As reported in a rambling article in the Wall Street Journal, Microsoft has announced the acquisition of Nuance Communications, Inc. The WSJ predicted, with its usual WSJ bravado, that this bold $16B move (paying 13 times Nuance’s 2020 revenue) will “kick start growth” in the “next big thing” — “healthcare.” But, this is not a bold move. It is weak because it is a shiny object “AI” acquisition. And, more importantly, it is weak because it conveys fear of health care process change, and fear of rejection by change-resistant physicians.
I’ve written a post about Nuance before, back in 2011, after spending time in their booth at the HIMSS convention, appropriately held in Las Vegas. Nuance became a leader in healthcare voice recognition (VR) when they acquired Dragon and its “Naturally-Speaking” technology years ago. It is a good product if the value proposition is to reduce the cost of manual dictation-transcription services without bothering doctors too much. Like other voice recognition companies, and their cousins, the optical character recognition (OCR) and natural-language processing (NLP) companies, Nuance tries hard to tell a more glamorous value-story. They portray VR, OCR and NLP as being about extracting the unstructured data in voice dictation, textual reports, handwritten notes, and medical images, so as to turn them into structured data that can be used by computers for all kinds of different purposes, such as medical billing, quality and outcomes measurement, reminders and alerts, drug and device effectiveness research, and more. This vision is very popular — including among Microsoft executives hunting for bold acquisition targets — because it seems to overcome a major barrier in efforts to modernize and improve the health care industry. That feared barrier is the reluctance of physicians to make changes to their daily practices, particularly changes that resemble tedious, clerical data entry. Nuance and its brethren have made impressive improvements in the ability to accurately capture data from voice, text, and images. But, after years and years of persistent hype regarding variations on this structured data theme, the results have always been generally disappointing, and the vision remains aspirational.
Why is data captured from voice, text and images so disappointing?
Many health care IT experts and most health care technology investors, are very clear about the difference between unstructured data (like voice, text and images) and structured data — data expressed in codes and numeric values that can used by computers as inputs for queries, logic, rules and algorithms. But these same people are very unclear about the difference between passively structured data and actively structured data. That is an unfair statement because I made up that terminology years ago and have so far been unsuccessful in promoting it into common use within our field. But it is fair to say that the concept of passive vs. active structure is poorly understood by people who have not spent a lot of time actually putting structured data to use in actual health care processes and to pursue actual lines of analytic inquiry.
The crux of the difference between passively and actively structured information is that actively structured information is proactively captured in a way that enforces the integrity and completeness of the data. If a doctor speaks words into a microphone, even if the voice recognition software does a perfect job of converting the sounds into text, and even if the natural language processor then does a perfect job of interpreting the meaning of the text to generate codes and numeric structured data, the resulting data is only as complete as the doctor’s words. If the doctor never vocalized a piece of information that is necessary for a particular downstream use, the data available for that downstream use is incomplete. This is not a big problem for some downstream uses, such as to generate human-readable summaries and to support search use cases. Google and Ping can deliver plenty of useful search results without having to guarantee that the results are exhaustive and perfectly complete. But for many of the most important uses for structured data in health care, missing values are a big problem. If I am trying to generate a quality or outcome metric that is a rate measure or a proportion (as are many such metrics), if I am missing some of the data needed to define the denominator or the numerator, my rate or proportion will be biased and untrustworthy. If I am relying on certain data elements that are required inputs to an alert rule, and some of those data elements are missing, I cannot guarantee that no patients will fall through the cracks. If I am trying to get numbers to add up, I need all the numbers. The inconvenient truth is that the most important purposes require actively structured data, no matter how annoying it is to ensure complete and accurate capture of the data. Nuance does have capabilities to configure voice-driven clinical document workflows in ways that capture some structured data, particularly in focused clinical domains such as radiology. But few health care organizations invest in developing the internal expertise to configure the platform to achieve a high degree of active structured data capture, and even fewer organizations take the time to integrate the resulting actively structured data into important downstream data uses.
My Advice to Microsoft
At the risk of being uncool, I remain a big Microsoft fan. My favorite device ever is my Microsoft Surface (a cherished gift from my favorite team of collaborators ever). I use Microsoft 365 on a 365 basis. Reward Health uses Microsoft Azure for cloud infrastructure. I wish Microsoft well and believe they can and will accomplish great things in health care, despite previous false starts. (Remember Caradigm?) But my advice to Microsoft is to avoid too much focus on two things. First, avoid getting twitterpated by shiny objects, like AI, ML, blockchain, big data, containers, microservices, etc. Second, avoid trying too hard to avoid doing things that involve changes in behavior by physicians. It is true that physicians are loath to change and that health care enterprise IT leaders cower in fear of refusal by physicians to adopt and use the clinical applications they deploy. But, Microsoft, please keep your eyes on the prize. In every industry where information technology has created great value, it is because it has enabled change in business processes. The process change is the source of value. So, please don’t continue to shape your acquisition strategy by a desire to sell technology based on how great it is at being deployed without any process changes required. Mr. Nadella, if you have any questions, please don’t hesitate to call.