Science feature

Machines and medicine

The future of AI in healthcare

8734 chest x rays
Ruizhi Liao of the Medical Vision Group hopes that AI can one day assist in the interpretation of chest X-rays such as these by quantifying pulmonary edema in the event of heart failure
Courtesy of Ruizhi Liao

Artificial intelligence is beginning to cement itself as a useful tool for researchers and engineers alike, and it has the potential to become a new addition to the health professional’s toolkit. There are a number of groups at MIT that are exploring the immense potential benefits that artificial intelligence can bring to the healthcare industry. Ranging from diagnostics solutions to making unbiased algorithms, MIT researchers across campus are working to provide new technologies and insights into the future of AI in healthcare.

The Clinical Decision Making group at the MIT Laboratory for Computer Science is one such lab. The group focuses on clinical and biomedical informatics using machine learning methods. Researchers study data taken during a patient’s stay in the hospital, such as blood pressure and heart rate. They then use this information to make predictions regarding whether or not a patient may need an intervention or if they are at risk of a certain disease. The group has also done work in natural language processing (NLP) techniques specific to parsing language in medical texts as well as many other applications in intensive care, security, and more. Matthew McDermott, a graduate student in the lab, believes the research will bring the “power of machine learning methods and data science to the clinical and biomedical space.”

The Clinical Machine Learning Group, which is part of MIT CSAIL and the Institute for Medical Engineering and Sciences, also hopes to use machine learning to make predictions about a patient’s health. One of their aims is to develop algorithms that can be used in any field, from healthcare to autonomous vehicles. Furthermore, they collaborate with clinicians to develop algorithms that enable precision medicine and guide clinical care. For instance, researchers in the lab partner with area hospitals to predict the progression of chronic diseases over years or decades in order to help patients and doctors better manage their condition. David Sontag, who leads the group, described one of his long-term goals as developing an “artificially-intelligent primary care physician.”

The Medical Vision Group, led by Polina Golland of CSAIL, takes a different approach to artificial intelligence in healthcare by developing algorithms for medical imaging. The group’s current work focuses on using machine vision to analyze how oxygen transport in the placenta affects fetal health, identify different conditions in chest X-rays, and use cardiac magnetic resonance images from pediatric patients with heart defects to aid in surgery planning. Ruizhi Liao, a graduate student in the lab, is working with the Beth Israel Deaconess Medical Center to develop algorithms for assessing heart failure patient status by quantifying pulmonary edema in chest X-rays.

Though AI provides many possibilities to better understand human health, it also presents many challenges. There are a number of barriers that must be overcome in order to deploy any of the models produced in the lab. A particularly difficult issue to tackle relates to data collection and data processing. “There are a lot of price concerns, data is expensive to gather, and data is gathered as a byproduct of treatment… It makes our job more difficult as machine learning scientists,” said McDermott of the Clinical Decision Making group.

Liao, of the Medical Vision Group, highlighted access to medical images as a significant barrier. “Lots of recent machine learning progress comes from larger datasets and better computational resources. We can solve the second problem, but the challenge is that it is still very hard to get healthcare data,” says Liao. He is also concerned that researchers are not targeting problems that are useful to clinicians due to the lack of communication between the two groups. Processing medical data presents an interesting and nuanced challenge, as clinicians use particular jargon and patterns in medical data can be confounded by a doctor’s choices. For example, a model might predict a higher rate of survival for a patient with a more severe illness because patients with more severe illnesses are treated first. These problems are difficult to overcome because care should come first, and data for machine learning is an offshoot of the primary task at hand. Researchers in the field are aware of these many complications and are optimistic about the future of AI in healthcare, but are also cautious as well.

Sontag echoed these concerns, stating that “many of the challenges won’t be solved by AI; they’ll be solved by better data.” In fact, he is currently working to develop partnerships with clinicians and faculty to expand the pool of available data. Sontag was particularly concerned about the relay of information between clinicians and algorithms. “We have to recognize that humans will always be in the decision-making process. The consequences of that are that patient-providers may have access to information about the patients that the algorithms don’t have access to.” After all, doctors often have an innate sense of what may be wrong with a patient that cannot be communicated to an algorithm (often called “clinical gestalt”). Data interoperability, where important information is not recorded in the medical record, is also an impediment to healthcare. Patients may have their own preferences on the types of information they’d want shared with such an algorithm as well.

Sontag also worries about maintaining fairness throughout the process of developing algorithms for healthcare. “What happens if your machine learning algorithms are very accurate for one group of people and very inaccurate for another group of people? Because there are limited resources, the money is often spent on the set of people your algorithm is most accurate for. That means that one group is getting much more of that resource than another group. And because health is so intimately tied to socioeconomic outcomes, that could have really long-term impacts on society, which I worry about,” he said.

With a cautious mind and an open heart, experts step into the future hopeful. There are certainly many reasons to be. As medical testing becomes more complicated due to the integration of better technology, the results these tests provide may be incomprehensible to clinicians who do not have time to be trained on every new procedure. “Medicine is getting so complex. It’s not really feasible for one clinician to take the output of a patient’s genetic sequencing and figure out what to do with that,” said Sontag. Instead, he believes that decoding these results is an area where AI may be able to gain traction in the clinic. Artificial intelligence can be used in fields such as telemedicine, where it can assist in making more scalable care so that patients all around the world can obtain care remotely.

While predictive models might not yet be deployable, machine learning is a rapidly expanding area of research, and this growth brings in many fresh minds. McDermott is hopeful about the future of AI in the medical space because “There’s a big potential for impact in terms of helping people, and we’re just now starting to reach the point where we can bring these machine learning technologies to bear on data at magnitudes that we’ve never really been able to do before. So it’s a very exciting time to work in the field.”