As humans, we naturally strive to minimize risks in our daily lives, carefully planning our paths and adopting preventive strategies to steer clear of illnesses, hazards, and emotional distress.
However, when it comes to managing the complex internal mechanisms of our bodies, our control becomes significantly more challenging.
Addressing this challenge, researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed an innovative approach to forecasting health outcomes: an artificial intelligence system capable of assessing a patient's likelihood of cardiovascular mortality by analyzing their heart's electrical patterns.
This groundbreaking technology, named "RiskCardio," specifically targets individuals who have experienced an acute coronary syndrome (ACS)—conditions characterized by reduced or obstructed blood flow to the heart. By examining merely the initial 15 minutes of a patient's raw electrocardiogram (ECG) reading, the tool generates a risk score that categorizes patients into various risk levels.
Patients identified as high-risk by RiskCardio—those in the upper quartile—demonstrated a nearly sevenfold increased probability of cardiovascular-related death compared to the low-risk group in the lowest quartile. In contrast, individuals flagged as high-risk by conventional risk assessment methods showed only a threefold higher likelihood of experiencing adverse events relative to their low-risk counterparts.
"Our research addresses two critical challenges: the data science problem of incorporating extensive time-series data into risk assessment models, and the clinical challenge of helping physicians identify high-risk patients following an acute coronary incident," explains Divya Shanmugam, principal author of the paper detailing RiskCardio. "The convergence of machine learning and healthcare presents numerous opportunities like this—compelling computational challenges with genuine potential for real-world medical impact."
The Challenge of Risk Assessment
Earlier machine learning approaches to risk evaluation have typically relied on either incorporating external patient data such as age or body mass, or leveraging domain-specific expertise to guide feature selection within the model.
RiskCardio, in contrast, operates exclusively using patients' raw ECG signals, without requiring any supplementary information.
Consider a patient admitted to the hospital after an ACS episode. Following admission, a physician would traditionally evaluate the risk of cardiovascular death or heart attack using comprehensive medical data and extensive diagnostic procedures before determining an appropriate treatment plan.
RiskCardio seeks to enhance this initial risk assessment phase. The system accomplishes this by segmenting a patient's signal into sequences of consecutive heartbeats, based on the premise that variations between adjacent beats provide valuable insights into future risk. The system was trained using data from previous patient studies.
To develop the model, the research team initially divided each patient's signal into collections of adjacent heartbeats. They then assigned a label—indicating whether the patient ultimately died from cardiovascular causes—to each set of adjacent heartbeats. The researchers trained the model to classify each pair of adjacent heartbeats according to patient outcomes: Heartbeats from patients who died were labeled "risky," while heartbeats from survivors were labeled "normal."
For a new patient, the team generates a risk score by averaging the risk prediction across all sets of adjacent heartbeats.
Within just the first 15 minutes following an ACS event, the system could extract sufficient information to predict the likelihood of cardiovascular death within 30, 60, 90, or 365 days.
Nevertheless, deriving a risk score solely from ECG signals presents significant technical challenges. These signals are extremely lengthy, and as the number of model inputs increases, learning the relationships between these inputs becomes progressively more difficult.
The team validated the model by generating risk scores for a cohort of patients. They then measured the increased likelihood of cardiovascular death among high-risk patients compared to low-risk individuals. Among approximately 1,250 post-ACS patients, 28 would die from cardiovascular causes within a year. Using the proposed risk score, 19 of these 28 patients were correctly identified as high-risk.
Looking ahead, the team aims to enhance the dataset's diversity to encompass different age groups, ethnic backgrounds, and genders. They also intend to investigate medical scenarios involving substantial amounts of poorly labeled or unlabeled data, evaluating how their system processes and manages such information to handle more ambiguous cases.
"Machine learning excels at pattern recognition, which is particularly valuable for patient risk assessment,'' notes Shanmugam. "Risk scores serve as effective tools for communicating patient status, which is crucial for making efficient clinical care decisions."
Shanmugam presented the paper at the Machine Learning for Healthcare Conference alongside PhD student Davis Blalock and MIT Professor John Guttag.
