AI-Based Environmentally Sustainable Cardiovascular Disease Diagnostic

All right, so hello everyone, My name is Sahiti Cherakuri and I'm a current high school sophomore at Santa Clara High School in California. And today I'm going to be presenting a research project I did this past year with a partner called AI based Environmentally Sustainable Cardiovascular Disease Diagnostic. This project actually won the gold medal at the International Genius Olympiad Fair. Starting off with a quick abstract, essentially what we were doing is that we were creating a reusable and household cardiovascular disease diagnostic. Essentially, CVD's or cardiovascular diseases are the number one cause of death, causing 32% of death. They also cause a huge amount of waste in the medical sphere. We want to create a reusable and environmentally friendly medical solution for this issue. We ended up with 83.8% diagnostic accuracy and a 93.3% overall functionality. A quick introduction to the topic. Firstly, every day an estimated 40,000 tons of medical waste are created globally. This is a huge, huge number. Cardiovascular diseases being the number one cause of death, leading to 32% of all deaths is a huge contributor to this. Now, why are there so many deaths in cardiovascular diseases? Well, firstly, medical practitioners are sometimes inaccessible and expensive to some. The average medical practitioner has a CVD diagnostic rate below 60%. This is just because humans have a hard time identifying the patterns which differentiate various CVD's. This can lead to delayed diagnoses and incorrect diagnoses, or even a lack of diagnoses. The ensuing result is more ICU beds, more disaster, essentially. So there's 138kg of CO2 emissions for every ICU bed and another 15kg of solid waste that are created. Most current CVD diagnostics used by medical practitioners are single use and so they take a high amount of vertical waste already. For example, if we think about the electrodes in ECGs or the syringes or needles that are used for administering blood tests, many of them are patient independent and also expensive. Now, talking a little bit about the current artificial intelligence models, they do show some promise in accurately diagnosing diseases. However, there's a stigma surrounding AI diagnoses in the medical world. And that is just because the AI just doesn't replace the doctor in that there's no support really that comes post, post diagnosis and it's often just like one diagnosis and that's the end. And sometimes these aren't even readily available AI devices like the software will exist, but it's just not found in your local vicinity. So I'm going to do a quick comparative analysis over here. Some key points to take away are that many current solutions are not created to be environmentally sustainable and the few that are often don't use AI. So they're generally done by real doctors who again have the issue of identifying efficient sounds. If there is support, it's not found in the artificial intelligence models. There's really no post diagnostic support or help or next steps or anything like that. So the solution that we proposed is called Telemedicine by Auscultation and it is an environmentally friendly and efficient auscultative audio based CVD diagnosis system. This is interactive and easy to use with explicit instructions and mental health and next steps are supported and offered to users post their diagnosis. So here I just summed it up real quick. Telemedicine by Auscultation is an accessible and interactive artificial intelligence based system that identifies and diagnoses the unique murmurs found in cardiovascular diseases early on reducing the trillions of kilograms of medical waste created globally. Some of the primary features are that firstly it is eco friendly, so we use very minimal hardware as you'll see later on. And we also had efficient power usage so that no power was being wasted. We had real time cloud based detection. Our model lives in the cloud and users can essentially upload that data to the model and then get a return. And this data is not shared anywhere else on the cloud. Like I mentioned earlier, we also have mental health support and next steps via a web app. So post detection there's mental health relief and what to do next and who to contact and all of these things are provided to patients. I want to talk a little bit about some of the background that exists. So I mentioned earlier how medical waste in landfills is detrimental. For example, you have a soil and water pollution and release of biogases like CO2 and methane and habitat loss. But surprisingly, the toxins found in these landfills can again cause heart diseases. So it's this continuous cycle of creating waste and then human death and like disease and all these things and we really need to put a stop to it and is what we saw. In addition, we have the second picture and this is of different heartbeats. So a normal heartbeat generally goes lub dub, lub dub. However, in some of these conditions, like aortic stenosis, you have lub, then you have some noise and then dub. So if we can train our AI model to identify each of these unique sounds, then essentially it will be able to differentiate these heartbeats. For this experiment we chose to have the normal heartbeat, aortic stenosis, mitral regurgitation and pulmonary stenosis. However, this idea can be applied for so much more, from respiratory to gastrointestinal to just so many different conditions. Talking a little bit about an LSTM autoencoder, this is actually the architecture that we ended up using. And LSTM autoencoders are just, they're amazing for time series or audio data. And autoencoders specifically are really good for looking at that anomaly detection side. So this was essentially the optimal model for us to choose. Talking a little bit about our hardware methods, so here I included a 2D and design as well as our final prototype. Some things to point out is that our, our microprocessor was a raspberry PI 4B and this actually comes with an inbuilt Wi Fi module which let IoT communication flow much better. We also had a micro SD card which held all of the data, the power which runs for 15 hours at 5 volts and then the PCB USB stethoscope. We laser cut a box made of acrylic to hold all of these components just to increase overall portability. Some things to think about is that firstly this device is 100% reusable. There's no waste product after the diagnosis. It's also long term reusable in that you can use them for decades and maybe as far as wear down you'll have to change them. But in general it's a pretty versatile design. Finally, there's a one size fits all. So an entire community can benefit off of just one device. That's how open it is. So it's not like user specific or it's going to get contaminated or anything like that. Here is a quick design of the inside of the PCB USB stethoscope. Looks like essentially what it's doing is it's taking this audio input, converting it into electrical signals and then amplifying this and then making them into digital signals, which is what's essentially in the end saved on the Raspberry PI. Okay, talking a little bit about the data collection. So we collected our data from accredited healthcare facilities like the University of Michigan's Michigan Medicine, Washington Department of Medicine, etc. After this, we labeled each audio clip with the appropriate condition, converted everything to have mono, to have mono signals and then we also made sure that everything was in a waveform because we used, we essentially reprogrammed all of this in Python and we use the Librosa library to do that pre processing. And Librosa works well with waveform audios. What we did in Librosa was we actually extracted Each heartbeat. So we took out the first and last ones because those could be incomplete in the audio, but each other one we found the non silent periods and that's how we split up each separate heartbeat. So each love dub was considered one heartbeat. We resample this so that we had 1,500 samples for all heartbeats. Normalize this data and then we shuffle this and created a train validation and test split which I will show below. And finally we converted the arrays which were this train validation test sets were initially arrays. We converted all of those into tensors which is what the model intakes. Let me go ahead and show you this. Here is our data distribution at the end of all of the pre processing. As you can see there was some disbalance. There's almost double the amount of data for healthy heartbeat than there was for pulmonary stenosis. We knew that the traditional multi classification model wouldn't work, but we also knew that it wouldn't work from the start just because if someone has more than one condition like aortic stenosis and mitral regurgitation, we wanted to be able to identify both. And most multi classification models only have to they have to select the most probable one. So instead what we ended up doing is we created four separate binary classification models which each evaluated for each of these conditions. Our split was we had 100 samples left for testing in each category and then in the remaining it was an 85 train and 15 validation split. Talking a little bit about the deep learning model, I just included a quick flowchart of what we wanted to do. So we wanted to import tensors that were uploaded to the cloud, that were taken by the user in real time essentially. And they were taking four recordings. And that reason we chose four is just because we wanted them to put the stethoscope at the part closest to the valve or the thing that we were identifying. So if it was pulmonary stenosis, for example, it would be closer to the pulmonary valve. So once it was matched up to its appropriate condition, we checked for that condition. So for example, the aortic stenosis condition would be matched up to the aortic stenosis model. If the probability was greater than 15%, then they are at a high risk. And if there was at less than 50%, they were at a low risk. We weren't necessarily saying you have this condition or you don't. We were just giving a higher low risk, which also helps to ease some of the pressure with these diagnoses. I've included our model architecture below. We use an LSTM autoencoder. LSTMs are great, like I said earlier, for looking at sequential or time series data. And the autoencoder is amazing for anomaly detection. So what it's doing is it's taking an input, it's compressing that input into H2, which is the encoder output. And for us that was only 64 neurons. And then that is then decoded and you get that decoder output. So in a perfect situation, the encoder input and the decoder output are the exact same thing. And so the model keeps training to get to that optimum, having that optimum difference. And we use the loss function of L1 loss, which compares the true to the predicted value to identify loss. And our optimizer, which we used to tune the weights and learning rate was Adam. So post training, here are some graphs of what it looked like. So the blue is the training and the yellow is a validation loss. So you can see that the losses went down, which show that overall it is finding those patterns and it is training. And also there was not really a point. This was we trained for 150 epochs because after that it was just training. Accuracy was improving, but not necessarily validation. So that was telling us it might be memorizing the data. So that's when we stopped. So now we have these values, but we want to be able to determine the high risk or low risk. What we did was we actually took the very last value for the validation loss over here and we put this into a variable called thresholds of name. Let me see if I can. We made a certainty formula which was essentially two to the power of a negative mismatch divided by thresholds. Thresholds. I just mentioned mismatch was a loss variation between the input and the trained data. So when, if we put a test data now and then we looked at that loss variation that would be stored in mismatch. In the ideal case, mismatch would be zero and then you would have that 100% certainty. If your mismatch and thresholds were equal, they both found that similar loss. Then we put it at around a 50% confidence. That's how we determine certainty or confidence. I'll go over the Raspberry PI program really quick as well. Basically what this was doing was it was recording the audio, saving it into those four audio clips I mentioned earlier, extracting one central heartbeat out of that, normalizing this, resettling it to 1500 samples, making this into the tensor to be uploaded to the model, and then sending these to the cloud. So we use a lot of Librosa functions here as well. But we also use PI audio to identify and record from the Ryzonic PCB USB stethoscope and we use Shootle in order to upload the audio to the cloud. I included some quick pictures of the web app here. The one page that wasn't included is the diagnosis page, which gives you basically the likelihood of having the conditions and then what the next steps are. So first of all we have usage instruction on how to use the product in an eco friendly manner. It's very cool and pleasing colors. And we made sure to keep most of the instructions and what we were saying in simple English so that a majority of the population can understand. We also supported empowerment through knowledge. So we talked about risk analysis and the genetic and environmental factors that you could use to evaluate your risks. We had mental health support and positive affirmations, comic symbols and then open contact to us for any help. We included next steps in future planning and we even included a map of like doctors in your vicinity as well as promoted environmental sustainability, sharing our positive impact and how users were helping that. Looking a little bit into testing. Let me just look at time. Okay, so essentially first we started off with a functionality test and this was like it sounds to evaluate the model functionality. So for each criterion achieved we gave one point and for each or not achieved it was zero points. We collected real, we checked if it was collecting real time data through the stethoscope, if it was running on the PI interface, pre processing data, converting this to a tensor, sending this tensor through the Internet of things, and then if this tensor was also loading into the deep learning model. So we did a total of five rounds each with 20 tests. And you can see how we improved over time in this graph. So you can see we did one round iterated, one round iterated and so on. In round five, which is our latest round, we had an average of 5.6 on six points, which shows very high reliability overall. And it passed a criteria which was to be above 90%, 90% functional. Talking about the diagnostic accuracy for here, we had 100 test heartbeats for each condition and that was a total of 400 audio clips. And we tested every single of these heartbeat clips in all four models. So that was a total of 1,600 tests. Here you can see the average certainties that the models had. And in general they were more most certain about this diagonal, which they should have been. And what we did was for each value that was above 0.5, we considered to Be a negative, sorry, a positive. And then for each value below that 0.5, we considered it to be a negative. Knowing that we were able to calculate the true positives, false positives, true negatives and false negatives. Here we were able to apply the accuracy formula. However, there was an issue and that was that we didn't really know if the specificity or sensitivity was more because we did have more data for negatives than we did for positives overall. So we wanted to make sure that we were taking into account both of these factors. So we started off with the sensitivity or recall, which measures the number of true positive predictions out of all positives predictions. And all positive predictions would be 400. And then we also did the same thing for specificity, where we're looking at the number of true negatives out of all potential negative predictions and the total negative predictions were 1200. So that's where you can see where some of that bias may have came from. However, when we looked at our results, on average, it was actually pretty close. So the average accuracy was 83.8%, sensitivity 84% and specificity 83.5%. So all these values are relatively similar. And they concluded that the accuracy metric wasn't necessarily biased in our case. And therefore we concluded with a 83.8% accuracy, which well past our criteria of 80%. So just kind of wrapping stuff up. Billions worldwide need an environment friendly, quick, accessible and early detection system for heart disease, the leading cause of death that takes more than 2,000 lives and generates over 40,000 tons of waste per day. Our solution, Telemedicine by Auscultation solves this issue as a fully reusable diagnostic test that analyzes heartbeats in real time for anomalies. This product is also interactive and easy to use. It aids in detection aftermath. It provides an immediate result. It also provides mental health support and guidance through a web app. It can be used outside of a medical setting and produces zero waste. However, there are some limitations, as in a majority of studies we do have that data shortage in that our models were not able. There's no way they could have taken into account the so many diverse conditions that people can have. And even within one condition there can be so many variations that our model may not be able to predict all of them. We also did have a lack of human testing for this project. We did not test on physical humans. We only used audios of prerecorded and anonymized audios. So we don't really know if we were to put this into the world, how would people react? How would they interact with it? That's still something that we would like to explore. When we were testing out the functionality and ability to record, we use a metronome because it was able to give us that beats per minute just like a heart. But it's possible that we could have blocked the audio or things like that, which could have made that not as accurate because we were just holding it up with our hand and listening to the audio. When resampling data, it's possible that we could have lost some key points. We used 1,500 samples. But it's possible that, I don't know, maybe there were just more points that we weren't necessarily seeing, or that the graph, the waveform that was created after resampling that data just wasn't clear. And those are all possibilities. In the future, we would like to experiment with more values and maybe even construct a artificial intelligence model to best determine which waveform is the most efficient. Some future advancements are that we would like to be more compact and have a better carbon footprint. We want to actually put the Raspberry PI microcomputer and battery into the stethoscope, like embedded on the stethoscope so that any extra carbon dioxide released or anything when manufacturing is just reduced. We also, like previously mentioned, want to see the real world application which includes reaching out to cardiologists, hospitals, various environmentalists, and getting feedback from them. Accuracy too. We would like to improve by having more diverse audios and also we want to have more detections and that like I said earlier, maybe using it for gastrointestinal conditions or respiratory conditions or anything that involves just audio and finding anomalies in that audio. And we can actually detect any other time series based data. So if you're thinking about ECGs or EKGs2, it technically can be applied and we've already been looking into that and how to apply that. We also want to apply this in various locations outside of hospitals. This is a remote device. It should be even like you should be able to use it in space. Even if you think about for astronauts and in rural communities. We really want this to be accessible. That was the goal that we were going for. We also were able to create a noise or sound detection in our audio files. So that was a software that we were able to create to extract that one heartbeat audio and that can be applied in other situations as well. And we also learned about pre processing audio and aiding that cloud communication, which additionally can be used for other projects and Other research as well. So Overall with the 93.3% functionality success and an 83.8% accuracy compared to a primary care physician's diagnosis success under 60%, this informative and interactive real time diagnosis shows promise in identifying and treating cardiovascular diseases worldwide while creating a sustainable and healthier environment for generations to come. So thank you all for listening. Let me go ahead and open up to questions at this point in terms of our time to market. First we want to actually go through those steps of talking to cardiologists and physicians and things like that. And once we get feedback from them and we're able to iterate accordingly, then we're thinking of opening it up to the general population. So it could take maybe a year, maybe two, just really making it that perfect optimal thing. And also just because of the stigma again that's surrounding artificial intelligence diagnoses at this time, we just want to make sure that everything that people feel comfortable using, that sort of thing. Any additional input, I would say that I would also like to see more conditions. So I actually have a family member who struggles with one of the heart conditions that we were diagnosing. So that was, it was really personal to me. And I know that there's so many other conditions that we were just for the proof of concert time, we weren't able to get to those or there was a lack of data or something like that. So I think really targeting more conditions and seeing more input on that in terms of data would be really great. Also just talking to the doctors and physicians, like I said earlier, but also just patients and generally just like talking to the general population and seeing what they have to say about something like this would be very useful as well, I think. Also, and I was talking about making the Raspberry PI and those components part of a stethoscope. The issue there is that as high school students we don't have access to manufacturers who are able to do that and who are able to take it to the next step. We're just kind of forced to work with what we have. So I think really being able to reach out to someone who works in the manufacturing field or can make something like that possible would be really useful in the future. Okay. In terms of the 60% misdiagnosis rate. So what we looked at was that there was a research study actually that looked at doctors cardiovascular diagnostic rate and it is a published paper. And I think they said that on average doctors have a diagnostic accuracy of 30 to 40%. And these are not like trained cardiologists but these are just like general physicians. However, the highest that it was, they looked at various different it was like an analyzation of various studies and the highest value there was 60%. It was just looking at cardiovascular diagnoses in general, but not any specific condition over another. I think that would also be something really interesting to look into the future if there's maybe certain conditions that are harder to diagnose or detect by doctors and others are. Okay. Yeah. Thank you guys all once again for coming. If you have any questions, you can feel free to email telemedicine by auscultationmail.com that's the email that we've been using for this project, and I would be more than happy to answer them. Thank you so much. Have a good day.