Do you see LLM based check-ins (for lack of a better term) as completely unsupervised? Similar to Dr. Attia's question below, would the future be via a medical device approval pathway and/or would providers be able to build their own agents and, perhaps, have to attest that they will supervise/keep a human in the loop as the agents interact with their patient panel? Really hard questions that we'll have to answer in the coming months and years.
Fantastic question, we're much more bullish on LLM-powered 'check-ins' in the near term vs. fully autonomous LLM-powered care delivery. We are full proponents of step-wise iteration, and automating relatively structured information gathering (in the form of a check-in) with the right feedback and escalation powered by LLMs seems more achievable (and less risky). Again, we feel that the winners here will not be those who do it first, but those who do it right (keeping their north start as always doing right by the patient).
I am unabashedly bullish on the necessary (and inevitable?) adoption of ML-powered tools in healthcare, and not just for front office functions like automated charting. These tools should start to touch care relationships directly, in my view. Two follow-ups come to mind:
First, what are some other data streams you all would like to see feeding into such ML-powered tools? Continuous engagement with an LLM itself would provide useful data, but I wonder if there's an opportunity to layer on additional insights from, say, the patient's smartphone, at-home wearables, etc.
Second, when, if at all, do you think we'll see an LLM approved as a SaMD to provide CBT for certain conditions, like mild/moderate depression? To my knowledge, healthcare still hasn't crossed this rubicon, but I wonder what your thoughts are on pros/cons.
Totally with you on both front! Would love to see additional streams of data feeding into these models to better inform on the patient's progress and help the clinician. A few that seem to make the most sense, as you mention, are data from wearables / connected home devices (Apple watch, Oura ring, Eight Sleep mattress, etc.), latent data from smartphones (RIP Mindstrong), and smart scales / blood pressure cuffs.
On the second, I think we're still some time away from LLMs being approved as SaMDs to provide autonomous care. The key here will be the collation of enough high quality data to train the models on, and then building a product that is evidence-based and with the right guardrails in place to ensure adherence to standards of care! We'll be watching excitedly to see who gets this right...
As always, enjoyed y'all's thoughts!
Do you see LLM based check-ins (for lack of a better term) as completely unsupervised? Similar to Dr. Attia's question below, would the future be via a medical device approval pathway and/or would providers be able to build their own agents and, perhaps, have to attest that they will supervise/keep a human in the loop as the agents interact with their patient panel? Really hard questions that we'll have to answer in the coming months and years.
Fantastic question, we're much more bullish on LLM-powered 'check-ins' in the near term vs. fully autonomous LLM-powered care delivery. We are full proponents of step-wise iteration, and automating relatively structured information gathering (in the form of a check-in) with the right feedback and escalation powered by LLMs seems more achievable (and less risky). Again, we feel that the winners here will not be those who do it first, but those who do it right (keeping their north start as always doing right by the patient).
Love this, thanks for a very relevant article!
I am unabashedly bullish on the necessary (and inevitable?) adoption of ML-powered tools in healthcare, and not just for front office functions like automated charting. These tools should start to touch care relationships directly, in my view. Two follow-ups come to mind:
First, what are some other data streams you all would like to see feeding into such ML-powered tools? Continuous engagement with an LLM itself would provide useful data, but I wonder if there's an opportunity to layer on additional insights from, say, the patient's smartphone, at-home wearables, etc.
Second, when, if at all, do you think we'll see an LLM approved as a SaMD to provide CBT for certain conditions, like mild/moderate depression? To my knowledge, healthcare still hasn't crossed this rubicon, but I wonder what your thoughts are on pros/cons.
Totally with you on both front! Would love to see additional streams of data feeding into these models to better inform on the patient's progress and help the clinician. A few that seem to make the most sense, as you mention, are data from wearables / connected home devices (Apple watch, Oura ring, Eight Sleep mattress, etc.), latent data from smartphones (RIP Mindstrong), and smart scales / blood pressure cuffs.
On the second, I think we're still some time away from LLMs being approved as SaMDs to provide autonomous care. The key here will be the collation of enough high quality data to train the models on, and then building a product that is evidence-based and with the right guardrails in place to ensure adherence to standards of care! We'll be watching excitedly to see who gets this right...