popyconcept - Fotolia

A physician-programmer experiments with AI and machine learning in the ER

One way for CIOs to start preparing for machine learning and artificial intelligence is to build a flexible architecture.

Steven Horng, an emergency physician and computer programmer at Beth Israel Deaconess Medical Center in Boston, is part of an elite team that's introducing machine learning and artificial intelligence (AI) to the ER.

"A lot of other people are talking about how to get a computer to diagnose patients or to replace physicians," he said. "But I believe that's the wrong approach. It's really about augmenting the physician so that you use the computer as a tool rather than a replacement."

One way Horng's team of 10 is introducing the technology goes practically unnoticed by colleagues. Machine-learning algorithms are buried deep into workflow processes, assisting medical personnel to perform in a more effective and efficient manner. Getting there has meant developing new data science, running hyper-agile experiments and relying on Beth Israel CIO John Halamka's flexible infrastructure, a component experts point to as a key for introducing machine learning and AI to the enterprise.

A data science project

Horng's team is using machine-learning algorithms to solve a data quality issue that plagues emergency rooms across the country: capturing a patient's reason for the visit or "chief complaint" in a structured, standardized manner. Medical personnel want to be exacting, and that makes data capture a struggle. Using text to document a chief complaint, they'll cite left-sided chest pain, for example.

Steven Horng, physician and computer programmer, Beth IsraelSteven Horng,
physician and
computer programmer,
Beth Israel

"Being able to capture chest pain as a discreet entity can be very valuable downstream for clinical care and in launching things like order sets and clinical pathways," he said. But this is difficult to do in most systems. "The struggle has always been in how to collect this type of structured data when humans want to be explicit and do their own thing," he said.

Horng's team and their computer science partners from New York University understood that if man and machine were going to work together to structure unstructured data in a practical setting, the system needed to do two things. According to their paper Predicting Chief Complaints at Triage Time in the Emergency Department, "the user must feel that the software actually saves them time, and that its results can be trusted."

Enter machine learning. "We've come up with a machine learning approach to turn those unstructured data fields into structured data fields, both on the back end and on the front end," Horng said. Now, when a patient is triaged by a nurse, the data captured is run through a predictive analytics engine, which determines the top five chief complaints a patient most likely has.

"The underlying data sets we've had to [build] have been broad -- from how to represent an ontology of chief complaints to natural language processing," he said. That includes deciphering "messy" parts of unstructured data: misspellings, double meanings and the tricky issue of "negation detection," or reliably identifying the status of a medical condition. Today, Horng said the program has improved collection rates of chief complaint data dramatically -- from 25% to 95%.

A dedicated R&D team

Horng credits Beth Israel's Halamka for giving his team "the freedom to explore this space, which allows us to have an operational role within IS," he said. Horng and his team are members of the IS department, abiding by the policies, procedures and governance structures of IS, but with direct control over their software and hardware. "That's very different from most organizations that have physicians who are consultants -- and afterthoughts," he said.

According to Kenneth Brant, an analyst at Gartner Inc., CIOs should follow suit when it comes to building a strategy for smart machines, a class of technology that leverages machine learning to do tasks humans would otherwise perform. Having a dedicated group for research and development is a good first step in culturally preparing the enterprise, Brant said. Plus, establishing a group like this will help CIOs start thinking in terms of the skills needed to push a smart-machine strategy forward -- skills that may reside outside of the IT department.

Kenneth Brant, analyst, Gartner Inc.Kenneth Brant,
Gartner Inc.

"How to build these smart machines or manage or deploy them is very different [from running] an ERP system," Brant said. "They will need people who have more of a view of artificial intelligence than an IT knowledge base."

Because smart-machine technology is still immature, Brant said organizations will need employees who can run trials and pilots and embrace the idea of failing fast.

Horng and his team are of that fail-fast mentality. "We believe in the physician-programmer approach, where you have the user who is the physician, the developer and the tester as well," Horng said. "This means we can have very fast development cycles."

When building for the future, think flexibility

Infrastructure is also key. Horng relies on the flexibility of Halamka's architecture, a tip echoed by experts as a way to ensure the infrastructure can keep up with the state of constant change. Beth Israel servers are virtualized with multiple fail-safes, Horng said. And the machine learning projects conducted by Horng and his team are run on a dedicated server -- separate from the rest of the clinical environment. "That means no matter what happens on the machine learning side -- if it crashes, if there's an issue, if the server goes down -- it doesn't affect the clinical side at all," he said.

Smart machine hype cycle

In July, Gartner Inc. published its first hype cycle on smart machines. Here's a sampling of the more than 30 technologies featured on the hype cycle.

On the rise:

  • Virtual personal assistants
  • Graph analysis
  • Deep learning

At the peak:

  • Predictive analytics
  • Autonomous vehicles
  • Natural-language question answering

Source: Gartner Inc.'s Hype Cycle for Smart Machines

Michael Harries, chief technologist at Citrix Startup Accelerator, refers to this as building "elbows into the architecture," a concept he believes is more important than ever. "We are at a point where the amount of change coming out of our industry has been pretty dramatic over the last decade, but the reality is we've not seen anything yet," he said at EmTech, an emerging technology conference hosted by the MIT Technology Review.

"This notion that I can centrally control all applications, whether they be mobile or desktop or whatever else is coming along or I can get to it from anywhere is very important," he said. "The delivery of content and the delivery of IT should be abstracted."

Part two of this SearchCIO story on AI and machine learning delves into the cultural and business process changes that AI, machine learning and robots will bring to the enterprise.

Let us know what you think of the story; email Nicole Laskowski, senior news writer, or find her on Twitter @TT_Nicole.

Next Steps

Find out how Gilt.com leverages machine learning to anticipate what customers are likely to buy before they've even seen the merchandise.

Meet your new colleague: The machine software tester

Learn how deep learning technology's infrastructure is beginning to change

Read about Loop AI Labs' push into deep machine learning

Dig Deeper on Enterprise artificial intelligence (AI)