Skip to content

Dead End: IBM Watson Health

Abstract

IBM won a quiz show in February 2011 and concluded it had a medical product. Watson Health, formed in 2015, was built by spending more than $4 billion on health-data companies and pointing the Jeopardy! question-answering engine at cancer. The flagship project at MD Anderson was scoped at $2.4 million for six months and cost $62.1 million before an audit killed it with nothing in clinical use. Internal documents leaked in 2018 showed the oncology product recommending treatments its own reviewers called unsafe and incorrect. In 2022 IBM sold the whole unit to a private equity firm for about $1 billion, roughly a quarter of what the acquisitions alone had cost. The technology was not fraudulent and the doctors were not obstructive. IBM sold a product it had not built yet, into the one industry where a wrong answer is a clinical event.

The Moon Shot

Watson beat Ken Jennings and Brad Rutter over three nights in February 2011 (see Fun Fact: Watson on Jeopardy). The system read a natural-language clue, generated candidate answers from 200 million pages of text, scored the evidence for each, and buzzed when confidence crossed a threshold. It was a real achievement in question answering, and its architecture was a hand-built pipeline of retrieval and scoring rather than anything that learned end to end.

IBM’s reading of the win was that a machine which could answer questions from text could answer medical questions from medical text. Chief executive Virginia Rometty called healthcare the company’s moon shot. The public timeline attached to it was 18 to 24 months: read the patient’s record, read the literature, propose the treatment, and take some of the load off a profession that was short of doctors and drowning in publications.

The premise was reasonable enough to attract serious hospitals. The trouble was the order of operations. As Robert Wachter of UCSF put it, “They came in with marketing first, product second, and got everybody excited. Then the rubber hit the road. This is an incredibly hard set of problems.”

Buying a Health Business

IBM did not have medical data, so it bought some. On 13 April 2015 it announced Watson Health and the Watson Health Cloud together with the acquisition of Explorys, a Cleveland Clinic spin-out with a large clinical dataset, and Phytel, a Dallas population-health company; the Phytel deal closed that May. Merge Healthcare, which handled medical imaging, followed later that year. On 18 February 2016 IBM bought Truven Health Analytics for $2.6 billion, the largest of the deals. Together the acquisitions came to more than $4 billion.

What that money bought was data and customer relationships. It did not buy a working clinical decision-support product, and the acquired businesses were mostly conventional health-IT operations that had nothing to do with Watson. For several years the unit’s revenue came largely from the things IBM had bought rather than the thing it had built.

MD Anderson

The University of Texas MD Anderson Cancer Center signed with IBM in October 2013 to build the Oncology Expert Advisor, a Watson application that would take in patient records, literature and treatment options and advise clinicians. The original contract was scoped at six months and $2.4 million.

A University of Texas System audit released in February 2017 recorded what it had actually become. As of 31 August 2016 the project had cost $62.1 million: $39.2 million to IBM and about $23 million to PricewaterhouseCoopers for consulting and project support. The audit found the procurement had bypassed the system’s normal rules. Two pilots had run, one in leukemia and one in lung cancer, both against MD Anderson’s old ClinicStation records system; the tool had never been made to work with Epic, the system the hospital had since moved to. There was no product in clinical use.

The technical detail underneath the money is the more useful part. The hard problem was not answering the medical question. It was reading the record. Extracting simple diagnostic concepts from clinical notes worked at 90 to 96% accuracy, but anything time-dependent, such as which therapy a patient had received when, ran at 63 to 65%. A machine that misreads a third of the treatment history is not a machine that can recommend the next treatment.

What Watson for Oncology Learned

The commercial product, Watson for Oncology, was trained with oncologists at Memorial Sloan Kettering, and the training method is the root of most of what followed. Watson did not learn from outcomes data across millions of patients. It learned the treatment preferences of a small group of specialists at one American cancer centre, encoded case by case.

Where the literature was thin, the doctors wrote cases for it. Mark Kris, the Memorial Sloan Kettering oncologist who led the work, described the mismatch: the information a physician takes from a paper “may not be the major point of the study.” When the FDA approved a cancer drug on the strength of four patients, there was no dataset to learn from, so the team built synthetic cases so Watson would produce the recommendation a human expert would.

The result was a system that reproduced Memorial Sloan Kettering, which is a defensible thing to sell in New York and a strange thing to sell elsewhere. Concordance studies made the problem visible: Watson’s recommendations matched the local tumour board about 73% of the time at a hospital in India and about 49% in South Korea, where it failed to suggest drugs that were standard practice there. Medicine is local, and the product had been trained on one locality’s judgment rather than on evidence.

The Internal Documents

In July 2018 STAT published internal IBM material, slide decks from Watson Health’s own deputy chief health officer, reporting “multiple examples of unsafe and incorrect treatment recommendations.” The documents laid the blame on how the system had been trained by IBM engineers and the Memorial Sloan Kettering doctors, and questioned “the process for building content and the underlying technology.”

One case in the documents involved a 65-year-old man with lung cancer and severe bleeding, for whom Watson suggested chemotherapy plus bevacizumab, a drug whose labelling warns of severe or fatal haemorrhage and which is contraindicated in exactly that situation. No patient harm was reported; the recommendations were reviewed by clinicians, which is what kept them advisory. IBM responded that it had shipped eleven software updates in the previous year and improved the product on client feedback.

Dead End: Why It Failed

Three separate problems, any one of which would have been enough.

The input was not text, it was clinical notes. Watson was built to answer questions from well-formed prose. A patient record is abbreviations, negations, copy-pasted history, dictation errors and timing that has to be inferred. The Jeopardy! engine’s core competence did not transfer to the corpus that mattered, and the MD Anderson accuracy figures are the measurement of that gap.

Expert opinion was substituted for evidence. Encoding what a group of specialists would do is a 1980s expert system with better natural language on the front (see Expert Systems and the First AI Winter). It inherits the same limits: it does not generalise past its authors, it cannot be validated against outcomes, and it degrades the further you carry it from where it was written. Martin Kohn, a former IBM medical scientist, put the missing piece plainly: “Prove to me that it will actually do something useful, that it will make my life better, and my patients’ lives better.” Peer-reviewed evidence that it did was scarce.

The sales cycle ran ahead of the product cycle. Hospitals signed on the strength of a television demonstration and a corporate promise, then spent years discovering what the system could not do. That sequence is survivable in enterprise software. In medicine it burns the institutional trust that the next, better system will need.

The Sale

IBM announced on 21 January 2022 that it was selling the Watson Health data and analytics assets to the private equity firm Francisco Partners. The deal closed on 30 June 2022 at a reported $1 billion, and the business was relaunched as Merative, headquartered in Ann Arbor, Michigan, under chief executive Gerry McCarthy. What changed hands was the acquired portfolio: MarketScan, Micromedex, Health Insights, Clinical Development, Social Program Management, Phytel and the Merge imaging products. Merative is a health-data company, which is what IBM had bought and not what it had promised to build.

The episode sits awkwardly in AI history because it was neither a hoax nor a research failure. Watson worked at what it was built for. IBM took a narrow retrieval-and-scoring system, described it publicly as a step toward machine medical reasoning, sold that description to teaching hospitals, and spent a decade and several billion dollars discovering the distance between the two. Nine months after the sale closed, ChatGPT made medical language models a mass topic again, on a technology with no relationship to Watson’s pipeline (see The LLM Race and The Rise of Artificial Intelligence). The claims being made for those systems in clinical settings are being met, this time, with demands for trial evidence, which is the one durable thing Watson Health produced.

πŸ“š Sources