Future-proofing AI: Embrace machine learning now because healthcare adoption is picking up speed
Artificial intelligence in healthcare is at the end of the beginning: It’s been researched, introduced, proven to work, and it’s been put to use in real clinical settings. But adoption is slow. Healthcare has only explored the tip of the iceberg and, of course, there is much work ahead to improve patient care.
AI in healthcare is beginning to emerge out of its infancy, said Ted Willke, senior principal engineer at Intel Labs.
“We’re seeing healthcare organizations and hospitals move beyond AI-based proof-of-concepts and program pilots into developing and adopting systems that work the best for their needs,” Willke said. “AI in healthcare, like in other industries, began as a way to help these organizations manage their vast amounts of data and simplify daily tasks, but we’re starting to see the emergence of truly innovative uses of AI in healthcare – from finding complex patterns in medical imaging to genomic sequencing to designing patient treatment plans.”
Additionally, right now, there’s a huge interest in predictive clinical analytics, or the process of inputting historical patient data into models to identify and forecast future events, such as the likelihood of a patient relapse.
Search, classification and reasoning
The most common uses for AI in healthcare today are for search, classification and reasoning, in that order, said Ajay Royyuru, vice president, healthcare and life sciences research, at IBM Watson Health.
“AI is able to ingest massive amounts of data using technology such as text analytics and Natural Language Processing,” Royyuru said. “With that, AI systems are able to detect patterns and similarities within that data that may unlock new insights for clinicians and scientists. A system that can reason not only can identify but make recommendations or suggestions based on its trained parameters.”
Both IBM’s Watson for Genomics and Watson for Oncology are examples of state of the art AI systems that are currently available to doctors and scientists. In research, the Watson team also is working on a system that will take this classification and reasoning system to the next level, by applying it to images. The project is called Medical Sieve and is ongoing work to use AI to identify instances of breast cancer and cardiac disease.
DeepMind is another example of state-of-the-art AI in action. It was founded in London in 2010 with the aim of building AI technologies and proving that they could have positive social impact. DeepMind Health is central to this social mission.
“One goal at DeepMind Health is to make a practical difference to patients, nurses and doctors and support the NHS and other healthcare systems,” said Dominic King, MD, clinical lead at DeepMind Health and an academic surgeon in the UK National Health Service. “We hope that our technologies will help to save lives, improve care.”
There remains a lot of hype about artificial intelligence in healthcare – the potential is very much there, but the reality is the industry is a while away from AI algorithms transforming everyday care. One reason for this is that most hospitals are still on a journey of updating their tools and processes for the digital age. Pagers and fax machines remain a common fixture in even the most digitally mature healthcare organizations.
“You can’t deploy an algorithm on a pen and paper or fax machine, you need to have a digital platform up and running,” King said. “So the delays in AI coming to nurses and doctors to help them improve patient care may be less about what’s technically possible in terms of the research, but more about the slow and painful process of digitizing and improving interoperability.”
AI today is really machine learning
AI is, of course, a huge research area, sort of like physics is a big research area, said Peter Lee, corporate vice president at Microsoft Research.
“In practice today, when people talk about AI, what is actually going on in almost every case is an application of machine learning.”
Peter Lee, Microsoft Research
“While there is a lot that we have learned through research, much of which has transitioned into practical technologies, there is a lot more that is yet to be discovered,” Lee said. “In practice today, when people talk about AI, what is actually going on in almost every case is an application of machine learning.”
Philosophers and scientists can argue about whether intelligence and artificial intelligence are the same thing. Lee believes they are, but said people really don’t know. But within this huge space of AI, there is a tiny part of it called machine learning, the study of machines that improve with experience.
And then within machine learning, there is a sub-part where the learning of models is based on data – often Big Data derived from the digital exhaust of human intelligence and activity. And then within that bubble is supervised machine learning, i.e., machine learning based on Big Data that has been labeled by people.
“From looking at this picture, it seems like we have barely scratched the surface of AI,” Lee explained. “And that’s right. There are still so many mysteries that consume the thoughts and actions of AI researchers around the world. The key point: Almost every practical application of AI today in healthcare is inside this tiny bubble; that is, AI in the real world today is generally based on supervised machine learning.”
The dependence of today’s AI on data has major consequences for healthcare. The industry is highly regulated, and so is access and use of medical data. So when systems are set up for accessing healthcare data in a compliant way, one will increasingly see the application of artificial intelligence and machine learning.
Microsoft’s developments in AI for healthcare have “followed the data.” For example, InnerEye is AI-powered computer vision designed to dramatically improve the productivity of radiologists, trained on thousands of 3-D images and radiologist inputs. Healthbot enables healthcare organizations to create conversational interfaces that provide always-accessible information about health, wellness, benefits and intelligent triage, making use of natural language and conversational AI models developed for Bing, Xbox, Cortana and more.
“The possibilities are truly limitless, once we have the ability to operationalize the access to key data sets,” Lee said. “The importance of compliant access to health data for training AI systems can’t be overstated. For that reason, we are working very hard to provide the best health data platform, as well.”
Machine-learned AI is not yet prevalent throughout healthcare, but its adoption is on the rise, said Katherine Chou, head of product for health, at Google Research.
“It commonly appears in the development of surgical robotics, predictive analytics for optimizing healthcare operations, computer-aided diagnostics for telehealth services, and assistive technologies that reduce the burden of clinical workflows and improve home care with patients,” Chou said.
So far, the Google Research team has been focused on increasing the accuracy and availability of healthcare by developing state of the art computer vision systems for reading retinal fundus images for diabetic retinopathy or detecting lesion-level tumors in giga-pixel microscopy images. The team also has been building scalable predictive models that can help detect patient deterioration from de-identified medical records.
Where AI is heading
With all of the work in progress in healthcare artificial intelligence, with AI at the end of its early stages in healthcare and cementing a place for itself in health IT, the question becomes: Where is artificial intelligence in healthcare heading? What will be asked of artificial intelligence by healthcare organizations in the years ahead? Healthcare AI’s top experts have a range of predictions.
“For AI to be successful, the models must be interpretable, not just intelligent. The more that people can understand how the models arrived at their output, the more credible and actionable they can be.”
Katherine Chou, Google Research
For one thing, there is a tremendous push to use AI to realize healthcare’s Triple Aim of better outcomes, reduced costs and better accessibility, Microsoft’s Lee said.
“For better outcomes, we will see AI-powered tools that empower clinicians; for example, by reducing the time and effort needed to do record-keeping or by automating the most time-consuming aspects of their work,” Lee said. “We will also see AI tools that make it possible for healthcare researchers and clinicians to stay abreast of the exponentially increasing amount of new medical science being published.”
AI also will enable new ways for people to access and take control of their own healthcare, Lee predicted.
“We already see the start of this with Microsoft’s Healthbot, where patients can, anytime and anywhere, converse with an intelligent health agent, go through an efficient triage, and then get intelligently handed off to a nurse or physician in a way that is more efficient, lower cost and more satisfying,” Lee said.
At DeepMind Health, the organization’s current research for the healthcare applications of tomorrow combines two AI approaches: deep learning and reinforcement learning.
“Our algorithms are able to interpret visual information in imaging scans, and the system learns how to identify potential issues,” said King of DeepMind Health and the UK National Health Service. “We believe these tools have the potential to recommend the right course of action to a clinician – and help tell the clinician why.”
As the algorithm processes more scans, it improves its understanding and interpretation of the information. Then it provides increasingly useful feedback about the data for expert clinicians to use for better diagnoses and treatment.
“It’s too soon for us to comment on the results yet,” King said, “but the early signs are very encouraging. This research will be subject to rigorous scrutiny and we plan to publish the results in peer-reviewed journals.”
AI is a huge priority for Intel moving forward, and the company is active in a number of areas as they relate to healthcare, from predictive clinical analytics to imaging.
For example, Intel helped Sharp HealthCare in San Diego build a predictive model with 80 percent predictability of patient decline within an hour of an event happening. This advance notice enables intelligently placed medical emergency teams – rapid response teams – to be located at key points in hospitals and to intervene before an event becomes life-threatening.
AI adoption to accelerate quickly
“In the future, as more hospitals begin to see the impact of AI and explore its potential, AI adoption and use-cases will greatly accelerate,” said Willke of Intel Labs. “When done right, AI in healthcare can create enormous efficiencies for organizations and greatly inform both doctor and patient decisions.”
During the next 10 years, IBM Watson Health expects to see more being done in the areas of machine learning, reasoning and decision support capabilities.
“With the inclusion of images such as X-rays and CT scans, audio, video, and continuous, real-time monitors of physiology and lifestyle, the amount of data to be leveraged will continue to grow, and AI systems will need new capabilities to unlock the insights within that data,” said Royyuru of Watson Health.
At Google Research, the AI team will be working on advancing the art and science of AI along four themes: augmented intelligence; assistive and ambient intelligence; accelerated engineering, analysis and workflows; and interpretability.
“Augmented intelligence can help expand the role of any domain expert, whether a musician or a doctor, so they can know more in order to do more,” said Chou of Google Research. “Assistive and ambient intelligence can help people with secondary operations so they can focus on what matters more to them.”
Deep learning research for everyone
Accelerated engineering, analysis and workflows can help expedite the processing of data or accelerate data-rich workflows, Chou explained. Google Research’s work here is not only in developing machine-learned AI processes, but also in opening up and improving its TensorFlow deep learning research platform to all researchers.
In healthcare, for example, the deep learning technology can help with harmonizing medical records data on top of the open standard FHIR, automating a process that historically has been extremely labor intensive.
“And finally there is interpretability,” Chou said. “For AI to be successful, the models must be interpretable, not just intelligent. The more that people can understand how the models arrived at their output, the more credible and actionable they can be. This is massively important in the healthcare space, as health is incredibly complex and full of confounding factors.”
How AI is driving forward-looking healthcare orgs.