AI outperforms human clinicians in treating sepsis

[lymphocyte]

Published in Nature Medicine. Researchers used machine learning to develop models for optimal fluid and vasopressor use. These models, on average, outperformed human clinicians; the further humans deviated from the machine model, the greater the mortality -- see the paper for further details.

I think the title is pretty sensationalistic, but, interestingly, machine models used less fluids and more vasopressor sooner, which seems to be the vogue strategy anyways.

The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care

---

Members don't see this ad.

 

[PubPhD]

Just a comment on the title as to why it's not crazy sensationalist, optimal refers to a set of actions that maximize the reward function. So it's not saying that this is the best possible choice of actions to treat patients, but it is given how you choose your reward. For example, the optimal policies can be dramatically different if the reward is a '-1 if patient dies' vs '+1 if patient lives' vs 'time factor component tied with recovery'.

A really challenging part of reinforcement learning is how to design this reward function, and the subfield of inverse reinforcement learning is just that. Given a set of state,action pairs made by an expert, can you infer a reward signal that would lead to the optimal policy fitting this expert?

If anything, the most interesting part that can be leveraged for future work is that patient states follow the markov property.

 

[SurfingDoctor]

Not really surprising. AI and machine learning can analyze real-time data and make implementations on it faster...

It’s certainly the wave of the future when there is a protocol to follow. Creative thinking will be a struggle for the foreseeable future though.

 

[PubPhD]

Not really surprising. AI and machine learning can analyze real-time data and make implementations on it faster...

It’s certainly the wave of the future when there is a protocol to follow. Creative thinking will be a struggle for the foreseeable future though.

I think protocol is the wrong word, as this type of work excels when there isn't a set protocol to follow, and requires dynamic decision making. If one can make a simple fixed protocol for a problem, then this type of work is significant overkill.

 

[SurfingDoctor]

I think protocol is the wrong word, as this type of work excels when there isn't a set protocol to follow, and requires dynamic decision making. If one can make a simple fixed protocol for a problem, then this type of work is significant overkill.

Well, there is a prescribed algorithm to follow in sepsis. The AI can be taught the variables and follow the plan. The machine learning aspects will improve its management based on the algorithm overtime.

However, I doubt the AI can figure out the downstream consequences of sepsis management (ie fluid overload, but maybe it can prevent them as in that article) or when the sepsis is atypical (ie HLH). But probably someday. I have little doubt AI can do a better job at straightforward management than humans and maybe someday, better at complex management.

 

[Gurby]

However, I doubt the AI can figure out the downstream consequences of sepsis management (ie fluid overload, but maybe it can prevent them as in that article)

This analysis took into consideration both in-hospital mortality and 90-day mortality, so I would assume down-stream effects are being considered?

 

[PubPhD]

Well, there is a prescribed algorithm to follow in sepsis. The AI can be taught the variables and follow the plan. The machine learning aspects will improve its management based on the algorithm overtime.

However, I doubt the AI can figure out the downstream consequences of sepsis management (ie fluid overload, but maybe it can prevent them as in that article) or when the sepsis is atypical (ie HLH). But probably someday. I have little doubt AI can do a better job at straightforward management than humans and maybe someday, better at complex management.

There is a lot of overlapping vocabulary so I apologize if I don't understand what you're saying perfectly. Is the currently prescribed policy a function given patient conditions, or more of a {if,else if,else} approach? And the machine learning algorithm learns a policy, and improves it over time based off of long term impacts to maximize patient outcomes.

And I think a consistent terminology might be useful discussing this:

Policy: A distribution (can be deterministic) of state-action pairs
Algorithm: A function to learn a policy
AI: a learned policy (I think that's what you mean)
machine learning: Synonymous with AI?

Lastly, I agree, optimal control problems are a pretty fun challenging task right now. I do a lot of research in stable RL, and I hate the hype RL has been getting in the media due to max sampling, information leaks, biases in evaluation, etc.

 

[SurfingDoctor]

This analysis took into consideration both in-hospital mortality and 90-day mortality, so I would assume down-stream effects are being considered?

Unless I’m reading the article wrong, it was testing the AI on a dataset. There is no prospective real-time application of the model. They matched the model to outcomes in patients in the dataset. I’d prefer to see prospective results before I would make any claim on mortality.

 

[SurfingDoctor]

There is a lot of overlapping vocabulary so I apologize if I don't understand what you're saying perfectly. Is the currently prescribed policy a function given patient conditions, or more of a {if,else if,else} approach? And the machine learning algorithm learns a policy, and improves it over time based off of long term impacts to maximize patient outcomes.

And I think a consistent terminology might be useful discussing this:

Policy: A distribution (can be deterministic) of state-action pairs
Algorithm: A function to learn a policy
AI: a learned policy (I think that's what you mean)
machine learning: Synonymous with AI?

Lastly, I agree, optimal control problems are a pretty fun challenging task right now. I do a lot of research in stable RL, and I hate the hype RL has been getting in the media due to max sampling, information leaks, biases in evaluation, etc.

Sorry, I don’t have enough knowledge on AI and machine learning to have much of a discussion other than to say, yes, I think it will be an important clinical asset in the future and especially when designed to follow and implement treatments based on established practice guidelines.

 

[PubPhD]

Sorry, I don’t have enough knowledge on AI and machine learning to have much of a discussion other than to say, yes, I think it will be an important clinical asset in the future and especially when designed to follow and implement treatments based on established practice guidelines.

Not entirely related, but this is something I've been curious about after a peer brought this up that I haven't seen discussed by people outside of my field:

Training data often comes from non uniform populations (maybe more affluent areas, which then suffer typical socioeconomic status biases). Is it okay for the model to then be able to treat patients better from these protected classes, and therefore produce reduced care for the protected classes underrepresented in the training set? Do we accept a lower overall performing model to ensure consistent treatment over all classes? Dp we make a reasonable effort to serve undersampled populations and then let this inequality exist after due diligence was done?

Right now, due to the publish or perish environment, we often end up reporting performance over the entire collection.


Edit: Just saw your comment after posting about wanting future work done, that's completely correct. They use an evaluation method with no provable bounds given this environment setting.