Was it the Data, or was it the AI that killed IBM Watson? What if AI Doesn’t Scale in U.S. Healthcare?

IBM announced Thursday that it was preparing to sell its healthcare business (IBM Watson Health). IBM had invested billions to develop tremendous processing power, and IBM Watson had acquired

huge datasets to educate its AI engines (e.g. Truven acquisition, 2016). No doubt IBM has developed incredibly powerful tools and datasets and it enjoyed some successes. The Wall Street Journal’s analysis points at problems applying AI to healthcare data, but I think the problem is much broader and deeper than either of those technology challenges. IBM’s Retreat From Watson Highlights Broader AI Struggles in Health 

Perhaps what evaded IBM wasn’t a particular dataset or processing technology innovation, but rather the capacity to efficiently apply technology to discrete, highly variable and complex (read: customized) solutions in highly idiosyncratic organizations for highly individual people. Healthcare tends to embrace complexity and to defy standardized solutions – which is the entire premise for technology.

Likely buoyed by its success in other industries, IBM had demonstrated extraordinary optimism and determination for over a decade as it looked for healthcare problems that its Watson technology could profitably solve. For at least the last five years, when healthcare industry analysts like me asked IBM to name major revenue-generating clients, the list was always short; the use cases seemed awfully niche and/or difficult to exploit. Now that Watson’s well of investment capital has finally run dry, the Wall Street Journal (WSJ) has consulted experts and identified three challenges that killed IBM’s project:

Gaps in knowledge about complex diseases whose outcomes often depend on many factors that may not be fully captured in clinical databases”

“Lack deep expertise in how healthcare works, adding to the challenge of implementing AI in patient settings...’You truly have to understand the clinical workflow in the trenches’”

“The lack of data-collection standards, which makes taking an algorithm that was developed in one setting and applying it in others difficult”

None of these challenges are new and none are unique to IBM’s healthcare business. Private and public sector experts have articulated and tried to fix these problems for decades - at least since the Clinton administration. Since the presidency of George Bush, federal initiatives have imposed technology-led solutions as these:

Despite public and private sector efforts (most of these are joint efforts) to improve access, cost and quality, there remains a huge opportunity to wring efficiencies out of the U.S. healthcare system. (Five years ago, I estimated that approximately one quarter of the ~$4 trillion U.S. healthcare cost is consumed by fraud, waste, and abuse.) However, the healthcare industry seems to add services and products constantly without showing much regard for cost reduction or overall productivity improvements. The drive to impose cost reduction and productivity improvements in healthcare (as in higher education) seems muted compared to that drive found in other industries. I believe it is likely that healthcare’s lack of drive to impose cost reduction and productivity improvements dampens the healthcare industry’s appetite for technology solutions in general.

If true, this might be reason that stepping into most healthcare environments in the U.S. feels like time travel: while paper charts are on the way out, in many cases you can still only pay by check or credit card during business hours. Viewed through a technology lens, healthcare presents huge opportunities to reduce costs and increase productivity. This perspective may have blinded IBM, whose successes in other industries may have seduced it to invest too much, for too long. It might be same story for Haven (the Amazon/ JPMorgan/Berkshire Hathaway initiative recently announced it was closing down). It might be the same story with Google’s healthcare initiatives (DeepMind).

IBM’s announcement might present a good opportunity to reassess adoption and obsolescence trends for technology in the healthcare industry generally. Why have technology adoption trends in the past proven so fraught, uneven, and surprising? How have healthcare adoption trends progressed compared to other industries for supply chain/ERP; operational workflow (EMR/EHR, practice management and telemedicine); finance and accounting (revenue cycle management); communications (pagers and mobile phones; faxes, email and chat). What exactly are the extraordinary challenges that healthcare presents to technology adoption?

Now might also present a good opportunity to revisit the non-technology related challenges that industry has faced over the last three decades. I am not very familiar with the Clinton- era proposals, but I would guess they identify problems that still plague the industry and reduce its efficiency and productivity, such as:

Poorly-defined and inefficient organizational and clinical processes 

Fractured and unscaled supply chains, procurement processes

Lack of transparency in incentives, contracting and cost accounting

Highly variable human responses to disease and care that defy rigid clinical pathways

Vested human interests that resist changes to industry and organizational structures

Large scale technology solutions may not always – or even frequently - be the primary lever of cost reduction and productivity improvement for healthcare organizations. If true, then the healthcare industry may never have been suited to the kinds of offerings that IBM has been seeking to identify and implement at scale. Instead, the U.S. healthcare may be better suited to embrace offerings from more specialized companies that address local, narrow solutions to discrete problems within the context of more long-standing and intransient challenges. “IBM Explores Sale of IBM Watson Health” WSJ (Feb 18, 2022)

“IBM’s Retreat From Watson Highlights Broader AI Struggles in Health” WSJ (Feb 20, 2022).