As soon as I spotted my patient in the waiting room, I knew that I’d be admitting him to the hospital.
His breathing wasn’t necessarily heavy, but there was something in the contour of his respirations that caught my eye. Or maybe it was the expression on his face — eyes more quizzical than crinkling, lips more decisively hewn than usual. I use A.I. all the time to help me diagnose and treat my patients. I also know where it falls short. There’s an ocean of distance between the “patient” that A.I. is analyzing and the patient that the human doctor is assessing.
Every doctor and nurse has had this experience — a glance at a patient and the instant recognition that something has run amok in his or her physiology.
This is especially common in primary care medicine, where we know our patients for years, sometimes decades. We know their gait, their heart murmurs, their blood counts. And we know when something is amiss.
Like most doctors these days, I’ve been incorporating medical artificial intelligence tools into my practice.
It’s become so easy to type in a quick description of an 86-year-old male with heart failure, diabetes and gout — toss in some test results, and see what the bot spits out.
I appreciate that A.I. can expeditiously outline next steps for the clinical evaluation, or provide suggestions for rarer diagnoses or spit out a feisty appeal letter for an insurance denial.
But the problem is that A.I. is evaluating only some statistical average of 86-year-old males with heart failure, diabetes and gout. It is not assessing that one specific 86-year-old man with these conditions whom I am looking at across the waiting room.
There’s an ocean of distance between the “patient” that A.I. is analyzing and the patient that the human doctor or nurse is assessing. …
THIS IS the inherent limitation of A.I. in medicine.
It’s simply impossible — at least for now — for these tools to truly see the multidimensional patient.
A.I. can’t know how the agony of a child estranged by substance use affects the blood pressure. It can’t factor in the economic and social crosscurrents that bear on medication adherence.
It can’t account for the simmering grief of a lost spouse that influences a patient’s health decisions far more than any clinical guideline.
So while A.I. is a useful tool, particularly for pattern recognition and data organization, the “patients” it manages feel like stock characters who share check-box traits.
A.I. might be quick to spit out a treatment plan, and it might even be correct, but a clinician must decide how to make that treatment work for the specific person sitting in front of her, or whether to even treat at all.
