Healthcare AI: Taking a Long Road One Pain Point at a Time

sagawa
Print Friendly, PDF & Email

SEE LAST PAGE OF THIS REPORT Paul Sagawa / Tejas Raut Dessai

FOR IMPORTANT DISCLOSURES 203.901.1633 /.553.9827

psagawa@ / trdessai@ssrllc.com

twitter.jpg @PaulSagawaSSR

January 25, 2018

Healthcare AI: Taking a Long Road One Pain Point at a Time

AI could greatly increase the usefulness and, thus, the use of computing in most elements of the complicated healthcare value chain, improving the efficacy and cost of predicting, preventing, diagnosing, treating and paying for treatment. However, there are obstacles. First, datasets are fragmented, poorly labeled, rife with errors, and in hard to interpret formats. Second, the complex web of stakeholders makes meaningful change fraught with conflict. Third, both AI and medical talent is in short supply. We believe AI will rise in many small pain points across the delivery chain, augmenting the effectiveness of physicians. Broader use cases may work where data quality is higher – i.e. Imaging, Wellness, and Drug Discovery – but will not precipitate major industry disruption. For the tech industry, we see AI accelerating a shift to the public cloud with greater use of computing, benefiting the top platform players as hosts, IT consultants as facilitators, and SaaS vendors with well-designed applications. Many startups will be successful, combining clinical and AI talent to tackle tightly focused opportunities. Amongst the platform players and consultants, IBM is furthest at building a real healthcare AI business, albeit with notable misses amongst its hits. Healthcare industry stakeholders will use AI partnerships to create advantage – we see none with the wherewithal to be successful alone. Over many years, we expect AI to permeate the healthcare landscape, helping to alleviate the cost and efficacy issues that plague it today, but one step at a time.

  • Healthcare AI has significant potential value. AI, offering unique insights into complex data, has the potential to address many industry pain points – relieving physicians of mundane tasks, improving predictive/diagnostic tools, optimizing treatment protocols, and extending the reach of medical expertise – thus, improving outcomes and reducing costs. While there are few places in the value chain where AI couldn’t be useful, we see the biggest initial opportunities in 1) Imaging/Diagnostic Support, 2) Wellness/Patient Compliance, and 3) Drug/Device Development.
  • Healthcare data quality is a major obstacle. AI learns best from huge, well-labeled, error free, consistently formatted, and easily accessed datasets. Healthcare datasets fall short on all dimensions – several high profile medical AI projects failed because ambitions did not account for the effort needed to aggregate data, accurately label it, correct errors, and convert it to consistent and easy to read formats. Progress is being made on two fronts: AI scientists are working to draw better inferences from small, poorly labeled datasets and to interpret inconsistent text inputs, while healthcare players work to improve the quality of their datasets and co-operate with holders of similar data. It is not coincidence that the opportunities we identified are rising in areas with better quality datasets.
  • Complex industry structure complicates AI adoption. With multiple, fragmented, and often competitively opposed stakeholders – patients, physicians, payers, providers, government, suppliers, researchers, etc. – it is difficult to aggregate data, to gain permissions, to get buy-in and participation, to implement solutions and to get paid. Furthermore, the inherently high stakes and public scrutiny further limit access to data and create regulatory hurdles to implementation. The “black box” structure of AI models also makes it difficult to precisely define the logic behind the recommendations it generates, leaving care givers and insurers potentially vulnerable in the highly litigious healthcare environment.
  • Talent is scarce. Experienced AI scientists are rare and expensive, with most still closely aligned with the biggest commercial platform players. Only these leaders, and perhaps some academic faculties, have the capability to tackle bigger, more generalized, problems addressable by AI. Startups and healthcare incumbents will have to stay much further from the AI cutting edge, perhaps relying on increasingly sophisticated AI platforms offered as a service by the big players. At the same time, medical expertise is also a scarce and necessary element to healthcare AI solutions, a substantial further barrier to completely “outside the box” industry disruption from Silicon Valley.
  • Evolution, not revolution. AI is unlikely to entirely disrupt healthcare nor will a single player become dominant – the barriers to adoption are substantial. We believe narrow, focused initiatives with clear benefits and aligned motives in areas of the industry with better data and fewer risks will be successful. Both AI leaders and well-funded startups are working in this direction, largely in the 3 opportunity areas we identified, and early wins will open wider opportunities. We expect slow progress on harder to reach pain points – easing physician paperwork burdens, building more reliable diagnostic tools, reconsidering treatment protocols, tackling compliance, etc. – particularly in the US market. Breakthroughs might be possible in less fragmented markets and/or where regulation is less restrictive.
  • AI will drive more computing in healthcare. Healthcare data will be scrubbed, new data will be collected under better protocols and new AI techniques will produce better results with less data and less than perfect data. Despite the complex industry structure and fragmentation, stakeholders will rally to high value applications as they emerge and build enthusiasm to tackle projects of greater and greater scope. AI talent shortages will lessen, and platform players will provide powerful, low cost infrastructure to empower 3rd party AI system development. With this, we expect healthcare to become more computing intensive, for medical applications to become more uniform, for new uses of computing to proliferate across the complex value chain, and for AI, running in the cloud, to be at the core of all the change. Globally, this could affect $50-100B of annual business for the tech industry within a decade.
  • Who will benefit? We believe startups, combining AI and medical talent to address specific pain points via SaaS applications, will be unusually successful, with less threat of a single dominant platform with overwhelming scale advantage. We also see substantial opportunity for IT consulting (e.g. IBM, ACN, etc.), bringing data quality discipline and AI development expertise to forward thinking healthcare organizations, none of which have the skills to pursue AI alone. AI will also hasten a transition to the public cloud (e.g. AMZN, MSFT, GOOGL, IBM), with much lower costs, better performance and superior AI support. While independent consultants and the top AI clouds will all participate, IBM has been the most active, acquiring promising startups and signing high profile partnership deals – albeit after a few public missteps. Importantly, it has the AI chops, the AI infrastructure, and the well-established industry relationships to lead the slow transformation of healthcare.
  • What does it mean for the healthcare industry? Because the changes are likely to be slow and incremental, it will take years to sort out winners and losers. Still, those companies that do not begin the journey will find themselves on the wrong side when that reckoning happens. Even seemingly sophisticated med tech names, like GE or SIE, lack the AI chops to go it alone, so effective partnerships will be critical.

Alexa … Read My MRI

At first glance, healthcare is a perfect match for AI. Overburdened physicians, ambiguous patient records, millions of important decisions made daily, and $3.5T in annual costs in the US alone, rising rapidly. As much as a third of US healthcare spending may be waste – unnecessary services, excess administrative costs, inefficient care delivery, inflated prices, fraud and prevention failures. Machine learning could improve almost every element of the value chain – i.e. predicting vulnerabilities, diagnosing illness, optimizing treatment, promoting patient compliance – while reducing costs, streamlining payments, protecting against fraud, eliminating bureaucracy, and automating mundane tasks.

In real life, the match is not so perfect. AI likes clean data, but healthcare data is mostly a mess. Fragmented datasets, inconsistent formats, frequent errors, often non-existent labeling, and other shortcomings are substantial obstacles. Even getting permission to use healthcare data is difficult. The industry is a complex web of stakeholders – patients, physicians, provider organizations, insurers, pharmaceutical and medical device suppliers, and governments – often with very different priorities and/or in competition with one another, while dealing with life-and-death risks and tight scrutiny. Given these obstacles, it shouldn’t be surprising that the healthcare industry spends less on IT than any other major sector of the economy. There are a million pain points in healthcare – anyone off the street could start a decent list – but there is no way to get the consensus support needed to make comprehensive change.

So, the change will come incrementally… and slowly. It will start where the data is best – Medical Imaging has the most attention, along with Wellness (all those Fitbits and Apple Watches), and Drug/Device Development. Healthcare organizations don’t have AI expertise, and the leading platform players that are hogging that talent don’t understand healthcare, so expect partnerships. IT consultants will scrub the data, developers will build applications, and AI cloud platforms will host them, all on behalf of providers, payers, suppliers and regulators. IBM may be the only company able to do all three things, and while it has a bit of egg on its face for underestimating the data obstacles in a few overhyped projects, its skills and positioning make it an obvious leader in the early days of healthcare AI. GOOGL and MSFT have expressed interest but haven’t yet made serious efforts to build the relationships that will be necessary to play further up the value chain. We see lots of room for startups able to combine a bit of technical talent with some medical expertise to address highly focused solutions for high value pain points. IBM has already acquired a few of these with promise, and both tech and healthcare players will be on the lookout for good fits from M&A.

We think good results – removing those pain points, improving outcomes, and reducing costs – will promote further investment. Some of those good results may come from countries where industry structure and government policy make it easier to take bigger steps. Payoff for tech will come not just from selling applications, but also from providing services, and from hosting, as IT budgets as a percentage of industry revenues trend higher. With time, some of these projects will take on more comprehensive change, pressuring healthcare industry participants to step up. Some will, and others won’t, and it’s much too early to judge what companies will follow each path. In a decade, we think AI adoption will drive an exodus of healthcare computing from PCs and private datacenters to the public cloud, all to the benefit of the big platforms – AMZN, MSFT, GOOGL and IBM – that we expect to dominate that business.

Healthcare AI – Help For Where It Hurts

Healthcare is nearly 18% of the US GDP (Exhibit 1 & 2). However, while it is hard to separate all the enterprises that participate in that revenue pool, it is very unlikely that the sector accounts for more than 10-12% of total IT spending. Of course, some of this makes sense – healthcare requires highly trained and well compensated medical personnel along with plenty of real estate and specialized equipment, while consuming expensive pharmaceuticals, devices and other very specific consumable supplies. Still, most consumers can relate to service failures that might have been avoided with thoughtful investment in computing – from finding a doctor and filling out the clipboard full of forms on the first visit, to chaotic trips to the ER, and to wrestling with insurance on the phone. Medicine often feels like a manual process.

Exh 1: US National Health Expenditure Distribution, 2017

It is in this context that we consider the likely path of adoption for AI technology in healthcare. Today, AI means Machine Learning (ML) and its particularly powerful variant, Deep Learning (DL). ML is a programming technique, which uses thousands of recursive algorithms that self-adjust to deliver ever improving results relative to established goals. Sufficiently complex DL models iterating through very large datasets, built and regularly tweaked by highly experienced and talented computer scientists and aimed at the right problems can result in powerful applications able to interpret ambiguous data and draw out subtle conclusions that can outperform human experts on the same tasks. This is the technology that allows computer programs to interpret spoken commands, to anticipate user needs, to make stunningly accurate predictions and to drive cars.

Exh 2: National Health Expenditure Forecast, 2010 – 2025

Exh 3: US Per Capita Health Care Dollar Wastage, 2017

AI solutions could be valuable in almost every facet of healthcare. Epidemiologists could use build analytic systems to better predict the spread of disease. Patient vulnerabilities to various conditions could be accurately assessed, and preventative programs better designed and more easily followed by patients. More precise triage could begin at home, avoiding wasteful office and hospital visits, while getting higher priority situations appropriate attention more quickly (Exhibit 3). Diagnoses might be made faster and more accurately, with less wasteful testing, and medical expertise levered more widely. Treatment protocols could be improved, and patient compliance more closely monitored, enabling better outcomes at lower costs. Physicians could be freed from mundane tasks– it is estimated that US doctors spend twice much time on record keeping as they do seeing patients. Payment processes could be greatly improved – eliminating fraud, hastening approvals, cutting bureaucracy. There are millions of pain points to be addressed (Exhibit 4).

Globally, healthcare is a nearly $8 trillion industry. More than $3.3T of that is in the US, growing at a better than 6% clip. The potential opportunity for companies that help AI deliver on its promise – to both reduce costs and improve the efficacy of care – is huge. But it will be very hard to get.

Exh 4: Key Areas for Artificial Intelligence Impact in Healthcare

Low Quality Data

AI needs data. It needs LOTS of data to get meaningful insights, and it greatly prefers that the data be error free, well labeled, fully representative, and in consistent and easily read formats (Exhibit 5). Healthcare data has problems. First, patient data is fragmented. Not only are the records tied to the specific caregiver organization and/or insurer involved with a specific patient, but elements of the record spread across multiple databases. Radiology may have the x-rays, pharmacy has the medication records, while the childhood history may be in a metal filing cabinet in the patient’s home town. Compiling a comprehensive picture of a person’s lifetime health is often impossible, as is building a decent sized population dataset of patients with any specific characteristic.

Second, there are a lot of errors – images misread, a decimal point entered in the wrong place, diagnoses botched, etc. – that have not been corrected. A physician looking at a patient may discount odd data in a file, or check it by asking a question, but for AI systems trying to learn by iterating through millions of

Exh 5: Factors Determining the Value of AI Data

 records, the errors can badly skew the workings of a deep learning model. Resolving the problem, even on a going forward basis, is complicated by a data collection process that favors expedience for the practitioner over accuracy and usefulness. Doctors are asked to fill out forms, and often simple multiple-choice menus are inadequate to describe the nuance of a patient’s condition, requiring plain-text written explanations. The standards for these prose blurbs are ambiguous and constantly changing. One dermatologist may describe a melanoma as she was taught in medical school, which may be different from the terms used by a colleague across town educated in a different place and era.

Exh 6: Typology of Data in Healthcare Useful for AI