I was at my employer's annual sales conference recently where many digital health vendors and their CEOs were present. In one of the sessions they all said that they use machine learning to personalize their product for users. The CEOs did not go into any details. I could tell that some of my colleagues and partners in the audience were skeptical, but they did not ask the leaders of those companies to elaborate. I suspect that most people in the audience did not understand what the leaders of those companies meant when they said they use machine learning. Unfortunately that might be the case with many enterprise software buyers. But it does not have to be that way. While designing machine learning driven products is tough and requires a highly skilled team of data analysts, data scientists data engineers, and product managers, the basic concept of machine learning is quite simple. In this post, I will take a few examples from the real world and my experience at Castlight, building machine learning driven products, to explain what a machine learning model is, what a simple rules based model is, and how those models are used in the real world to benefit employees of companies that buy software driven by such models.
At Castlight Health, we make it simple for employees of American companies to navigate their healthcare, benefits and wellness programs. We provide employees with a web application and a mobile application where they can see health benefits information personalized to their needs. To do this, we use a combination of rules based and machine learning based models. This is a hypothetical example of a rules based model. "If a pregnant woman is over 35 years of age, place her in a segment called high-risk-pregnancy". This model is a not a machine learning driven model. It is driven by clinical rules written by experienced clinicians. Software developers simply listen to clinicians and turn that rule into a software program. This information is then given to our personalization engine. Once our personalization engine understands a woman is in the high-risk-pregnancy segment, benefits programs that are relevant for a woman with high risk pregnancy are promoted to her via our web, mobile and email channels. When she logs into our application, instead of viewing information about all the benefits her employer provides, which could be exhausting, she will see benefits that are relevant for her. The hypothesis is that such personalization makes employees aware of relevant benefits, engages them with the benefits providers and wellness programs in a timely manner, improves their health, and reduces healthcare costs for them and their employers. This actually works. That is why hundreds of employers pay us tens of millions of dollars every year. The example we saw above is rules based prediction and personalization. You may be wondering about an example where machine learning is used to predict and personalize. To understand that we need to first understand what machine learning is.
What Is A Machine Learning Learning Model?
A machine learning model is where software developers do not program the computer (the machine) with an explicit logic. Instead they make the machine learn by training the machine with historical information. For example, if we want to check if an email is spam or not spam we can create a machine learning model. To do this data scientist will take several known spam emails, label them as 'spam' and feed those to a suitable computer program, 'the machine'. Data scientists will not say why the email is spam. They will simply say "My dear machine, this is a spam email. I want you to look at the email and recognize that this email is spam". So the machine learns what a spam email might be like and, after looking at enough spam emails, gets better and better at identifying a spam email. After it gets really good at identifying spam email, the model is deployed and goes to work identifying spam email in the real world and putting them in the junk folder. It is important to note that data scientists in most product teams don't invent the computer algorithms they use. They simply use an existing computer algorithm and build a model. It is a bit like this. An electric car engineering team does not have to invent the electric motor. They just have to design it for the particular type of car and build it.
A Machine Learning Model use Case For Benefits Navigation
We just looked at a real world example of a machine learning model. I now want to give you a hypothetical use case of a machine learning model in a health navigation product such as the one Castlight Health provides. Let's say that employers want employees who get an unnecessary back surgery to get a second opinion before they decide on the back surgery. This is because clinicians know from experience that many back surgeries do not improve the condition of a person's back. Instead they cost a lot of money for the employer and the employee and cause a lot of pain and suffering for the employee. In most cases, a surgery also results in weeks of time off from work and, in some cases, lost wages for the employee. So there is a big incentive to identify people who might get a back surgery and make them aware of second opinion programs as well as inform them about the costs and benefits of back surgeries. The problem is there is no simple rule to find out who might be considering a back surgery. This is where machine learning comes handy. Castlight data scientists have access to de-identified information about the medical history of people who had a back surgery. They can feed that information to a computer algorithm (the machine) and tell that machine "My dear machine, this is the medical history of people who had a back surgery in the past. I don't know why they got a back surgery. But they all did. I want you to look at this data and learn to identify people who are likely to get a back surgery." With enough data the Castlight machine learning models gets really good at identifying people who are likely to get a back surgery. The model is then deployed to analyze medical data of employees and predict if someone is likely to get a back surgery. Once it identifies such people, the model informs the Castlight personalization engine about this. The personalization engine then goes to work, promoting second opinion programs and educational information to those identified via web, mobile and email channels. Once again, the hypothesis is that even if the product prevents only a few unnecessary back surgeries a year, the cost savings could be in the hundreds of thousands of dollars and the health improvements could be significant too.
I avoided technical terms on purpose in the above examples. Data scientists reading this post, will recognize that what I described above is supervised learning. There are other types of machine learning, which I did not go into in this post.
If you are an enterprise buyer or a sale person competing against another product claiming to be machine learning driven or artificial intelligence driven, ask the product manager or the sales person to explain the use case. If they are not able to explain in simple terms what they are using machine learning and how it turns into real value, you should be very skeptical about their claims and verify before buying their software. A product does not have to be machine learning driven to be good. Simple rules based engines can do a good job to address many problems. However, it is important to understand the difference.
Please note that for business reasons, I did not use actual use cases. I also did not go into more details about our personalization engine, and our system of intelligence, which are far more sophisticated than what I outlined here. If you would like to learn more, contact me or my colleagues at Castlight Health and we will be glad to share more. If you are in the San Francisco Bay area drop by at our San Francisco or Mountain View offices and I will be glad to give you a demo of our products. If you like this kind of work and want to join Castlight, give me a call. We are always looking for good data scientists, data engineers, clinicians and product managers who understand machine learning and data driven products.
At Castlight Health, we make it simple for employees of American companies to navigate their healthcare, benefits and wellness programs. We provide employees with a web application and a mobile application where they can see health benefits information personalized to their needs. To do this, we use a combination of rules based and machine learning based models. This is a hypothetical example of a rules based model. "If a pregnant woman is over 35 years of age, place her in a segment called high-risk-pregnancy". This model is a not a machine learning driven model. It is driven by clinical rules written by experienced clinicians. Software developers simply listen to clinicians and turn that rule into a software program. This information is then given to our personalization engine. Once our personalization engine understands a woman is in the high-risk-pregnancy segment, benefits programs that are relevant for a woman with high risk pregnancy are promoted to her via our web, mobile and email channels. When she logs into our application, instead of viewing information about all the benefits her employer provides, which could be exhausting, she will see benefits that are relevant for her. The hypothesis is that such personalization makes employees aware of relevant benefits, engages them with the benefits providers and wellness programs in a timely manner, improves their health, and reduces healthcare costs for them and their employers. This actually works. That is why hundreds of employers pay us tens of millions of dollars every year. The example we saw above is rules based prediction and personalization. You may be wondering about an example where machine learning is used to predict and personalize. To understand that we need to first understand what machine learning is.
What Is A Machine Learning Learning Model?
A machine learning model is where software developers do not program the computer (the machine) with an explicit logic. Instead they make the machine learn by training the machine with historical information. For example, if we want to check if an email is spam or not spam we can create a machine learning model. To do this data scientist will take several known spam emails, label them as 'spam' and feed those to a suitable computer program, 'the machine'. Data scientists will not say why the email is spam. They will simply say "My dear machine, this is a spam email. I want you to look at the email and recognize that this email is spam". So the machine learns what a spam email might be like and, after looking at enough spam emails, gets better and better at identifying a spam email. After it gets really good at identifying spam email, the model is deployed and goes to work identifying spam email in the real world and putting them in the junk folder. It is important to note that data scientists in most product teams don't invent the computer algorithms they use. They simply use an existing computer algorithm and build a model. It is a bit like this. An electric car engineering team does not have to invent the electric motor. They just have to design it for the particular type of car and build it.
 |
| Image Courtesy: Europeana Collections |
A Machine Learning Model use Case For Benefits Navigation
We just looked at a real world example of a machine learning model. I now want to give you a hypothetical use case of a machine learning model in a health navigation product such as the one Castlight Health provides. Let's say that employers want employees who get an unnecessary back surgery to get a second opinion before they decide on the back surgery. This is because clinicians know from experience that many back surgeries do not improve the condition of a person's back. Instead they cost a lot of money for the employer and the employee and cause a lot of pain and suffering for the employee. In most cases, a surgery also results in weeks of time off from work and, in some cases, lost wages for the employee. So there is a big incentive to identify people who might get a back surgery and make them aware of second opinion programs as well as inform them about the costs and benefits of back surgeries. The problem is there is no simple rule to find out who might be considering a back surgery. This is where machine learning comes handy. Castlight data scientists have access to de-identified information about the medical history of people who had a back surgery. They can feed that information to a computer algorithm (the machine) and tell that machine "My dear machine, this is the medical history of people who had a back surgery in the past. I don't know why they got a back surgery. But they all did. I want you to look at this data and learn to identify people who are likely to get a back surgery." With enough data the Castlight machine learning models gets really good at identifying people who are likely to get a back surgery. The model is then deployed to analyze medical data of employees and predict if someone is likely to get a back surgery. Once it identifies such people, the model informs the Castlight personalization engine about this. The personalization engine then goes to work, promoting second opinion programs and educational information to those identified via web, mobile and email channels. Once again, the hypothesis is that even if the product prevents only a few unnecessary back surgeries a year, the cost savings could be in the hundreds of thousands of dollars and the health improvements could be significant too.
I avoided technical terms on purpose in the above examples. Data scientists reading this post, will recognize that what I described above is supervised learning. There are other types of machine learning, which I did not go into in this post.
If you are an enterprise buyer or a sale person competing against another product claiming to be machine learning driven or artificial intelligence driven, ask the product manager or the sales person to explain the use case. If they are not able to explain in simple terms what they are using machine learning and how it turns into real value, you should be very skeptical about their claims and verify before buying their software. A product does not have to be machine learning driven to be good. Simple rules based engines can do a good job to address many problems. However, it is important to understand the difference.
Please note that for business reasons, I did not use actual use cases. I also did not go into more details about our personalization engine, and our system of intelligence, which are far more sophisticated than what I outlined here. If you would like to learn more, contact me or my colleagues at Castlight Health and we will be glad to share more. If you are in the San Francisco Bay area drop by at our San Francisco or Mountain View offices and I will be glad to give you a demo of our products. If you like this kind of work and want to join Castlight, give me a call. We are always looking for good data scientists, data engineers, clinicians and product managers who understand machine learning and data driven products.
