Awais Bajwa founder and CEO Ophthalytics, an applied artificial intelligence company specializing in artificial intelligence-based medical diagnostics.
This article features a Q&A from the Data Chats Podcast episode featuring Awais Bajwa with a quick look at the challenging problem Ophthalytics is working to solve.
The problem:
In the U.S. there are approximately 37 million diabetic patients–or more, broadly speaking, about 10% of the population. Conversely, there are only 3,000 retina specialists. That means that for every retina specialist there are approximately 12,000 patients.
Retinal disease is dangerous, and it is simply impossible for these retina specialists to examine all the patients.
Additionally, in the United States, around 2.1 million patients lose a portion of their vision to this disease every year. Of those, about 60,000 people will go blind.
The Solution:
Ophthalytics uses deep learning methods for retinal disease detection. The primary disease they detect is diabetic retinopathy.
Their artificial intelligence technology detects diabetic retinopathy within 3 seconds and can be used in any primary care or other diabetic centers.
It is a game-changing technology, and we’re excited to bring that to market.
What inspired you to start this business?
It’s a personal story and I want to share it. Both my parents are now in their 80s. They are also diabetic patients with Type 2 diabetes. I watched them struggle to schedule an eye exam.
After I moved to Atlanta, they visited me for the first time in my new house. They were trying to get an exam at the same time while they were in the U.S., and even here we had a hard time getting an ophthalmologist appointment. There was nearly a three-month wait to get in, and that’s not to mention the insurance challenges and other costs.
In 2017, I was heavily involved in machine learning and computer vision was one of my specialties. I was hunting for a use case. Luckily, I had a doctor in my house; my wife is a physician.
I told her, “I will develop a technology and you are my domain expert.” That was the origin of our startup company.
Artificial intelligence is data-hungry, especially deep learning and neural networks—they require a lot of data.
Within one year, we developed decent software and started working with the FDA.
Where did you get the collected data since you needed so much?
That was our greatest challenge because in the healthcare industry, there are a lot of regulations.
Luckily, we had some connections offshore. I’m originally from Pakistan and over there the data regulations are not as challenging. Also, I had connections in Africa. So we collected some of the data sets from those locations and included whatever we could collect here in the U.S.
For our first minimum viable product, we certainly had a hard time collecting data, but we were able to do it.
Then, we got aggressive in modeling and labeling. In artificial intelligence work, especially supervised machine learning, you need to be accurate in your training. So apart from getting quality data, we had to also be excellent at labeling and training.
What did you learn from that early experience that could help other organizations looking to leverage AI?
I would like to generalize a little bit because there are different techniques and domains in artificial intelligence.
If we talk about traditional machine learning, there is a different perspective of the data and there are different best practices. If you talk about deep learning, that is sometimes what we call the most innovative part of AI because it’s all about data and how you train.
In general, we need to understand from the business side what is the business requirement and where the existing information lives.
Whenever we talk about information, it’s all about data. But the best practice is, what is the accurate data? What is the source of truth? What is the correct data? What is the data quality? Who is providing this data?
These are the most fundamental things that sometimes people miss. Even the smartest people, just jump into the data and they start doing accelerated data analysis, not understanding the strategic value of what they want to achieve.
How did you set a target and then go about collecting and making sure that the data was as useful as possible?
I will tell you a little bit about the data techniques that we use. We use different types of deep learning architectures and neural network architecture, especially convolutional neural networks.
The number one thing is on the architectural side. So there are two approaches that we are leveraging. One is that we have some of the existing histories. That is one pipeline, and it is where we can leverage Natural Language Processing (NLP), which isis mostly text-based.
And then we have purely visual data where we have just images. On the image-based data, you have to have a different strategy. You have to have some knowledge, and have trained some artificial intelligence models on some specific knowledge.
But I think at the end of the day, the most important thing is how you generalize.
Now, in real-time, when you put into production what you’re looking to handle for patients in the real world, your model needs to perform; that is the real test.
The secret I have seen is that you have to have a very diverse set of data sets and then you have to be in active learning mode.
If you have any doctors, they have to get certifications to keep their knowledge up to data. It seems similarly, the same thing goes for the machine learning models.
There are concepts of data drift and model drift. But in short, I think active learning is key to success.
This approach really helped us. We always try to generalize the data set. We put a lot of variations in the data set and we try to have a bigger population handled at the start of the modeling process as well.
How do you measure success?
One aspect is commercialization, but the other aspect for our company is to just help the community.
In January, we signed an MoU (Memorandum of Understanding) to provide free services in underdeveloped countries where they don’t have ophthalmologists available.
We have had tremendous results in the outcome of using AI. So that was a five-month pilot, and it’s finishing next month.
One of our KPIs is how many patients will we detect in a given month. We’re not looking for money over there, we’re looking for how many people are getting diagnosed and then how many get treated. Also, we evaluate how many really had the disease.
If my software is diagnosing patients who need treatment. I’m feeling very lucky, and that is one of our KPIs.
Now coming to the US market and others, I think there can be different aspects. One of the things I would need to label our models successful is sustainability.
Sustainability for artificial intelligence models is very important because in a lot of businesses you can maybe push the products that are very shiny, but once it gets to production, and once it gets to the market, they do not perform very well. Then it loses its shine.
So you have to make sure that there is continuous feedback coming from the customer. And make sure you are accommodating if there are any gaps within the product performance.
How do you get feedback to implement continuous learning?
In our existing clinical implementation, we are having a one-weekly rhythm because this is our new build software that is in production now.
Feedback is everything for the product once you bring it to the market. We get new feedback every week, which includes the patient, the condition and the diagnosis.
We actually have a feature called explainable AI. So it’s not a simple yes or no diagnosis. Instead, we go into a bit more in detail. So, this is a kind of feature that we are very proud of. It is very helpful for the physician community and regulatory professionals because they rely on AI. This feature tells you what decision was made, and more importantly, why it was made.
Can you tell us more about the importance of explainability?
So I think I want to be very sensitive to one aspect of explainability, especially in deep learning.
If anyone says that it is explaining what is happening, it is not the right thing. Because deep learning is totally a black box when it comes to knowing how it’s making a decision.
But there are different ways to manipulate those decisions and make it explainable.
So in deep learning, there are different ways to classify. For example, if you have a chest X-ray, you can use AI or deep learning as a method of pure classification. Meaning it is a binary yes or no.
This is a more advanced option, though. It is more on the segmentation models or object detection models. We also have different approaches where we can go to the exact location of the chest X-ray and we will pinpoint the area where the disease is.
Now, how the neural networks learn is a process of forward propagation and back propagation, then adjusting the weight and then calling the activation function. This all is very complex. Even with the latest research, we cannot give exactly why and what is that black box.
But we can somehow manipulate the system. In a sense, we have one classification, then we have another segmentation and then object detection models that will actually pinpoint what is the area of the diseases.
That actually helps a lot. For example, if you are a physician, this is the image, and I think this patient has cancer. Maybe 90% of the time it is. But the doctor will be more interested in finding out exactly where it is to give a more accurate diagnosis.
The same is true for retinal diseases. If you look at the image of the retina, we can say if the patient has diabetic retinopathy or not.
This is yes or no. But again, if we kind of go into segmentation and localization, it will be more explainable. We cannot call it an explanation of deep learning, but it is just explainability that we are trying to manipulate the technologies to make it more explainable.
What Are Some Best Practices for AI?
I have seen a lot of technologies come and go. Artificial intelligence grew exponentially. This is one of a kind technology that we have in our hands right now. However, it comes with a lot of challenges.
One of the challenges is we still need to figure out artificial intelligence and its best use and the best adoption in multiple bigger enterprises. The advice I want to give is that we cannot buy AI – many companies try to just buy it like any new technology.
I would also recommend all the high-level directors, senior directors and CIOs should know the basics and the data foundations of this AI. It’s also important to know why they are looking for this technology and what it’s going to do in the future. At the end of the day, the data is within the organization.
Want to Find Ways to Better Use Data at Your Company?
By enrolling in Data Science for Business Leaders, you’ll learn how to partner with data professionals to uncover business value, make informed decisions and solve problems.
This 7-hour course teaches how to:
- Focus Data Projects on Business Impact
- Drive Better Outcomes Through Stronger Partnerships
- Put Insights into Action
- Champion Data-Driven Decision Making
