By Anders Gundersen, Research Assistant, HGHI
The frequency with which artificial intelligence (AI) currently pervades conversations within various disciplines is at an all-time high. It seems to be everywhere, but one place where its potential use is of particular interest is in medicine. The prospect of drastically improved patient care or health outcomes with the implementation of innovative technology is enticing, and even sceptics may warm up to the use of AI if it’s in the business of saving lives.
We are quite some time away from full scale AI implementation into our healthcare system, but that doesn’t stop scientists such as Leo Celi from imagining and designing uses for it. Celi is a researcher, a professor, and a physician who works at MIT, Harvard Medical School, and Beth Israel Deaconess Medical Center, and shared some of his work during a talk at the Harvard Global Health Institute in December.
As the director of the MIT Laboratory of Computational Physiology, a group that aims to bring data scientists and clinicians together to improve the quality of data that is routinely collected in intensive care units (ICUs) and make it available for research purposes, he envisions a medical context guided by objectivity, and believes that AI can play a significant role in its realization.
Celi’s group is known for creating the Medical Information Mart for Intensive Care (MIMIC), a rigorously deidentified, publicly available dataset that provides minute by minute medical and medication information from Beth Israel Deaconess Medical Center’s intensive care unit. This is a massive and widely available dataset that can provide great insights, but the data need to be cleaned-up and codified to be useful for research. Celi works with volunteers and collaborators, and via monthly “datathons,” to develop machine learning models that can do this.
A point that Celi made was that, despite tremendous technological feats in healthcare over the past 30 years, we are still left with the same big picture healthcare issues as were experienced in 1989. This indicates that while aspects of healthcare have been improved by technology, healthcare as a whole has not been drastically altered by it.
Technological innovation for its own sake won’t improve care, or health outcomes. It must be implemented in contexts where it is capable of making a difference and is designed to do so.
Does this mean that technology in healthcare shouldn’t be pursued? Certainly not. But we may have to reconsider to our vision of how it can be applied. The idea of AI as an all-encompassing force in healthcare currently captures the imagination of many; the reality will likely lean more toward computer code that endlessly digs through masses of data.
For Celi, AI can be a driving force in moving medicine further towards objectivity. For example, with AI we may no longer have to base lab test standards on the reference ranges provided by only 200 people within a community (as often happens). Instead, a dataset large enough to be generalizable to the population can be analyzed via machine learning, and the most appropriate ranges for a variety of medical measurements can be used to guide diagnosis and treatment. This will not only improve patient care but may also cut costs by reducing overdiagnosis and overtreatment.
The care variation seen in open-heart surgery serves as an example. Celi presented research in which 50 cardiologists had been provided with the same patient case, including the necessary information to choose the best-practice care process. The surprising result: the 50 doctors were evenly split between two possible treatment options, and there was little agreement on what constitutes best-practice for the case presented.
It is this kind of variation in care protocols and advise, prevalent in many medical fields, that Celi and others hope to overcome with a combination of large enough datasets and machine learning. Based on tenths of thousands of cases, the thinking goes, best-practice will be objectively determinable.
Celi also introduced other projects that use machine learning in medicine, including a collaborative effort in gathering data on tuberculosis and the likelihood that a patient will fail their first TB treatment. In fact, much of Celi’s efforts are focused on using AI to improve care in developing nations. One immediate obstacle for this work is the fact that creating such large databases that are actually useful for research is arduously complex, while having a useful database is the cornerstone of creating machine learning models that are capable of producing clinically viable information. Clearly, there is no shortage of potential applications for AI in medicine, and Celi has no shortage of ideas of how and where to implement them.