Deep learning offers lessons for healthcare
Artificial intelligence and machine learning are helping revolutionize healthcare, but a newer and more powerful digital application is entering the limelight.
Deep learning is a subset of AI and machine learning, which is a broader category in the field of computer science. It offers benefit advisers promising new ways to help their employer clients improve healthcare services and consumer experiences.
The focus on this advanced technology, which is becoming an integral part of healthcare data analytics, is on processing massive amounts of information and spotting data patterns. By developing more reliable prediction models, the hope is that earlier intervention will help halt the progression of chronic and costly diseases.
Experts see tremendous potential in terms of streamlining both the delivery and operational efficiency of employer-provided healthcare benefits to improve outcomes and reduce costs.
Deep learning capabilities include analyzing medical imaging and diagnoses, as well as decreasing both the cycle time and cost in drug discovery, says Kerrie Holley, a technical fellow with Optum. He says other applications include reducing out-of-network care, improving auto-adjudicated claims, and creating risk and premium models that use a wider range of data than current actuarial models.
Alex Ermolaev, director of AI for Change Healthcare and one of the most frequent speakers on AI in healthcare in the Silicon Valley, notes that deep learning also can improve the accuracy of reimbursement paid to physicians and hospitals, member identification and retention, fraud detection and population health management.
While both machine learning and deep learning use artificial neural networks based on algorithms inspired by the brain, the latter involves more prolific or dense layers than the former, as well as more math and computer power. Driving forces of this technology include a single-chip processor used mostly to manage and boost the performance of video and graphics, known as a graphical processing unit, and backpropagation, a technique used to train classes of neural networks.
“Everything that we can do with machine learning we can do better with deep learning,” Holley explains. “Deep learning uses algorithms that have set new records in accuracy like in image recognition, sound detection and tone detection.”
Using deep learning to better predict conditions for high-risk patients and make personalized treatment recommendations can bolster early intervention strategies, improve diagnostics and outcomes, and eventually reduce costs.
Deep learning techniques provide a powerful way to extract hidden correlations among vast amounts of data that may combine in complex ways to affect health outcomes without having to define features of the data upfront, observes Bonnie Ray, VP of data science at Talkspace.
“For example,” she says, “from examining the unstructured language a client uses in messaging with his therapist, we can determine when a risk assessment may be needed and guide the therapist in conducting the assessment, potentially averting adverse incidents and facilitating a quicker recovery.”
This, in turn, enables researchers to identify low cost ways to improve health outcomes that may have been previously unknown.
Brave new world of speech recognition
Some of what deep learning can potentially do may sound like medical science fiction. For example, Holley says it can be used to screen for certain chronic diseases simply by analyzing nurseline voicemails. Such advanced predictions require unprecedented levels of big data that are the bailiwick of deep learning, and they may be particularly valuable for treating multiple chronic conditions.
As the U.S. healthcare system transition to electronic health records, the marketplace is ripe for deep learning. Just a single hospital stay, for example, “might generate 100 pages of data,” according to Ermolaev. He says the larger the amount of data and computing power used, as well as the size of the AI model, the more powerful AI applications for deep learning will be.
Speech recognition technology also can be mined to improve customer service by reducing the amount of time health plan members spend on the phone and improve satisfaction, Holley explains.
This is done by helping identify and authenticate, for instance, nurseline callers, understand the reason for their call, optimize call routing and personalize their interaction. Speeding the phone tree queue can significantly reduce the long hold times and multiple transfers associated with interactive voice response technology, which he says fuels caller frustration. Smart recommendations also can be used to improve healthcare system navigation for healthcare consumers.
One of the most promising healthcare applications for deep learning is with pharmacy benefits management for analyzing clinical data to help determine critical points to intervene with patients in their care path, he says. The technology also can predict non-adherence in various disease categories such as diabetes, hypertension and depression, as well as automate prior authorizations for pharmacy claims.
There’s a self-care component to deep learning. For example, Holley says it can be used to help health plan members better understand their own health risks. When patients are armed with more meaningful information to understand specific disease features, comorbidity factors and what caused their condition, it can improve clinician workflow. Unencumbered by so many clinical notes, which can be properly transcribed, he says doctors and nurses will be more insightful and strategic with their patients.
In terms of employee wellness programs, Holley notes that ambient computing can be used to create a contextual environment that senses the presence of people who speak into a voicemail system or virtual assistant like Alexa and responds. One potentially important application is for elderly populations. “We can begin to not only detect if you’ve fallen, but that you’re not going into rooms as often as you should,” he observes, noting that intervention strategies can help suggest clinical care pathways.
The rise of data science
When Holley joined Optum about three years ago, he would eventually work with “more data scientists and machine learning models in production probably than any other healthcare company in the world.” His initial mission was to build an AI infrastructure, and once the hardware was in place, test the predictive modeling capacity.
Large employers have increasingly shown interest in the strategic use of AI to eliminate waste and friction, as well as create a compelling customer experience, according to Holley. As such, he stresses the importance of organizations setting goals and establishing performance measures for the expertise that is imparted.
“You’ve got to have an AI infrastructure and real experts,” he says. “Sometimes people think that AI is over-promising and under-delivering, and I would argue the reason for that perception is this conflation that’s occurring with AI.”
But that’s expected to change as deep learning overtakes machine learning and powers the next wave of AI, according to Holley. While not a silver bullet, he considers it vital tool for employers to elevate clinical outcomes and rein in claims.