Artificial Intelligence + Machine Learning & Pharmacy

[fewaopi]

This is a broad topic but I was wondering what people's thoughts were on AI & ML's potential impact on the pharmacy profession. And not just retail but hospital, industry, managed care, academia, all areas where pharmacists are employed.

AI/ML is being explored in drug development and at Pill Pack, there was a camera scanning every bag verifying the drugs in each individual bag. That's much more efficient and with advances in ML, it's easy to imagine a time where machines are doing the verification, not the pharmacists, and that they're more accurate. Already AI/ML is making headways into diagnoses by ML and imaging. It's making inroads into professions we all thought would be immune such as law, writing, business, trading, manufacturing, medicine. It seems every time we say AI cannot, it does. So why not pharmacy?

At its core, AI/ML is information processing. I feel pharmacy is low intensity in information processing relative to other jobs like engineering where there isn't always a standard. Part of pharmacy is just following guidelines, and there's little critical thinking I feel in hospitals/retail and work can seemingly be done by AI with enough data. I could see some tasks being difficult to outsource but I can imagine the overall workload being dramatically reduced.

I'd imagine pharmacists in large sectors like retail and industry will be cut first. Areas where there's a lot of fat and waste that could be cut out to justify developing. Some really custom jobs like compounding might be preserved as AI development wouldn't be justified maybe and there's not enough data for AI/ML to harness. What areas would be hit first, what would the future be like and will it be more challenging for future pharmacists than before? What challenges will grads face? Pharmacy today is different than 20-30 years ago. It can change 10-20 years later. Work/labor has always evolved. What skill sets would be most needed in pharmacists? What are strategies people are thinking of in a future that could be dominated by machines? Anyone see examples where AI/ML is already making impact in pharmacy?

---

Members don't see this ad.

 

[WVUPharm2007]

I figure we got at least another 10-15 years before tech could even begin to completely replace the pharmacist from the equation.

That's enough time for me to save enough to retire on...good luck, youngins'.

 

[Jbrl]

Timely post, given the discussions recently.

I figure we got at least another 10-15 years before tech could even begin to completely replace the pharmacist from the equation.

That's enough time for me to save enough to retire on...good luck, youngins'.

I agree with this. Healthcare is a lumbering giant that is always behind the shifts occurring in tech. It's my belief that we are at least 10 years away from anything seismic, for a few reasons:

1. Legal, regulatory, liability, and bioethics
2. Customer/patient preferences
3. Feasibility and logistics

For pharmacists, retail's (and perhaps hospital's) largest barrier to automation is the 1st. For industry, one could argue it is the 1st or the 3rd. Much of the work done by pharmacists in industry is team-based, people-facing and project-oriented, the detailed analytics shipped off to 3rd party contractors or done by other individuals. An MSL's job wouldn't be replaced by AI, for example - at least not until MDs no longer prescribe. Pharma has larger issues to deal with, including public perception, cost of capital, and drying up of low hanging fruit, the latter of which AI/ML can actually help to mitigate through more targeted drug development.

The most immediate thoughts that come to mind in terms of progress made using AI/ML are precision medicine via genomics, metabolomics, and proteomics, and smart medical devices. For the former, companies like Grail, Illumina. For the latter, big players like JNJ (you can check their advancements in surgical tools and digital health) as well as some startups. These developments are still at least a few years away, if not more, to fruition though.

I've thought about what it would take to survive in a labor market where machines play a greater and greater role.... I think it would be a good idea to pursue a mid to end-stage in either:

1. Generalist knowledge focused on people leadership, management, and/or persuasion (interaction between you and me is something machines can't take away... yet...)
2. Deep subject matter expertise focused on developing tech solutions and integrating them into healthcare (e.g. a regulatory expert on the team)

Programming, if not now, in the future will be synonymous with reading, so it would be to a pharmacist's benefit to at least be familiar with the tools used.

 

[Charcoales]

Gonna be funny with all these ai refusing to dispense a zpak because the md put #1 tablet in the qty field lol.

 

[WVUPharm2007]

Gonna be funny with all these ai refusing to dispense a zpak because the md put #1 tablet in the qty field lol.

Or when the AI gets the insurance rejection requiring contact with physician and documentation before filling. It will contact the physician...wait for response...annoy the consumer.

A real life pharmacist? M0, 1G. Resubmit.

Physicians offices are going to be livid with all the extra work. How many rejections are miscoded or coded in confusing ways? Medical offices will be getting prior auth notices for refill too soons.

I'll chuckle from my villa in Medellín when it happens.

 

[lord999]

Being at this for over a decade, I am always deeply amazed at the arrogance of AI researchers and the constant, ridiculous marketing that has always been a part of this field. AI will change the situation, but behind the buzz are 50+ year algorithms that have never been that generalized in the first place as in a tool in search of a problem. We've had attempts at Robodoctor ever since Homer Warner started HELP. By the 1970s, Ted Shortliffe got the Turing Award (the same one as Kurtzweil and Knuth) for what was thought to be a solved problem in getting diagnostic robotics to work.

We're not 10 years from anything seismic, because we refuse to confront the same problem that every generation of AI researcher refused to acknowledge until their funding was cut, what problem are you trying to solve and how does it fit within the worker context? Sure, if the problem is a calculation one or something of a guess, AI and Machine Learning can help if someone monitors the models and actually programs it to the business scenario. But AI does not work the way Malcolm Gladwell says it does, the algorithm does not get better, it just adapts to the training without context. That usually never happens, but there's more.

One of the primary reasons why AI does not work in healthcare like anywhere else is for the lack of good data to go into a model. A model is always broken up into two parts: what is predicted ("model") and what is not predicted ("error"). The not predicted part ("error") can be categorized as noninformative (in math terms, iid, in human terms, random and no underlying pattern) and informative (there is a reason, but we haven't put that in the model). For most of the analyses done, informative errors damn the analysis. The classical case is try to predict cancer outcomes. Sure, you have SEER which tracks primary cancer outcomes, but did you know that we collect no data about secondary tumors (which is what nails a good proportion of people that aren't counted because we don't collect that data?). For pharmacy, it would be the reconciliation of the MAR from the nurse administration to the actual basement processing for shrink reasons.

But what really angers me is some management consultant who comes into my office and says that we can replace your radiologist with better PACS technologies or something similar. PACS AI technology will automate away the diagnostic portion ideally, so you won't need your radiologist. My reply is "great, now, the radiologists need to get on doing the other parts of the job, like talking to my hospitalists and getting those reports written as well as deal with their QC matters." Much of what AI automates away in healthcare does not replace the actual job functions of the person. And for right now and for the foreseeable future outside a miraculous advance in both hydraulics and miniature power supplies, robotics will not replace humans at complex physical tasks. AI can automate away a part of the job, but it has never been able to deal with context issues. How much of your actual job is automated away, and my answer for pharmacy or radiology (two occupations that have fairly high potentials for automation and information processing) is not enough to make a meaningful difference at the scales already in place.

For AI to actually make a meaningful difference, it's going to have to integrate into the workflow. In that sense, AI is going to have to learn some sense of causality (X cause Y) and even more in the sense of context (to think through counterfactuals) to satisfy the weak criteria. These are within the theories of cognition that make human thinking complicated to reduce. I am not saying that AI cannot get there (I am on the side of Strong AI being possible), but we are not actually trying to solve that problem in general right now. If we are going to go down this road in pharmacy, you have to actually capture your workflow in a way that you can analyze it. Small firms can do that. Big firms...well, there's a reason why Oracle, SAP, Tata, and all of those other companies fleece their clients.

We are going to have another AI Winter at this rate due to the overpromising and underdelivery for the cost of setting up the AI/ML, but I am definitely fine with the posers and marketing hype bankrupting themselves hiring inferior talent to provide negative value products. As for industry, what you call AI is called high throughput, and that and other performance chemistry/chemical engineering approaches have been a part of PhRMA ever since X-ray crystallization. You want to hear a great story about how that was suppose to revolutionalize the PhRMA industry, there are probably people at your school who would know. As for genomics, what drugs have come out of the HGP as of right now? Investigators hype technologies to get funding, but the actual delivery is so disappointing most of the time.

Then again, I get paid to fix these problems, so I hope there are more Elon Musk wannabes who screw up a company's enterprise data warehouse or data architecture for me to get night work.

 

[lord999]

Or when the AI gets the insurance rejection requiring contact with physician and documentation before filling. It will contact the physician...wait for response...annoy the consumer.

A real life pharmacist? M0, 1G. Resubmit.

Physicians offices are going to be livid with all the extra work. How many rejections are miscoded or coded in confusing ways? I'll chuckle from my villa in Medellín when it happens.

That's AI for you, and a present day form of Minority Report. There are fairly sophisticated SAS PL/I and COBOL programs (that actually no one alive can read in Blue Cross or Prime Therapeutics, please submit your resume to me if you can read and write cogently in both languages and I can get you a $250k job today) that read and try to do claims analysis and rejects. We just feed the machine at this point for some of those claim analyses.

I just can't wait for the DEA's ATHENA system to get online that will basically tell pharmacies that given a patient ID, whether or not to fill opioids for them taking to account their relatives and their other contacts.

 

[WVUPharm2007]

I just can't wait for the DEA's ATHENA system to get online that will basically tell pharmacies that given a patient ID, whether or not to fill opioids for them taking to account their relatives and their other contacts.

That's a thing? Well tell them to hurry up with it. I'd love to not have to make that judgement call ever again.

 

[Jbrl]

Being at this for over a decade, I am always deeply amazed at the arrogance of AI researchers and the constant, ridiculous marketing that has always been a part of this field. AI will change the situation, but behind the buzz are 50+ year algorithms that have never been that generalized in the first place as in a tool in search of a problem. We've had attempts at Robodoctor ever since Homer Warner started HELP. By the 1970s, Ted Shortliffe got the Turing Award (the same one as Kurtzweil and Knuth) for what was thought to be a solved problem in getting diagnostic robotics to work.

We're not 10 years from anything seismic, because we refuse to confront the same problem that every generation of AI researcher refused to acknowledge until their funding was cut, what problem are you trying to solve and how does it fit within the worker context? Sure, if the problem is a calculation one or something of a guess, AI and Machine Learning can help if someone monitors the models and actually programs it to the business scenario. But AI does not work the way Malcolm Gladwell says it does, the algorithm does not get better, it just adapts to the training without context. That usually never happens, but there's more.

One of the primary reasons why AI does not work in healthcare like anywhere else is for the lack of good data to go into a model. A model is always broken up into two parts: what is predicted ("model") and what is not predicted ("error"). The not predicted part ("error") can be categorized as noninformative (in math terms, iid, in human terms, random and no underlying pattern) and informative (there is a reason, but we haven't put that in the model). For most of the analyses done, informative errors damn the analysis. The classical case is try to predict cancer outcomes. Sure, you have SEER which tracks primary cancer outcomes, but did you know that we collect no data about secondary tumors (which is what nails a good proportion of people that aren't counted because we don't collect that data?). For pharmacy, it would be the reconciliation of the MAR from the nurse administration to the actual basement processing for shrink reasons.

But what really angers me is some management consultant who comes into my office and says that we can replace your radiologist with better PACS technologies or something similar. PACS AI technology will automate away the diagnostic portion ideally, so you won't need your radiologist. My reply is "great, now, the radiologists need to get on doing the other parts of the job, like talking to my hospitalists and getting those reports written as well as deal with their QC matters." Much of what AI automates away in healthcare does not replace the actual job functions of the person. And for right now and for the foreseeable future outside a miraculous advance in both hydraulics and miniature power supplies, robotics will not replace humans at complex physical tasks. AI can automate away a part of the job, but it has never been able to deal with context issues. How much of your actual job is automated away, and my answer for pharmacy or radiology (two occupations that have fairly high potentials for automation and information processing) is not enough to make a meaningful difference at the scales already in place.

For AI to actually make a meaningful difference, it's going to have to integrate into the workflow. In that sense, AI is going to have to learn some sense of causality (X cause Y) and even more in the sense of context (to think through counterfactuals) to satisfy the weak criteria. These are within the theories of cognition that make human thinking complicated to reduce. I am not saying that AI cannot get there (I am on the side of Strong AI being possible), but we are not actually trying to solve that problem in general right now. If we are going to go down this road in pharmacy, you have to actually capture your workflow in a way that you can analyze it. Small firms can do that. Big firms...well, there's a reason why Oracle, SAP, Tata, and all of those other companies fleece their clients.

We are going to have another AI Winter at this rate due to the overpromising and underdelivery for the cost of setting up the AI/ML, but I am definitely fine with the posers and marketing hype bankrupting themselves hiring inferior talent to provide negative value products. As for industry, what you call AI is called high throughput, and that and other performance chemistry/chemical engineering approaches have been a part of PhRMA ever since X-ray crystallization. You want to hear a great story about how that was suppose to revolutionalize the PhRMA industry, there are probably people at your school who would know. As for genomics, what drugs have come out of the HGP as of right now? Investigators hype technologies to get funding, but the actual delivery is so disappointing most of the time.

Then again, I get paid to fix these problems, so I hope there are more Elon Musk wannabes who screw up a company's enterprise data warehouse or data architecture for me to get night work.

Perhaps you should write a book. Seriously. Great read.

So, in your mind - how would the next 10 years play out for pharmacy in this context? What would change?

Secondarily, how could the core issues you laid out for AI in healthcare be tackled, considering the problems and outcomes aren't well defined to begin with?

With respect to retail, what I fear is a future where a pharmacist’s role is minimized to escalated/edge cases and almost all of the DUR/verification work is completed by a computer (moreso than now), the technicians doing the manual labor in the meanwhile. Taking regulations out of the equation, how reasonable is this, really?

 

[lane one]

Gonna be funny with all these ai refusing to dispense a zpak because the md put #1 tablet in the qty field lol.

Thought you were gonna say "refuses to fill because the data shows that your constellation of symptoms is likely due to a self-limiting viral infection". Imagine that rejection as a robot voice. There really are endless hilarious scenarios.

 

[fewaopi]

Perhaps you should write a book. Seriously. Great read.

So, in your mind - how would the next 10 years play out for pharmacy in this context? What would change?

Secondarily, how could the core issues you laid out for AI in healthcare be tackled, considering the problems and outcomes aren't well defined to begin with?

With respect to retail, what I fear is a future where a pharmacist’s role is minimized to escalated/edge cases and almost all of the DUR/verification work is completed by a computer (moreso than now), the technicians doing the manual labor in the meanwhile. Taking regulations out of the equation, how reasonable is this, really?

I'm not the one being asked but I imagine automation (the bigger, more current threat) with AI and ML will continue to threaten retail so my imagination of retail is similar to yours. The advances in image recognition and with understanding language may automate or outperform retail pharmacists in certain parts of prescription processing such as interpreting, verifying, writing up Rx's...eventually. I'm not sure how insurance claims are tackled, maybe pharmacists just all become insurance claim specialists. If all prescriptions are fully electronic as the trend is going, I imagine that's even easier. Part of me feels there's overhype and cash grab. I can't see it happening so soon, as easy as it sounds, I think it's still daunting to develop some parts of that. Even if the technology is there, the time to adopt, accept. We've had automation for a good while now and we still have techs counting and filling. Mail order has been around but there's still brick and mortar. But technology will continue to advance, so I do feel whether 5 or 50 years from now, eventually retail will shrink dramatically.

People thought TKIs and precision medicine were going to cure everything 18 years ago...well we're making progress but not quite there yet nor has it met the high expectations. Similar to how microbiome, CRISPR are overhyped, my guess with AI/ML is it'll be overhyped with some small steps being made. I do wonder who is going to regulate if going to be regulate, all the software, algorithms. FDA I believe has had to approve some software (not AI/ML) accompanying a drug. Would there be an AI/ML approval process? How does that factor into regulatory?

I'm not fully aware of an MSL's role (MDs asking an MSL deeper questions on a drug) but isn't part of an MSL like Asking Siri/Google/Alexa? I could see some deep, unanticipated, unaccounted questions still handled by a human but some questions a person could just look up online, could in theory be replaced. Unsure whether it's justifiable though, there are bigger fish to fry atm. I'd think in industry, the growing glut of pharmacists trying to enter industry, competition, the challenges of drug approval, drug costs, political pressure, and its overall instability, are bigger concerns than AI/ML for now. I also worry the decreasing number of drug targets left will probably be a bigger threat than AI and our current understanding of disease processes isn't deep enough to make new drugs. If there's a plateau to how far human health can be improved (there are some scientists who do believe death is purely a biological problem that can be solved), how close are we to approaching it? I just wonder if at the current approval rates, can drug industry be sustainable without the high drug prices? The challenges are daunting but work being done now I think is exciting

 

[deleted103644]

Thought you were gonna say "refuses to fill because the data shows that your constellation of symptoms is likely due to a self-limiting viral infection". Imagine that rejection as a robot voice. There really are endless hilarious scenarios.

www.theguardian.com/technology/video/2018/may/09/new-google-assistant-mimics-human-voice-video

 

[Chrish]

Visual verification process can be replaced by robots in theory. If chains start utilizing these robots with 99.99% accuracy compared to less accurate human eyes, even boards would approve that since to them patient safety would always come first. So, around 40% of rph workload and hence job hours are gone just like that..

 

[RxPreceptor]

AI seems far from perfect. In data entry, there are roughly 15+ variables/entry fields a human usually has to enter when sending. Upon receiving, I find myself having to correct an average of around 3 or more (mostly incomprehensible sig code/dirty translation, unspecified quantity/day supply, pkg rounding the qty, insurance limitations, etc.).

Seems like AI would only work if all MDs were perfect in e-scribing AND (big AND) insurance was not an obstacle. There is always the idiot/err factor of a human being involved with filling machines &I being involved in any data inputs. Not to mention RPh being necessary as a check and balance to MDs blowing through all the overrides for potential/less than ideal interactions (many of which are moderate/minimal, few of which really do require intervention/major or contraindication)

 

[lord999]

Perhaps you should write a book. Seriously. Great read.

So, in your mind - how would the next 10 years play out for pharmacy in this context? What would change?

Secondarily, how could the core issues you laid out for AI in healthcare be tackled, considering the problems and outcomes aren't well defined to begin with?

With respect to retail, what I fear is a future where a pharmacist’s role is minimized to escalated/edge cases and almost all of the DUR/verification work is completed by a computer (moreso than now), the technicians doing the manual labor in the meanwhile. Taking regulations out of the equation, how reasonable is this, really?

I have and you can easily figure out who I am from the public writing. I'm kind of well-known in the industry for being hired to fix broken models or rightsize biostatistics departments getting rid of the incompetents and unproductive (the Up In The Air version for Biostatisticians).

The real use for AI is to feed AAI (Artificial Artificial Intelligence), basically do all the rote stupid stuff that would require human review anyway but reduces it to manageable terms (so the second level of Artificial Intelligence). Things like calculating financial ratios in a buyout situation come to mind in finance. In pharmacy, it actually is used to apply to drug images rather than sending the bottle to the pharmacist. The pharmacist has to check anyway, and unless there is an input error, the images by and large or correct.

Amazon product

The New Ruthless Economy is a pretty good guide to how things work. The real use of AAI is higher throughput and dirty work objectivity. Where AI is used best is to find a supposedly objective way to do the dirty work that you've always wanted to do. Let's take @ancienbon 's current predicament. If I were his/her supervisor, and I wanted him/her fired, I'd put @ancienbon through the HR ringer, then cherrypick from his statistics all the bad ones, and then make a case to HR that warn and written have been done, let's toast this one! Information is almost always weaponized and objectivity in statistics is a myth, everything depends on some fundamental assumptions that people continually gloss over or do not understand (which is why stock market plays against people work better in general than plays against fundamentals unless there is a major systemic fundamental shift). @ancienbon is totally screwed, because he/she cannot marshal any supporting data that's actually quantifiable (you cut my hours, I lost all my good techs, and I keep getting floats) to answer for those bad AI metrics.

That's how AI is going to be used, to put seemingly "objective" data in support of soft, gut decisions. Belief always trumps frequency.

There's a second, and actually achievable goal, which has been stated earlier. In QC or statistics terms, sigma reduction. In human terms, variability reduction. There's some processes in pharmacy (verification, when to fill proactively) that you can pretty much write classical AI to manage which would not be a big problem in the workflow, and would actually be a good reminder. The problem with implementation is alert fatigue (think of that pharmacist feeling guilty over early opioid refills from punching through those noisy warnings, I'd honestly do the same and not give a damn) and exception handling for when the AI gets it wrong. The question for the public is going to be how much do I care that AI runs my life? Most AI functions (like credit checks and return allowances) are based off companies that no consumer interacts with and is for all intents and purposes unaccountable. It's not clear how AI would with a direct accountability chain situation, which is why I don't see sigma reduction AI applied to most customer facing matters, because who is responsible for when the AI fails?

That's a thing? Well tell them to hurry up with it. I'd love to not have to make that judgement call ever again.

Yes, that's a thing. My dig on West Virginia's opioid problem relates to testing the system out using the entirety of West Virginia's opioid data alongside all distributor quantity invoices. ATHENA basically outrageously denied more than 90% of the fills based on the programming criteria (and the real offender was the same provider being a superprescriber, which you can't help in rural WW), so obviously it had to go back to the drawing board. I think this system is far too ambitious, it really should be used in AAI situations as described above to let DEA know which pharmacies and prescribers are worth a visit just to see how things go. The best AI systems are hypothesis-generating (ones that narrow the noise that the likelihood of taking appropriate action is higher than not) rather than hypothesis-proving (this pharmacy is a pill mill AI is going to have false positives which are going to be a regulatory mess). I've basically called ATHENA the No machine, because if it cannot justify a patient getting something or has high variance, then the answer is no.

 

[John Detter]

While I think AI could certainly help enhance some aspects of a pharmacist's job, I am not afraid of it taking over my job. You have to remember that in order for any type of automated system to work, it has to be fed the correct data. Just take a look in any EMR and you'll find it riddled with erroneous or omitted data since nurses and physicians do most of the "data entry" so to say. Most healthcare providers can't even be bothered to enter the correct information to be interpreted by another human being down the line, let alone an automated computerized system.

 

[lord999]

While I think AI could certainly help enhance some aspects of a pharmacist's job, I am not afraid of it taking over my job. You have to remember that in order for any type of automated system to work, it has to be fed the correct data. Just take a look in any EMR and you'll find it riddled with erroneous or omitted data since nurses and physicians do most of the "data entry" so to say. Most healthcare providers can't even be bothered to enter the correct information to be interpreted by another human being down the line, let alone an automated computerized system.

That part of the job will never go away, cleaning up other people's messes. That is the real metajob.