As a physician who does lots of bioinformatics, this seems like a very optimistic article which is likely just overfitting the existing data.
Patients don't come to the hospital saying "I have pneumonia". The come for a cough or a fever and it takes these tests to figure out the diagnosis. A small hospital like this probably doesn't have the data to really build a robust model that will help all its patients. Even major academic centers have relatively limited data for the number of permutations of health possibilities. Sure they have lots of pneumonia data, but how many cases of pneumonia do they have in cystic fibrosis patients who also have COPD?
I certainly think AI can help cut costs, reduce over-testing and reduce length of stay, but decreasing length of stay by 2 days would be a shocking advance. (The average simple pneumonia LOS of stay for a medicare patient without co-morbidities* was 3.3 days and with major co-morbidities* was 5.7 days in 2017). [0]
Thanks to the weather and lack of income tax, Florida is a retirement destination for the US. A quick google implies that St. Augustine probably has a lot of retirees. Older people are particularly prone to having intractable problems requiring repeat hospitalizations.
the median age in St. Augustine is 43 which is approximately 3% higher than the Florida average of 42... 37% of St. Augustine residents were born in Florida, 54% were born out of state
I see no mention of cystic fibrosis in the article. Most older people with COPD likely have a diagnosis on file. I'm not personally aware of people with CF being prone to also having COPD. My general impression is that COPD is typically a disease of the elderly and people with CF tend to die young.
> probably doesn't have the data to really build a robust model
It is both possible to build robust models on small datasets (and that data is shared, or more abundant than you expect).
> Even major academic centers have relatively limited data for the number of permutations of health possibilities.
This is one of the problems they (the IT supplier) solved. They are able to build regulatory-proof models from 100.000's of permutations. Another, more recent, advance is in counterfactual analysis (what would have happened if this pneumonia patient also had COPD?).
Listen I'll be thrilled if I'm wrong, but the article says they 'expect to save' $xxxx.xx. Not that they are saving that much. Come back to me when they actually have reduced average length of stay for pneumonia patients by two days over a long time period. Just because they can find a most efficient pathway for patients, doesn't mean it will be followed, or that it is even possible to be followed clinically.
My hospital currently uses Allscripts, we have care pathways, and when I want to order something for a patient of mine (because I think the patient needs it) that isn't in the care pathway, I just order it. It seems that what they are trying to do is 'Nudge', in the Thaler sense, providers to do what they suggest is the most efficient pathway. Would love to see a randomized trial of giving some providers this nudge and others the current pathway to see which is most effective. But what I often see is that providers have learned behaviors from years of training where they 'need' tests X, Y, and Z for disease Q, so without extensive education, they still tend to order those things.
edit: just saw that ballenf beat me to this punchline already.
They actually reduced the average length of stay for knee replacement from 3.3 to 2.4 days. They verifiably saved costs of a middle-sized hospital of 10m$ a year. If we assume professional integrity, then they won't intentionally inflate projected savings, so we have no reason to doubt their expectations. (Just like I don't have any reason to doubt that you don't give inflated life expectancy projections to your patients, or receive kickbacks from needless prescriptions). It is, on a professional level, like telling an engineer who projects a bridge to be safe, that they likely build an unsafe bridge, and you'll only believe it when at least a 1000 people have crossed it (trust me, I am a lorry driver who loves driving over bridges).
The hospital in the article also used Allscripts.
> The next step for Flagler was to review the findings with the Physician IT Group (called the PIT Crew) and to make the necessary changes to AllScripts. Physician buy-in is critical...
> There are two interesting anecdotes from this process that bear repeating. The first is that once doctors became aware of the work that was being done, requests for membership in the PIT Crew skyrocketed and attendance at the bi-weekly meetings doubled. Doctors want access to data.
> The second is that one of the more accomplished physicians remarked that the care process model for pneumonia was far lighter than what he would have used, but upon looking at the outcomes, readily agreed that it delivered the same or better care in almost every case – and that what he was doing was essentially unnecessary, or wasteful. Presented with the evidence, he committed himself to rethinking his approach.
Edit:
> but decreasing length of stay by 2 days would be a shocking advance.
If you look at the industry average, maybe. But these 2 days were for this specific hospital (being a community hospital they get many different patients, and you can't have experts for every area). Perhaps removing those 2 days brought them closer to industry average, which seems like a very reasonable advance.
Sorry for the shameless plug but check out what we're doing at Emerge. (Savinghealthcare.net)
We are offering better, cleaner amd aggregated patient data at the point of care. We use ML not to predict patient outcome or progress but to make sense of the data.
For the record, I'm working on a topological data analysis project that is in some ways quite similar to this, and I'd be astonished if a single community hospital has the depth needed. We're working with thirty-some.
The key here is that TDA is packaged into an application that is designed explicitly for use by practitioners. All of the underlying math (and you know there is lots of it in TDA) is abstracted. What is shown is the groups and the atomic level explains (this group is here for these reasons e.g. they received albuterol upon admittance). Your instinct is correct, but that is what is interesting about this case - the hospital, without a single data scientist, was able to to achieve this with only slick SQL skills and engaged doctors.
HIT news is paid articles by the vendors. It's an industry rag and all their stories read the same, with the vendor name getting dropped about paragraph 3.
> It expects to save $1,356.35 per pneumonia patient in direct variable costs (35 percent savings) versus the status quo, while reducing length of stay by two days. The new sepsis pathway has also been deployed.
So the headline was BS. This is the projected savings estimate used by the vendor, who totally coincidentally got a bump in their contract from 4 case studies a year to 12 on the basis of this totally objective guess.
Reminded that episode from Greys Anatomy where they, using similar approach [manually, not AI though], discovered that pulling kidney surgery drainage couple days earlier (on the 3rd day instead of 5th) minimizes complications like fistula.
The AI produced recommendations in the article doesn't seem to have any rationale other than statistical correlation ("topological data analysis"). Patients who had that care path did better - i hope they controlled for the possibility that those patients may have been the "better" ones to start with.
>“After looking at the data, a cardiologist on the Pit Crew said, ‘Oh my, the Goldilocks group of doctors did less than I would have done but achieved as good or better outcomes. Therefore, when we add anything outside the CarePath, except regulatory requirements, we are adding cost without any benefit to the patient.’ For him, that was a ‘light bulb’ moment.”
I hope there was another cardiologist near by when this one was having his light bulb moment.
Overall i suppose it is a nice (by the metrics of healthcare IT ) sales push piece. I'm kind of wondering why such basic analytical practice hasn't been a routine for decades already ( may be because anybody able to do basic statistics would make much better money on wall street than being a lowly number cruncher at a dusty office in the hospital basement ... right to the moment of course when the thing gets renamed to AI and starts to be shiny and glorious and the hospitals start to pay good money for it).
>Flagler expects to save $1,356.35 per pneumonia patient
Yes, of course, but this article seems to be neither unsupervised nor supervised learning, just some data analysis and statistics.
It's not that it is completely different from AI, but you don't need the term "AI" to describe it, except if you want to hide the methodological weaknesses. They say they haven't really studied if the recommendations work in practice, it's just the prediction of a statistical model.
I am glad that they are lowering costs but is this really about AI? Usually someone can sit down and quickly find ways to reduce costs but the problem is usually the power to actually implement them.
It's like dealing with consultants. The company's employees already know what's wrong but don't have the power to change. Then consultants come in with CEO backing and suddenly change gets made. Consultants (or AI) get the credit but it's really about being to implement.
This is more of a puffy PR piece for the IT supplier.
The methods they use are very advanced mathematics / graph theory. If we take the modern usage of the word "AI", then this qualifies. These methods are beyond what anyone can do in Excel. Usually data is so big and complex that a manual analysis takes months. It does not scale to sit down and find ways to reduce costs. Using unsupervised pattern detection does scale.
About the consultants vs. engineers and the power to actually implement systems: Many IT systems that call themselves "AI", are not. They require tons of labeled data, a human making assumptions, lack justification and transparency, are not embedded properly in the business (lack UI, documentation, buy-in, data processing), and don't continuously learn.
You need an entire ecosystem of processes and software tools to have the power to change (identifying problems is easy, of course engineers already know what's wrong with their daily work, if they could implement a solution, then a proper engineer would, but they really need a consultant and CEO backing to actually get something actionable).
The article cites a specific example of how a model spat out a positive relationship between pneumonia with existing COPD and the speed of starting nebulizer treatments, which they stated was non-obvious to staff.
Shortened stays, less expense, and fewer tests sound nice, but and independent verification after some period of time seems imperative.
It also makes me wonder about some of the outlier patients. Those tests may be very beneficial for them. However, I'm sure doctors may feel pressure to not order them due to the cost savings.
As mentioned elsewhere in the comments, small/tiny hospitals rarely keep anything complicated and ship everything vaguely complex to tertiary/quaternary center.
I don't understand how their readmission rates are so low at 2.9%, I was just checking some stats on that at the link below, typical readmission rates are 15-20%:
"Tiny hospitals" have low readmission rates, because the complicated patients with lots of readmissions get transferred to tertiary care centers.
This also came up in things like Consumer Reports and their analysis of hospital acquired infection data. Small community hospitals with uncomplicated patients end up looking really good.
when such terms as "unsupervised learning" are coined I doubt that there was something achieved, common sense wouldn't have achieved, too. As an AI-expert once said to me: unsupervised learning does not work
Patients don't come to the hospital saying "I have pneumonia". The come for a cough or a fever and it takes these tests to figure out the diagnosis. A small hospital like this probably doesn't have the data to really build a robust model that will help all its patients. Even major academic centers have relatively limited data for the number of permutations of health possibilities. Sure they have lots of pneumonia data, but how many cases of pneumonia do they have in cystic fibrosis patients who also have COPD?
I certainly think AI can help cut costs, reduce over-testing and reduce length of stay, but decreasing length of stay by 2 days would be a shocking advance. (The average simple pneumonia LOS of stay for a medicare patient without co-morbidities* was 3.3 days and with major co-morbidities* was 5.7 days in 2017). [0]
*as defined by medicare.
[0] https://www.cms.gov/Medicare/Medicare-Fee-for-Service-Paymen...
the median age in St. Augustine is 43 which is approximately 3% higher than the Florida average of 42... 37% of St. Augustine residents were born in Florida, 54% were born out of state
https://www.areavibes.com/st.+augustine-fl/demographics/
I see no mention of cystic fibrosis in the article. Most older people with COPD likely have a diagnosis on file. I'm not personally aware of people with CF being prone to also having COPD. My general impression is that COPD is typically a disease of the elderly and people with CF tend to die young.
I would be more careful with your assessment. Or you are overfitting to your limited 5-minute-understanding of what this does. Telling mathematicians and statisticians they have overfitted is akin to a grave professional insult (you are basically accusing them of foolish behavior at best, fraud/unethical behavior at worst. As a bioinformatician, you may be interested in these articles, which go a little deeper on the technology: https://www.nature.com/articles/srep01236 https://www.sciencedirect.com/science/article/pii/S240547121... http://journals.plos.org/plosone/article?id=10.1371/journal....
> probably doesn't have the data to really build a robust model
It is both possible to build robust models on small datasets (and that data is shared, or more abundant than you expect).
> Even major academic centers have relatively limited data for the number of permutations of health possibilities.
This is one of the problems they (the IT supplier) solved. They are able to build regulatory-proof models from 100.000's of permutations. Another, more recent, advance is in counterfactual analysis (what would have happened if this pneumonia patient also had COPD?).
My hospital currently uses Allscripts, we have care pathways, and when I want to order something for a patient of mine (because I think the patient needs it) that isn't in the care pathway, I just order it. It seems that what they are trying to do is 'Nudge', in the Thaler sense, providers to do what they suggest is the most efficient pathway. Would love to see a randomized trial of giving some providers this nudge and others the current pathway to see which is most effective. But what I often see is that providers have learned behaviors from years of training where they 'need' tests X, Y, and Z for disease Q, so without extensive education, they still tend to order those things.
edit: just saw that ballenf beat me to this punchline already.
The hospital in the article also used Allscripts.
> The next step for Flagler was to review the findings with the Physician IT Group (called the PIT Crew) and to make the necessary changes to AllScripts. Physician buy-in is critical...
> There are two interesting anecdotes from this process that bear repeating. The first is that once doctors became aware of the work that was being done, requests for membership in the PIT Crew skyrocketed and attendance at the bi-weekly meetings doubled. Doctors want access to data.
> The second is that one of the more accomplished physicians remarked that the care process model for pneumonia was far lighter than what he would have used, but upon looking at the outcomes, readily agreed that it delivered the same or better care in almost every case – and that what he was doing was essentially unnecessary, or wasteful. Presented with the evidence, he committed himself to rethinking his approach.
Edit:
> but decreasing length of stay by 2 days would be a shocking advance.
If you look at the industry average, maybe. But these 2 days were for this specific hospital (being a community hospital they get many different patients, and you can't have experts for every area). Perhaps removing those 2 days brought them closer to industry average, which seems like a very reasonable advance.
We are offering better, cleaner amd aggregated patient data at the point of care. We use ML not to predict patient outcome or progress but to make sense of the data.
Screenshots for the app and videos can be found here: https://www.ayasdi.com/solutions/clinical-variation-manageme...
Given the conflict of interest in this article, I'd say that's the least of what the authors should be accused of.
What is the conflict of interest for a company doing or hiring a PR? Is it fraud or unethical to promote your company?
Your logic/conclusion makes no sense to me at all.
So the headline was BS. This is the projected savings estimate used by the vendor, who totally coincidentally got a bump in their contract from 4 case studies a year to 12 on the basis of this totally objective guess.
The AI produced recommendations in the article doesn't seem to have any rationale other than statistical correlation ("topological data analysis"). Patients who had that care path did better - i hope they controlled for the possibility that those patients may have been the "better" ones to start with.
>“After looking at the data, a cardiologist on the Pit Crew said, ‘Oh my, the Goldilocks group of doctors did less than I would have done but achieved as good or better outcomes. Therefore, when we add anything outside the CarePath, except regulatory requirements, we are adding cost without any benefit to the patient.’ For him, that was a ‘light bulb’ moment.”
I hope there was another cardiologist near by when this one was having his light bulb moment.
Overall i suppose it is a nice (by the metrics of healthcare IT ) sales push piece. I'm kind of wondering why such basic analytical practice hasn't been a routine for decades already ( may be because anybody able to do basic statistics would make much better money on wall street than being a lowly number cruncher at a dusty office in the hospital basement ... right to the moment of course when the thing gets renamed to AI and starts to be shiny and glorious and the hospitals start to pay good money for it).
>Flagler expects to save $1,356.35 per pneumonia patient
better than if i switch to Geyco.
It's not that it is completely different from AI, but you don't need the term "AI" to describe it, except if you want to hide the methodological weaknesses. They say they haven't really studied if the recommendations work in practice, it's just the prediction of a statistical model.
It's like dealing with consultants. The company's employees already know what's wrong but don't have the power to change. Then consultants come in with CEO backing and suddenly change gets made. Consultants (or AI) get the credit but it's really about being to implement.
The methods they use are very advanced mathematics / graph theory. If we take the modern usage of the word "AI", then this qualifies. These methods are beyond what anyone can do in Excel. Usually data is so big and complex that a manual analysis takes months. It does not scale to sit down and find ways to reduce costs. Using unsupervised pattern detection does scale.
About the consultants vs. engineers and the power to actually implement systems: Many IT systems that call themselves "AI", are not. They require tons of labeled data, a human making assumptions, lack justification and transparency, are not embedded properly in the business (lack UI, documentation, buy-in, data processing), and don't continuously learn.
You need an entire ecosystem of processes and software tools to have the power to change (identifying problems is easy, of course engineers already know what's wrong with their daily work, if they could implement a solution, then a proper engineer would, but they really need a consultant and CEO backing to actually get something actionable).
Shortened stays, less expense, and fewer tests sound nice, but and independent verification after some period of time seems imperative.
It also makes me wonder about some of the outlier patients. Those tests may be very beneficial for them. However, I'm sure doctors may feel pressure to not order them due to the cost savings.
https://data.medicare.gov/Hospital-Compare/Hospital-Readmiss...
This also came up in things like Consumer Reports and their analysis of hospital acquired infection data. Small community hospitals with uncomplicated patients end up looking really good.