There is no shortage of enthusiasm in healthcare technology today. We hear endlessly about promising early results that will “transform clinical workflows” but for the people actually delivering care, work hasn’t changed much if at all. The day-to-day provision of care has largely remained the same, and the system remains overwhelmed. ED volumes are constantly rising and have surpassed prepandemic volumes. More than half of the physician workforce reports burnout. We are facing a physician shortage crisis in healthcare.
The bulk of health AI investment today is dominated by largely administrative problems that are well-defined and low-risk to solve. Meanwhile, the problems that make delivering care so challenging remain untouched. We must find ways to improve the daily reality for our clinicians. The next generation of health AI should move to solve deeper, more complex issues that we need help with the most.
Thus far, AI has found real traction by targeting largely administrative problems. These are genuine wins, and these areas — documentation, radiology assistance, and sepsis prediction — matter.
Ambient scribes are perhaps the clearest success story in health AI so far. Products under this umbrella have matured quickly and are being deployed broadly. These tools listen to natural conversation between patient and clinician and autonomously compose clinical documentation, reducing time spent writing notes and reducing after-hours work time. This automation reduces cognitive workload and improves documentation efficiency. Ambient scribes are an excellent case study of healthcare AI identifying and addressing a visible and obvious burden of patient care.
With respect to actual direct clinical impact, AI in radiology has been perhaps the most influential application in medicine to date. Already, there are many FDA-authorized commercial radiology AI products, spanning a variety of medical imaging applications. Radiology AI is performing well in narrow diagnostic tasks, and in some cases it is even outperforming humans. In the future, health systems may even start to consider whether in some instances it may be more optimal for AI to perform alone.
AI-based sepsis prediction has also generated promising results. For instance, the Sepsis ImmunoScore has demonstrated accuracy in identifying sepsis and in-hospital mortality. TREWS is a sepsis alert system that has been shown in a multicenter study to reduce in-hospital mortality. eCART is another sepsis tool proven to predict in-hospital deterioration. These are well-designed systems that maintain sensitivity in identifying high-risk patients while alerting on fewer total encounters, reducing alert fatigue.
But while these improvements are helpful, they are only a component of our day-to-day tasks. Seeing patients is still rewarding for physicians; the problem is that those interactions only account for 27% of their time. Every patient we see turns into an overwhelming mountain of downstream tasks that we know is coming. This is the daily reality of practicing medicine, and AI doesn’t move the needle on managing that burden, at least not yet.
In almost every other line of work, when work becomes overwhelming, we tell people to focus on the big picture and to avoid getting stuck in the weeds. In medicine, we do the opposite. During busy shifts and increasingly overwhelming work environments, we demand physicians spend even more time going through menu items and checking boxes, often on individual tasks that make no meaningful difference to the patient’s care. An alert tells you to put in a HEART score on patients who didn’t come in with chest pain. It gives you a pop-up saying that your healthy 26-year-old with upper respiratory infection symptoms and a fever might be septic. It reminds you to order elopement precautions on the 96-year-old patient with a hip fracture who hasn’t walked in years. Every clinician can think of at least a few irritating alerts they blindly click through just to move on with their work. This is where solutions like Doximity Ask offer a departure from this new norm. By highlighting the specific source passages for its answers, Ask allows physicians to verify information in context. In doing so, physicians gain a transparent information synthesizer that supports decision-making, rather than one that forces clinicians to navigate yet another layer of complexity.
Clinicians almost always know what the right decision is for the patient in front of them. The issue is that they face incredible amounts of structural friction in almost every setting. Insurance authorization delays block urgent procedures and timely clinic appointments. Bed availability and subsequent ability to transfer to higher levels of care is often unpredictable and declines during worsening capacity crises. Complex, fragmented consultant coverage makes it hard to identify who needs to be called for a given patient. Methods to reliably tell a patient how they can best follow up with the right specialist at the time of discharge from the hospital are variable at best. We ask physicians who have trained their whole lives to be experts in patient care to now be experts in logistics as well. And when the logistics fall through, it’s the physicians who are to blame.
And while managing logistics is exhausting, time pressure is the single most important driver of emotional exhaustion in medicine. It is associated with chaos, stress, burnout, and intent to leave. This extends across care sites, from the outpatient clinics, to the ED, to inpatient settings, and surgical care. Clinicians feel the constant need to sprint, and yet it’s never enough. The other conventional lever to pull has been hiring more staff to open new care spaces. This cannot be the only answer. With thinning margins and worsening retention of clinicians, this is not a viable plan. We now see care sites and groups engaging in patient care with shrinking supervision of board-certified attending physicians in attempts to keep up both in volume and finances. This is a symptom of a larger problem. There needs to be a paradigm shift to help our staff manage these workloads.
Today we face an inflection point in healthcare. For the first time in recent memory, healthcare is at the forefront of enterprise AI adoption. It’s not because we are more advanced, far from it. It’s because our systems are broken, and we are desperate for help. The tools being adopted, where AI is working, target the lowest-hanging fruit, and the billions flowing into them reflect what’s easiest to measure rather than truly transformative.
The opportunity is significant. A common concern in healthcare technology is clinician adoption, or lack thereof. Much is made of clinician resistance to change. Clinicians are not technology-averse — if AI can reduce the structural frictions that clinicians face and make their lives better, they will adopt it. But before we ask clinicians to use AI to be smarter, we should ask AI to undo the damage that decades of poorly designed systems have inflicted on our caregivers.
