Information technology (IT) has been promising for decades, largely since the advent of electronic medical records (EMR), to improve and streamline health care as it has multiplied productivity in countless other industries. In addition to the long-standing problems with EMRs, more recently there have been early disappointments with the latest iteration of IT focused on artificial intelligence (AI) and machine learning (ML), as big players like IBM Watson and Google have tended to over-promise and under-deliver with algorithms that are poorly matched to the data or the patient need.

Major research hospitals, such as Johns Hopkins and the Cleveland Clinic, now are focused much more closely on their own internal efforts to develop ML systems that can improve patient outcomes. Johns Hopkins has rolled out a sepsis detection AI algorithm that's in use currently across several hospitals and has been the subject of several publications along the way; it aims to put to work additional specific algorithms in the clinic in the coming months.

Past the hype

Cleveland Clinic is looking to debut its own AI algorithms in the coming months for use by its physicians, as well as to launch a clinical trial that could serve as a template for future studies and aid in facilitating how the FDA can work to validate and approve these continuously self-improving algorithms.

Aziz Nazha, who heads the Center for Clinical Artificial Intelligence that launched this spring at the Cleveland Clinic, noted that the current implementation of AI and ML at most hospitals, including major research hospitals, is almost entirely nonexistent. But he remains an optimist, expecting that AI and ML will become utterly routine in health care within the next decade or two. The iterative process to get there, however, must ultimately result in improving patient outcomes and gaining physician trust.

"The biggest barrier is overselling; you don't want to oversell because then you lose trust," Nazha told BioWorld MedTech, citing problems with some of the larger tech players in the field. "The way we approach it is: we have a problem; how will we solve it? It's a tool in your toolbox to solve your problem. If you can solve your problem using linear algebra, use linear algebra. But if you get a big and complex problem, you probably need to use AI technology. Think about it as a tool to get you to where you want to be. But your focus is on the outcome; you're focused on providing the value, not the AI."

"Unfortunately, we're at the top of the hype cycle now," he continued. "You can raise some funding, but I hope in the next few years we will pass that – so we see our way through to innovation that is all around us."

Cleveland Clinic currently has about 24 AI-based health care projects. These are focused on analyzing a broad range of problems, including medical readmission, cancer diagnosis and prognosis, pathology and image analysis. Cleveland Clinic expects to start using the first one of these in the next few months, although implementation has been difficult requiring wending through myriad physician, legal and applicability issues.

It's also finalizing a clinical trial for an AI-based personalized patient prediction model that's designed to identify who is most at risk for hospital readmission and why, enabling physicians to offer preventive treatment and follow-up. Nazha expects that this study could offer a model for future clinical trials of AI algorithms in health care, which the FDA has thus far encouraged but will also need to become increasingly sophisticated in assessing. He noted that when it comes to research, it can be difficult to even publish papers on AI in health care in journals, since peer reviewers are often profoundly unfamiliar and uncomfortable with the subject.

Educating the next generation by bringing together software coders and physicians is another emphasis for Nazha. To that end, he helped to start the first classes this semester at the Cleveland Clinic for medical students and at Case Western Reserve University for programmers to train each in the language and needs of the other. He himself is trained as both an oncologist and a software programmer, having worked to build a model to offer personalized myelodysplastic syndromes treatment recommendations.

From pattern recognition to complex problems

There are some AI startups that are starting to make headway in pattern recognition, particularly in imaging. These include coronary disease detection from Heartflow Inc., ophthalmology analytics from Idx Technologies Inc., large vessel occlusion identification from Viz.ai and various AI analytics from cloud-based imaging companies such as Zebra Medical Vision and Arterys Inc., according to Suchi Saria, who heads the Machine Learning and Healthcare Lab at Johns Hopkins University.

But these have yet to be widely adopted and often even if physicians are using them, they may not know that AI underlies these tools. That's an explicit strategy for some startups that recognize information technology and AI have been overhyped in health care and prefer to lead with patient outcomes data and ease of integration into physician workflow.

Google sister company Verily Life Sciences is ambitious on the corporate side of delivering AI to health care. It's already been infused with almost $2 billion in outside capital and more from parent company Alphabet since it was spun out in 2015, but hasn't delivered much yet into the clinic or onto the market.

It sees pattern recognition as starting to breakthrough in AI applications in health care. Verily and Google have published amply on using AI to detect diabetic retinopathy in retinal images. Early this year, they partnered with a hospital in India, Aravind Eye Hospital, to start using the technology for screening.

"I'm starting to see the biggest change with recognizing different patterns like with fundus images of the back of the eye," said Verily Chief Medical and Scientific Officer Jessica Mega. "Machines can be really good at that, also at looking for anomaly detection. If you show a human 100 slides very quickly, or 100 images, we may not be tuned to find the one that has an anomaly. But when you're looking for pattern, whether it's in the back of the eye, or a pathology sample or an electrocardiogram. I'm starting to see the biggest impact in noticing patterns, noticing anomalies and surfacing those for clinical action."

Beyond imaging and pathology, Verily also is looking toward AI in surgical robotics and for plumbing complex, interrelated datasets to improve the understanding of when health veers into disease. On the former, it partnered with Johnson & Johnson's Ethicon business to form joint venture Verb Surgical, which aims to employ AI to make surgical procedures more accessible globally despite a shortage of surgeons. On the latter, it's running the massive Project Baseline observational study that's integrating and analyzing all sorts of data, such as electronic health records s, claims data, lab test, as well as genomic and wearable data.

Outcomes and efficiency

Saria's efforts at Johns Hopkins have been focused on AI analytics aimed at interpreting multiple data streams that require a complex perspective, one that may even vary from physician to physician. The first AI tool that she has worked to implement is in sepsis, with more algorithms slated to follow soon.

A year ago, Johns Hopkins deployed the system at scale, and it is now in use at five hospitals by 4,500 physicians. The software has been in iterative development since 2015, taking four years to get this far. She noted that other AI algorithms are emerging from various major research hospitals, such as those at Duke University and the University of Pennsylvania.

Saria declined to disclose the details of upcoming analytical deployments, but she has published on chronic disease applications. Her group is dedicated to identifying indications where AI approaches can make a difference and designing a ML approach that can achieve a measurable result.

Like Nazha, among her foremost concerns is that the right AI tools are applied in the correct way to the data. She's seen too many projects where the tools are a poor fit to the data and so the results aren't useful or transferable to other environments.

"There's a large amount of work being done right now, where people are naively taking AI tools that were developed for other data modalities or other application areas and just applying this to clinical data without paying attention to the sources of data," explained Saria. "The complexity in this kind of data is very critical in getting the output right. To me, this is like 95% of the problem – taking off-the-shelf tools that showed promise in other applications without adjusting and accounting for health care data. But we really desperately need tools that do the latter; there is a lot of progress that will be needed."

Both Saria and Nazha agree that AI and ML applications for diagnosis and prognosis are where the field is headed, but that is greatly complicated by the fact that physicians often can disagree and protocols at hospitals can differ. How can a machine learning analysis be expected to be definitive, if the field of medicine itself is not itself in complete agreement?

"In images, you can see it and it is apples to apples, there's nothing to argue about. But in medicine, the challenge is that even experts often disagree," said Saria. "So if you did not use the right kind of data to develop algorithms or you validate against some very coarse measure and get 95% accuracy, you're going to think it works really well, but physicians aren't going to find it very useful. That's sort of a big challenge and something that's unique about medicine."

Another point of agreement between them is on the so-called 'black box problem,' the fact that the elements considered may not be transparent to a physician. Both insist that the factors that underlie an algorithm determination should be transparent to the practitioner, and, perhaps most importantly, the next steps in the care of the patient that are implied from the results should be as clear as possible. How the determination was arrived at and what it means for continued patient care must each be entirely clear to a physician.

"There are many high impact areas of medicine that can target who gets what kind of care and when that can really eliminate waste and get the right kind of treatment to the right kind of patient at the right time," summed up Saria. "Now, with having the electronic infrastructure be more ready, the most exciting work in this space is at the start. Devices and drugs were where the biggest action was in terms of benefit in medicine; the next decade is going to be all about software and software-related inventions bringing therapeutic and efficiency benefits to medicine."

No Comments