That the main problem is not in lab model performance, but in data availability and quality. Most useful applications could be very dumb models pertaining to trivial things. But we need a solid base of quality data for that, that we absolutely do not have apart from minuscule and hyperspecialized niche domains.
The model itself does not have to be incredibly performant. It absolutely has to, on the other hand, make the clinical process of which it is part more efficient. That mainly means: "efficiently automate the most trivial tasks", currently.
That's why the urgent action to be taken pertains to policy and not to complex tech. We need policy to encourage routine automated data gathering. No data, no ML.
As a clinician, I don't effing care if your shiny new toy can give me a shitty estimate of some parameter extrapolated from some random population not including my current patient. Just getting accurate trends on vital signs would be stellar. This to say that what interests clinicians is workplace integration and not having hundreds of monitors all over the place that aren't even interconnected.
The model itself does not have to be incredibly performant. It absolutely has to, on the other hand, make the clinical process of which it is part more efficient. That mainly means: "efficiently automate the most trivial tasks", currently.
That's why the urgent action to be taken pertains to policy and not to complex tech. We need policy to encourage routine automated data gathering. No data, no ML.
As a clinician, I don't effing care if your shiny new toy can give me a shitty estimate of some parameter extrapolated from some random population not including my current patient. Just getting accurate trends on vital signs would be stellar. This to say that what interests clinicians is workplace integration and not having hundreds of monitors all over the place that aren't even interconnected.