03:06
C1 The Importance of Baseline Models in Sepsis Prediction
-
Video
-
Table of contents
-
Video
- Description
- Transcript
- Discussion
About the talk
00:00 Intro
00:11 The Importance of Baseline Models in Sepsis Prediction
01:17 The Models
02:40 Conclusion
About speakers
I work on translating deep learning methods to medicine.I am an MD/PhD student splitting my studies between two Universities of Texas. In 2020 I completed my PhD at UT Austin. My study of deep learning complements my MD in that it positions me to be at the forefront of machine learning advances in medicine. I study the application of deep learning to medicine but switch to a theoretical study of deep networks as problems arise.
View the profileI am currently a data scientist for BCG Gamma, and interested in machine learning applications related to time series forecasting, reinforcement learning, and computer vision.
View the profileDr. Sriram Vishwanath is a Professor in the Electrical and Computer Engineering department at The University of Texas at Austin. Dr. Vishwanath received his B.Tech. from the Indian Institute of Technology (IIT), Madras, M.S. from CalTech and his Ph.D. from Stanford University, all in electrical engineering. His research interests include information theory, wireless communications and coding theory. His industry experience includes work at the National Semiconductor Corporation, CA and at the Lucent Bell labs, NJ. He has won a NSF CAREER Award and the 2005 IEEE Joint IT/Comsoc Best Paper Award.
View the profileHello, my name is Christopher Snider. I am an MD PhD student. Focusing on deep learning research. This is a half-track that I wrote with Jared to attract and Sharon vishwanath at the University of Texas in Austin. It's not strict about the importance of using Baseline simpler comparison models as to contextualize more complicated, neural network models in clinical medicine. So we wanted to do was to take an existing model that was published and available in the literature and then try and see if we could run some simple experiments and comparisons to train untangle. Exactly.
What was making that model. And she has a performance at it does. So we looked at a substance prediction test on the mimic dataset. This model by Chi, Chi at all. What the model tries to do is to take the previous sequence of clinic, visits them, predict hospital stays and then predict weather on the next day. The patient is going to develop sepsis or not. And so in the tension, lstm model was employed by taiji. And this is a model which uses a lstm on on top of a Time dependents. And then uses it attention on top of the nasty. And so has a lot of bells and whistles lot of moving
parts. And without these comparisons is not clear whether these moving Parts contribute to the performance that it actually achieve its, right? You just know that the performance is good. So I wanted to do was to take some time and depending models and see how those ferrets. So we'll just use a random forest. For example, I'm on the bottom rack at the poster random Forest using the previous day's features to make predictions about whether to have sepsis and what the sex of any of these metrics. For example Roc you get very comfortable results with his time in Independence. So then you can
you come away with a completely different picture of why your model was working? It's not because of the time defendant's, right? And then, you can ask another question. What about with with respect to a very simple predictor. Was just predict sepsis tomorrow. If they had it yesterday and never on the first day, will call the satellites predictor. This is on the top of the chart and they're like predictor Chiefs, worst performance for sure. But it's it's interesting to see how well it can actually do on the right hand side. We get into the more. There is a lot of correlation between, you
know, having sepsis today and tomorrow and that you can do really quite well just just using that at ignoring everything you know about the patient. I'm so if you can do fairly well with ignoring all of the differences between patients and just understand whether they had such as yesterday, then you need to to know that to properly, understand how good your ass at this model is. So, I hope we can use these experiments to do to fund of knowledge about the phenomenology of modeling with neural networks to hopefully improve and the future. Thank you.
Buy this talk
Ticket
Interested in topic “Artificial Intelligence and Machine Learning”?
You might be interested in videos from this event
Similar talks
Buy this video
Conference Cast
With ConferenceCast.tv, you get access to our library of the world's best conference talks.
