Health data is most often associated with vital signs: height, weight, blood pressure, and heart rate immediately come to mind. But in recent years, the reality of what health data encompasses has become so much broader. Today, a multitude of data is being collected on patients. This data tells the story of the patient – not just of their physical health or even their genetics and medical or family health history but also of their social and environmental health.
The data generated in the healthcare industry each year is truly astounding, and it’s increasing by 48% each year. The amount of data produced in 2020 alone could exceed 2.3 zettabytes. We’re talking about trillions of gigabytes here. That is equal to the time it would take to someone to binge-watch 262 million years straight of HD movies. And no single piece of data presents a full description of a patient.
Where a patient is born, how they are raised, where and how they work, and many other social determinates of health help to paint this picture. The below plot compiled by Dr. Isaac Kohane Et. al displays this ecosystem and the different opportunities we have to combine these pieces to generate a more wholistic view of each patient.
In the past few years, artificial intelligence in healthcare has slowly transitioned from an emerging opportunity to a true game changer. The capability of artificial intelligence to combine data sources to reach conclusions makes the technology uniquely suited to healthcare’s needs. But it is no secret that while artificial intelligence has been noticeably present in other fields, it has remained somewhat elusive in its application within the healthcare ecosystem by comparison.
Although the past decade has shown some powerful solutions in early development, initiatives launched in 2019 such as the CMS-sponsored AI Health Outcomes Challenge and the $250M Health AI Lab in the U.K. indicate that we are moving quickly into what is going to be a decade of investment of artificial intelligence in healthcare.
But why – with all its promise – has artificial intelligence not taken off in the healthcare industry when it appears to be such a good fit?
While most industries – like the field of IoT – have seized on technical electronic standards, the health sector still faces high variation in operating technology. Artificial Intelligence models primarily operate on widely available and consistent data to make decisions. So, when one hospital is using handwritten records and faxing them, and another hospital is using electronic health records and transmitting them through the internet, creating artificial intelligence solutions that suit both of these hospitals is difficult.
Fortunately, we are seeing increasing adoption and incentives for the use of technology standards and information sharing. FHIR (Fast Healthcare Interoperability Resources) and Blue Button 2.0 are two programs that are aiming to open and standardize data formats, allowing for more patient control of information in a single way. Uniquely, hospital claims data used for payments is already quite standardized, and for those with access, it offers an incredible opportunity to assess even a single patient’s needs even today.
Creating and testing artificial intelligence models relies upon having data available in large volumes. Unique to healthcare, HIPAA rules prevent widespread dissemination of any patient’s records for their own safety. But this is changing in a positive way. In the past, academic teams have often faced difficulty in securing patient data for research and have instead worked with an individual hospital’s data or with social media data. While these offer some information gain, when compared to true patient information there is a much greater gain garnered from diversified and broad patient data.
Last year saw new guidance from the U.S. Department of Health and Human Services around patients permitting the viewing of and sharing of their own data beyond care use cases. Data is also moving more easily when these requests are approved. With the common formats described earlier, this is resulting in providers better able to care for their patients and researchers more empowered to build new solutions.
Most artificial intelligence models operate under the assumption that the user retains some control – as in the case of a self-driving car – or that the task may be reattempted if an undesired outcome is achieved, such as in the case of a smart speaker identifying a command. With healthcare, we must remember that our decisions are less reversible and can have very impactful effects. As a result, empowering clinicians by communicating a model’s reasoning is essential in healthcare applications than other fields where artificial intelligence is applied.
The technology surrounding this is evolving. Neural networks (i.e., deep learning) thus far have suffered from lack of the necessary transparency that is required in most cases to communicate to a clinician why a model is making the decisions it is making. While research in this space is immense, and there is still a lot of ground to cover, imaging models such as CheXNeXt are able to successfully indicate areas of interest to radiologists. Machine learning models, alternatively, have largely achieved justifiable transparency within the last few years.
One final point of interest is environmental stability. When a model is trained, the assumption is that the environment in which is it operating will remain largely unchanged. For example, if we are building a model for detecting breast cancer for radiologists, we assume that the make and model of the imaging system used is identical every time. If we try to feed a new image taken by a different X-ray system into the same artificial intelligence model, we are unlikely to get the same results.
While this problem is not exclusive to healthcare, the rate at which technology and standards across healthcare is evolving is more rapid than most fields, making it a challenge to know specifically what technology is worth developing, deploying, and supporting and when you should do so. Advancements in online learning, or “reinforcement learning,” as well as new strategies for how artificial intelligence is deployed and monitored over time is enabling teams to overcome this hurdle. Ease of artificial intelligence development software is playing an integral role here.
Despite these challenges, there is industry-wide interest in driving forward adoption of artificial intelligence in healthcare. Driven by the right data, at the right time, in the right setting, today’s artificial intelligence in healthcare solutions are poised to dramatically change the game in the next few years.