By Deborah Borfitz

May 7, 2019 | Clinicians have largely overcome initial fears and doubts about the use of artificial intelligence (AI)—specifically, machine learning (ML) applications—and are now wanting to know what real-world clinical problems it can solve and how to trust it, observes Karley Yoder, senior director of product, Edison AI. "No one is excited about the driverless car because of AI," she offers as an analogy. "They're excited about the prospect of less traffic and accidents... the outcomes that it drives."

In the everyday world of medicine, AI remains a novelty, Yoder says. A raised-hand survey at an AI workshop hosted last August by the National Institute of Biomedical Imaging and Bioengineering found most of the national and international experts in attendance had used AI tools in their research. But only about 5% had ever used AI in their clinical practice. That likely won't be the case for long.

GE Healthcare actively employs ML to create more intelligent devices enabling precision health and intelligent applications focused on improving productivity and accuracy, while Change Healthcare focuses on "less sexy" but practical applications for improving provider workflow and administrative cost savings. From a data science tools perspective, Amazon's ML model SageMaker is now doing most of the heavy lifting behind the scenes for both companies.

SageMaker has been incorporated into GE Healthcare's Edison, one of the largest AI platforms in healthcare built to connect data from millions of imaging devices, says Yoder. Among its outputs is a recently FDA-approved precision diagnostics tool, AIRx, launched at the 2018 Radiological Society of North America annual conference. AIRx uses deep learning and anatomy recognition to learn from a database of 36,000 MRI brain images, reducing a manual step by technicians of setting plane properties during longitudinal assessments for neurological disease, she explains.

The minutes saved per exam quickly add up, and the patient experience improves because they're less likely to be recalled due to an incorrect slice placement, says Yoder. Clinicians can also more easily diagnose change over time and make their treatment plans. She predicts this "reproducibility and quality" AI frontier will be changing rapidly over the next three years.

AIRx launched around the same time as a digital radiography system, still under FDA review, which employs AI algorithms to identify cases of pneumothorax (collapsed lung) at the time of image capture so clinical teams can prioritize image review, Yoder says. Already on the market is another AI-powered tool that automatically segments and measures lesions to help ensure consistency among different ultrasound users, or even the same user, for documentation and follow-ups. GE Healthcare has similarly automated the process of measuring the fetal brain, which an internal study finds can help reduce keystrokes by more than 75%. Then there's the CT app that facilitates image readings of the spine so radiologists can stop manually counting vertebra to indicate the precise location of suspicious spots.

Edison is likewise enabling precision therapeutics, and the latest in development is an AI-based clinical decision support bot called Virtual Collaborator being jointly developed with Roche. It aims to integrate data from electronic medical records and other clinical systems to provide an early detection system for sepsis by inferring the reason behind clinicians' questions and providing additional context to help them better treat patients, she says. No reliable biomarkers or diagnostics currently exist.

GE Healthcare has a long history of bedside monitoring, Yoder adds, so the company is also looking to use AI to individualize the experience, which is increasingly happening at home rather than at the hospital.

The Three Vs

It's all very much a "man and [not versus] the machine," says Yoder. "Technology is changing but the final responsibility and clinical context of what is happening still lies with clinicians. A lot of what we're doing is removing some of the drudgery from their roles by helping them do what they're already doing better, faster and easier." The onus is on the industry, she adds, to "lighten the black box of this technology and the traceability of the data science."

Deep learning approaches don't vary greatly, Yoder notes. The data is what matters and "high volume alone is not sufficient." Equally important is that the data adequately represent the patient population being addressed and truthfully relates to the issue at hand. In 2016, Microsoft famously proved the "garbage in, garage out" adage when its Twitter bot was quickly corrupted by racism.

Given "data volume, variety and veracity," says Yoder, deep learning may be able to succeed where rule-based computer-aided detection (CAD) systems have failed to help radiologists interpret mammograms. Higher reimbursement, rather than improved screening performance, became the only reason some clinicians ever used CAD.

Not everyone is following these "rules of the road," she adds, because of fast-changing developments in the field of AI and the difficulty in obtaining high-volume, high-variety data. Partnerships with the University of California San Francisco and Partners HealthCare are helping to fill those needs for GE Healthcare.

AI for the Rev Cycle

Change Healthcare has been using AI to beef up internal capabilities to identify claims-related behavior, ingest and understand medical records, and better tie quality assurance efforts to diagnostic and procedure codes, says Adam Sullivan, senior director of artificial intelligence.

On the provider front, its AI efforts are focused squarely on reducing billing-related costs that account for a jaw-dropping proportion of the nation's healthcare spending, he says. The company announced its Claims Lifecycle Artificial Intelligence service in February, which integrates with the company's Intelligent Healthcare Network and financial solutions to help prevent claims denials. This comes on the heels of last summer's launch of Dual Enrollment Advocate blending AI and behavioral science to expedite Medicare Advantage dual-eligibility identification and enrollment.

Dual Enrollment Advocate pinpoints individuals with the highest likelihood of qualifying for full or partial dual eligibility with up to 93% accuracy, Sullivan says. Of the nation's 58.5 million Medicare beneficiaries, an estimated one-third are living at or below the poverty line and eligible for Medicaid but may not be identified as being potentially dual eligible.

The newest AI tool was trained on what happened in 2018 with over 205 million unique claims on the network, Sullivan says. Had submitting providers been following all the rules, $6.2 billion in denied claims could have been flagged for preventable denial at the point of service. Claims Life Cycle Artificial Intelligence alerts providers to likely denials and pinpoints reasons with specific codes—related to eligibility, prior authorization and medical necessity—so corrections can be made before claims are sent.

"If a hospital system is submitting to a few hundred payers there is no way a human could understand every procedure code in every scenario," says Sullivan. To capture missed savings using the app, providers would need only change up their workflow a bit so any flagged issues get cleaned up prior to submission. "It pays to get it right the first time," he adds. An extra minute or two of a biller's time on the front end can prevent having to rework a claim on the backend at a price of $30 to $118 per instance.

Large health systems stand to gain "hundreds of millions of dollars," largely by eliminating the human resources currently devoted to retroactively reworking claims and appealing denials, which could be redirected to patient care, Sullivan says. The tool has already been embedded into two existing customer applications within Change Healthcare—Assurance Reimbursement Management and Revenue Performance Advisor—and Medical Network Solutions will soon join the list. In the future, it will be available to customers directly as well, he adds.

Savings should also accrue to payers when claims come in clean and the adjudication process doesn't needlessly drag on over missing or inaccurate documentation, says Sullivan. "At the end of the day we're talking about reducing errors and people's care in general." In talking to customers of all types, he adds, "I've heard the word 'gamechanger' quite a few times."

Knowing everything about claims at the point of submission impacts patients as much as payers and providers, continues Sullivan, by taking away one of the everyday complexities of navigating healthcare. Claims Life Cycle Artificial Intelligence brings the revenue cycle one step closer to a normalized experience where coverage is understood, claims get appropriately submitted and patients don't get any unwelcome news long after they've left the doctor's office.

A Tool for All Skill Levels

SageMaker was introduced to the market in 2017. As a fully managed service, SageMaker is particularly useful to customers when it comes to factoring in social determinants of health when building, training, and deploying algorithms, and "with access to some of the most accurate deep learning models available,"  says Taha A. Kass-Hout, M.D., senior leader in health and AI at Amazon. Users don't necessarily need much ML training to solve concrete problems.

Machine learning offerings of Amazon Web Services come in a stack to accommodate different skillsets, explains Kass-Hout. The bottom layer is meant for ML experts who need a framework programming language such as MXNet or TensorFlow that they can scale in a cloud infrastructure, enabling them to build new techniques, products and models. Within the healthcare industry, such experts are rare except at institutions with an academic affiliation. It’s a customer group that includes Beth Israel Deaconess Medical Center, he says, which has used MXNet in addition to SageMaker to optimize operating room operations—including triaging patients to beds.

The middle layer, which includes SageMaker, is for statistician and software developers who want to build better predictive models and need an optimized environment for training their novel algorithms, Kass-Hout says. This layer is also best suited for clinicians who want to annotate some data.

Last fall, Amazon launched a natural language processing service for medical text called Amazon Comprehend Medical to help customers use machine learning to extract relevant medical information from unstructured text, including protected health information, diagnoses, medications, medical procedures, or treatments in a string of text to, for example, find suitable patients for clinical trials. In a 48-hour period—the window of opportunity for oncology patients to switch from one treatment regimen to another—Fred Hutchinson Cancer Research Center used the software to run through 10,000-11,000 documents per hour trying to match patients to the right trial, Kass-Hout says.

Almost 90% of patient data needed for trial matching is in a text format, creating a search problem that has become humanly impossible without the aid of machines, says Kass-Hout. Amazon Comprehend Medical, which sits at the top of the ML stack, puts a framework around multiple unstructured parts of a medical record so all relevant data can be queried at once and in a scalable manner. New ML models for improving operational efficiency, trial recruitment, and population health analytics can also more easily ingest their fuel.

Predicting Patient Readmission

Amazon has collaborated with Cerner using SageMaker on perhaps one of the largest cohorts ever—over 216,000 congestive heart failure (CHF) patients—which could help providers better predict when they’ll return to the hospital after discharge, says Kass-Hout. As explained in a soon-to-publish, open-access article, Cerner was able to predict the onset of CHF 15 months in advance utilizing demographics, medical diagnosis and procedure data.

The predictive model is based on "long short-term memory," an artificial recurrent neural network architecture used in the field of deep learning that proved more accurate than other known linear baselines and deep-learning-based models. The authors state that they expect the model will support efforts to implement risk factor reduction strategies and help researchers evaluate interventions to delay or prevent development of the costly and often-devastating disease.

Cerner, like other major EHR vendors, has built a population health analytics platform with clinical decision support tools to help its customers in their journey to more value-based care, says Kass-Hout. Predicting and preventing readmissions for a long list of high-frequency, high-cost conditions is at the top of everyone's to-do list.