Duration 40:38
16+
Play
Video

Fairness in health AI: ensuring equity in healthcare cost. By Miguel Martinez, Data Scientist, Optum

Miguel Martinez
Data Scientist at Optum
  • Video
  • Table of contents
  • Video
Video
Fairness in health AI: ensuring equity in healthcare cost. By Miguel Martinez, Data Scientist, Optum
Available
In cart
Free
Free
Free
Free
Free
Free
Add to favorites
12
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Miguel Martinez
Data Scientist at Optum

Miguel Martinez is a Data Scientist at Optum Enterprise Analytics. Relied on as a tech lead in advancing AI healthcare initiatives, he is passionate about identifying and developing data science solutions for the benefit of organizations and people.

View the profile

About the talk

With the widespread use of AI systems in our everyday lives, it is important to factor fairness while designing these systems. In this work, we discuss fairness and bias related to healthcare AI: it’s impact, source, and actionable methods for evaluating and mitigating it. We demonstrate a hands-on fairness analysis of an algorithm used to predict healthcare cost, revealing fairness in model predictions across demographic groups.

Share

Hello everyone. My name is Miguel Martinez. I am a data scientist at Optum. And today we will be going over Earnest in health AI ensuring equity in healthcare cost predictions. So I'm first we will give you an overview of darkness in LA and then we will walk you through a house on demo of a furnace analysis on health care costs predictions. To talk to you a little bit more about our company optune. We are a health services Innovation company. Our mission is helping

people live healthier and helping make the health system work better for everyone. You might hear a little bit of background noise on my street. I think everything should be fine just now apologize for that. So I work core competencies are around data and data and analytics, you know focused on clinical expertise and better things and consumer experience and all the power products and services fall within our main business units of them inside Optum Health Optum RX

We aspire to improve the experiences and outcomes for everyone Reserve while also reducing the total cost of care. We connect and serve members across the Healthcare System. We are very grateful. We get to serve nine out of 10 1400 employers around 80 Life Sciences organizations for Auto V Health Plans, 9 out of 10 US hospitals and government across the 50 states and DC overall empowering more than a hundred and twenty-five million consumers to live healthy lives. So if you

look at it that you're on an Ethics, you will see some general principles for ethical and you might see princess accountability and sharing that AI practitioners teams and organizations responsible for the are responsible for the design development and implementation at Arkansas VA Solutions is related to what are the policies in practice of a company and honor they clear and actionable. How are you? How is the company thing at keeping track of all the records for

processing and making decisions related to A second point is related to privacy. So AI Solutions should be should protect personal privacy interest ensure consistency with applicable policy stander's United States showing in healthcare related to Toshiba in the in the in the US as well. As any other regulations in that a company has in practice and especially protecting taste like they use and abuse. Another Bitcoin is expendability. So AI Solutions should allow humans to understand Monitor and review its decision processes. So when the model makes a prediction why is why is it making

that prediction on a what are the important features in wide is the Pacific resource for a Ford Alistair City prediction and what kind of takeaways can we get from the from the weather machine learning model is is learning. And finally fairness near this place is very to AI Solutions should minimize or avoid introducing historical bias and ensure distributive justice for the benefit of all groups. So think you know, how can we make sure that the Hour Air Solutions are not picking up in justices that are happening in the world right now

and how can we benefit all o patience especially those who are part of disadvantaged groups? And this will be the focus of a bistok furnace in the context of healthcare and AI. There are many opportunities for health care and may I hear some prominent areas of healthcare and some examples? We see that in terms of self-care and provide information on your whereabouts right now like the Apple watch has a personal issue G-Eazy and they can they can run some analyses have to do it on your day today. We have a diagnosis such a system that

check your symptoms as nothing such as medical diagnosis 45 images MRIs and x-rays and silver. Can you call the Station Sports Chat prediction the the occurrence of acute injuries trading appearance? And you know, those are some prominent areas of butter in our company up to me. It goes across the entire Healthcare Spectrum ringtones, including the administrative part of Healthcare Management. I'm feeling all these different types of healthcare and really be

helped by the use of and yeah, I will continue growing it is it's opportunities and its Health after feel so give you know that I've given that is becoming more prominent in healthcare. How can we ensure that a i is for everyone? Are Healthcare AI Solutions in a student's true? Turn it across all patient populations all these benefits and all these applications. They should be helping all they all the users that are that are seeking to use them. Harvey's is avoid to avoid perpetuating

or excessive masturbating in health disparities when using AI Solutions, so if there are some discrepancies are our differences and disparities in the greenhouse in the current health system, how can we avoid transporting those into the into our AI Solutions? Anesthesia for vulnerable vulnerable groups, how can we proactively evaluate outcomes of our AI systems in order to ensure that they are not being negatively affected by these Technologies. You say you know,

we were talking about well, we don't want you to have bias in an hour in our AI Solutions and you know in order to understand that it is helpful to see what what is bias. How can we think of bias and he was a very good definition of algorithm bias in the context of these are the incentives when the application of an organism compounds exist in inequalities in socio-economic status raise in background revision gender disability or sexual orientation to I'm amplify

them anniversary impact inequalities in Health Systems in our current economic status in sibelius Swan. You know, how come you how can we make sure that all of our algorithms are not compounding those discrepancies? negatively impacting those groups in the house system Okay. Well now we we have a definition of what what bias is here is a diagram of you know, giving the definition given that that we know what what is going to be what are the different points across the email lifecycle where we could find by starting in the generation?

Of course, we have historical bias. Those are all the discrepancies and disparities that have her in the past that may be occurring that right now in our current Healthcare System. We also have representation bias. So if maybe we are selecting a population to train a model 2 to perform some analytics. How is that a population related to the overall Society? Are we grabbing a random sample could possibly be an excuse sample from a population and maybe we are not taking into account in a

minority groups in that sample. Then we can look at other aspects like measurements. So if we have a given variable that we are trying to measure like this is progression. How are we measuring those variables that what kind of measurements are we doing to make sure that I never using any process that may contain bias. After that, I'm we have already decided and we have power Inn and Suites and pre-processing. We have our training data and a work evaluation data. There were looking to the action

model building and implementation and this could be you know, this could contain evaluation by us. How are we evaluate in the performance of our model? Are we are by The Waiting. Based on a on a population that represents our whole society. Do we have a bias an evaluation group for Portugal wedding the performance of power model. Maybe we're comparing tomorrow is some sort of external Benchmark. How is that a stem to stern out benchmarks determine? How does he compare to the overall population? Those are some areas where to buy a food critic or

aggregation bias. So, how are we using the tomorrow for making predictions can the Moral Moral correctly predict across various groups, or is it mean to perform only in certain population and maybe for all the populations you might need different models. Then once you have your model output and you might you might see the point in bias. So how is Howard the model predictions used and interpreted is the model being used for what it was intended to be used

for and I know this kind of this kind of deposit point in applications that we would watch out for rent. Make sure that they contain no bite us. And now we're going to take a closer look into historical bias in the context of healthcare. Health disparities can be affected by multiple interrelated factors. Here are two very good and definitions of disparities. We have Healthy Start. He's so a higher burden of illness injury disability or mortality experience by one group relief another Healthcare disparity, which are the differences between groups in health insurance coverage access to

and use of care and quality of care. So we ended up seeing that a complex and interrelated set of individual providers Health Systems societal and environmental factors affect disparities in health and Healthcare. So, you know for any given two populations, we can look at different factors and different result and we can see where I'm at with my City princess princess in their health insurance in the quality of care. You might see differences in propensity of off of the seas in mortality. I know those differences might

be affected by a complex at Rancho factors such as you know starting from the individual level 2, what kind of provider they have access to to the health system and so on so understanding of the spectres and how they are related to other Elko to get get a better understanding of health and Healthcare disparities and how they are related to model thermos. A big part of furnace and unbiased are Ann and disparities in healthcare outcomes are social determinants of Health the World Health

Organization Nashville, social determinants of Health as the conditions in which people are born grow leave work and age so they truly impact health and Healthcare outcomes and disparities. 13 reason I'm economic stability employment income expenses that medical bills and support. The neighborhood. Is there housing are there Transportation options available in the crown are they are there parks and playgrounds available to walk around? Educational what up? What about like in literacy and language that people getting childhood vocational

training access to higher education. Food quiz, very important factors such as hunger and access to healthy options. social context such as social integration for assistant Community engagement discrimination Kern Family Health Care system after its adjust health coverage provided availability provider bias cultural and linguistic competency and quality of care. And you know, I am toast all of those are our present that we have taken you have an impact on

on on the different outcomes of health and Healthcare and healthy now, we're model predictions then no friends that I have. I haven't spoken with positions here in Boston who tell me they make they are not enough there. There are not enough providers and officials who speak Spanish do to treat the the Spanish population here in in the Boston in the Boston Massachusetts area, and there are many examples of different populations and geographies. And information another factor of another factor of of these of Health disparities the

results of the generic and biological aspect. So different populations may have different genes and different biological characteristics that are determined by biology Sol health and Healthcare are one of those areas where L Comes My Baby naturally difference a princess. If you are measuring disgusted on levels between men and women, you will see natural differences. So in this thing as you know furnace doesn't mean that everything might be equal for everyone in every Asterix, but rather you not giving you a specific needs even your specific

genes and biological factors. How can we eat every 20 treatment and the terror that they need? And to illustrate this point of healthcare disparities, here are some concrete examples of what is Paradise look like in our society in the US. Renaissance and black women are 42% more likely to die from breast cancer than white women. black American Indians Alaskan Indians and Eve and conditions including half and diabetes Between 2010 and 2018 blacks remain 1.5 times more likely to be an easier than what in the Hispanic miniature

red Romaine or 2.5 times higher than they were then under braids for white people. Especially in this year with the pandemic health and Healthcare disparities have shown to have a big impact on the on the outcomes of different communities and groups for example over 55% of covid-19 cases and 50% off of Kuwait XIX Roman racial and ethnic minority groups. So even though they they represent a smaller percentage of the population. They have were there the majority of the cases of covid-19.

And if it has been shown also that Community, 19 mortality rate among black the black population is twice the rate than the normal amount white. So they're there their significant and I'm big differences in how different than it has affected different demographic groups specially specially mine are vulnerable groups. Then we nuke highest points in representation and measurement. So what kind of population do we have for our model and how we're measuring different features and labels that we are using for our ml solution? Taking a look at these

parties in healthcare data collection is cute to wear those who can afford and access quality and temperature there is incomplete data prevent us from having a holistic view of patient population report population and especially vulnerable groups, you know, those groups may not know how many might have found more my half hour less complete. They that they might be lacking carrying the pronostic so we have incomplete or those vulnerable groups. There's also the case of a small sample sample

size. So vulnerable populations are not correctly represented in healthcare data. So I'm gonna be here population in the in the training size for a birdie in real life. But for those minorities we might have a disproportionately small amount of data and therefore the models might not learn as much given that they don't have the right data. Albino, historical inequalities embedded actress like a higher burden of morbidity. I'm very stupid to accessing care. So all the different factors

that we previously saw like the social determinants of Health. They might be contained the data. And you know, I'm looking closer into those aspects and there are gaps of cool when disadvantaged groups experience barriers in access to to Carousel. If you are not able to go to the hospital or get a certain treatment, they may not have that complete their outer that those records when when they're used for for model training and testing. Socioeconomic status influences were patient access Care on data collection varies across care size.

So they might not have access to points were most of the Darius collected there for the day that might be lacking. Remind me of subjectivity. Can you find what that is captured during a visit to the different notes in the different diagnosis that a provider may may may may may vary across population. And finally health-record systems and processes may not be billed to collect holistic patient data. So especially with things like social determinants of Health there might be too damn there. Codes for 4/4 in putting them, but they hold this down the

whole processes are not really created for holistic view of the patient. Then we look at the use of process proxies in healthcare data. It's good to carefully evaluate. The use of measurement is a proxy for a true outcomes of Interest even that processes might be affected by social inequities. So, you know, it's very very important to 2 to look at this option that we're making some we we we assume that a proxy data measurement is a proxy for a real outcome. Is that the case that the we know for sure that they might not be any bias or

any them in the qualities that are embedded in that proxy. Some examples of Prophecy sorry hospitalization as a proxy for TC severity and healthcare costs as a proxy for healthcare outcomes. And you know, there have been stories that showed that for a given health care cost different populations may have different Healthcare outcomes. So it is very important to re-examine your cross proxies and ensure that no bias is I know he would look at the Apollo Mission by us. So after we train the model,

how are we about waiting tomorrow predictions? And how are we ensuring that the model is predicting? Well across different purplish populations. Here's a very good diagram for deciding the model A valuation metrics of what kind of metric are you using to see how well your model is performing you can have as a first question is based on history presentation or ml system errors. So maybe you look at representation. So you when you and that you wanted to be that has a population for your model is correctly represented according to the societal representation

from Dad. Would you want to select an equal number of people from each group or proportional to the percentage of the population? So then you could have equal Paradiso you have equal numbers or proportional you have a proportional Authority. Danny for looking at the errors are the interventions from your NFL predictions 2018 or a C+. So are they would it be something that could help people or would it be something that could harm people in this case percentage of the population? Yes or

no, and then we have different metric Sardar recommended. So this is my percentage. We can use false teacher that they bought Scurry rate is steady and equal across a different populations. If it is something that is no helping people. I get a winter beanie with a very small percentage of the population if yes, we will look at 3:30. So, you know of the people who need this help if intervention. Which group how many are we correctly identifying from our ml

Solutions? So for instance, if you are looking at hospital Grand Nationals were when we are helping a very small percentage of the population. We will look at the Ricoh party. So we would be the more we would be pretty in Hospital admissions. So we went to make sure of the people who need a hospital. I need an intervention for hospital admission and each group how many of our what percentage of them are we correctly identifying on which demographic group and that way we can shoot at the recall is is steady across the groups under the model is performing. Well

across our demographics. There are different tools for a viable wedding furnace and different companies and different. Groups and Academy groups have created multiple tools that help us make sure that our models are fair. We have Google created Microsoft created IBM created the iPhone SE 60th and University of Chicago created all of them have some very nice pictures and there might be there are many others that are not listed yourself or choose to explore

around and choose the one that you see best fit your needs for evaluating furnace for the purposes of today's top. We will be using Equity has to have all the way the furnace cross different demographic groups for a organism that performs that reduce Healthcare cost. So now we will go to the demo of ensuring equity in healthcare costs predictions. I can share my screen. Alright, so this is a demo of Exodus doing a furnace analysis. And so this is to show you what a furnace and all these would look like in healthcare data again,

you know when we have here to help their competitions Equity as he saw us and help us make sure that our morale is pretty well across all of our demographic groups and show some very nice tables and illustrations for model performance. Mesh wreaths across a different populations for these health care cuz we have demographic attributes for a race gender and age for the purposes of his analyses. We have selected populations of black and white and we have the target variable of the

outcome of healthcare costs prediction, so it can be an actual we have they are the trueresult we have the actual Then we have to have been us for healthcare costs and these guys are healthcare costs East evaluated as a binary variable. Yes or no, so it would be the people who are in the right help her calls or low in healthcare costs and buy High healthcare costs near me in the top 2% of the health care cost. And we have my dear friend with the following columns against sex and gender the age the member age there pretty tight cause and the actual high cost

and 4-H. We are bucket in the edge into multiple categories into into different ranges for age. We first conduct some initial experts right now is he received that they are / 2.4 million members in our dinner. We are looking at their bank accounts for each population group so we can see the population is around 90% white and 9.3% black. It is around 53.3% female and 46.6% Now here we can see the difference between 6040 and 56 to a hundred and you can see that they are around evenly distributed. So 27%.

6% starting a 4041 25% of 726 and 20% starting at 56. Take me to look at actual versus Predator versus actually traveling demographic categories. So here you have for the black population in the white population the predicted to send to know for the black population. We have around 2%. The high score for why it is is around 1.5% and 4 wide we have around 1.5 1.5% Then we look at those distribution Us by gender to female. We have a predicted high cost of 2.1% for male. We have a pretty good cause of a 1.8% hike or sweetheart perfume, 1.5% and

4 male around 1.5% Then we look at the dogs distributions according to the age categories and you can stay here to 25.05% and 26 to 40. We have a run 1% starting at 41 around 2% and 4% for the actual one. We have first Buckhead .5% 2nd Buckhead .8% and the last pocket around 3% Of course, we see that age is a big factor in health and Healthcare outcomes. So you can see the people of a older age tend to be more than to have a higher participation in actual and predicted high cast. Then we will perform an hour from now. Fairness and I see if we think we

have further than the bias of their looks at those groups and those metrics and calculate disparities between those groups metros play some dusty Francis across groups groups. We import all the packages from Exodus on the all the different objects. And here you see what the data frame looks like before we have the same about you. You have the race. Why black you have the same X male female. You have the different Edge pockets. And then the score will be there the prediction if this person to hike out yes or

no and then the label is that the actual the actual outcome of deceased person has yes or no. I'm here some documentation from the aqueous website. You can see the baby. Very good job defining all the different terms Supreme sensor false positive on the prediction is a true when the when the prediction is true brain real life is is a negative and a recommended that you check out. These gas station is very helpful as well as a bunch of very good. Very good metrics definitions for the different model performance metrics. So this is our different ways

of evaluating the Motor Performance. If you're using the false-negative rate and then swollen and you get some very good definition of why do they mean and they also have a very nice decision Tree on the website. I think you saw a little bit of the slides, but you could you go to their website and see that they're tired the entire tree in your body. But I know you have some sort of like completions like, you know, this is the one that were using for for this case. So maybe you want to help people Court higher cause you to be and then something goes you

will look at a party and so he knows we can only provide assistance to a small fraction of people always need attention sure. It is distributed in a representative away. So we're helping all the people who needed a demographic groups and their menu that metrics that you can look into and you can lose their definition and they do a very good job explaining. Here you can see some actual metric that they have like the true positive rate and they file false negative rate and here you can see all dysmetric theme compute.

So you have been raised and sex and age and negative radon deposition as the one for the for the different. Are the different groups and and do you know you get out very nice to you as well as they're in the prevalent like the one we can have earlier so you can see across groups. You can get across groups are very good idea of how is the model performing across the different different metrics after you have an again the three give you a very good notion of which metric you should you should use and it is very important that shows

to choose the ride at their iMessage can come in different populations. They might have different distributions of naturally, they might have different distributions of actual and predicted. So it is something very important to to really make sure that the day the message that we choose provider in the Middle East is the correct one. So here we have some for you can make a plot of the true positive rate. So you can look at their two different to positive. Right so you can see how you know, you

know, how you can use equity to see the capacity of Rights or another metric across different different groups illustrations of different and different metric that you can compute a so you could have, you know, I'll discover a false negative thoughts, and so on. So there are there many many different metric that you might you might use. So something interesting. So you couldn't Franklin you could see like what disease is the predictive? Fred for the different positions and you

can see here that you know, the prevalence for the white population is is 90 vs that the prevalence for the black population in stands to you that is a a big difference. But in this case the prevalence comes from the fact that the population of India cineworld in the world did the white population represents 90% of the population versus the black population represent represent 10% of the population. So because this because naturally 90% of the people are are white and I were there said in the deputy prevalence which unites is based on the distribution is going to be going to follow that

so it's going to be a 90% chance for white and 90% black. So in this case, you know, we know what that what that is. I met you that works that way so that is something to be expected. So that is and that is not that that might not be the best metric that we want to choose for portable. writing by Ashley mattress and Metric like a Ricoh printer and then we enter the BIOS class. We have found a reference group. In this case. We have the race of white the sex with male and ages 26 2:40. So we use that as a reference group and then we compare

all the different groups to the tobacco group. And then we see how many how much are they buried and then you can see here the different attributes like race race or black or white and then you can see by how much they they vary so I can because why is our reference class they have one and you know, all the other ones are buy it by how much they did the earlier. We got to hear that for the race of the position is 1.05 times greater than that for the rights of wife. So their position is actually a little bit better for for the

black population than done for today for the white population. And you know you can you can and you can computer all day all those meds reason for the different differences for the different metrics on the different populations. And you also have some very nice like that for you. Hopefully the station so you can see by age that a 27-day Falls I'm straight and then you can have a different color or so if it's very big and you you might be like, you know, if you get smoothies color is very important to tell you these things into concept that so, you know, we know that maybe hire people

tend to have naturally hired hired if we see that this has created for the older popular population versus the younger one, you know, if we we see that there is a natural difference that is not something that that is really affect. You can have all these blowouts and finally for the furnace class. You see the dog is difference is that you have calculated and you have some sort of dress called that you can choose so you can say well, you know if it is different, but you delete the difference is greater than in a 20% And so and then you can flag some then you can Flex on

some some Since I think that's fair or unfair, so we here we see like to find out the final conclusions from Ecuador. So I think yeah for this is a specific case. We are looking at it through positive right or recall so that the greater this value is the better because it means is capturing more people that needed so you could see here that for example for Race by population has a supposed to be afraid of 0.35 X population has a positive rate of 0.43 Sol de Mona Lisa actually believe in better on the black population

compared today to the white population and so on for the different metric that are you see here that is the end of the presentation. I think we're just at the time. Thank you very much for your time and for being here. I hope this is how fo you know, this is no way I can truly have a big impact on you know with your participation. And all of your work, you can really eat. We we can really make sure that it is fair, and I need is helping at Patients across all the different demographic groups. So, thank you very

much and I hope you have a good rest of you.

Cackle comments for the website

Buy this talk

Access to the talk “Fairness in health AI: ensuring equity in healthcare cost. By Miguel Martinez, Data Scientist, Optum”
Available
In cart
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Global Artificial Intelligence Virtual Conference”
Available
In cart
Free
Free
Free
Free
Free
Free
Ticket

Similar talks

Shahidul Mannan
Head Of Data Engineering & Innovation at Mass General Brigham (Partners Healthcare)
Available
In cart
Free
Free
Free
Free
Free
Free
Sanji Fernando
VP Technologies at Optum
+ 4 speakers
Matthew DiDonato
Senior Staff Product Manager - Artificial Intelligence at GE Healthcare
+ 4 speakers
Nataraj Dasgupta
VP Technologies at Rxdatascience Inc.
+ 4 speakers
Slava Akmaev
Chief Technology Officer at Scipher Medicine
+ 4 speakers
Shahidul Mannan
Head Of Data Engineering & Innovation at Mass General Brigham (Partners Healthcare)
+ 4 speakers
Available
In cart
Free
Free
Free
Free
Free
Free

Buy this video

Video

Access to the talk “Fairness in health AI: ensuring equity in healthcare cost. By Miguel Martinez, Data Scientist, Optum”
Available
In cart
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
561 conferences
22100 speakers
8257 hours of content