The inconvenient truth about "The 'inconvenient truth' about AI in healthcare

Drs. Panch, Mattie, and Celi recently published an article in Nature’s Partner Journal, Digital Medicine titled, “The ‘Inconvenient Truth’ About AI in Healthcare.” It’s a thought-provoking piece and I recommend taking the time to read it. They identify a huge problem in healthcare. But as brilliant as the three authors are, and as much as I hate to disagree with them, I think it’s important to offer second opinions on the cause of the ailment and a subsequent course of care.

Despite its title, their article isn’t really about AI. It takes on a more important issue; the challenges presented by the lack of access to electronic medical record data (EMR). AI is used as an example of a capability hindered by the lack of access. But of course, lack of access causes greater harm than just slowing AI adoption.

Anyone that has attempted to gain access to EMR data for research, quality improvement, clinical decision support, population health, etc, knows how frustrating, endemic, and detrimental the lack of access to data is. 

They go on to lament that, most AI we read about is not executable at the front lines. Two reasons why are offered:

• AI alone doesn’t change entrenched financial incentives
• Healthcare organizations lack the data infrastructure

Here’s where I disagree. The root cause is the lack of incentive — not the inability of AI to change entrenched financial incentives and not the lack of data infrastructure. This problem affects any effort at learning and improving. AI, data infrastructure, population health, interoperable EMRs, clear discharge instructions — these are all tools for keeping people healthy and out of the hospital. None are widespread realities due to the lack of incentive.

Fortunately, there’s some progress in the shift of incentive. While the majority of care is still paid for with FFS mechanisms, the number of value-based contracts is rising. According to a recent report, 34% of healthcare payments were tied to value-based contracts in 2017, up from 23% two years prior. Many of these contracts still represent minimal downside risk, but it’s moving in the right direction.

So too are investments in digital infrastructure and tools, including AI. According to the CB Insights 2019 Global Healthcare Report, startups working on AI in healthcare reached an all-time funding high in the second quarter of 2019.

I’m the founder of one of these startups. Lack of data infrastructure makes projects more complicated but we’ve come a long way in what we can do with heterogeneous, distributed data. On the other hand, my business could not exist without this shift in incentive.

At one Accountable Care Organization (ACO) the Cyft team is combining claims data and clinical data from three different EMR systems to prioritize those most likely to benefit from care management and palliative care programs. The reason why the impossible is now possible? Up to 15% of the ACO’s revenues are dependent on their ability to reduce preventable cost.

The problem with treating the lack of data infrastructure as a cause versus a symptom is that it affects the authors’ recommended course of care.

They call for a “secure, high-performance data infrastructure to make use of this data for AI applications.”

Their “two possible routes to building the necessary data infrastructure” include:

1. Follow the example set by some research datasets such as MIMIC
2. A government mandate that all healthcare organizations store their data in commercially available clouds

There is a precedent for prioritizing investment in infrastructure versus changing the incentive that would lead to changes in infrastructure. In 2009 President Obama introduced $19B for the adoption of EMRs in the American Recovery and Reinvestment Act. It was, in effect, a carrot and stick mandate that led to greater than 95% EMR adoption in just a few years.

At the time, I argued in JAMA that we were at a crossroads. I worried it was irrational to expect investment in any technology to lead to lower cost and higher quality care while healthcare reimbursement policies reward volume and complexity.

Ten years later, according to Panch, Mattie, and Celi, “The typical lament of a harried clinician is still ‘why does my EMR still suck…?” We know why. Well-meaning people, rightly frustrated with the state of electronic health data, mandated infrastructure rather than create incentives for keeping people healthy.

I stand with Drs. Panch, Mattie, Celi and the thousands of others demanding access to data that will help us keep people healthy. I would, however, advocate that our efforts recognize an inconvenient truth — without aligned incentives, no technology, mandated or otherwise, will help us meaningfully improve the value of healthcare in the US.

Leonard D’Avolio, PhD

CEO, Cyft

Asst. Professor, Harvard Medical School & Brigham and Women’s Hospital

[email protected]

@ldavolio