Episode 26: Snorkeling, Artificial Intelligence, … and Healthcare?
Summary Written by Dagny Reese
In this week’s episode of the Monday Science podcast, listen to Dr. Bahijja Raimi-Abrahram chat with Brandon Yang, a machine learning engineer at Snorkel AI, about AI and its applications in healthcare.
Dr Bahijja: To start, could you tell us a little bit about yourself?
Brandon: Definitely! I am currently a machine learning engineer at Snorkel AI — we focus on making AI more practical for a variety of business applications.
I have been very interested in healthcare and AI since I was doing my undergraduate degree at Stanford, where I had been looking at applications of AI in radiology. The field was relatively new at the time, so there were a lot of challenges, but it was a really exciting project. After that, I moved to Google Brain, where I did some research on deep learning and computer vision, as well as on self-driving cars. Now, at Snorkel AI, I focus a lot on creating these practical applications and more practical value for AI.
Dr Bahijja: What are deep learning and natural language processing and how does it relate to artificial intelligence?
Brandon: That’s a great question! AI is quite a broad term and it usually refers to the building of intelligent machines, that can think or act similar to people. Machine learning is a sub-section of AI focusing on interpreting and processing natural language — such as learning from text and speech, and responding to that kind of data. Compared to normal machine learning, deep learning creates a network where the model almost learns features from the data itself — a lot of the focus is on data collection, rather than model design.
Dr Bahijja: When discussing the application of deep learning technologies in healthcare, how are you defining healthcare?
Brandon: I think that it’s quite broad. It’s easier to think about it in the terms of where and how the technologies are used, rather than defining healthcare itself. Broadly, I might say healthcare is anything part of the healthcare delivery pipeline, which is pretty long. There is a lot of work, especially in imaging, such as with radiology and automatic diagnosis. There are similar applications for pathology, such as with slide grading. There is also quite a bit of EHR mining — which helps extract important data from electronic health data, and can improve diagnosis, as one potential application.
To get even more specific — with clinical trials, with pharmaceuticals, it can be used to create exclusion or inclusion data. It can be hard to specifically define, as there are so many applications.
[Within Pharmaceuticals specifically], there is also quite a lot of usage of AI within the process of drug discovery, such as predicting properties of molecules […].
Dr Bahijja: What is the history behind the application of deep learning technologies in healthcare — do you know what the first ever application of AI in healthcare was?
Brandon: Off the top of my head, I’m not exactly sure about the first usage, but I definitely have a few historical tidbits. AI has a long history in healthcare, which is why its hard to pinpoint. In the 1960’s, there were systems to help in in diagnoses, that were mostly based on handcrafted rules or criteria given to an AI system. In 2016, there was another really exciting update where Google created a system to grade diabetic retinopathy, that was seen to be on par with board certified ophthalmologists in accuracy. There has been recent usage of AI for triage — helping people get rooms in hospitals quicker, and more effectively. There have also been a lot of recent publications about the boundaries of AI usage or guidelines in Nature (a scientific journal). […]
What is diabetic retinopathy? Diabetic retinopathy refers to damage caused to the blood vessels in the back of the eye, as a complication of diabetes.
Dr Bahijja: Do think given how AI in healthcare has progressed, that its too late to start getting into the field?
Brandon: Definitely not — it’s not like bitcoin and there’s definitely a strong trajectory. There’s two kind of stages, firstly where we sort of figure out the AI and get it working well, and then start working on specific data sets, applications, and functions. I think it’s important to recognise problems where AI approaches would actually help, which often requires subject matter expertise. For example, being able to recognise what parts of a problem could be solved by AI automation and which wouldn’t be. [It’s definitely not too late].
Dr Bahijja: What are the practical challenges when applying deep learning technologies in healthcare?
Brandon: There are tons! I can think of 4 or 5 off the top of my head.
First and foremost, problem selection is a big challenge. It’s important to choose a meaningful question that AI can effectively be used to solve, that will improve care or complement existing workflow. A great example of that is the triage AI system I mentioned earlier. I think the application of CAD modelling to mammography is an example of a failure of application, as millions of dollars were invested into helping identify cancers with almost no increase in successful diagnoses. Problem selection and analysis is a huge one, but there are also many problems that are more AI specific.
One is that hand labels are hard to get and require a lot of expertise, especially in a healthcare setting. For example, with radiology and identifying abnormalities with AI, that would take months of hand labelling by board certified radiologists. Out of domain generalisation is also a big issue — if an AI system is trained with one data set, there’s the question as to whether it will work with other, similar data sets. For example, from one hospital to another an AI model might perform worse. This applies to data subgroups as well — some AI models might be better at working with one subclass of data than another. In the example of a chest X-ray AI that helps in diagnosis, it may be better at identifying one abnormality than another — this can be clinically useful in some ways, but can also cause problems itself. […]
From a performance standpoint, or patient comfort standpoint, there are a lot of questions about the implementation of AI. Right now, there is mostly a focus on creating complementary system that kind of supplement where humans don’t work as well, such as with triage, case prioritization for radiologists, etc. There is currently a lot more impact in making the system more efficient, rather than trying to overhaul the entire healthcare system.
Dr Bahijja: Do you think there is a risk of AI models replacing medical professionals?
Brandon: I think we are pretty far off from any potential of AI replacing medical professionals, [like doctors]. From a performance standpoint, or patient comfort standpoint, there are a lot of questions about the implementation of AI. Right now, there is mostly a focus on creating complementary system that kind of supplement where humans don’t work as well, such as with triage, case prioritisation for radiologists, etc.
“There is currently a lot more impact in making the system more efficient, rather than trying to overhaul the entire healthcare system.”
Dr Bahijja: What are the ethical challenges when applying deep learning technologies in healthcare?
Brandon: That’s a good question! There are several — one issue for sure is bias. AI systems can often reflect or even amplify bias that was present in the data set they were trained on. There was an interesting paper I recently came across, looking at nature language models trained on clinical texts, that would predict words. It would predict different likely courses, and provided the learning system a “fill in the blanks” exercise. What they found is that patient race affected the course of action that robot would suggest, and that there was racial bias found within the prediction. It’s really an example of one of the worst things that could go wrong with AI implementation in healthcare.
I think this is becoming more and more societally relevant — AI models are trained on data and they rely on the accuracy of that data. More work definitely needs to go into studying this and fixing this — unlike humans, AI doesn’t really self correct as it only has access to the data provided to it. [The data provided needs to be unbiased]. With the data provided, if there is more data provided for example for men, as is seen in many clinical trials already, then there would also probably be better performance by the system for identifying traits within men. [The way the data is collected could also be changed to remove this bias].
This data bias can also manifest in other ways. For example, if a system is trained to diagnose TB in America or the United Kingdom, where they have very sophisticated imaging systems and data collection methods, and its then brought to another country, where they maybe are just using simple X-rays with lower resolution, there will definitely be gaps in performance.
There is also the question of mistakes — who is responsible for when AI makes a mistake? For healthcare, where this [could result in life-or death situations], it’s unclear who would be responsible, whether that be the developer of the AI, the person who approved its usage, the medical professionals themselves, etc. […]
Dr Bahijja: What are the risks of hacking or foul play with deep learning technologies in healthcare?
Brandon: There are quite a few risks, especially with hacking. For example, with adversarial examples, you can see AI that has been trained to make poor decisions, usually not on purpose, or in the case of healthcare, to make misdiagnoses. I am not sure if this has really presented a huge issue in healthcare, but with self-driving cars there was been research into how this could affect recognition of stop-signs. The best way to combat this is often to build robust systems that are resistant to making those kinds of mistakes.
Additionally, as AI are trained on patient data, there is the question of the usage of that data and the storage. This is an issue oftentimes with any electronic storage of patient data, but this is oftentimes a blind spot.
Dr Bahijja: Should people be scared of AI usage in healthcare?
Brandon: I think the answer is no — both on the aspect of AI potentially taking over the jobs of doctors, and the practical and ethical issues. It does need to be carefully studied and regulated before being widely deployed, but there is a lot of potential value and promise within AI. People should be excited, while proceeding with caution.
Dr Bahijja: Do you have any take home messages?
Brandon: I think the goal is really to convey there is a lot about excitement about AI. While there are a lot of practical challenges, there is a lot of promise and potential applications, such as within pharmaceuticals, drug discovery, diagnoses, etc. We can really help empower subject level experts to help build AI platforms in the future.
Want to learn more about Snorkel AI?
Dr Bahijja: I just wanted to ask as well, where did the name snorkel come from?
Brandon: Snorkel really goes back to when it started at Stanford as a research project. Before snorkel, they were working on a project called “Deep Dive” related to data processing, and Snorkel was an attempt to make that kind of processing easier — just like how snorkelling is easier than deep-diving!
Dr Bahijja: What does Snorkel AI do?
Brandon: Snorkel was a research project for four years, and is now a company called Snorkel AI. It’s goal originally was to allow AI to be trained without hand-labels. To do this, one of the most important concepts was programmatic data labelling — for example, having a program label your data for you. It’s often about having enough data with AI, to be accurate, so hand-labelling can often be an exhaustive process and we are really trying to tackle that issue of programmatic labelling?
What is hand-labelling? Hand-labelling refers to the process of data labelling that allows AI to differentiate different pieces of information or data. For a self-driving car, this might be something like labelling images of stop-signs as stop signs, so the system can recognise them and differentiate them from other signs.
Dr Bahijja: What sectors does Snorkel work in and where do you think your systems could help?
Brandon: We have quite a few advantages, particularly in areas where obtaining hand-labelled data is difficult. A great example of this is radiology, where hand-labelling can take months of work by radiologists, and Snorkel can speed this up and help utilise the expertise of radiologists to create programs to label the data instead.
Another example of how programmatic labelling is useful, is for when you want to add another category. In the case of healthcare, let’s say you’ve programmed a system for two diseases, but suddenly discover or want to add a third. With hand-labelling, you’d have to manually go back and make all of these changes, but with Snorkel, you can just change the programmatic labelling rules.
[I think Snorkel also has great applications in data privacy] — with hand-labelling, patient data must be sent to hand labellers, so there is a privacy concern there. With Snorkel, it’s just a program, so there is no need to have all of these people looking over patient data. […]
If you have any questions you’d like to be answered by Dr Bahijja, feel free to send them in via the website chat, or email MondayScience2020@gmail.com.
You can also send us your questions as a voice message via https://anchor.fm/mondayscience/message.
We love to hear your thoughts!