As medicine has grown in complexity, the amount of data a single patient can generate — even during a brief hospital stay — has skyrocketed. This big data challenge is reflected in the array of clinical measurements that clinicians can gather at patients’ bedsides, such as blood pressure readings, electrocardiograms (ECGs), and electroencephalograms (EEGs). For critically ill patients or those whose conditions require constant monitoring, these routine readings can quickly swell into vast oceans of data, complicating physicians’ efforts to make timely, sound decisions.

Click here to watch Disruptive Dozen: 12 AI Technologies That Will Reinvent Care

Consider the case of EEGs. ICU patients with neurological conditions, such as a ruptured brain aneurysm, typically receive constant EEG monitoring for about two weeks. Neurologists now examine this data by eye, searching for changes in the size, shape, and frequency of the waveforms. Yet due to the massive volume of information that comes with EEG, it is feasible for doctors to review these patients’ recordings only twice a day, searching for specific signatures that can forecast a seizure — or other signs of neurological illness. But what if EEG data could be mined continuously in an automated fashion to preempt adverse events?

Researchers, including a team in Boston, are working toward this goal by harnessing artificial intelligence (AI). Using machine-learning (ML) approaches, the team has developed an algorithm initially designed to address another important and vexing problem: Deciding whether or not to withdraw care from comatose patients. Indeed, these patients represent a deeply challenging group. Doctors currently base their recommendation to continue (or discontinue) supportive care on multiple variables, both clinical signs and electrophysiological signals. Yet the tools for assessing these variables generally lack precision.

To help bridge this gap, the Boston team developed a method to automatically quantify patients’ brain activity in response to external stimuli — a measurement known as EEG reactivity, which can often help predict comatose patients’ outcomes. Typically, clinicians visually inspect the EEG outputs and compare them pre- and post-stimulation — an approach that may miss subtle variations. With ML-based, data-driven methods, it is possible to detect these and other EEG changes automatically. In a proof-of-principle study, the group’s method showed significant promise, performing at least as well as the consensus opinion of three expert EEG readers. Now, researchers are working to hone and improve this tool so that at any point in time it can assign a probability score that accurately reflects a patient’s likelihood of recovery in six months, helping to bring greater precision to the care of comatose patients.

Using similar techniques, researchers are leveraging AI to help analyze and interpret ECGs, another form of physiological monitoring that strains humans’ capacity to discern slight variations, particularly in high volume data. In contrast, automated computer algorithms can swiftly process these readings, resolving and evaluating each individual heartbeat. Researchers have begun to leverage these algorithms to help predict patients who are at high risk of acute cardiac events.

For example, in a 2011 study, researchers in Michigan and Massachusetts identified a set of so-called “computational biomarkers” — signals derived from AI-based analyses of continuous ECG data — that are strongly associated with cardiovascular death following acute coronary syndrome (ACS), an umbrella term describing conditions under which blood flow to heart muscle is blocked. Collectively, the identified biomarkers reflect distinct aspects of cardiac function, such as the amplitude differences between successive heartbeats, and can be used to evaluate differences in one patient’s ECG readings from those with a similar condition. When combined with existing models for prediction, these ECG-based biomarkers significantly improved risk stratification of patients with ACS. Now, another Massachusetts-based team is harnessing this approach to scrutinize ECGs from hundreds of thousands of patients to uncover biomarkers linked to sudden cardiac death (SCD). This condition arises when the heart abruptly stops beating, typically stemming from electrical disturbances that interrupt the organ’s normal rhythm. SCD is the leading cause of natural death in the U.S., responsible for more than 300,000 deaths each year — often adults in their 30s and 40s with undiagnosed genetic predisposition for heart disease. The researchers’ goal is to develop an ECG-based tool that can serve as an early warning system, helping doctors identify patients who are at high risk of SCD.

For more information about Dr. Westover’s research, please contact Partners HealthCare Innovation by clicking here.