While healthcare traditionally approaches new technology with caution and are rarely first adopters of the latest craze, attitudes are starting to change.
Healthcare providers have begun to seek enhancement in what is typically perceived in other industries as customer service. Patients are consumers of healthcare services and now expect more from their chosen providers. In turn, these providers use technological solutions to enhance their services, improve efficiency and retain patients.
Now that medical records are in a digital format, the door is open for a variety of improvements in administration, practice management, health monitoring, and of course, diagnostics. These advances make for a competitive market where continuous improvement is the ultimate goal.
“It took the healthcare domain many years to realize the necessity of becoming more data-driven and it is now finally realizing data is more valuable when it is coupled with AI,” said Fred Rahmanian, chief technology officer at Geneia, a healthcare analytics solutions and services company.
What innovations are possible for healthcare diagnostics? Is AI useful in healthcare environments?
One recent example of AI and machine learning comes from a team from MIT’s Computer Science and Artificial Intelligence Laboratory. They have developed a solution that studies existing medical data to suggest cancer diagnoses to doctors based on lymphoma type. As their data volume grows, they hope that easier diagnosis will be possible.
Innovation Set to Continue
There is no doubt that the use of AI in healthcare is sparking a lot of interest.
“We are finally seeing a general willingness to explore what AI has to offer in this domain as is evident by the amount of investment we are seeing in healthcare-related AI startups. Healthcare-focused AI startups with funding deals increased to nearly 70 in 2016, up from less than 20 in 2012. Flatiron Health, a company that uses AI in oncology, is one of the many startups that secured a significant funding deal in 2016,” said Rahmanian.
It is not just clever algorithms that make AI possible. Big Data, the cloud and use of high-performance computing (HPC) methods, such as swarm computing and clusters that enhance computing power, are major contributors as well.
“Computer-aided diagnostics have been around in healthcare imaging for a long time What we are seeing more of now is diagnostics using other modalities. I believe this is because of a combination of the sheer amount of data available today as well as the increase in computing power. Our algorithms also have improved, but I feel it is the availability of data and computing power that are much greater contributors to these advancements. We will see many more of these diagnostic solutions in the next few years,” said Rahmanian.
As Rahmanian pointed out, the MIT innovation is designed to work with data contained in multiple medical reports. However, a different technique is needed for image-based diagnostics, whether that be photos, X-Rays or MRIs.
“I tend to believe deep networks will have more success with visual diagnostic problems. In the recent Data Science Bowl 2017, in which teams were asked to identify lung nodules, the majority of top teams used a deep net approach,“ said Rahmanian.
Deep learning networks are part of the technological pursuit of the artificial brain and are considered a subset of machine learning. Let’s say machine learning is used for a chess program. Any ‘learning’ that takes place is based on evolving algorithms that require direct implementation by data scientists. Deep learning, on the other hand, is a subset of machine learning. This type of learning requires more processing power and speed, which is really only possible thanks to the supplemental use of graphical processing units (GPUs). In a deep learning environment, the computer could play itself over and over again at high speed, learning from every game until it eventually becomes unbeatable.
When this is used on a large data set of images, deep learning allows analysis, presentation and predictive capabilities that were previously considered impossible.
Given the ongoing rise in AI applications, what paths look most promising?
“We’ve seen major advancements in diagnostics related to imaging and we will continue to see major advancement there. But I believe there is greater promise in diagnostic methods that mine patient medical history, various omics (like genomics,[the study of genetic components] proteomics[proteins and their functions] and metabolomics[systematic study of the unique chemical fingerprints that specific cellular processes leave behind]) and exogenous[External, not part of an organism] data,” said Rahmanian.
He believes that HIPAA is of little concern, but that regulations could present a barrier to innovation. However, this barrier can easily be solved with the correct methodology.
“Patient privacy will always be front and center but with proper de-identification of data, we can and do preserve patient privacy. The combination of regulations and the need to gain clinicians’ trust will be much bigger hurdles,” said Rahmanian.
The lack of a central database for medical records is also an issue.
“Having access to a complete patient history is certainly a limiting factor in all of these [use cases]. Healthcare in the US is very local and fragmented. Interoperability continues to be an issue. I am hopeful advancements like blockchain will result in better interoperability,” said Rahmanian.
While AI and deep learning in healthcare are promising, similar to autonomous driving, future ubiquitous adoption will depend on a variety of factors. These elements include data privacy, removal of personally identifiable information (PII) during analysis, data security, regulations and the necessary computing resources to churn the data quickly in a useful manner. Will future healthcare visits involve an android doctor? Probably not. But, there is no doubt that AI is set to play a bigger role in the near future.