Following a cardiac event or stroke, medical professionals frequently rely on risk prediction models to steer treatment decisions. These sophisticated algorithms calculate a patient's mortality probability by analyzing various factors including age, symptoms, and distinctive health characteristics.
Despite their widespread utility, these models frequently generate inaccurate predictions for numerous patients, potentially leading physicians to select ineffective or unnecessarily aggressive treatment approaches. This critical limitation in patient risk assessment has prompted researchers to seek innovative solutions.
"Every risk model undergoes evaluation using specific patient datasets, and even when demonstrating high accuracy rates, they never achieve perfection in real-world clinical settings," explains Collin Stultz, who serves as both a professor of electrical engineering and computer science at MIT and a practicing cardiologist at Massachusetts General Hospital. "Inevitably, some patients will receive incorrect predictions, and in healthcare, such errors can have devastating consequences."
In response to this challenge, Stultz and his collaborative team from MIT, IBM Research, and the University of Massachusetts Medical School have pioneered a groundbreaking methodology enabling healthcare providers to determine whether a specific model's predictions can be trusted for individual patients. This innovative approach promises to revolutionize clinical decision-making by facilitating more personalized and effective treatment strategies.
Stultz, who also holds positions as a professor of health sciences and technology, a member of MIT's Institute for Medical Engineering and Sciences and Research Laboratory of Electronics, and an associate member of the Computer Science and Artificial Intelligence Laboratory, serves as the senior author of this pioneering study. MIT graduate student Paul Myers leads the paper, published today in the prestigious journal Digital Medicine.
Advancing Risk Prediction Through AI
Computer models capable of forecasting patient risks for adverse events, including mortality, have become integral tools in modern medical practice. These predictive systems typically employ machine learning algorithms trained on comprehensive patient datasets encompassing diverse health information and outcomes.
While these models demonstrate impressive overall accuracy, "minimal attention has been devoted to identifying when a model is likely to fail," Stultz observes. "We're attempting to transform how healthcare professionals perceive and utilize these machine learning models. Understanding when to apply—or not apply—a particular model is critically important, especially when incorrect predictions could result in fatal outcomes."
For example, a high-risk patient mistakenly classified as low-risk might not receive sufficiently aggressive intervention, while a low-risk patient incorrectly identified as high-risk could undergo unnecessary and potentially harmful procedures.
To demonstrate their methodology's effectiveness, the researchers focused on the widely implemented GRACE risk score, though their technique can be adapted to virtually any risk prediction model. GRACE (Global Registry of Acute Coronary Events) represents an extensive dataset used to develop a risk assessment tool evaluating patient mortality risk within six months following an acute coronary syndrome—a condition characterized by reduced blood flow to the heart. The resulting risk calculation incorporates age, blood pressure, heart rate, and other readily available clinical indicators.
The researchers' innovative approach generates an "unreliability score" ranging from 0 to 1, where higher values indicate less reliable predictions. This scoring mechanism compares predictions from the specific model being evaluated (such as the GRACE risk score) with those generated by a different model trained on the same dataset. When these models produce divergent results, it suggests the original prediction may not be trustworthy for that particular patient.
"Our research demonstrates that patients with unreliability scores in the top 1 percent receive predictions that are essentially no better than random chance," Stultz notes. "For these individuals, the GRACE score cannot distinguish between those who will survive and those who won't. It becomes completely ineffective for this patient subgroup."
The research team's findings also revealed that patients for whom these models perform poorly tend to be older individuals with higher prevalence of cardiac risk factors.
One significant advantage of this methodology is that the researchers derived a mathematical formula enabling prediction disagreement assessment without constructing an entirely new model based on the original dataset.
"Our approach doesn't require access to the original training data to calculate unreliability measurements, which addresses critical privacy concerns that often limit clinical dataset accessibility," Stultz explains.
Enhancing Model Reliability Through Retraining
The research team is currently developing a user interface that would enable clinicians to evaluate whether a particular patient's GRACE score is reliable. Looking ahead, they also aim to improve risk model accuracy by facilitating easier retraining with data that includes more patients similar to the individual being assessed.
"For sufficiently simple models, retraining can be accomplished rapidly," Stultz envisions. "We can imagine comprehensive software integrated into electronic health records that would automatically indicate whether a specific risk score is appropriate for a given patient and then potentially retrain models in real-time to create more accurate predictions."
The research received funding from the MIT-IBM Watson AI Lab. Additional paper contributors include MIT graduate student Wangzhi Dai; Kenney Ng, Kristen Severson, and Uri Kartoun from IBM Research's Center for Computational Health; and Wei Huang and Frederick Anderson from the University of Massachusetts Medical School's Center for Outcomes Research.
