Cheap things like breathalyzers + techniques like compressive sensing and machine learning/large and well-maintained data are being paired to create super cost-effective and reliable ways to test for everything from a cold virus to alzheimer's.
This makes for a win-win scenario where medical professionals and patients don't have their confidence considered because even if the machine is wrong, it saved you massive amounts of time and money 99% of the time when it got a diagnosis correct.
If you have to rule out three things to get to one lab test rather than 4 for each theory you have (say it's in between infection, messed up heart valve, anxiety, or cardiac arrhythmia), you can use a machine to quickly test based on symptoms and at least send you off to take the most important test first, and maybe save the doctor time before running into real diagnostic questions that require a human like feeling for lumps, moving a echocardiogram to the right place in your chest to see your heart, etc. things that a robot cannot do but hopefully less humans will have to do on the wrong patients who likely have something else wrong with them.
Machine learning will fix the problem of being sent multiple places to find out what's wrong before it ever fixes your actual disease. rather than have to get a blood-test, an x-ray, and a long visit with the head doctor to repeat everything you've already told nurses, you'll likely be sent one place because they've found out that 99%+ of the time, the robot knew what was wrong with you the first guess. It won't know how to fix it. But it'll know what the doctors need to do next.
Maybe ten years later IBM will be a healthcare company and they will be able to prescribe you medication based on these tests before you have to consult with a doctor because they'll know your allergies, your other medications, etc. will all be on file and shared among these companies and you'll be in a network of other successful diagnoses from today used as training data.
Most of medicine remains more qualitative than quantitative. There is a lot of "this is what is done in practice" aka "best practices" which are happened across due to happenstance without much quantitative evidence or systematic strategies for optimizing patient outcomes. Medicine needs to continue to become more quantitative in terms of what data doctors are using to base their actions on, before machines can be fed the corresponding recorded data to infer optimal actions that can be optimized according to (future discounted) patient outcomes.
As with many other machine learning applications, we have the necessary algorithms (more or less), but not the necessary data sets (in terms of relevance, quantity and quality) to train machine learning models that can surpass doctors.
There's a big push for evidence based medicine, but it's really hard to do a prospective rct for every choice and every disease presentation. That's before you even get to the interactions of various comorbidities, atypical presentations, etc. That's why medicine is often a parallel process: what are the most likely and most deadly diagnoses for this presentation? What exam, history, lab, and imaging information is reliable for including/excluding those diagnoses? What treatment can (or needs to) be started with relative safety before confirming the diagnosis
For most diseases, AI has a long way to go before it can diagnose patients even reasonably close to doctor level accuracy. But that is okay. We don't need an AI to replace doctor's -- rather we need an AI that serves as a clinical aid. In other words, the opportunity for machine learning in the medical field lies where doctors need the most help (i.e diseases where doctor to doctor agreement is low) or the doctor's can't interpret the results. Maybe you measure the expression levels for 100 genes for a patient. Your typical doctor won't know how to interpret that data. And maybe the biologist who measured the expression levels has an idea but not in a rigorous or formal manner. This is where machine learning can come in. Algorithms can make sense of this data given a trustworthy label set (which is also a nontrivial task depending on the disease or event you are trying to predict or screen for).
If you have to rule out three things to get to one lab test rather than 4 for each theory you have (say it's in between infection, messed up heart valve, anxiety, or cardiac arrhythmia), you can use a machine to quickly test based on symptoms and at least send you off to take the most important test first, and maybe save the doctor time before running into real diagnostic questions that require a human like feeling for lumps, moving a echocardiogram to the right place in your chest to see your heart, etc. things that a robot cannot do but hopefully less humans will have to do on the wrong patients who likely have something else wrong with them.
Machine learning will fix the problem of being sent multiple places to find out what's wrong before it ever fixes your actual disease. rather than have to get a blood-test, an x-ray, and a long visit with the head doctor to repeat everything you've already told nurses, you'll likely be sent one place because they've found out that 99%+ of the time, the robot knew what was wrong with you the first guess. It won't know how to fix it. But it'll know what the doctors need to do next.
Most of medicine remains more qualitative than quantitative. There is a lot of "this is what is done in practice" aka "best practices" which are happened across due to happenstance without much quantitative evidence or systematic strategies for optimizing patient outcomes. Medicine needs to continue to become more quantitative in terms of what data doctors are using to base their actions on, before machines can be fed the corresponding recorded data to infer optimal actions that can be optimized according to (future discounted) patient outcomes.
As with many other machine learning applications, we have the necessary algorithms (more or less), but not the necessary data sets (in terms of relevance, quantity and quality) to train machine learning models that can surpass doctors.
For most diseases, AI has a long way to go before it can diagnose patients even reasonably close to doctor level accuracy. But that is okay. We don't need an AI to replace doctor's -- rather we need an AI that serves as a clinical aid. In other words, the opportunity for machine learning in the medical field lies where doctors need the most help (i.e diseases where doctor to doctor agreement is low) or the doctor's can't interpret the results. Maybe you measure the expression levels for 100 genes for a patient. Your typical doctor won't know how to interpret that data. And maybe the biologist who measured the expression levels has an idea but not in a rigorous or formal manner. This is where machine learning can come in. Algorithms can make sense of this data given a trustworthy label set (which is also a nontrivial task depending on the disease or event you are trying to predict or screen for).
Comments
Post a Comment