Jan 21 2019
A newly developed artificial intelligence program may soon enter clinical studies in London. The program is better than human physicians at proposing treatments for sepsis, a type of blood infection.
Leo Celi, a researcher at MIT and a physician at Beth Israel Deaconess Medical Center, leads a global series of health care hackathons, or datathons, to bring doctors and data scientists together and encourage hospitals to make use of their electronic medical records. (Image credit: Inka Pavelka)
The new machine learning model is part of the latest method of practicing medicine that collects electronic medical-record information for more effective means of detecting and treating difficult medical issues, including sepsis that kills approximately six million people across the world every year.
The findings of a potential treatment method for sepsis did not emerge the standard way, through carefully-controlled and prolonged experiments, but rather materialized at the time of a free-wheeling hackathon held in London in 2015.
In a competition that brought together health care professionals and engineers, one group came up with a better method for treating sepsis patients in the intensive-care unit, with the help of an open-access MIMIC database from MIT. Matthieu Komorowski, a team member, would continue to work with the MIT team who supervises MIMIC to design a reinforcement learning model that forecasted higher survival rates for patients given higher doses of blood vessel-constricting drugs and lower doses of IV fluids. The study results have been reported in Nature Medicine.
The study is part of a series of studies to emerge from the “datathons” conceptualized by Leo Celi, a MIT researcher and staff physician at Beth Israel Deaconess Medical Center. In January 2014, Celi conducted the first datathon to trigger partnership among Boston-area doctors, nurses, data scientists, and pharmacists. After five years, a datathon now takes place every month somewhere around the globe.
After months of preparation, participants assemble at a sponsoring university or hospital for the weekend to search through a local database or MIMIC to look for better ways for diagnosing and treating critical care patients. Among the participants, many continue to publish their work, and in a new landmark for the program, the reinforcement learning paper authors are now preparing their sepsis-treatment model for clinical studies at a couple of hospitals associated with Imperial College London.
Celi noticed a wide difference in patient care—a fact that troubled him. The young doctor also observed that for the average patient, the optimal treatment usually appeared to be rather ill-suited for the patients he met. By the 2000s, he came to know that personalized care can be provided to patients by examining electronic medical-record data with new, robust tools. He subsequently quit his job as a doctor and studied for a dual master’s in public health at Harvard University and biomedical informatics at MIT.
After graduation, he joined MIT’s Institute for Medical Engineering and Science and eventually identified two major obstacles to a data revolution in health care: medical engineers and professionals seldom interacted, and the majority of hospitals, anxious about liability issues, preferred to keep their patient data—everything from laboratory tests to physicians’ notes—out of reach.
Celi believed that such barriers could be overcome through a hackathon-style challenge. The physicians would brainstorm queries and answer them with the aid of the MIMIC database and data scientists. In the process, their efforts would highlight the value of their untapped archives to hospital administrators.
Ultimately, Celi believes that in developing countries, hospitals would also be motivated to develop their own databases. Scientists who cannot afford clinical trials could comprehend their own patient populations and treat them in a more improved manner, democratizing the development and corroboration of new knowledge.
Research doesn’t have to be expensive clinical trials. A database of patient health records contains the results of millions of mini experiments involving your patients. Suddenly you have several lab notebooks you can analyze and learn from.
Leo Celi, Researcher, MIT; Staff Physician, Beth Israel Deaconess Medical Center
To date, several sponsoring hospitals—in Paris, London, Tarragona, Madrid, Beijing, and Sao Paulo—have started plans to create their own MIMIC version, for which MIT’s Roger Mark and Beth Israel tool seven years to create. At present, the process is relatively faster because of tools, which the MIMIC researchers devised and shared with others to de-identify and standardize their patient data.
Long after the datathons, Celi and his group continue to be in touch with their foreign associates by hosting scientists at MIT and reconnecting with them at datathons worldwide.
We’re creating regional networks—in Europe, Asia and South America—so they can help each other. It’s a way of scaling and sustaining the project.
Leo Celi, Researcher, MIT; Staff Physician, Beth Israel Deaconess Medical Center
Italy’s largest hospital, Humanitas Research Hospital, is conducting the next datathon—the Milan Critical Care Datathon from Februarys 1st to 3rd, 2019—and Pierandrea Morandini and Giovanni Angelotti, MIT’s recent exchange students, are assisting to put it on.
Most of the time clinicians and engineers speak different languages, but these events promote interaction and build trust. It’s not like at a conference where someone is talking and you take notes. You have to build a project and carry to it to the end. There are no experiences like this in the field.
Pierandrea Morandini, Exchange Student, MIT
Thanks to tools like GitHub, Google Colab, and Jupyter Notebook, the speed of the events has picked up, allowing teams to delve into the data instantaneously and team up for months after, thereby reducing the time to publication. Currently, Celi and his group are teaching a semester-long course at MIT, HST.953 (Collaborative Data Science in Medicine), designed after the datathons, producing a second pipeline for research of this kind.
Apart from standardizing patient care and making artificial intelligence in health care accessible to all, Celi and his team see another advantage of the datathons: their integrated peer-review process could avert more flawed studies from being published. The researchers highlighted their case in a 2016 piece in Science Translational Medicine.
We tend to celebrate the story that gets told—not the code or the data. But it’s the code and data that are essential for evaluating whether the story is true, and the research legitimate.
Tom Pollard, Study Co-Author and Researcher, MIT
Pollard is part of the MIMIC team.