The approach is based on the notion that greater variability between heartbeats reflects greater risk. Scientists trained the machine learning system using historical data for patient outcomes. If a patient survived, their heartbeats were deemed relatively normal; if a patient died, their heart activity was considered risky. The ultimate risk score comes by averaging the prediction from each set of consecutive heartbeats.

There’s plenty of work to be done, including refining the training data to account for more ages, ethnic backgrounds and genders. It clearly needs to be accurate when mistakes could have dangerous consequences. If RiskCardio does enter service, though, it could prove vital to health care. Doctors could quickly assess a patient’s health and decide on an appropriate level of treatment. CSAIL also hopes it can help understand less-than-clear scenarios by running poorly-labeled data through the system.