AI is spreading rapidly across the healthcare sector, with applications huge and diminutive finding their way into workflows across the industry.
Whether it’s assisting physicians during telemedicine visits, transcribing entire conversations between physicians and patients, writing notes for nurses in response to questions in the patient portal, helping patients triage issues via chatbots, or a host of other applications, AI is proving useful to many healthcare entities.
Narinder Singh has been involved in AI for years. He is the CEO and co-founder of LookDeep Health, a company that offers virtual sitting, virtual nursing, and virtual care. Previous roles include working at Accenture’s Technology Strategy Center, a position in the corporate strategy department in the office of the CEO of SAP, co-founder of Appirio, president of Topcoder and vice president of engineering at webMethods.
sat down with Singh to discuss how AI can support expand the potential of telemedicine, the risks generative AI poses to hospitals and healthcare systems, how healthcare organizations can overcome these risks, and the role AI plays in healthcare technologies.
Q. You note that telemedicine obviously removes the burden of distance from healthcare interactions. But you say that it doesn’t boost their capabilities. How do you see AI helping here?
AND. Let me start with some context about why this is a critical question, and maybe a question about the future of hospital care. Every week we talk to hospitals that are seeing either increased acuity or staffing constraints, and most of them are reporting both.
US Population Over 65 Between 2010 and 2020, the population grew five times faster than the total population – the fastest pace in more than a century. This is part of a long-term trend and underscores the increasing age and associated acuity of the patients that hospitals will care for in the future.
At the same time, we observed repeated disturbing forecasts devastating for nursing and other roles in the hospital – and that’s on top of financial pressures that make it almost impossible to boost staffing levels.
Now we have a generation of telemedicine in the hospital – from eICU to teleconsultation and now virtual sitting and virtual nursing. At the project level, there have been many successes, but at the macro level, collectively, telemedicine in the hospital has had very restricted impact on care, with one huge exception – COVID.
During the pandemic, we learned that seamless access via telemedicine creates flexibility that allows the system to adapt. However, it has not increased our resources. Tele-opportunities can bridge huge distances, but they do not change the basic units of work necessary to provide care.
Now, AI can mean many things, but let’s start with what it means for telemedicine—the ability to expand our observational capabilities (not how they affect decision-making). Right now, a nurse treating six patients will be in each patient’s room for one to two hours. Doctors will be in a patient’s room for typically just a few minutes a day.
That is why in the immense majority of cases the patient is deprived of the watchful eye of the doctor. This is despite the fact that so much of what is happening to the patient can only be assessed and understood at the bedside.
Are they less dynamic, are they struggling to get out of bed, does their breathing seem more labored, did the alarm go off because the sensor slipped off their finger or the endotracheal tube came out of their neck, etc.?
One branch of AI, computer vision, allows us to keep an eye on every patient at all times. This can support to more appropriately allocate the scarcest resources in a hospital – the clinical care of nurses and doctors.
We have decades of evidence that increasing clinical throughput has a positive impact on patients. Video alone – even in legitimately attractive areas like virtual nursing – will simply repeat the disappointments of the past. With AI, we can better leverage the time and expertise of our greatest constraint.
Imagine a world where AI acts as a guardian angel for patients and their caregivers. It identifies potential problems and alerts healthcare workers before a diminutive problem becomes a substantial one. This isn’t just about efficiency; it’s about fundamentally changing the way we deliver care.
AI can provide an extra layer of support, ensuring that no patient is left without care, even for a moment. It is not about replacing human touch, but augmenting it, making our healthcare system more responsive, more resilient, and ultimately more human.
Q. You warn that generative AI poses real threats to hospitals and healthcare systems. What are they?
AND. Generative AI has the potential to streamline prior authorizations, patient coding, and intricate interactions between insurers and healthcare providers. But it could also spark an epic civil war between them.
This productivity could lead to a faster but more complicated dispute landscape, ultimately requiring more human arbitrators to resolve them. Rather than reducing administrative work, it could actually boost it. Generative AI could scale the most cynical stereotypes of overuse and aggressive claim denial indefinitely.
AI tools are making progress in reducing the time doctors spend on paperwork, especially outside the hospital setting. But in hospital settings, the complexity of care and the lack of defined “visits” mean these tools are not yet as effective.
We have had years to learn how arduous and specific the development and application of machine learning algorithms are in hospitals. The temptation to magically remove this tedious demanding work and integrate it into clinical workflows is tempting but naive.
“Generative” designs are vital for many parts of healthcare operations, but they are not the silver bullet. They do not yet address the need to synthesize defined sets of information and repeatedly draw the same conclusions from them. Predictability of inputs and outputs is crucial for judgment and confidence in clinical decision-making.
Q. How can hospitals and healthcare systems overcome these threats posed by generative AI?
AND. On the first point about insurer-provider battles, I don’t see an immediate solution. You simply can’t afford to have humans trying to cope with the volume of requests or responses generated by AI, so participation in this arms race is inevitable.
However, engaging in a way that creates a foundation for evaluating and incorporating generative models into workflows provides leverage for the future. Key steps include securing PHI, providing checks and balances on results, evaluating models within and outside their scope, and not scaring employees away with premature claims to replace their roles for a few dollars an hour.
This is just the beginning.
We are already seeing insiders like Sequoia and Goldman questioning the hype and benefits of generative AI. We will go through the valley of despair; however, focusing on pragmatism and not falling in love with a broad proclamation will save many innovation teams from being cut. Hospitals need two antagonistic mindsets.
First, experimentation is necessary. Generating non-clinical content (emails, messages), assessing EHR context summaries, improving language translation and transcription – these are all areas where generative AI can be safely tuned and targeted for improvement. These applications can free up valuable time for healthcare workers to focus on more vital tasks.
Second, hospitals must enforce tough evaluation and require repeatability. For clinical scenarios, evidence of any claims of capability should be expected. Furthermore, there should be an approach to continuously assess the AI’s capability within the solution. Specific claims must ensure that the same set of inputs yield the same results, maintaining consistency and reliability in clinical decision-making.
In other industries, technologists, as Norman Vincent Peale once quipped, “shoot for the moon and settle for landing among the stars.” In healthcare, we’ve seen the disastrous consequences of such strategies, setting industries back a decade or more (Theranos for blood testing, Watson for artificial intelligence for cancer).
You can be pragmatic without being snail-paced – the right leaders will provide that balance.
Q. You’ve seen more than a half-dozen transcription companies raise more than $30 million in the last few years. Why is that? And what role does AI play in these scribe technologies?
AND. There are over a million physicians in the United States. Their time is incredibly valuable, and a generation treated as entry-level experts and data scientists has led to massive burnout.
The math is basic, and now the technology is more accessible than ever. The narrative that “the time is now” is nothing recent, but it may finally become a reality. It’s a great operate of technological advancement.
AI plays a key role in these scribe technologies, radically improving the accuracy and efficiency of transcription. With AI, transcription can be done in real time, with greater accuracy, and at a fraction of the cost.
The challenge is that AI advances have continued to advance at a breakneck pace in recent months—redefining the starting point for building such solutions. It’s clear that transcription solutions are not, in and of themselves, AI foundational models; rather, they are solutions built on top of AI foundational models.
Competitive development costs have likely fallen by 95%. Better integration into clinical workflows, unique go-to-market models, and novel derivatives remain critical differentiators. However, the qualitative differences between the best solutions in the AI aspects of transcription itself will become essentially zero.
Ultimately, in this future, only inertia will prevent prices from falling dramatically, which should be great for healthcare providers. Lower costs will make these advanced transcription solutions available to more practices, further reducing the administrative burden on physicians and allowing them to focus more on patient care.
The surge in investment in transcription companies is evidence of the transformative potential of AI in healthcare – and there is a risk that the commodification of the category will lead to desperate, over-promising efforts to meet investor expectations.