June 11, 2018
Healthcare is one of the most exciting areas of artificial intelligence research, and is currently ongoing at some of the largest companies in the world, including Intel, IBM, Google and Apple. There are also numerous well-funded start-ups in the space, including Ayasdi, Digital Reasoning, and Sentient.ai.
This makes perfect sense, given the magnitude of the market and the extremely broad range of problems to be addressed.
In the European Union, there are currently 55 separate medical specialties officially recognized. We could conservatively estimate that each of these could benefit from custom-built artificial intelligence powered diagnostic, therapeutic and management tools, putting the eventual number of potential stand-alone artificial intelligence solutions in the hundreds.
We are just now scratching the surface of what is possible.
When we speak about artificial intelligence research in the area of healthcare, what we are usually talking about is a subset of machine learning called deep learning, which involves the application of various types of deep neural network models. When trained using extremely large data sets, which might take the form of patient medical records, or images that are the output of various scanning devices, or photographs of patients, these models can successfully distinguish between healthy and diseased states, categorize disease types, and identify treatment options.
The promise of these artificial-intelligence powered solutions is that they can improve accuracy of diagnosis, improve speed and efficiency of care, improve the use of time by professionals, and relieve problems arising from a lack of qualified professionals in some areas.
The problem areas that artificial intelligence applications in healthcare are able to tackle could go a long way towards solving the global healthcare crisis.
One of the major problems in healthcare is human error. A 2008 study claimed that human error was responsible for “80 percent of adverse events that occur in complex healthcare systems”, while other studies showed that adults in the United States receive only 55% of recommended care, and that 30% of care in the United States could be unnecessary. These errors have major negative impact in terms of outcomes and cost, and given their size, it is possible that technologies with the ability to eliminate or reduce error could go a long way towards solving our health care expenditure problems on a global scale.
Speeding the process of research and diagnosis would also allow physicians to spend more time with patients, allowing them to have more direct contact and to employ their emotional intelligence by interacting in a personal and holistic way.
Besides diagnostic assistance and treatment identification, artificial intelligence solutions are also in development for behavior analysis — including ensuring whether or not healthcare professionals wash their hands, and tracking how much time is spent with patients. This information can be used to improve quality of health-care and guide recommendations for improvement.
But that’s not all. The pharmaceutical industry and researchers are also utilizing deep learning to identify potentially interesting molecular targets for drug discovery.
There is also very promising research being conducted in the area of precision medicine, which involves predicting potential future health risks based on analysis of genomic data and crafting personalized treatment plans based on this information.
And in the future, we may even have fully robotic surgery powered by artificial intelligence. Already, companies like Intuitive Surgical have released the Da Vinci Surgical System, a robotic surgical system designed to allow surgeons to perform minimally invasive procedures. They also produce simulator environments that allow surgeons to train with their tools. In the future, this technology may allow for remote operations to be performed and even for fully automated procedures.
There are several problems presented by artificial intelligence solutions in healthcare, however.
One of these is known as the “black box problem”. To date, deep learning has proven very effective at providing correct answers to problems, provided it has enough data that is of sufficient quality to train the model with. What these models cannot do, however, is to provide an answer to the question of ‘why’ they are making that particular choice. This is the essence of the black box — the decision-making process that the machine is using to make the decision is opaque to the user. This means the user has to simply trust that the AI is correct, given its history of returning correct answers in the past. For most healthcare decisions, this lack of transparency is not suitable to providing life or death health care advice to patients.
For this reason, tools which provide diagnostic and treatment recommendations are typically designed only to assist medical professionals, not to replace them. At the end of the day, a trained physician is still required to understand and either agree or disagree with the recommendations being made by the artificial intelligence software solution. Even so, the AI can still be of great help by reducing the number of unlikely options a physician may have to work through. By pointing the doctor in the right direction, error rates can be reduced and time and efficiency can be improved dramatically.
Another major problem with artificial intelligence solutions in healthcare is the data problem. Actually, this is more of a family of problems related to data. First, access to the amount of data needed to train accurate models for specific use cases is extremely difficult to come by. Second, this data needs to be accurately labeled, which is an expensive and time consuming process. Third, a major issue when it comes to health-related data is privacy. We can also call this the “trust problem”. Providing a software company, hospital, or even a doctor with usage of private health data is a very difficult issue — and it is understandable that some patients are extremely reluctant to do this.
Finally, certain disorders are extremely rare and therefore may be difficult to find even in the largest sample sizes. And given the amount of data required to train an AI solution effectively, it is unlikely that existing traditional data sets will prove sufficient for rare disorder use cases.
The result of all of these issues is that cheap, effective, reliably labeled data that patients have given permission to use is very hard to come by.
Some areas of data, however, are easier to collect, such as output from Internet of Things (IoT) devices, output from scanning devices, or video gathered in hospital environments. Or a hospital may use its own written records — including billing, admission and treatment records in an anonymized way. This data is frequently not ideal for training neural networks, however. In fact, most of the healthcare data collected is unstructured data that is very difficult to use in artificial intelligence applications.
If we look at the largest projects currently in healthcare, we can see examples of how each of these initiatives is in fact utilizing creative solutions to avoid these problem areas.
IBM and Google are focusing on drug discovery, which does not require patient data. They are also working on therapy recommendations, known as ‘clinical decision support’. This avoids the black box problem, by requiring a trained physician to interpret the data. In fact, IBM states that their solution will not conduct diagnosis at all, and will only work based on a doctor-supplied diagnosis.
Start-up Ayasdi is using patient billing records and insurance claims which can be more effectively anonymized. Another start-up, Sentient.ai, is focusing on evolutionary computation, which is a different branch of machine intelligence entirely from deep learning, and may prove promising in the future.
Neuromation is perhaps unique in this space in its focus on creating highly varied yet highly specific synthetic data sets for training neural networks for image recognition tasks for specific medical disorders.
Neuromation can create accurate imagery of hard to find physical characteristics, visual signs of disease, and can vary their occurrence in an incredibly wide range of situations, producing extremely large and accurately labeled data sets. In computer vision applications, this has resulted in impressive real world results in diagnostics.
Examples of this include improved cancer detection in digital pathology, and more accurate lesion detection in MRI.
In February, Neuromation announced a partnership with Longenesis, which uses blockchain technology to safeguard the privacy of personal medical records. In the future, Longenesis may provide a method for patients to monetize their medical records by selling access to them to researchers or AI-powered predictive analytics companies — thus solving a crucial aspect of the healthcare data problem.
Neuromation also recently announced a partnership with Monbaby, one of the world’s leading producers of baby monitors to provide a truly intelligent computer vision baby monitoring application, with the potential to improve infant health, reduce parent stress and even save lives.
On the robotics front, Neuromation has partnered with TRA Robotics on training artificial-intelligence powered robots with synthetic data. Results to date have been extremely encouraging with robots being successfully deployed in real world environments and performing with a high degree of accuracy.
Neuromation looks forward to continue making great strides by partnering with major healthcare providers around the world — and providing creative solutions to some of the problems faced by artificial intelligence researchers in the space.
Senior Analyst, Neuromation