A machine-learning model that combines patient health history with cardiac MRI findings outperformed traditional statistical methods for predicting major adverse cardiovascular events (MACE), investigators have reported.
“[Our results suggest that] repository historical data prior to cardiovascular MRI holds prognostic value through non-linear interactions which the standard linear Cox regression model is unable to capture,” a team led by Mathias Permlid of Lund University in Sweden wrote. The group’s findings were presented at the recent International Society of Magnetic Resonance in Medicine (ISMRM) meeting.
Accurate risk prediction for coronary heart disease helps clinicians guide patient therapy and follow-up, Permlid and colleagues explained, noting that this kind of risk assessment often uses survival models to predict time-to-event outcomes and noting that “cardiovascular MRI perfusion imaging has prognostic value but is commonly modelled in isolation or together with few clinical variables, possibly limiting the diagnostic potential.”
The team sought to evaluate if risk assessment of cardiac events after a perfusion cardiovascular MRI exam could be improved if patients’ clinical history was taken into account. It hypothesized that a random-forest model capable of handling a non-linear clinical data would be superior to a standard linear Cox regression model for this task.
The investigators conducted a study that included data from 2,159 patients referred for adenosine perfusion cardiovascular MR imaging; they incorporated information from a national health registry regarding diagnoses and procedures related to cardiac health for the 10 years preceding the exam. The study’s primary outcome was a “composite endpoint” of major adverse cardiovascular events (MACE) — defined as cardiovascular death, myocardial infarction, unstable angina, or coronary intervention. Permlid and colleagues compared two sets of data: one that consisted of cardiovascular MRI and demographics data and another that consisted of cardiovascular MRI, demographics, and historical health registry information.
Mean age of study participants was 65.4; mean BMI was 27.6; mean follow-up time was 2.5 years, and MACE occurred in 9.1%.
The group found that, when the models used data from demographics and cardiovascular MR imaging alone, Cox regression achieved a C-index of 0.80, while the random-forest model reached a C-index of 0.79. But when patients’ historical data was added to the cardiovascular MRI and demographics information, the C-index for Cox regression significantly decreased to 0.77, while the random-forest model significantly improved to a C-index of 0.81 (p
C-index for Cox regression and random survival forests models using two feature sets. Adding historical registry data decreases the C-index for Cox regression but increased it for the random survival forests model, consistent with the non-linear structure in the registry data.Mathias Permlid and ISMRM
The results suggest that administrative health registries — data that is already routinely collected — could be a resource for improving cardiac risk stratification when paired with the right analytical tools.
Check out our full coverage of ISMRM here.