IBM Information on Demand Conference (Part 2)

By Ellen M. Martin

In Post 1, I discussed the IT-focused sessions. On the other hand, there were the “business” sessions. I spent one day sessile in the healthcare track room listening to a series of “use cases” that amounted to hype and vapor. One wag in the back of the room asked every speaker the same question: “What value has your enterprise achieved with this installation?” The answers were all variations on “Uh, nothing yet, but we have high hopes for transforming medical practice.” We are clearly still on the exponential track of the hype curve, with substantial real-world results yet to be gained. Not to discourage much-needed experimentation, but for hospitals and clinical practice the journey from hype to hope has barely started and seems to be still a few years away (not counting the fallout from ACA/Obamacare!).

The areas of greatest progress seem to be focused on foundations: IT infrastructure, migrating legacy systems, and transforming cultural practices, including convincing management to invest in expensive IT projects and physicians to use them. The sub-sectors that are furthest along are payment and reimbursement; not surprising, given nearly half a century of financial IT, transaction processing and data mining, thus experience in demonstrating value.

The most interesting presentation was on using natural language processing to fill in the gaping holes in EHR/EMR data on vaccination and routine screening status in a large hospital. Two observations:
1. electronic records are woefully incomplete from the data-entry stage and the quality of data is still very poor (missing input, out-of-date entries, duplications, etc.). There’s much cleaning up needed before this data can be useful in guiding clinical decision-making or improving patient outcomes.
2. Improving data quality through capturing and structuring the enormous amount of unstructured data (where, alas, most of the valuable information still seems to be trapped) using NLP and other techniques can create rapid value for those organizations that undertake it, as well as for the consultants and vendors that can provide such services.

The Watson presentation was particularly disappointing. This is the fourth Watson-on-healthcare presentation in the last 4 years I have seen myself or virtually through Bonnie’s attendance at meetings. Nothing appears to have advanced. The IOD speaker was an MD/Engineer, obviously a smart and articulate expert who had given his spiel many times before, but he rambled and hit no new talking points, the slides were few and uninformative. And the answer to the wag’s question, was “not yet.”

Reading between the lines, there is something not right about the Watson model of AI-aided diagnostics and clinical guidance. That model assumes that medical knowledge is a lot like Jeopardy (where Watson has excelled) and thus, a machine that can consume all of human medical knowledge and deliver statistically ranked Dx suggestions to practitioners can improve accuracy, reduce errors, improve outcomes, save time and lower costs.

But if Watson merely presents a list of obvious diagnoses with statistical weights (which is what I have seen in every example presented), it’s hard to see how that helps practitioners much, and easy to see how it might insult their expertise. It seems more valuable to offer non-obvious diagnoses—the ones physicians typically overlook—with algorithms to suggest next steps (questions, procedures, tests) that could tease out the few cases of unusual disease that now are frequently missed from the statistically common diagnostic possibilities.

The most “out there” talk was about using advanced statistical methods to produce artificially intelligent decision support for clinicians. Although the presenter was jet-lagged to near-incomprehensibility, he conveyed a vision of how AI could help determine which patients are at high risk of re-hospitalization, and guide providers to offer particular treatment programs to mental-health patients based on behavioral pattern recognition and expected values.

Perhaps a combination of the Watson encyclopedia and more sophisticated AI algorithms could improve on both models. A major goal for many is reducing re-hospitalization: it’s a financial hot button under regulation and reimbursement and has the potential to save money and improve patient outcomes through intervening in a relatively small proportion of cases, which can be targeted through simple algorithms.

Many presenters repeated the observation that healthcare companies are in the data business whether or not they know it or like it. Today’s challenge is not generating data (a tsunami of data is already upon the industry), but cleaning up, securing, managing and analyzing the data they already have to move towards the vision of personalized healthcare.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: