How Deep Learning Is Changing Healthcare Part 2: Prevention

Share This Article

0 0 0 0

Last week we talked about how AI is changing the way doctors diagnose illnesses and disorders.

But, as important as a fast, inexpensive, and accurate diagnosis is, there’s one thing that’s even better: prevention.

This week, we’re diving into how AI is transforming the way doctors predict and prevent disease and hospitalizations.

Timely predictions will help prevent disease

Every year, U.S. hospitals admit 4.4 million patients unnecessarily, costing $30.8 billion, according to estimates from the U.S. Agency for Healthcare Research and Quality.

Just two ailments—heart disease and complications from diabetes—account for half of all unnecessary hospitalizations.

Heart disease infographic (Source: Huffington Post)

Here are some examples of how deep learning is already helping to predict and avoid negative health events related to heart disease and diabetes:

  • Researchers at Boston University’s Center for Information and Systems Engineering have been working with local hospitals to monitor patients with heart disease and diabetes and predict which of them will require hospitalization. If healthcare providers can predict who will need help before it’s needed, they can prevent many of these hospitalizations. The deep learning model the researchers are using can predict with 82% accuracy who will need hospitalization about a year in advance.
  • Researchers at Sutter Health and the Georgia Institute of Technology can now predict heart failure using deep learning to analyze electronic health records up to nine months before doctors using traditional means.
  • Frans Von Houten, Chairman and CEO of Royal Philips, told CNBC in May that his company now uses AI to accurately predict whether a patient will have a heart attack hours before it happens.

But AI isn’t just helping prevent sudden healthcare events. It’s also helping to thwart ongoing degeneration.

For example, diabetic retinopathy is a leading cause of blindness among working-age adults.

Diabetic retinopathy diagram (Source: news-medical.net)

Such diabetes-related complications arise from spikes and drops in blood glucose levels, so accurately predicting blood glucose levels is key to preventing drops and spikes in the first place with well-timed snacks and insulin injections.

A July 2017 paper shows that deep neural networks, which perform deep learning, can learn from one set of diabetic children how to accurately predict blood glucose levels (in order to prevent these drops and spikes) in a larger group of children.

Understanding of how genes lead to disease will deepen

Another way to prevent disease with AI is to predict who will develop certain disorders based on their genetic makeup.

According to Gartner healthcare analyst, Richard Gibson, genes are “the biggest thing to hit healthcare maybe ever, definitely since the advent of antibiotics in 1950.”

Specifically, as researchers collect genomics data at unprecedented levels, and deep learning models make analyzing that data and drawing connections easier than ever, we’re learning an incredible amount about how genetic factors such as mutations lead to disease.

These advancements are leading to personalized or “precision” medicine, where the goal is to tailor treatments to each patient’s genomic makeup.

Your genome is the complete set of chemical instructions for building a “you.” Though genomics is still in its infancy, there are projects making strides. For instance, a team of researchers at the University of Toronto is working to build a genetic interpretation engine to quickly identify cancer-causing mutations in individual patients.

Also in Toronto, a startup called Deep Genomics applies a deep learning model to huge data sets of genetic information and medical records to match genetic variations with corresponding disease.

Both organizations use AI computing platform, Nvidia GPU for their models.

Prepare for deep learning with the right software

While GPUs such as those made by Nvidia are essential for running deep learning algorithms, you also need specialized software to make healthcare AI a reality.

The Boston University group was able to predict who would need hospitalization with much greater accuracy than doctors alone, because they used deep neural networks (DNNs).

DNNs can analyze up to 200 factors, such as health history and demographic information, to identify those that are associated with future disease. However, in order for the DNN model to work, it needs data from EHR records.

One potential challenge is that EHRs typically store this type of data in large blocks of text. For instance, an EHRs may have a record of a patient’s history of depression in a “Notes,” section, where a physician writes “Patient’s mother suffered from depressed moods” along with current complaints, issues, etc.

But, in order to work, AI models need well-structured data. It’s easier for a machine to parse that a patient has a family history of depression if there’s a column called “family history” and a check box next to “depression.”

Soon deep learning will be “mandatory for people building sophisticated software applications,” Andreessen Horowitz Partner Frank Chen tells Fortune.

Most venture capitalists, including those who invest in SaaS startups, didn’t even know what deep learning was five years ago. Today, investors “are wary of startups that don’t have it,” Chen says.

Likewise, you should be wary of EHRs that don’t create and store the kind of well-structured data that works with deep learning models. You could even look for an EHR that has embedded AI into its clinical documentation functionality, for example, Epic partnering with Nuance.

However, most EHR systems won’t have AI embedded for a while, according to Anil Jain, M.D., FACP, and the vice president and chief health informatics officer for IBM Watson Health. The option in these cases is to integrate AI functionality into your existing EHR. From now on, most healthcare systems will have to develop and deploy AI as add-on functionality.

That’s what Intermountain Healthcare did with their EHR, building more than 150 protocols into Cerner. With each protocol, Cerner throws up an alert when it receives patient information that indicates a certain medical condition and then guides clinicians through suggested further examinations and potential treatments.

Building these protocols used to require 12 doctors, nurses, and analytics experts, and would take more than a year. But, by partnering with Intermountain, they can be built in 10 days with no human labor.

When you talk to software salespeople, whether you’re looking for EHR software or medical practice management software, it’s important to know which questions to ask.

Take a cue from VC partner Chen and ask questions such as:

  • “Where’s your natural-language processing version?”
  • “How do I talk to your app so I don’t have to click through menus?”

Next steps

Currently, large research centers and healthcare systems are developing deep learning models that can predict and prevent disease and hospitalizations and discover which genes are associated with future diseases and disorders.

When comparing EHR software, ask the vendors on your shortlist about any AI functionality or integrations they offer. For example, is the data stored in a block of text or is it more structured?

Ideally, you should choose an EHR that has AI functionality built in, or one that can integrate with a deep learning model.

Looking for Medical Practice Management software? Check out Capterra's list of the best Medical Practice Management software solutions.

Share This Article

About the Author

Avatar

Cathy Reisenwitz

Cathy Reisenwitz is a former Capterra analyst.

Comments

No comments yet. Be the first!

Comment on this article:


Comment Guidelines:
All comments are moderated before publication and must meet our guidelines. Comments must be substantive, professional, and avoid self promotion. Moderators use discretion when approving comments.

For example, comments may not:
• Contain personal information like phone numbers or email addresses
• Be self-promotional or link to other websites
• Contain hateful or disparaging language
• Use fake names or spam content

Your privacy is important to us. Check out our Privacy Policy.