Closing the gap in diagnosing and preventing heart disease would have an enormous effect on global health. Researchers at the Imperial College London are using virtual 3D models of the heart and machine learning to do just that.
Heart disease is the leading cause of death around the world. Around 24 million people will die each year from heart-related ailments by 2030, according to American Heart Association estimates.
The research team at ICL combined image analysis and machine learning algorithms to model heart contractions — you can see the mesmerizing work in the video below. They then match those models against past patient outcomes to suggest better treatments and potentially improve outcomes in future patients.
In a recent research paper, lead author Dr. Declan O’Regan made the case that the solution is more accurate and faster than current methods and, he says, GPUs are a big reason why.
“Machine learning has been used to identify the structure of the heart previously, but GPU-accelerated deep learning approaches enable this to be performed in seconds rather than hours, and have also boosted accuracy,” O’Regan said.
Tapping Underused Data
O’Regan is a senior clinician scientist and consultant radiologist at the MRC London Institute of Medical Sciences (LMS) who heads a research program using machine learning to predict outcomes in heart failure.
He says that clinicians often have difficulty making evidence-based predictions about individual patient’s conditions. And that clinicians are challenged with effectively using the growing array of available data, including imaging and genetics test results.
His team’s strategy was to use automated image analysis and machine learning algorithms to detect the most discriminating heart function features that predict outcomes. To do this, they set out to train the system on the results of hundreds of patients with known outcomes, replicating the way 30,000 points on the heart contract with each beat.
This involved creating a virtual three-dimensional heart for each patient and learning which features were the earliest predictors of heart failure and death.
“The computer performs the analysis in seconds and simultaneously interprets data from imaging, blood tests and other investigations without any human intervention,” said co-author Dr. Tim Dawes, of the London Institute of Medical Sciences, who developed the algorithms that underpinned the software. “It could help doctors to give the right treatments to the right patients, at the right time.”
The research team has been using NVIDIA Tesla K80 GPU accelerators, as well as CUDA and cuDNN, to train its models. It’s now exploring convolutional neural networks, with the goal of using clinical data to comprehensively analyze a patient’s heart, and making survival predictions based on that evidence.
Exploring Next Frontiers
While using machine learning to measure the size and function of the heart can improve diagnostic efficiency, O’Regan believes the research would take another major step forward if the team could integrate imaging results, genetics tests, blood pressure results and the like into the process.
Such a system could autonomously interpret them and recommend actual treatments.
“It’s feasible already in a research setting,” O’Regan said.
In the end, said Dawes, it’s all about improving outcomes, and extending lives, in the face of the world’s biggest killer.