Like heart attacks, pulmonary embolisms have subtle symptoms and potentially life-threatening consequences. Doctors consider the diagnosis in a large number of patients who report shortness of breath or chest pain. But unlike heart attacks, there is no easy test for a pulmonary embolism. CT scans have risks – and costs – meaning that every patient cannot be scanned.

This was the quandary physician-scientist Ziad Obermeyer, MD, MPhil, of the Department of Emergency Medicine, sought to resolve. Nearly every patient with suspected pulmonary embolism already gets a chest X-ray in the early stages of their workup, and Obermeyer determined that the ideal solution would look a lot like an X-ray – quick, low radiation, no IV line needed and low cost.

However, there was a considerable problem with using X-rays as a diagnostic tool here: While pulmonary embolisms sometimes appear on an X-ray, the signs are so subtle that they have essentially been written off. Radiologists may have noticed vessels ending suddenly or faint opacities, but these images were rarely definitive and hard to identify in a two-dimensional image.

“But what if there were hidden signals for pulmonary embolus in X-rays – a signal too subtle for the human eye to perceive, but accessible to machine vision algorithms trained on thousands of patient studies?” Obermeyer asked. “In short, this method has a lot of conditions that make the task of diagnosing pulmonary embolus hard for physicians – but also a lot that make us think this might be much easier for a machine to see.”

One question sparked another: Where could he access the necessary computing resources to figure that out? That was when Obermeyer reached out to the team at what was then known as the MGH Center for Clinical Data Sciences (CCDS) – a facility at Massachusetts General Hospital with pools of powerful servers that researchers and clinicians were using to advance the use of artificial intelligence (AI) in health care.

This informal collaboration was one of several the CCDS had with BWHers soon after its launch last year. Recognizing that the Brigham and Mass General could share their resources to create, promote and commercialize even stronger opportunities for using AI in health care, the hospitals teamed up to form the MGH & BWH Center for Clinical Data Sciences in March.

“The combined power of both the Brigham and Mass General allows the CCDS unprecedented access to the data and clinical expertise required to create real-world applications that empower clinicians and enhance outcomes,” said Giles Boland, MD, chair of the Department of Radiology. “We’re harnessing the power of data so we can put it to work to develop smarter, more efficient ways to care for patients and run our systems.”

The ‘Golden Age’ of AI and Machine Learning

If you’re not sure what artificial intelligence is, you’re not alone. For many outside the tech world, it brings to mind science-fiction movies with sentient cyborgs or IBM’s Watson, the supercomputer that competed in Jeopardy! and beat two prior champions in 2011. The term, also known as AI, refers to a branch of computer science in which machines are trained to perform or simulate human tasks and behaviors. But learning for these machines, like humans, is an iterative process – rather than being trained once, they receive feedback and correction to improve performance going forward.

In health care, this technology is being used for everything from improving the accuracy of diagnostic readings to recognizing patterns of diseases to identifying new candidates for clinical trials. At the CCDS, scientists from BWH and MGH are working on more than 20 projects, including ways to use artificial intelligence to identify cancer cells in pathology images, classify bone age based on X-rays and recognize brain tumor mutations from MRI scans. These projects require providing powerful computers with massive amounts of data that can be organized and analyzed.

The more data available, the more likely computers will be able to, for example, identify patterns and make predictions. This is a type of artificial intelligence known as machine learning, an area where the CCDS is currently focusing its efforts. These applications are overseen and validated by a human expert.

“We’re in the golden age of this technology,” said Mark Michalski, MD, the CCDS director. “There are great investigators at the Brigham already doing work in this space, and we’re happy to be able to facilitate that so we can start to look at all our data comprehensively. It’s a tremendous opportunity to take two of the best hospitals in the world and make machine learning part of both.”

Machine learning is about more than analyzing large volumes of data, noted Joshua Moore, head of Strategy and Operations at the CCDS.

“In a more traditional approach, you would use data to predict an outcome based on inputs you clearly define. You’re making predictions based on what you assume the right inputs to be,” Moore said. “In machine learning, you say to the computer, ‘Hey, you figure out what the important variables are,’ because computers are infinitely more capable at piecing together information than the human brain. You allow the computer to train itself and develop its own predictive algorithm.”

One example where machine learning can make a big difference is analyzing low-radiation chest CT scans to detect pulmonary nodules that may indicate the presence of lung cancer, explained Michalski. For a radiologist to perform this analysis, the process is time-consuming, manual and prone to human error.

“Machine learning can prescreen these studies, scrolling through both sides of the lung field and identifying where the nodules are so that a human reader knows to look exactly where they would be,” he said. “Beyond that, the computer can start to measure the nodules and hold onto those readings as structured data.”

Obermeyer’s collaboration with the CCDS on pulmonary embolisms is ongoing, and it is one of on several studies he has worked on with the center. His other investigations range from projects that simply make use of the CCDS’ server farms to those that his researchers and the CCDS team have designed and executed together.

“Overall, I think we are benefiting enormously from their expertise as well as the data and computing resources, and it’s a real privilege to be working with them,” he said.

Learn more about the CCDS.

This story is an example of Partners 2.0, a system-wide initiative to find opportunities to enhance efficiency across the Partners system. Read more about Partners 2.0 on BWH PikeNotes.