 Okay. Good morning everybody. I am Tathagatha, one of the co-founders of Siktappal. We founded Siktappal back in 2015 with an idea of applying the latest advances in AI towards the healthcare domain. And incidentally neither me nor any of my co-founders had any background in healthcare at all. The last time I studied biology was in class 10. So it has been an interesting, if not difficult journey so far. So in today's talk I'll be presenting a few challenges that we have come across and how we have solved them and also the challenges that face healthcare industry as a whole in general. So our motto is data driven intelligence for healthcare. What we want to do is to make sure that every medical decision taken is guided by data and AI in the future. The outline of the talk is as follows. I will start with the challenges that face the healthcare industry today and then followed by what are the possible areas of application of AI and ML in the healthcare industry. And finally we'll give you a few glimpses of what we are doing at Siktappal in this area. The challenges and opportunities in healthcare today. Now healthcare as we know it today has several problems. The first problem is the lack of doctors. Now if you search for the number of doctors in India in Google, the results that come up are as follows. All of them actually talk about the lack of doctors in India. And the rural India houses 68 percent of the Indian population. Whereas only 10 percent of our doctor population lives in the rural areas, which makes it extremely difficult for people in rural areas to access quality healthcare. And this infographic, I don't know whether it's fully visible on the screen or not, gives you a doctor to patient ratio across the world. The redder areas are where the problem is more acute, whereas the greener areas where things are still better. As you can see, much of India is in the yellow and red zone and almost the whole of Africa is in the red zone. So the problem now is that we have an abysmal doctor to patient ratio and as a result, a lack of reach of quality healthcare. Now the only way to solve this is through technology. You can neither grow new doctors overnight nor can you force doctors to leave their comfort, their urban comfort and relocate to rural areas. This is a meme. I mean, no, no disrespect to doctors here. But at many times, what the job that doctors do is mentally taxing. And this leads to variability, inaccuracy and inefficiency, especially when it comes to diagnosis from visual medical data. Let us take a very common example of malaria. Now malaria is not very common in Bangalore, but if you just go outside, maybe to Bangalore or the coastal areas, it's very endemic over there. And till today, the gold standard for malaria detection or malaria treatment decision is manual microscopy. WHO mandates that a single sample, which is suspected to be of malaria be examined for 20 minutes by a pathologist. Now a typical lab in Bangalore gets around 300 samples a day suspected malaria cases. Now do you think they actually get to spend 20 minutes examining that sample? The answer is probably no. So again, technology can help this. And finally, it's a very simple problem. It's the cost. The quality medical equipment costs a lot of money, which means that even the biggest hospitals or the biggest hospital chains can only have the best instruments at the central locations only, which means again that the people at the peripheral centers do not have access, ready access to the latest technology. And this is again augmented by the problem that technology changes very fast, but upgrades are only possible through replacements. So if I bought an MRI machine today, and if a new version or a better version comes up tomorrow, then what will happen is that I have to throw away my existing MRI machine and buy a new machine. There is no simple way to upgrade. This is one of the challenges that we are trying to solve at Citadel with a special emphasis. So we have talked about the problems. Now this obviously means that there are a lot of opportunities for us to apply AI and ML into the health care industry. So what are the typical opportunities that come to mind? One is medical signal recognition and segmentation for disease detection, which I think Kiran is also going to talk about after this is that you get a medical image, maybe an MRI or a CT or a X ray, and you automatically detect whether this person has a tumor or a cancer, some kind of cancer has growth. The next is emergency medicine. Emergency medicine is a completely data driven discipline. The doctors say that I do not treat the patient. I look at the numbers, not the patient. Now machine is much better at looking at numbers than a human being. So this is actually a area where AI can make a lot of difference. The next is holistic health care. Now when we go for a treatment, there are multiple tests that are done. I go to the laboratory to get my blood test done. Then I might be going for an ECG, maybe a chest X-ray or whatever. And then all these data needs to be collated together to make a final diagnosis for other decide the course of treatment. Again, AI can help in a big way in this to collect data from multiple sources and suggest the best course of treatment. IBM Watson is talking about this in a big way these days. Drug discovery drug discovery has a lot of experimentation. They try out with different molecules. There are first of all, there is a text based data, which is analyzed. Then there is an image based data as well where they actually take thousands of molecules and see how they react with the different kinds of cells. And this is again done through high, what is called as a high content screening. Due to the manual nature of the process, they can only sample parts of the choices that they make. AI based solution can actually help it make it much more comprehensive and bring down the cost of drug discovery. And this list is ever growing every every day. We make new advances in healthcare. There's a potential application of AI into it. Now today I'm going to talk about mostly about medical signal analysis and now what is a medical signal? The medical signals come in multiple shapes and forms. So we have signals from ECG, which is a one dimensional signal to a CT MRI or an OCT, which is a three dimensional and people are talking about four dimensional images as well where you're looking at a fetal ultrasound. And this interpretation of signals is very, very subjective. It varies a lot from one doctor to another. We have seen that the inter-observer variability between what is normal and what is abnormal or what type of cell this is to what other type of cell this is. If you give it to two doctors or three doctors, the inter-observer variability can be as high as 30%. And AI can, so therefore AI can help standardize the abnormality detection process. It can enable even a junior doctor to actually diagnose or treat patients with the same level of expertise at, as a senior doctor. In fact, I myself have seen this happening in a lab where a junior pathologist will come running to the senior saying, ma'am, I found a cell which looks like a cancerous cell to me in this blood sample. Can you please come and come and have a look? Now the junior pathologist at the central hub of a laboratory has that luxury, but a junior pathologist sitting at a rural hand center or a rural hospital does not have that luxury. This is where AI can help. So I'm going to talk about a few advances in AI towards the medical diagnosis, medical signal analysis in the recent years now. So we'll start with ECG. Now ECG is a, on the face of it, it's a very simple thing. There are a few lines which go, there are patterns and abnormalities in there and it has, it is a very effective and not only in terms of detecting heart failures, but also in terms of cost and it has been perfected over decades. Yet the automated algorithms that have been built so far for ECG signal analysis fall movefully short. The G ECG machine or many of the ECG machines actually come with some inbuilt indicators. They say, okay, this might be a heart attack and all, but the false positive rates are so high that the doctors do not pay any attention to it at all. But this has changed recently. Now the Stanford, there's a ML group at Stanford came up with an ECG analysis platform, which actually they claim is detecting arrhythmia, which is a type of early indicators of heart attacks with the level of competence of a pathologist. What they have done is they have actually used a single lead ECG. So when we normally do an ECG, we have 12 leads fitting into different parts of our body, but they have used a single lead ECG, which could be a wearable device as well and use that to detect potential chances of arrhythmia. Now this actually opens up a new, a new avenue altogether. Now I might be wearing a fitness tracker on my hand and if that keeps sinking my heartbeat data to a cloud, I can actually say that, okay, this person is at risk of heart attack in the next 48 hours. And they have used, they have some 64,000 samples. They could collect it through this wearable device and they trained a 32 layer deep neural network. There is another company called cardiologs in France, which has recently got FDA approval for their ECG analysis algorithm. They have trained on some 200,000 samples. The problem is that both of these solutions, but target only one particular type of disease, which is arrhythmia. So what is missing still till now is that a comprehensive ECG analysis platform, which can detect all types of diseases. Next, I will come to hematology. Hematology is the basically the study of blood cells. So this is what a pathologist sees in a blood, blood smear slide under a microscope. The, the blue things are called white blood cells, which give us immunity. The others are the red blood cells and the small blue dots over here are called platelets, which help in clotting of the blood. What they do is by looking at this, these pictures is they classify these white blood cells into different types, neutrophils, lymphocytes, monocytes, eosinophils and so on. They look for abnormal morphologies. When a disease occurs, these white blood cells undergo more morphological changes, which is indicative of the diseases. And they also look for parasitic infection like malaria in the white blood cells. Now the question is, can AI automate the process? The answer is yes. There, there are actually quite a few solutions for hematology and automation in the market today. We have companies like Medica, Cellavision, Vision Hema and there's a machine which Roche is also coming out with, which actually automate, automatically detect, go through these images and come up with decisions. We have our own solution called Shonith, which we are working on and I'm going to talk about it next. So what they do is basically they have an automated microscope, which takes pictures of the blood from multiple areas in the sample, and then they have an AI or NN based algorithm to detect and classify the different cells. But classification of blood cells is a phenomenally tough problem. I mean, there is so much of interobserver variability that you can hardly be sure. And there is also no public data set on which kind of, I would say, comparison can be made of different solutions. And all of these have varying degrees of accuracy. And most of these are actually pretty costly as well. For example, the Cellavision machine costs, I think close to a crore, which is makes it difficult for most laboratories in India to actually adopt that. And therefore it is yet to these solutions are yet to see white widespread adoption in countries like India. There are several more applications. I just rushed through them. One is detection of diabetic retinopathy from fundus images. Now, diabetes is becoming a more and more critical problem in our society today, mainly due to lifestyle and sometimes due to some other these reasons as well. And the diabetics affects the eye in a big way. It is one of the most common causes for blindness and which is also preventable with proper treatment. Now, Google recently came up with this paper where they actually trained an algorithm with 128,000 images and came up with a F score of some 0.95 towards detecting diabetic retinopathy versus normal. Baidu went one step ahead. They said, Okay, I will not only say whether image is affected with retinopathy or not. I'll also point out which areas of that image are showing the changes due to retinopathy, which makes it more explainable to the doctor as to what the algorithm is saying. This is a 2017 paper itself and there is a company in China called he'll go who are also working on the same problems. Other applications include cancer detection, histopathology images. In fact, this paper on from 2013 is what prompted a lot of companies in this world to actually take on this challenge of applying AI towards the medical domain. These people were not doctors by profession. They were just software engineers or AI researchers, but they took on this problem and they came up with a algorithm and a solution which were actually competing with the level of accuracy of an expert human. So in six couple, I'll just rush through the remaining of it. AI plus what we are trying to do is building smart screening solutions which address the health, the base of the healthcare pyramid, which means that we want to target the tests which are voluminous and which differentiates between a healthy individual versus a diseased individual. We have this platform called Mantana, which is a continuously learning AI platform. It takes in visual medical data of all types, which are ex annotated by your experts. In some cases, the visual medical data is already available. For example, in the diabetic funder scans or x-rays in certain other cases like clinical pathology, the digital data itself is not available. So to counter that problem, we have built our own smart microscope, which is a as you can see from the image. It's a complete sugar has a cell phone camera as as the eye of the microscope and it has a robotic attachment which automates the movement of the sample under the lens. And our first product is called shown it. It is a platform for automated analysis of peripheral blood smear. So the way it works is this. The blood smear slide, which is prepared as a normal course of action in all laboratories is scanned by the smart microscope and the images are uploaded to Montana in the cloud. The month are then actually analyzes these images. It takes out the individual cells, classify them and comes up with a host of statistical parameters about this blood. And the report is now available to the reviewing doctor anywhere in the world on any web connected device, which means that the doctor could be sitting here in Bangalore and the patient could be in Hoobli, but still the diagnosis can go out, which was not possible till now. And after review by the doctor, the report is sent to the patient. This is a sample of the report. We have actually classified some of the cells and group them together. These are called lymphocytes and these are the statistical parameters like what is the cell ratio, cell diameter, what is the cell pallor ratio and so on and so forth. And I'll just briefly talk about a problem that is interesting problem that we are solving. It's I call it the camel problem. Now if I look at this image, which is the camel, the black ones or the white ones? It's actually the white ones. This is taken from the top and these are the shadows of the camels. Now, let us try to see if you can estimate the height and weight of each camel. It's possible if you know what is the spatial resolution. If you know where the sun was, I can do some trigonometry and figure out what is the height of the camel and and I can probably figure out what is the ratio of height to weight and all and do that. What we have is a picture like this. These are real blood cells and what I want to do is find out what is the mean what volume of the cell? What is the mean hemoglobin content of the cell and how many of these cells are there in one microliter of blood? How is that possible? These are two dimension. These are not even though they look circular. They are actually not spheres. They're toroidal in shape. How do you estimate the 3d parameters from 2d images? This is one of the problems that only one company to my to the best of my knowledge is working on in the world, which is Rosh and we are the seconds. So I'll be happy to talk about it later. And we have multiple other solutions based on Mantana. One is Adi, which is the solution for semen analysis. Shrava, which is the solution for urine analysis. Vaksha for chest x-rays, Drishti for retinal scans and Shonet I've already talked about. Thank you. Questions. We're almost about out of time. One question, two questions, maybe. Hi, thanks. So this is an area where when someone's nailed it, it's going to be highly scalable, right? But when someone has nailed a problem in any medical diagnosis, it's going to be highly scalable and providing it crosses enough of an accuracy. It's something that people will probably take up globally. So when you look at these companies or the examples you gave like the Stanford ML project, they seem very laser focused on one particular problem. Yes. Right. And getting that problem, right? Right. And it's going to be winners going to take all at the end of the day. Why are you as a company doing so much? Yeah, it's basically to leverage the power of AI. Now, what is AI? Essentially, you feed it data. If you have enough data, building an algorithm or a building a solution on top of it is not that difficult. So we have been actually leverage able to leverage the power of our platform and the data sources that we have access to to be able to simultaneously work on multiple problems. And in the at the end of it, when you're bringing a solution to a laboratory or a hospital, it is always more appealing if you actually solve a broader range of problems than rather solving one particular problem with a laser focus. We're about to we're out of time. So please continue. He's available.