The Robo-Doctor Is [In]
The combination of deep learning algorithms and big data are having transformative impact on a range of industries, including the medical field, where AI-based physicians called Robo-doctors will eventually oversee healthcare treatment and dispense medical advice to citizens around the world.
We’re not there yet, but the time is rapidly approaching when communicating with a chatbot or other artificial digital presence about the state of your health will seem as routine as handing over your insurance card at the pharmacy counter.
One of the people who is trying to make robo-doctors a reality is Walter De Brouwer, the founder and CEO of doc.ai, an AI startup that just emerged from stealth that’s using deep learning technology to crunch biological and genomic data with the goal of creating an AI that generates personalized medical advice.
“Ten years ago we were looking in a book to find a restaurant. Now we’re going online, we’re looking up Yelp results,” De Brouwer tells Datanami. “In banking, we first put in ATMs then robo advisors. We’re building all this complexity in a full stack of intelligence… I think it’s the same for doctors.”
One of the factors driving us toward the robo-doctor fate is the relative scarcity of medical professionals. Earlier this year, the American Association of Medical College (AAMC) published a study that found the United States will have a shortage of between 40,800 and 104,900 primary care and specialty doctors by 2030. Worldwide, there’s a 7 million shortage of doctors, nurses, and other medical professionals, according to the World Health Organization. Relatively inexpensive robo-doctors could perform some of the more rudimentary tasks, freeing human doctors for more difficult cases.
Another factor is what German psychologist Gerd Gigerenzer calls the SIC syndrome, which stands for self-defense, innumeracy, and conflict of interest. Because so many doctors practice defensive medicine, do not understand health statistics, and pursue profit instead of virtue, Gigerenzer is optimistic that robo-doctors can actually do better jobs than many human physicians.
“Why not resort to a radical solution: thinking machines? Robodoctors who understand health statistics, have no conflicts of interest, and are not afraid of being sued by you?” Gigerenzer writes in a 2015 Edge story.
Early Days for Robo-Docs
While the medical field currently uses software to help make decisions in some situations, we are a long way from handing medical decision-making over to a robo-doctor. But thanks to the digitization of medical records and the progress made in other fields, we are moving solidly toward a future when preventative health and personalized medicine are common.
In fact, we’ve already started down road to relying on computers to control some aspects of our medical care. “More and more of our [medical] assistants are quantified systems and boxes,” De Brouwer says. “I think robo health is centrally here to stay.”
Doc.ai is currently working with Deloitte Life Sciences and Healthcare on a beta version of its first product, called Robo-Hematology, which was unveiled last month at Deloitte University in Dallas, Texas.
The solution is a conversational AI that allows patients to receive answers to basic questions about the health of their blood at any time of day. The patient can ask questions like “What should be my optimal Ferritin value based on my iron storage deficiency?” or “How can I decrease my cholesterol in the next three weeks?”
De Brouwer, whose background is in computational linguistics, expects Doc.ai to unveil three additional products over the coming months, including Robo-Genomics, Robo-Hematology, and Robo-Anotomics. All of the solutions will be focused on preventative aspects of medicine, as opposed to assisting with clinical care.
The goal of doc.ai is to develop software that “gives people a very multi-dimensional view of their heath and how they can improve it, how they can improve their quality of life, how they can live longer and understand what they’re doing,” de Brouwer says.
Deep Learning Inside
De Brouwer decided early on that Doc.ai would use deep learning techniques and neural networks to train the various models behind the products.
Specifically, the company uses long short-term memory (LSTM), a type of recurrent neural network (RNN), to analyze language and text-based data. The company also uses convolutional neural networks (CNNs) to work with visual data and text-based data at the same time, De Brouwer says.
Doc.ai relies heavily on Google’s TensorFlow framework to define the neural networks, which run on water-cooled NVidia GPUs in the Amazon cloud. The company has a Python bias, and recently started working with the PyTorch framework. “I refuse to do anything that’s not Python,” De Brouwer says. “We don’t use R or Matlab or random forests anymore.”
Analyzing blood was relatively easy with Robo-Hematology because the results of blood tests are so well defined. “Blood results are regulated,” De Brouwer says. “There are 800 markers with clear ranges, so we know what’s good and not good, and can feed that into the machine. We can actually teach the machine that triglycerides are linked to cholesterol and your cholesterol is linked to LDL and HDL and that the value should be this.”
Currently under development, Robo-Genomics will add people’s DNA into the mix and enable doc.ai to advise clients based on their genetic predispositions. The hope is that as more models are added to the mix, that the recommendations get better and better.
“Basically you scrape everything you can then do a word [transformation] so that you make the machine understand it,” de Brouwer says. “If you then do genomics separate from blood results and then afterwards you combine it, you get some intestine results.”
Doc On the Edge
Eventually, the plans calls for putting all this power into a mobile app that can detect if you’ve gotten your prescribed 10,000 steps for the day, or tell if you’ve actually gone to Walgreens to pick up the high blood pressure medication. This is the ultimate goal of a robo-doctor, but we’re not there yet.
“Distributed deep learning is one of the most challenging and revolutionary things you can think of, because in the end you have to push it to the edge device, so that all these predictions happens in the palm of your hand,” De Brouwer says. “Our technology is not yet there to do this.”
Doc.ai will be conducting a hackathon at Stanford University, where it’s been recruiting students from, to get more ideas on distributing the workloads to edge defaces. “You see, the cradle of innovation is always rocking,” he says, “and it’s now rocking towards … decentralization, which means a lot less in the cloud, and a lot more in the edge devices.”