A Machine Learning System to Guide Clinical Procedures in Real-Time

Robert Blake | 18-LW-078


Most sudden cardiac death is caused by abnormal electrical patterns (arrhythmias) initiated by portions of cardiac tissue that erroneously or spontaneously activate during an inappropriate point in the cardiac cycle, resulting in a self-perpetuating constant contraction (ventricular fibrillation) that stops the heart from pumping enough blood to the brain. Using a standard electrocardiogram (ECG) as input, we built an "electrical stethoscope," an imaging technology observing cardiac electrical phenomena that can reconstruct detailed anatomical maps of electrical activity. Unlike existing industry solutions, this approach is non-invasive, more detailed, and dramatically less expensive. The technology consists of a machine-learned network trained on thousands of detailed cardiac simulations that offers the opportunity to stratify patients, both retrospectively and prospectively, based on metrics currently available only through invasive clinical procedures. A large database of cardiac simulations will be released for the greater scientific community so other researchers may reproduce and improve upon our work.

Impact on Mission

This project combined Lawrence Livermore National Laboratory's core competencies in biosciences and bioengineering with Laboratory expertise in machine learning to advance two of the Director's initiatives: 1) predictive biology, enabling data-driven solutions to human health challenges, and 2) cognitive simulation. To meet many of Livermore's mission areas in the future, as well as the Department of Energy's goal to bridge science and technology innovation, our physical-simulation-trained machine learning methodology can be generalized to problems for which scientists have detailed physical simulations but not enough raw experimental data to drive a pure machine-learning approach.