When it comes to healthcare technology solutions powered by artificial intelligence, speech-driven workflows are poised to have an outsized influence on the care experience for the most essential stakeholders — patients and clinicians.
Here Thomas Polzin, PhD, director of natural language understanding with 3M Health Information Systems (3M HIS), answers six questions on the evolving use of conversational AI technology in healthcare.
Question: So much of medical communication happens with speech but is all speech recognition created equal, more or less?
Dr. Thomas Polzin: I wouldn't say so. In the early days, the conversation centered on recognition of the spoken word. The value proposition was simple: you spoke a word and the software turned it into text in real time. But from the very start, we realized that for the healthcare industry, this would not be enough. The goal was to create technology that helps physicians capture the complete patient story without getting in the way of the patient-physician interaction or being a burden on the physician. At 3M HIS, which acquired M*Modal in early 2019, we knew that in order to achieve this, speech would have to be understood, not merely recognized. By that I mean, as the doctor speaks into the medical record, the technology must understand not only the content of that speech but also the context. Of course, it must also know the subspecialty lexicon, and be aware of what's going on in the overall medical record. Only then can it bring true meaningful respite to the physician and reduce the resistance to widespread adoption with high out-of-the-gate performance.
Because of this, we refer to our cloud-based technology as speech understanding because it includes natural language understanding for that ability to also interpret and contextualize the spoken word. This may seem like a small thing, but it makes all the difference in the immediate user experience and helps lay the foundation for cooler things that can then be done.
Q: You mentioned the medical record. How important is it for this technology to be integrated into the EHR?
TP: It is critical. Without this integration, any innovation in this industry would not be truly meaningful in moving the needle on what healthcare organizations need from technology. These physician-assistive technologies need to complement the EHR and help doctors get all the right information in the right places within the record more easily, efficiently and naturally. For 3M HIS partnering with EHR is a big focus — our solutions are compatible with over 250 EHRs and that number is continuously growing. But it's not just about more EHR partnerships, it's also about deeper EHR partnership to continue delivering incremental value to joint clients.
Q: With AI, it's really hard to tell what is fact and what is fiction. What are your thoughts on this?
TP: It's true that it's sometimes hard to tell where the healing ends and the hype begins. But AI has true potential for healing some of what ails our healthcare system, especially when you think of the administrative and cognitive overload on clinicians. We expect doctors to not only be great doctors, but also documentation experts, compliance officers, coders and risk mitigators — all while staying engaged with the patient. This is where AI can play a significant role. But to me, it's not a black box that can do everything right out of the gate. Conversational AI is a really difficult challenge, one that requires an incremental approach and I prefer to think of it as explainable AI. We started by understanding what was being said by the clinician in context to the patient record and made it easier for the physician to interact with the EHR for a superior user experience. Then we took that contextual understanding to also provide closed-loop nudges in real time to the doctor regarding high-value clinical insights. Incrementally innovating along this continuum, AI provides greater automation in the physician documentation process and in EHR information retrieval. Similarly, for backend coding and quality workflows, AI drives meaningful prioritization and automation to improve efficiencies.
Q: What you said about our unrealistic expectations of doctors is so true. Can you talk more about this overload and how technology can help with real-life scenarios?
TP: Physician burnout has truly reached critical proportions now, exacerbated by the COVID-19 patient crush and bad clinical outcomes for hospitalized patients. It's not a buzzword or a cliché. It requires extreme empathy from all of us and a targeted commitment from those of us who can help alleviate it. At 3M HIS, our mission is to create time to care for the well-being of both patients and physicians. Focusing solely on physician efficiency so that doctors can see more patients can have a negative impact on the user experience, quality of care and patient outcomes.
A recent study in the New England Journal of Medicine talks about the "Ecology of Attention" and how detrimental task shifting is to patient care and physician wellness. For instance, physicians could be queried on a patient record they may have documented a week ago. Not only are they then asked to shift their attention from the task at hand, but they are also compelled to recollect details of that patient encounter which happened days ago and/or sift through the EHR to find that nugget of relevant information needed to satisfy the query. Instead, what if we could deliver the right clinical insights to physicians proactively as they are documenting the patient encounter, nudging them with context-specific insights to close any gaps in patient care and the clinical note — all this in real time and within the normal EHR workflow? Such computer-assisted physician documentation (CAPD) is a good example of "optimized attention" that leads to all sorts of downstream benefits, including physician satisfaction and well-being.
Q: How is this technology customized for healthcare?
TP: Our approach to understanding clinical documentation is to derive an independent meaning representation using what we call information models. These information models rely on ontologies like SNOMED, LOINC, or RxNorm and also include, for instance, information about certainty and temporality that is abstracted from the clinical narrative. This goes beyond traditional NLP and makes it a more robust, adaptable, extensible and configurable platform. And that's why 3M HIS refers to it as natural language understanding. We have built a large content library with associated information models so we can cover a large number of medical conditions and procedures.
Take heart failure as a use case: we have built a very complex information model to get the most accurate clinical picture around heart failure, even if the physician doesn't explicitly mention heart failure in the note. Our information models link conditions to things like symptoms, procedures, medications and laboratory results. Ultimately, this allows us to incrementally advance and refine our natural language understanding and deliver ever-increasing value to the physician and the healthcare organization.
Q: Where do you see this technology going in the next five years?
TP: Natural language understanding, in and outside the healthcare industry, is making great strides every day and becoming more data driven. It will soon be at a tipping point where things that have not been possible for only technology to achieve will be fully automated. But until that day comes, we must continue to make steady but strong progress towards our ultimate goal and let some human intervention carry us over the last few miles until technology alone can get us all the way there.
3M HIS, and M*Modal as well, has always believed for technology to be extraordinary, it needs to be quite ordinary to the user. Technology is never the end but merely a means to an end. Therefore, it should be unobtrusive and present when you need it. In the CAPD example I mentioned earlier, the user sees an interactive nudge appear and then maybe disappear, but in the background the technology is working hard in real-time to normalize the data, contextualize it, analyze it, apply clinical and sematic reasoning, summarize all the documentation in the encounter, bring best practices to light and report on the physician engagement with the system. Technology should work to enhance the user experience and the patient-physician relationship without getting in the way. To really do that effectively, it needs to be quite extraordinary.