an important problem if you are trying to do something that requires knowing the phenotype of individuals in a population (for population management, knowledge discovery, etc.)
The most interesting paper in the session, in my opinion, described “semi-supervised” machine learning for phenotype identification from free-text notes. In traditional (“supervised”) machine learning, a system is given a set of documents and manually-applied labels as to their contents (the “answers”). Based its analysis of the associations between the contents of the documents and the labels, it develops an algorithm that it can use to infer the appropriate labels for an unlabeled document.
In semi-supervised machine learning, following the supervised process, the system refines its algorithm based on its own inferences on the contents of the data. To my knuckleheaded family physician brain, it’s as if you teach someone that an AMIA attendee with a backpack is more likely to be a grad student than one without a backpack, and then they notice that the AMIA attendees with backpacks are more likely to be wearing sneakers than those without backpacks, and then that person starts inferring that AMIA attendees wearing sneakers are more likely to be grad students. In other words, after learning from being taught explicitly, the computer starts to be able to learn just from what it’s seeing. Intriguing stuff.
Following the session was the welcome reception in the exhibit hall. Among the folks I chatted with was John, the medical director of quality for the Medicaid program of a Midwestern state. It was his first AMIA. He was excited about the potential of sophisticated data analysis for assessing quality, but also mentioned that at present, the only data he has to work with is claims data — he has no way to get any data from EHRs. While we’re making great strides in thinking about how we might use healthcare data in positive ways, the options for much of the real world are limited.
Stepping outside the San Francisco Hilton, the realities of human misery are stark and obvious. The Hilton is right in the middle of the Tenderloin district, full of individuals who are clearly mentally ill and/or intoxicated. It is an important reminder of the urgent need to expand knowledge about human health and how to improve it, in which informatics has a critical role to play. As we dive deep into the intellectual challenges of our field, we must never lose sight of whom we’re doing all this for.