>> Eric Horvitz: Okay. We're honored today to have Joel Robertson, founder and CEO of Robertson Technologies. Joel is a scientist and what is seen as biosocial entrepreneur. That's a nice phrase. He's developed quite a few healthcare programs using computing as its foundation. Robertson Technologies is dedicated to making -- having impact on healthcare around the world, and it's a -- I guess it's a network of nonprofit and for-profit organizations with a kind of a global perspective. Joel is the -- also the visionary behind NxOpinion, written N-X Opinion, a healthcare platform that is based on probabilistic methods for doing diagnosis and recommendation. That application over the years has been known as an application built on Microsoft technology that's been featured in a variety of ways, and it earned Robertson a 2004 Computerworld Honors Award nomination. And I always like bios that end like this. I think it's really nice. Dr. Robertson founded his organization on the principles of compassion, respect, and hope. Three nice words. So today Joel will talk a bit about NxOpinion: A Novel Integrated and Predictive Solution for Global Healthcare Delivery. Joel. >> Joel Robertson: Thanks, Eric. The first thing that I'll do is apologize because when you take an hour and a half and you have to talk about vision you have to talk about science and you have to talk about applications. You can't go deep into all of them. And so for those of you who want more information, more than willing to share with you, so we're going to try to do that in an hour and a half, talk about all of those various aspects of NxOpinion, so that you can see this is a product that has gone from R&D and definitely we usually say you can see David Heckerman all over it. David has been actively involved from the beginning in helping us to design the mathematics, which also go back to Eric Horvitz' and David's philosophies that have come into this application. And it's very definitely has a Microsoft stamp on it because of that aspect of those guys and their research and thinking. They're the minds behind it. We kind of put it into making it work in the real world. Robertson Technologies is a -- we refer to it as a Robin Hood theory of medicine. You make money where there's money and you give it to the poor. And so we have profits and nonprofits, 501(c)(3) public charities. We work with various organizations such as the Bill & Melinda Gates Foundation, Intel World Ahead, Johnson & Johnson, Project Unite, a lot of those nonprofits. As well as in the profit world, you're going to see this fall a launch out of UPG with the OneApp. We'll be on the OneApp platform most notably towards South Africa and India. So that the whole theory is how can we make a significant difference in world medicine is what we look at. And it comes through partnering. And partnering with people such as Microsoft in order to be able to deliver a solution. There is -- all of us, when you look at it in the world of the developed country, there is an ethical responsibility for us to care for the poor. In 2002 is when, you know, I was kind of setting there and had a successful publishing career in behavioral medicine area and began to look at, you know, what about the rest of the world. And I've done enough international traveling to say we do have a responsibility for us to look at the poor around the world and especially if the solution lies within technology, which we do believe, at looking at how can we get there. In India alone there's 512 million who will never see a physician. And I'll say very simply: Should everyone see a physician? Yes. But the reality is it can't happen. So how can you take medicine, safe with clinical supervision, to the poor without having quacks delivering the care but having good, solid medicine and getting good metrics out of it? The only way you can do it is through technology. And so that was our solution. That's what we looked at. And we said we have to have it innovative. One of the most important things about a solution for the poor is it has to be more sophisticated than a solution for the wealthy. When you take people who don't know enough to ask the question, you have to be smart enough to ask the question that they don't know to ask. So that was one of our challenges, to say how do you get software that's that smart that it can get into people who are untrained and yet get a right answer. So in 2002 we started NxOpinion. And we started it in a nonprofit environment for two reasons: Another guy and I invested in it and we didn't know if we could do it. In fact, we had doubts, so we at least wanted a tax write-off. But we had three questions, you know: Could we do it, could technology do it, and does anyone want it. And so we did those three questions in a nonprofit, and when we came up with the answer to that solution, then we kind of moved into an area that said let's add the profit to be able to move on. So could it be done. One of the things that was very important is there had been a lot of programs out there, QMR, a lot of programs that had been out there in the area of clinical decision support and things of that nature. But one of the first things that we had to do is we said is there the ability to be able to be smart enough again, if you don't know a question, to be able to get to the answer. And let me use an example of how this works, and I give you a big picture how it works, and we'll take it from there. Let's say you're a 19-year-old female. It's February. You're in Banaganapalli, Andhra Pradesh, India, and you have a cough. And you have a health worker that that's all they know. This software has to be smart enough to know what are the most likely diseases at that time of the month in that location that are going to show up with a cough and begin to ask the question to differentiate between the top diseases. Then it has to give you a confidence level that you have that disease. And then if it's a disease, one you're able to treat because you've been permitted to treat it, such as malaria may be able to be treated by a rural wealth worker, and you have a confidence level, then you get a green light to move to the treatment. And that treatment might be chloroquine. Next what's important is a few days later you ask the person are they taking their medication and not are they getting better but are the chills better, are the intermittent fevers, because you're looking at the symptoms, so you have to have a findings-based diagnostics, and then you have to have a follow-up based reverse thinking, asking findings as to whether they're clearing up. That's the general gist of what this is all about. So we found that we needed multiple diagnostic engines. If we started just with Bayes, for example, which is probabilistic, you'd run down a lot of bunny trails. We also found that there wasn't a database that existed to support that kind of data. There's millions and millions of relationships that have to be in there, relationships that are regionalized, what is -- exists in the U.S. is not at all relevant to something that might exist in South Africa or Rwanda, is might be very different than the mountains of Uttaranchal in India. So you really had to have a database. And there wasn't anything there. So we said, well, I guess we have to develop it. And that was fun. When we began to test it, we tested it through a lot of different dynamics, and it's been well tested. It's in the commercialization phase. And I think Bill Crounse here, director of worldwide health, in a recent program that I did with him said: The result of this work, NxOpinion, is one of the most significant, accurate, responsive and intuitive diagnostic programs I've seen. And that's pretty much what we get. Could technology support it in 2004? Steve Ballmer said that we cut our costs by 90 percent, our development time by 58 percent, and we need like 35 to 40 percent as much code. So better solutions delivered quickly, more maintainable, supportable, in a very, very mission-critical environment. That is essentially what we see ourselves. Our space is what Steve Ballmer had talked about. That's what we do best. We partner with people that do other stuff better because the solution on such a vast problem has to be a working and cooperation between various organizations. Good technology. We had lots of questions. You know, what can you do. There's connectivity issues and, you know, you've got all kinds of mobile platforms and how do you get data in on a mobile platform and how do you make it be able to integrate all the way up through a hospital HIS system, all communicate together. And so what we decided is that we developed what's called a platform. It's a database management platform. Multiple ways of getting data in, multiple ways of getting data out, all using an integrated system. Needs to be modular. Modular, for example, a level-one worker is somebody who doesn't know much. So why ask them the same questions you're going to ask a specialist physician. So modular in that when you publish that out to a phone only the questions that they can ask are asked of them for usability issue. And if they don't have labs, don't ask them about labs. But when they get up into a PHC, a primary health center, and they may have the availability of labs, what labs are available and what can we put the data in to enhance the diagnostics. And more importantly could we tell them what would be important to help them differentiate the diagnosis. So we really had to have a multiple platform. And I think the biggest challenge it has to be agnostic. There are so many databases out there that we have to be able to read to, pull from, and write back to those platforms because those databases, if we're going to do global medicine, people already have various -- you know, Africa is strong in OpenMRS. We've got a lot of different systems. But if we want to look at global medicine, we have to be able to agnostically look and be able to jump into the health vaults, jump into the open MRSs in many different areas and to be able to read in and out. Question: Would anyone want it? What happened when we finished our tests and we went through various tests around the world, 90 different test sites in 14 different countries. And basically it came back and our conclusion at this particular point isn't does anyone want it but how in the world do we partner with people to support the enthusiasm of what has been developed. One of the examples that we talk about, does it work for them. And we recently just got an article where the physician has diagnosed in a rural village in India that the -- I mean, the multipurpose health worker diagnosed indigestion, and they ran it through the NxOpinion. It came back and said it looks like it could be an MI. And they referred the person up and it was an MI. That's an example of the type of stuff that that patient would not have been treated because there was not the information available. So we have thousands of ongoing clinical encounters. So one of the questions comes up, how did we do it? And so because we're here with Microsoft Research, we want to talk a little bit about what's the mathematics logic behind it as well as our vision for the use of it. So the first thing is we had to develop an engine. And as I said, we developed four engines. One bombed. And so we ended up with three that work. And they all work simultaneously in one diagnostics. So it's rather interesting in how we were able to do that. But the key thing is it needed to mimic clinical reasoning. It needed to think like a doctor. And it needed to be predictive. And then the database had to be modular, it had to be global, and it had to be regionalized. In order to do this, if you want to do it on a world basis, you have to regionalize. Regionalize includes language, but it includes terms. There are four different terms for diarrhea in the state of Andhra Pradesh alone in India. So how do you do that? What are the terms that they refer to? Because information in -or diagnostics out is only as good as information in. And then most importantly taking a complex, extremely complex program, needs to be user friendly. So we started out with Bayes. And I think that's why you hear terms such as Heckerman and Horvitz, because when you think Bayes, that's who you think of. So we started out with Bayes and we said I like this concept. It says Bayes' theorem can then be understood as specifying how an ideally rational person responds to evidence. Now, we're making an assumption that physicians are ideally rational, which is stretch sometimes. But if we use that concept, we say how do we do that. So it really comes back, our first thing in saying how do we take medical facts, put them in a clinically weighted. You need to know how important is that symptom to that disease and how important is that symptom related to that. How do we do that. So we have a probability factor, and we went through a lot of discussions with David. We talked about sensitivity, specificity, positive predictive values, negative, and we came up with something that said there is a clinical predictive value. And so what we use, and I'll talk about it later, is what we call the CPV. Very important. A clinically predictive value. Prevalence data, which you also need in order to run Bayes. You're going it's scattered, it's spotty, it's slanted. You know, you may get data on a hospital in the U.S., but where's the relevance elsewhere. And it's really looking at what is the clinical presentation probability. And so we use this clinically weighted prevalence. And what it really says is clinicians agree with it, you can rapidly regionalize it, and most importantly at this particular point in science, that data is subjective, until you get enough data back to turn it into empirical. And so it needs to be a learning process that says this is what we believe to be the clinically weighted prevalence. But in time we'll validate that or change it. So this is our best guess. That's what you have to look at it, because that's where we're at in this field. Sensitivity. You know, is this a test, is it a symptom, is it a sign. So we went into the clinical data descriptive aspect and we said, you know, physicians typically think in terms of I often see this, I rarely see it, or, you know, in our experience. So we had to take this clinically semantic approach and do what we called a CPV, which is a clinical predictive value. And that really means how much does X evoke a suspicion of Y. So now you've got those two terms that we use, CPV and CWP. And we developed what's called buckets. And this, again, we tried to go through, and there is no specific numbers that exist. So what are the buckets that exist. So with the CWP you'd say is it a very rare? Is it a rare? Is it common? Very common? You can get consensus on that. CPV. You can get consensus on this is rare, it's often, it's typically associated with it. The strength of using the bucket is that you've got a broad clinical agreement and you can also regionalize fairly rapidly. You can say these are the top ten diseases that we see in rural medicine globally, but here in Sub-Saharan Africa would you add some to that bucket or would you take them away. And then the same thing you can do with findings. These findings are often associated with these diseases. Would you add something to that bucket or take it away. The next thing that was very important was leaks. One of the things that I always struggled with clinical decision support systems is you have to know what to put in. It's almost like you need to know the answer before you put the data in. And what if we were going to take data, for example, an electronic health record, and we want to incorporate all the data into the diagnostics? Everything that's there. Your disease-to-disease relationships, your preexisting diseases, whether you have a village that has mosquito-infested malaria diseases or whether you have bad water or whether you have previously had an MI. All of that's important in a physician's mind, but they can run -- make Bayes run on a bunny trail. If I'm putting in that I have a headache, I'm dizzy, and I have a sore foot, there's three facts. But two of them are related and one is not. So the challenge of how do you get irrelevant data not to mess you up and take you down a bunny trail if the assumption is we need to make this work with nontrained individuals who don't know the answer as to what's relevant. And that includes our consumer. That includes sometimes a physician who's not specialized because a disease is specialized. And it often occurs in very rare diseases that are not seen very frequently. The physician may be the one who says I don't know enough to ask that question. In order to do that, we had to create what's called a multiple engine. And the first thing when you enter in data in this particular system is it uses a rules-based engine on templates. What it is is we are gathering up and building prior probabilities. Because at that particular point by asking questions directly to get prior probabilities prior to Bayes kicking in, you have taken related symptoms and build up enough that Bayes can then start into prediction. But if you did not do that, you'd have a multiplicity of data and never be able to get Bayes to work. So we created chief complaint templates that are related things. So if you say cough, you may want to know is there sputum production, you may want to know whether that's cardiovascular oriented, or you'd want to know, you know, whether it's respiratory related. So you ask a few questions to get the related probabilities together and then let Bayes kick in. In addition to that, Bayes is a great probabilistic engine, but sometimes there are possible diseases that Bayes will miss, so we ran simultaneously an occurrence-based engine, so that when you see our software, it will say this is probably what's in existence, but you can kick in and say what is possible. And then Bayes can take a possible disease and make it into a probable disease with a specific finding. So you can go to possible, click on it, and then enter in the findings that make it into a probable disease. So that was the creative thing. So what we had to do is we had to create suggestions. We call it one more question. It needs to look through, and when you get to a certain point in Bayes that you've got five, or in our case we use ten top diseases, let's ask the questions of you that will differentiate those top ten diseases. Because you may not know enough to ask it. So we'll ask that. In clinical use it's rather fascinating because you can put a finding in. Let's say you can put a result of a CAT scan in and see what it changed in your probability before you spent the money on the CAT scan. So as you can see, there's a lot of uses for this sort of thing. Let me describe something as you go through. I want you to connect this. We feel what we've built is a combustion engine that can be used in a weed whacker or it can be used in a Ferrari. And so the application of where it is used is varied upon the user, where they're looking at. Are they looking at it from a healthcare cost containment predictive, are you looking at a diagnostic, are you looking at a rural health in order to be able to do that. So we developed this platform. This is a way that we did some early engine testing where we just would put in probabilities and we look at diseases and we put it in front of expert witnesses and we'd -- expert witnesses -- expert panels and we put it in front of -run use cases that may be in literature or something of that nature to see if we were on the right track. The database then became rather interesting. Because now you're talking about when you add lab and when you add imaging and you add findings and you add various locations and extraordinarily complex database. So when you look at the data, you can start out and you can say in general textbooks I can find some relations. And if you look at most of the clinical decision support, you'd be able to see them, you know, connect in with Griffith's 5-minute clinical diagnosis, things like that. It's nothing more than these symptoms follow with this. There's no prevalence, there's no sensitivity, but there's a relationship. And then the next thing that you do is you have to go through and you have to author certain specialty texts. And in our particular case, because we felt that we were on the road of building something rather interesting, every single finding has to be documented of where that finding came from and what piece of literature, and you have to have a standard on what literature is legitimate and what is not. So all of that comes with our -- what we call our content management team. So then you go general, then the next level you go specialty text, and then the next thing you have to do is go to literature review. And there's where you have some real standards of saying what is legitimate information, what is a legitimate study, and then the criteria before that data can be considered a positive relational and what we call published data. Then we go through medical expert reviews. And these reviews, what we did was originally on a global aspect just what are the disease relationships, what are the finding relationships that goes through these internists, specialists, and then you go next to the international. So let's say that there are approximately 850 diseases that you may look at outside of psychiatric and orthopedics. And so you'd look at that. And then there are, for example, 212 diseases that are seen frequently in rural medicine. And so you have a global what we call rural medicine diseases of 212. Then you might go to India and take that specialist and they'll say these are the 212 within India. And then you might go to Africa and say these are the ones that fall within there and add and subtract to it. So you're taking this data and not only on the relationship that exists with the original diagnosis but the relationship that exists within that region. So we developed, then, what we call the NxOpinion knowledge management platform. And early on we met with the team from Microsoft from content management and some technology people to describe our content because we said can we buy a content management system. And the conclusion back in 2004 or '5, I believe it was, was there isn't anything that exists that can do this. So we had to build it on our own. And with a lot of, again, support from Microsoft. So we developed these things that are called a rapid regionalization. And we have a finding name, a regionalized name, a disease name, and then various ways that we can very quickly go in and regionalize a disease. After we got the clinical decision support and we began to do some studies on the accuracy and efficiency and the quality of the diagnostics, we found that there are a couple of things that were very important. One of them is that we found that a clinical decision support without certain infrastructure and electronic health record storage, data, the ability to do metrics on it was somewhat -- there wasn't a purpose for it. It was great technology, but as Eric and I were talking prior, technology putting it into the rubber meets the road and making it work, we found that we required that every encounter that a person had has to build on a health record. If you see a person and, for example, you see a pregnant woman and you put her in the database and she's eight months pregnant, you have two people in the database. You have a woman who's pregnant and you have a minus-one-month child who's going to be needing to be asked were they born alive or not in a month. So you need to have the data that says every encounter creates it, and you have to be able to monitor the metrics of that. So we included a suite. So we kind of look at the platform. And you're going to see when I describe the platform a diagnostic engine, an event engine, a search engine, but you're also going to have an electronic health record and a reporting in metrics that are all associated in this suite that is a platform. How you access that could be through, for example, our Microsoft UPG OneApp, might be a PC, it might be Web based, whatever it is you access this. Where it's stored could be proprietary. And we'll talk about some of the challenges that come in working globally. So NxOpinion version 1, the database took us four years and more than $6 million just to put the numbers in. And if you look at that, that's actually a very, very efficient number, for those of you who have been involved in this field. And we did it by using a very, very skilled group of people. And because we are a small company, low overhead, you're very agile. And so that was what we did. The next thing we had to do was look at clinical UIs. And Microsoft actually sent a team with us to the Dominican Republic when we first started, and this was kind of the first time I'd gone from the theory into the practicality. And we'd take the NxOpinion and put it in front of doctors, and no one would tell them how to use it. And Microsoft team would watch and see where they got stuck. What I can tell you today is it takes less than an hour for a person to really use this software. Now, there's a lot of features that will take longer, you know, metrics and stuff, but it's a very, very user friendly. As said, we had become one of the top five finalists in the world in 2004 as the best medical solutions as nominated through the Computerworld awards. We've done some academic and clinical studies. I just put a couple of them there. We did Improvement of Diagnostic Performance of Indian General Practitioners. We found that general practitioners increase their diagnostic accuracy between 25 and 41 percent as compared to the specialist physician diagnosis. And that was in GI diseases. We of course ran it against some of the powerhouse, and we'll show you some of those studies. But we also ran it with who can use it so that you looked at use case of can [inaudible] use it, what can [inaudible] do, what are the limits of these people. Can some of them only gather data. Can some of them make diagnostics. What are the multipurpose health workers come in. Did you improve not only their data collection and integrity of the data, did you improve their diagnostic skills. So all of those tests have been done, and I won't bore you with those. But here's an example, for example, of side-by-side comparison of some of the diagnostic programs. And I'd like you to look at the line 4 where NxOpinion's about a 94 percent, and you take some of these like Epocrates and Isabel and they're around 83 to 84. But one of the most important things is that's if they put the data in, and you have to know the data to put in. If you spell it wrong, it doesn't get in. If you don't know what to enter, it doesn't get in. So even though you put all the data in side-by-side were more accurate. What's more important is because we ask for data from an end user, most of those programs will be around 60 percent accurate, where we'll still maintain our 90 percent or so, because we're asking questions, where those aren't. And you can put in multiple system symptoms. You could put in cephalgia and headache in most of them and get two facts. We will not. The way that we've develop it is that's one fact. We did some pilot studies. Well, many pilot studies. Dominican Republic, one of the things that we did is just threw it into a college student to see if they could gather information. The quote/unquote back was this was easier than falling out of bed. That was very important for us to be able to say can data collection, which is very important -- because, you know, in global medicine, we have to know who exists, you know. And getting data in, very easily done. In Congo we worked at it with various teaching hospitals and in Uttaranchal and various places, Andhra Pradesh, done a lot of the pilot studies initially to see if we were on. Looking at the current features. One of the things that I think is very important is that NxOpinion has a personal health record. Now, I'm going to use the term differently, personal health record, than most people. And I come from a behavioral medicine background, and so from that standpoint I'm going to make a statement that comes from that background. And I say, you know, personal health records, when you think of us as consumers, where you're getting data in, you're typing data in, I make the comment sometimes that says, you know, that data is probably not going to be entered unless you have a little bit of an OCD in you. If you have a little bit of obsessive compulsive disorder, you're going to keep your health record. Because why am I going to keep my health record? We use the phrase I will keep a health record if I can interact in it, if I can ask good questions. You take the technology of NxOpinion, you could go into your health record with more technology. But I'm saying the engine does this. I could say can I drink Red Bull, and it will go in and say you have hypertension, you're on this antihypertensive, it is not indicated -- it is contraindicated. No, you can't. Are you looking for energy? This software right now with the supporting systems can do that. That is a health record to me. That's an interactive personal health record. That is something that will be used universally, because I now have a reason to have that and interact with that. So we look at it and we say we needed to be able to access this health record from a cellular device. Our first cellular device, of course, has come through Microsoft UPG in the OneApp, and it is very, very cool. And that platform what it can do is really slick. And so that's something that will sometime around November be released. Personal PCs, obviously. But it also needs to be able to be incorporated into the public health and into the hospital information systems. We have an HIS dashboard that I'll show you that uses all of this sort of stuff. Your health record also has to have vital data registration and integration. To give you an example of how simple it is, let's say using the UPG phone OneApp, if you were data collecting and you went into a village, the first thing it might ask the rural health worker is a village questionnaire. You know, might be asking questions about water quality, it might be asking questions about airborne diseases. And that questionnaire is filled out on that phone one time. The next time you go and you enter in a patient and create a patient record, all that data is entered into their health record because that's important in making diagnoses. So the ability to be able to integrate. The ability to be able to go in and say if I have a mother with a history of diabetes and I create a son, that son has a family history of diabetes, I don't need to enter that in. So that we're looking at being able to get in so the data entry is so easy that we're going to get good data and vital statistics. The other thing important with this clinical decision support is it needs to be used with skilled or nonskilled user interfaces. We use three different levels. Level one is a person who really can only ask symptoms and only in a limited basis on a handful of diseases. And there are five major diseases, for example, that create a significant number of deaths around the world, and there are ways to be able to get a fairly high confidence level with symptoms. They might be a level 1. Level 2 says I have some rudimentary laboratory and imaging. I might have a chest x-ray and I have a few labs available. And I can ask those next-level questions. That might be an accredited nurse, midwife, or it might be a multipurpose health worker or something of that nature. And you need to be able to plug in various devices, blood pressure monitors, et cetera, so that you take away the training of the individual in various organizations or companies have those. And then you have level 3, which is highly sophisticated. I've got a specialist, I've got a lot of information. So each interface and each question has to be very different. Every template has to be different. For example, with the lowest level, you go through a chief complaint. The next level you might go I can go to disease and I can start entering my findings very quickly and then find out my predictive diagnostic. Various ways you can go in when you get up to the higher levels. You need to be able to customize management and treatment for each of those levels. And when you take a rural health worker and they can't -- they're away from a hospital and they can't have the access to the medication, then you have to give them various ways of stabilizing disease and transferring. We believe what we have provided is something that is extremely important. You now have a documented clinical supervision of rural health workers. Every event is documented. Their decision is documented. The findings that put in, the treatment they put in, the follow-up to their treatment is all documented electronically now in a very user-friendly fashion. And so now you to have for the first time physicians can look at these people and statistically find out in very easy reporting systems do they need more training, are they treating their diseases, are the diseases being treated properly, is there a resistance occurring. And all of that on data that right now is not captured. If it is, it's written down, but it's not able to be done on the fly. And it provides a standard of care. The monitoring and evaluation piece is very interesting because even on the mobile phone one of the really -- again, I keep talking about OneApp -- is that we're able to customize just our ability to monitor and evaluate. And it can be done on the fly with the forms and immediately uploaded into the phone. Just very simple. So you can do a monitoring and evaluation. One of the cool things is let's say you're in a village in this particular organization, perhaps like an organization like World Vision wants to know something as simple as is there electricity. You can add that to the form and gather that data and immediately get other information. And then it comes in and NxOpinion knowledge management can run stats. Obviously that data is important in diagnostics, necessarily, but it does allow us to do some monitoring and evaluation. We can monitor and evaluate not only follow-up with medications, but pregnancies. You can through, for example, the OneApp. When a child is born you can send out a note that says to the health worker see if the child is able to nurse. Is the child drinking water. You can ask these questions and if all the answers are hitting the standard that is appropriate, you got a green light and it just shows up green: Keep doing what you're doing. It shows up yellow, then what happens is immediately an alert is sent to a -- the next higher level supervisor that says we need intervention in this area or we need an immunization or whatever it is. All of that is programmable very easily on this system. One of the things that we found very early on is the diagnosis accuracy -- not only in U.S. medicine, there's various studies that say there's 7 to 20 percent misdiagnoses, but in the rural medicine is people make diagnoses based upon their level of knowledge. And so they allowed and make a diagnosis and it's considered a diagnosis and all of our public health systems count that as a diagnosis. What we've done is we said we need to be more granular. We need to go down to the finding. We need to gather the findings because the grouping of findings may validate a diagnosis or it may validate that this is not a diagnosis. This is not the appropriate diagnosis because they didn't have enough information to make that diagnosis. So now I think that we can do better early outbreak. We can tolerate misdiagnoses. We can look at various reporting functions because technology has now allowed us to get so granular and run these reports. One of the things using the architecture that we have is it provides a common medical data architecture. It doesn't matter how the data is stored. It allows us to be able to pull it in, interpret it, and dump it back into your system. Very important from that standpoint. It allows the whole broad data accessibility, whether you want to pick it up on a phone, whether that person goes through multiple levels of treatment, their records are still available on every event in which they were seen and can be reported back. We also believe in a transparency of data analysis. Our job is to be able to get data in. Our job is to share that data in the public health system, to allow other people to be able to go in and figure out what solutions exist. And so we are very much in the public health data is so important to be available to all and be transparent. The regionalization of NxOpinion was one of the greatest challenges we had. And so I think that what we looked at is we took global and regional. And what I -- I refer to it as what I call a fallback system. India, for example, has some diseases that we can overall say these diseases exist or Sub-Saharan Africa, but then India may be broken down into regions that have winters -or mountains with winters, mountains without winters, coastal and plains. So as we gather data and as we learn, we can go more granular on our prevalence and sensitivities, but it always falls back to what we know. So that that data is not yet there, it goes back to what is known. And then as we learn, we can add the data in. And so it uses that ability to regionalize. What that means is, for example, the creative group at UPG, we can do a data collection in a new region in a matter of hours. Takes hours to be able to take it into another area. To be able to do the diagnostics and reports is a matter of just days to be able to convert it regionally. Now, to do the full diagnostic in a specialty language takes a lot longer. But we have the ability to say that this finding shows up in this fashion, and you can put it into a different language. The database doesn't care the language; it cares that that finding get filled into that field. So it's very interesting. And if you use voice-directed technology, not voice recognition, you can ask questions in a native tongue in a very fast method using NxOpinion's clinical decision support system. One of the other challenges that you have in global medicine is publish set. Everybody wants to own their data and nobody wants to share with anyone else. Not only do countries do that, but organizations do that. And sometimes even within organizations they won't share from one department to the other. So one of the things that we decided to do was to create what's called published set. You can go into our database and create a subset of data that says for this organization for this location, this is the data that is published in as your database. We keep a mirrored D-identified image. They can own that data. They can use it for financial support. They can do whatever they want with it. We with the mirrored data can run all of the statistics, all of the public health information and can pinpoint it back to an individual. They can connect it to the individual. So by creating this we've allowed people to say we don't want your data, we don't want to make the money off of the database; we want a global database that is what is necessary to really do public health. So they own it. They control the data. We manage the data in a D-identified. As I talked about before, regionalization has a user skill level and it has a structure of being able to fall back. Let's talk about NxOpinion as a platform. I've given you a little technology. Let's start moving into what do you do with this. So as a platform we look at four different areas. There's -- yes, sir. >>: What's the scalability of what you've built? >> Joel Robertson: That's exactly where we're going. Exactly. Because that is a huge challenge, scalability. And that's exactly where we're going. >>: I'd like to know what are some of the interfaces to organization to prevent, and the third part is what about -- I mean, you mentioned briefly that if someone's collecting imaging data, that that can prove a whole set of [inaudible] ->> Joel Robertson: That's right. >>: Right? Are you looking just at x-rays? Are you looking at just being able to take visual example with the cell phone camera or being able to do ultrasound or ->> Joel Robertson: Yeah. There's -- again, based upon the interface and the level of the program being used, cell phone or PC. For example, if you're looking at a cell phone and you have an x-ray finding, you'll only have the result showing. If you go into our HIS system, you're going to see the x-ray come up, and you're going to see what that looks like. So it goes through. And you're absolutely right. With the cell phone, you frequently can take a picture, get that data. There are some -- some areas where they're beginning to look at automatic diagnostics through the imaging. But it can always go back to the PC -- the PHC or higher and get it and put it back in. So you're able to transmit. And the OneApp is really cool. That's one of the things we liked, is we added imaging to that piece that they had not at UPG previously -- I mean, picture imaging, for that various reason. Dermatological. It's a great way to do that in rural areas. So -- and I'll show you, for example, in our HIS system you hover over a finding, it will pull up an x-ray right in front of you or whatever imaging you need. But scalability is a huge issue. Do we have an answer to all the problems? The answer is no. And one of them is, has to do with why I'm here, why I'm interested to be here, is there's still problems to be solved, and that needs to be solved by some people brighter than I am in those areas. And so what we're looking at is saying this is our domain knowledge, but there are other domain knowledges that have to build in this if we are going to change world medicine. So we aren't the solution; we are part of a solution. So like you said here, we've got a whole data collection system that you have to talk about. Data interpretation, data storage and data reporting. Here is the platform that we have built. I want you to think of this. You've got this global databases of all the aggregated data. You've got a personal health record which whoever wants to own the her. The platform that says the diagnostic engine, event engine and search engine, that's what we do. The diagnostic engine is I'm going to help to diagnose a disease. An event engine is I'm going to take such things as follow ups, I'm going to ask for alerts, I'm going to ask for various events. And the search engine has the ability to be able to go in and get the data all the way down into granular basis. Out of that we have the monitoring and evaluation. Messaging engine, UPG, for example, with the OneApp, is a messaging engine. That's using that. There may be an SMS methodology that may come out of there. There maybe various aspects that come in. We look at ourselves as a platform. We say other people build. You can build some interfaces on us. We've got this piece put together that is a huge effort. And how it's built and how it works in. Monitoring and evaluation is another area we're weak at. We need lots of people to be involved in that. Epidemiological studies. So I think the whole ability to dump into this system and to use this system to come back out, there are some aspects. >>: So if I understand your vision correctly, it's almost like you're making a case for publishing KPIs to your [inaudible] ->> Joel Robertson: Absolutely. >>: Being pretty open. >> Joel Robertson: Absolutely. >>: [inaudible] business as well as a technical problem. But essentially you're talking about an open source. >> Joel Robertson: I'm talking ->>: [inaudible] >> Joel Robertson: -- you're absolutely right. Now, it's very interesting because what's fascinating is -- like I said, we have a 501(c)(3) and we have a business model. And essentially what we're saying is, you know, we're not going to give away our code to our engine, but we are really building on the open source concept. We are really building on saying, you know, this is what we want to do and we're going to do this. We're going to give you a lot of APIs to be able to commit. We have to do that. This -- you know, that's why we kind of used open MRS as our first mapping. Our NxKM database is actually -- can do more and has got a lot of better things. But you know what? It makes it proprietary. But if you got open MRS, you can come in and do that sort of stuff and get to us. So, you know, we really are looking at that aspect. So you're absolutely right that we felt, again, when Eric described my mission, my mission is not to build a proprietary database diagnostic system; it's to build a platform that other people can use because unless we all work together, unless it is open to people who are brighter than I am and our company is in other areas, the ultimate solution is saving lives. And that's the only way you can do it. >>: Do you have an STK, then? >> Joel Robertson: Not yet. We aren't at that point. But you're absolutely -- that's our next -- we are close. December maybe. That's where we're at right now. This is just the Cojax. This is the OneApp application which is out of the UPG. And it's just various ways of -- there's two elements to what we've built with UPG. One is called a health manager, which is accessible by consumers to be able to go in on a cellular platform, to be able to get regionalized diagnostic data, health tips, all of that, without using the term -- we're kind of the UPG's WebMD of developing countries. So that's like a November release. And we'll be looking in South Africa and India. And that's our role. We've got 15 countries or so that we're geared towards. And that's the consumer base. >>: [inaudible] >> Joel Robertson: We're running all the back end. Yes. So when you go to OneApp, you'll see whatever the name is, but Robertson Health Manager or NxOpinion Health Manager, et cetera. And then in the back end we're doing all the health tips that are regionalized, we're doing the diagnostics, et cetera, storing the data. The other part of that is data collection, which we look at with a lot of humanitarian efforts. Gathering data at the HEW level, pulling data in using the cell phone. And so those two components are both ready to come out of -- in November. >>: [inaudible] who are your carrier partners? >> Joel Robertson: What was that? >>: Who are your carriers? >>: Operators. Telecom. Telecom operators. >> Joel Robertson: I'm not -[multiple people speaking at once] >> Joel Robertson: Oh, yeah, yeah, yeah, yeah, yeah. You know, like I think Microsoft has relationships with blue label telecom and Oxygen in India. We also are going to be looking at independently we can work with companies such as Tata in India and all of those things from a distribution standpoint. So company based, corporate based, as well as we've got distributors. We've got some meetings on humanitarian groups to release and use it also. Sorry. This is the Microsoft OneApp, just kind of a -- this is a mid-level phone. And just kind of the way that it might work out where you just can kind of go in and click a few buttons and put a few things in and the data comes in. One of the things that we thought was kind of an interesting challenge is in dealing with rural India, for example, names, you know, are rather interesting because you have some very difficult names to put in on a cell phone. One of the simple solutions is you take a paper, write your name on, hold it up, snap a picture, and a person that's on a keyboard can enter the data in. You've already got a unique identifier with it. So there's many different ways that you can find some practical solutions to be able to get some of the problems. >>: But what exactly [inaudible] one of the biggest problems for healthcare, particularly in India, is that the government doesn't even know what their population count is. So how do you make sure that everybody is uniquely identified? >> Joel Robertson: What we have been able to do with our NxOpinion is to have an identifier that is associated -- if it is a particular ASHA or health worker that has a phone, that phone is connected to a region and to a village. And so what you have is a beginning of a numerical solution to this is the village, this is the state, this is the region. That numerical, which they don't have to enter because it's automatically done, and then when you create the record you have an additional unique identifier. That unique identifier can be mapped to a national ID or anything. Because they don't have it, we create it and will map it to however they want to do it, so that we have had to create the unique identifiers that can be mapped. You're absolutely right. >>: [inaudible] >> Joel Robertson: Yeah. And we do work with the governments as far as seeing if -what's their -- do they have a plan or what is the plan, et cetera. This is an example of data gathering. Now, this is just the pilot study. Now, don't think of this as the user interface, but this is kind of how NxOpinion works. On the left side is where you have patient data. On the right side is general reference. So that means anytime I go in and do a search, where NxOpinion is very different than your typical WebMD Mayo Clinic searches, is if I am on the right -- left side and I am a 52-year-old male with a history of diabetes, et cetera, and I go search the database, you won't find me able to get diseases that are only common to females. You won't find me finding diseases that are unique to pediatrics. It filters your findings based upon what is existent in your health record. On the right side is generic. I can go in and find out anything. So that's part of the search process. You have some various ways of getting in epidemiologically. Then what we have on even at the PHC is you have what's called quick picks. On the right-hand side here, you'll see at the bottom you have all the different diseases, and it's either do I have it or does my family have it. And it has a smart search. But on the top are quick picks. Quick picks are what are most likely to exist in your area the most common diseases. And it's a user usability interface. So that what comes up is it's very much -- they can stay up in that top area most of the time, click it in, and it's very much of a user. And that -- usability issue. And that is part of our regionalization. You can go through and all of the findings, whether there's symptoms, whether there's labs, whether their imaging and all the various things, you not only have a description but you have a link that they may have with other diseases and/or findings. Now, data interpretation. I'm going to just show you a couple of screen shots of how NxOpinion works. And, again, this is the research tool, not the user interface. On the phone it works different, et cetera. Here is a series of symptoms were entered, and at the bottom you'll see that there are 11 pieces of information and evidence. And those 11 pieces, you know, being male or female, age, are pieces of evidence. And various aspects, former history, whether you take -- whether you've had an immunization or not is in evidence. But all of a sudden we've run in some findings through a chief complaint and we have -whoops, I skipped a slide, didn't I? No. Well, we have a mis-slide. What you'll see here, though, is I've got malaria and I have a confidence level. Now, with the end user level 1 and 2, that's a dot: green, yellow, or red. We don't give them the number. With a more sophisticated user, we may give the number. So that it comes this way. A green disease says the group of physicians feel at your level you're qualified to treat with the finding or the treatment that's here. For a disease to reach green, it has to be you're qualified to treat, we're confident you have it, then you get a green. If you're qualified to treat, we're not confident you have it, you have a yellow, which says you need more information. Over here on the right is what are the different information. Those are suggestions that help us differentiate your disease. If it is a yellow disease, meaning you aren't qualified to treat at your level but we are confident, then you have a different statement: It's likely you have this but you need to refer the person. A red disease says: There's no way, get him in for help immediately. So we're able to use green, yellow, red not only on diagnostics but on alerts, on follow-ups, on all of those various things. >>: Did you say there were two kinds of yellows? >> Joel Robertson: Yes. Yellow is I'm not qualified but I'm confident I have it. Or I'm -- it's you're not qualified and I don't know if you have it. >>: You are qualified. >> Joel Robertson: Or you are qualified but I don't know that you have it. Both of those would be yellows. Both would have a different statement. There is a qualifying statement that says, you know, you probably have this disease, but you aren't qualified, you need to have somebody treat it. That's a good telemedicine call type of disease. This, going back to your imaging, is a dashboard that we created that integrates with any HIS system. Now, we've created a clinical dashboard in which to put in perspective our phrase that we use is doctors should be information interpreters, not information gatherers. You need the data when you want it, where you want it, and how you want it. So we have a dashboard where basically using this engine, if you are an infectious disease doctor, there will be a template that comes out and it says most infectious disease doctors would like to know platelet counts -- I mean WBC counts, temperature charts, and whatever it is, this is what you want. NxOpinion, when you go in and pull up that -- your patient, it will show you your graphs of temperature, it will show you the medications, it will show you any alerts of anything. This is a chest x-ray. And if you hover over it, it shows the chest x-ray. You don't have to go get it. It shows up. What's most important is let's say you prescribe a gentamicin, for example, for an infection and we know that it has a high side effect of renal disease. NxOpinion in a suggestion will say you might want to run a creatinine or a BUN. So this is something that will be ready in January. But it goes back to that core technology of different use. This is kind of a Ferrari where the OneApp is a weed whacker. Although it's a pretty powerful weed whacker. So, again, using it from that standpoint. You'll also notice that in the corner or in the right-hand side -- and these are Widgets, so you can move them around wherever you want them, you're going to see NxOpinion setting there with the probability of the disease. What we feel is it needs to be sitting in the background -- this is mostly a hospital information system; this is at a secondary hospital or things of that nature -- is it should be running in the background. It shouldn't be interfering. But if all of a sudden you look and you say I've only got an 80 percent diagnosis, I'm ready to go, you can go over there and it will say here's some ways to get you up to 90 percent. It will also be able to go -- you could say I want to run a CAT scan or an MRI, and it will give you which one is most likely to add to the diagnosis. And in our version 2, it not only will say in that hospital how long will it take you to get a CAT scan run and an MRI, but what are the different cost effects, or the different effects of that cost. So you might be able to get something done in two days less expensive and it adds almost the amount of diagnostics to it. So it has a little bit of workflow to it. That will be ready in June of next year. NxOpinion is a product. You know, I really look at NxOpinion, one, as a plug-in to HIS systems. Once again, to steal a phrase from another company, BASF, "we don't do anything ourselves; we make other people better," that's kind of what we see ourselves. We look at, for example, there be Almalga HIS system and say, you know, we can make you better. There may be another system out there, we can make you better. Because what we say is this is our particular strength. This is what we do. And we want to stay within that realm, but we want others to work with us. We especially have spent a significant amount of time into the rural and remote healthcare. Our passion is to take technology which is essential to take quality medicine into the rural area. And by using NxOpinion you have clinical supervision, you have standard of care, you have diagnosis in the area of the nonphysician yet you have the supervision of that. You have a treatment standard. You have follow-up, you know it's done, and you've monitoring and evaluation and the ability to provide alerts. Many of the challenges we that we've had -- my next slide is going to talk about the challenges we haven't met -- but at this particular point. It allows us to do telemedicine. Yes, sir. >>: So are you just running one instance of this gigantic thing that everybody feeds into? >> Joel Robertson: No, no. >>: Or lots of instances? And what kind of business relationship do [inaudible]? >> Joel Robertson: We are in the process of -- you know, you asked the question are we running one instance, and the answer is in each of these pilot areas, the answer is yes. The technology that we're doing now is moving that out of that. So that will be I think October is coming out of that system. We, again, philosophically believe the cloud is the answer. There are some hesitancies with various organization groups, et cetera. So in that case you run instances on their local server and then put it back into the cloud on a D-identified data. So our business model, that's what we look at. That's also what we look to Microsoft to say, okay, we've done this, help us with this problem. That is a problem that we come and say we need you to answer that. You know, one of the reasons why I'm really pleased to be here is there are problems that are above our skills. That's one of them. What is the infrastructure to support this realtime? And I think it's a great question. So when we're ready to roll right now, we have to do it as we can at this particular point. But if we're going to do a really worldwide global database, where does HealthVault fit, where does Amalga fit with -- you know, and I'm going to go into that. Where do we go with the infrastructure? >>: So currently someone would just license the shrink wrap [inaudible]. >> Joel Robertson: Absolutely. Absolutely. >>: And you do it as a nonprofit. >> Joel Robertson: It depends on who it is. You know, for example, an area in which we have gained a tremendous amount of demand of small hospitals in the U.S. for an HIS system because they can't afford them. And it's very affordable. That comes through our profit system. India hospitals are very much liking us because we have a regionalized clinical decision support. That's out of the profit. On the other hand, if, you know -- sometimes even go government contract can be profit, depending. But if you're setting there in a -- especially Sub-Saharan Africa or in areas where there's just -- you know, certain states in India, Chhattisgarh, there's no money. So then we run it through our nonprofit. But it is our ability for us to be sustainable to say nonprofit, profit, various things of that nature. I believe strongly that we -- the next movement, I personally would like to see that because I have a relationship pretty heavily for years with Microsoft to take something such as HealthVault and make an interactive personal health record. The technology exists. Google's playing with it with predictive. They've got one guy running around doing some stuff. We are so far beyond that if we are ready at this particular point for medicine to go to interactive personal health records, that is an extraordinarily, I believe, important aspect. Also case management resource utilization. We've been contacted by insurance companies that are saying can you keep healthcare costs down. Well, obviously, if you can pinpoint a specific test to be run to help you differentiate versus a battery of tests, that helps. Various ways of dealing with that. Here you go. Scaling. Big problem. This comes from a book that Kristin Tull [phonetic], David Heckerman, Delda Hart [phonetic], a chapter that Kristin Tull, David Heckerman, Delda Hart and myself wrote. And it's called "How Do You Scale Healthcare." How do you take this and scale it. Well, I'm going to give this to -- this phraseology to Kristin, because Kristin created this. We call it scale up. How do you take medical databases that are hundreds of thousands and turn it into millions. Problem No. 1. I don't have a solution necessarily. I think you guys do. Problem No. 2 is scale out. How do you regionalize to multiple country and regions with language, regulations, organizations. We have a good solution there. Scale across. Multiple diseases and programs, how do you coordinate HIV, you know, diarrheal diseases, maternal care with the Gates Foundation, et cetera. We've got the technology. It's a matter of those joint cooperations to put those programs in. Scale in. One of the things that I think is extremely important about what we do is being able to get down to a granular level. I can tell you having a great deal of experience in rural health, if you can't scale in and you can't get down to that patient finding, you're just making decisions on inaccurate data without any validity. We need to validate the data we're making decisions on. So when I look at that, I say one of the major problems that we see is this scaling up. The scaling up. It's an infrastructure system that's far beyond our small company. It's an infrastructure that I think ties in several companies and several solutions. And that is one of the challenges. We can run this and we can run it to a point, and then we will stall without that scaling up. What I have here is a global database and some touch points and where we look at Microsoft, for example, and we look at other partners and say where -- what can we do and where do we need help. You have this user and then you do have a method of data collection. OneApp, for example, is a great data collection system. I think OneApp is fantastic. I love the way it goes. It's limited from an SMS text standpoint. That means certain areas of the world can't use OneApp. So the question is where does that technology come from that's going to be necessary. I can argue -- I can argue the OneApp platform as being a superior platform for data collection. But if you don't have the infrastructure to use that mobile phone system, you can argue all day that it's superior, but if it doesn't have the ability to work in that system, then you need another solution. So there's an area where we need a solution. By using the API system that we have, many people can use data collection systems in through our system. So we say, you know, we've created a foundation; let's get other people to be able to do that. We can build for them on OneApp. We can build for them, et cetera. But there are other people that need to be able to look at that. Data storage, I tell yeah, this is a nightmare for me. This is an absolute nightmare from the standpoint of we can gather data. We can store data. We have to deal with the regulatory and most important we have to deal with the global aspect of this. Has anyone ever tried a database system that could potentially be this large and manage it. So I take that one and I say you know what? I can brainstorm with you. I have no solutions because you are far over my head on that one. I do not know how to do that. And that is our weight limiting step at this particular point. Data interpretation. You know, I look at Amalga. And I'm intrigued by that technology, for example. And I think of what we do and then with that, the Amalga system, that there's some ability to be doing even greater data interpretation. Especially because they have similar languages, probably because we have a couple of guys that are pretty heavy in that area. Data feedback. I think not only gathering data but going back and reporting data back and being able to use that, whether it's through the mobile OneApp platform or whether it's through a HealthVault system. We are in a position that I want what our fall release is is I want to give you a new term. It's called push medicine. We believe in what's called push medicine. Pull medicine is -pull medicine is what people do now. They go to the Internet and they say I have diabetes; I want to find out about diabetes. Or they go and they say my doctor said I have this; I want to find out about it. Push medicine says if I take the database that NxOpinion has plus one of our companies called Robertson Wellness, which is another technology that we do, I can go in and say because you're diabetic, obese, and male, here's a new piece of information you need to read. I can push it to them. I can push information to them, only relevant information to them. To me, that is a whole new way of doing medicine. Pull medicine means I have to have something that asks me to go dig it out, whether it's all of a sudden I got a lab test back, I may be months late. Push medicine says I'm going to give you new information, pay attention to it. Here's a new way of doing it, here's new research, here's something that we know about your medication. It's a whole new way of doing it. NxOpinion can do that. But it needs the data feedback system. OneApp we will be doing that. Data learning. Again, I kind of look at the Amalga system in saying ooh. I salivate with taking the data that we have, our data learning, but Amalga's data learning, for example, or some system like that can really help us change medicine. And, again, the analytics that are provided from Amalga. You keep seeing -- hearing me say Amalga. I think it's the future. It's a semantic database. We're a semantic database. So there's a lot of fascinating things that I find. So when I kind of look at this in conclusion, and then I'll kind of open it up for a little discussion, you know, we don't have all of the answers. We have done a lion's share of and met some of the solutions and some of the challenges in dealing with global medicine. We now are in the position of saying the solution needs to deal with some very, very important infrastructure issues, with being able to work with other people who do things specifically better. But it is fascinating because from my perspective, and it's been with the help of Microsoft, for the first time, you know, you take -- Eric, how long have you been dealing with Bayes? I mean ->> Eric Horvitz: Since birth. >> Joel Robertson: Since birth. So 25 years, huh? >> Eric Horvitz: [inaudible] >> Joel Robertson: I mean, you take -- you take guys like Eric and you take guys like David that have just been digging into it, and it's kind of neat that -- you know, I feel that we're very fortunate to be able to take those pioneers and take their stuff and put it into a system that offers a solution. But we are only a piece to the puzzle. And I think one of the things you'll find about our attitude is we need help. We are not the solution. We are a piece to the puzzle. And from that it's very interesting that we look at and say this is what we're not. My father was very wise when he said to me know what you're good at and do it, and then find people who are better at things and ask them to do it. We need help in epidemiology. We need help in metrics. We need help in infrastructure. We need help in analytics. We can't be everything. And for that we need to have good, solid partners with research-valid information. Any questions? Eric. >> Eric Horvitz: [inaudible] catch up with David on the qualitative [inaudible]. >> Joel Robertson: CWP. >> Eric Horvitz: It would be kind of nice to take for a set of diseases [inaudible] go to the population and do some sort of qualitative or quantitative evaluation, whether it be an actual set of runs with real patients or you have a proxy database coming in from its reporting model in India or any place you want to -- New York City, and see how well the system is actually working. And that would be a really nice slide 17 to have in your presentation. >> Joel Robertson: We have three studies that I did not put in because I didn't want to -I, you know, again, trying to get vision, trying to get technology, et cetera. The first study that we did had to do with 300 patients in which we were looking at did we make general practitioners better and were we accurate. So that was the basis of that first study. The second study was what level of skill -- and this was a Gates Foundation-funded study, was what are the skill levels of the people who can even be involved in clinical decision support and who can be involved in data collection. And we were somewhat surprised that that is a major hurdle that some of the people that are the front-line workers don't read -- either they're -- can't write their own language and so therefore have a difficult time. So voice directed helps. If you can ask a question in their native language and they can answer yes or no. But that was a challenge. We have a partial solution. The third thing was to take 2,000 multipurpose health workers and look at two things. One is the collection skills of their present system versus does NxOpinion get what I refer to as valid data. Because I'm a stickler that says when you make conclusions on data that's subjective is a difficulty. So we looked at that in just two pieces of data collection capabilities. And the other thing is how did they rate with those 2,000 patients compared to the physician. And that study has not yet been published, but it's complete, correct? That's been complete. Those were done using Nizam Institute of Medical Sciences as our supervising. We went through all of the standards that are appropriate for India medical services and -- or medical society and all the protocols so that it could be, quote/unquote, a publishable article. And so from that standpoint we used all of those criteria on all three of them. But there's a lot more stuff. I mean, it's just continuous. Continuous. >>: Is the study on data collection that you did in India for the Gates Foundation, has that been published? >> Joel Robertson: They have not published it yet. We finished it in December. I have not seen it published. It's something that's available. We can give you a summary of it, I believe, without getting in trouble. >>: [inaudible]. >> Joel Robertson: Because we were somewhat surprised. When we first -- you know, of course, Microsoft owns -- I don't know if they own totally, but have a relationship with Tellme. And Tellme being voice recognition and we said, you know, that's not a -- that's a great application perhaps with HealthVault interactive personal health records, which is phenomenal, but voice directed, using our rules-based -- what's really cool is using our rules-based template to get predictors in, you can ask questions. You can set it up where somebody just has to ask the question and you answer yes or no before Bayes kicks in. Once Bayes kicks in, you have a series of questions that can be asked and answered. So it's really adaptable way of being able to get to those people who are illiterate ->>: [inaudible] along the same lines. >> Joel Robertson: Yeah. I've heard about that. And, you know, ->>: Is it only funded [inaudible]. >> Joel Robertson: Okay. See, this is -- this is where, you know, our philosophy, of course, is -- and this is what we do. But there are so many solutions that can build on this. And for us to try to find all of those things is ridiculous. We're just, again, kind of in the attitude of saying we're really -- we really think we have something cool here. It's been tested enough. It's got some solid theory and testing. And now it just needs to be made better by partners. >>: [inaudible] opportunity at the [inaudible] health summit for discussion and interaction in a lot of mobile health solutions from around the world. And so I do hope that you'll have an opportunity to come and participate as well. >> Joel Robertson: I was just invited to September 15th to a meeting in India by Intel just today. And ->>: Big notice. >> Joel Robertson: Yeah. And similar with the India public health system, and so that I will see if I can get there, because that -- it is a matter of cooperation. >>: I will not get there. That's too short of notice. >>: One of the things that we were talking about, too, is that when you use these kind of things, and you have developed some plug-and-play analytics for unskilled users [inaudible] or whatever. So even someone [inaudible] validated tools, validated data, and start doing some studies on their own, both publishable, because [inaudible] models that you know about. >> Joel Robertson: I think there's another thing, too. I know you guys at Microsoft Research had a maternal care database or -- I mean, those are the types of things that we say, you know, if we get to the point where people think what we do is legitimate, we need data. And grabbing that data and putting that in in a format that will be used, those are very important things that we look at from that standpoint. >>: [inaudible] scaling up [inaudible] but just from a business point of view, is it this -is it a way for people to sort of put sustainable businesses on the map? Right? So people have to start making money. If there isn't money being made or providing some services, then there's an incentive to sort of keep things up-to-date. >> Joel Robertson: You know ->>: [inaudible] >> Joel Robertson: The practicality of it is is it has to be sustainable and it has to be something that people work with. Blue Label Telecom, which is one of your ventures, I talked with one of the guys last week, some really creative ideas of how to use a micro business with this process. We have some strong corporate relationships in India that provide some of those things. So I think there's a whole strategy like you say of not only data collection but business management, maintaining a business model, infrastructure model. There's just so much that needs to wrap around it. You're fortunate that you're the first people that have heard our vision -- well, not the first people that hear our vision, but the first group from the standpoint of us saying this is where we're at now in our deployment and here are some of our problems. And I can't think of a better place to throw our problems than at Microsoft Research and see if you have any solutions. >> Eric Horvitz: I think we should stop there [inaudible]. Thanks very much. >> Joel Robertson: Great. Thank you. [applause]