Clinical phenotypes among patients with normal cardiac perfusion using unsupervised learning: a retrospective...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Myocardial perfusion imaging (MPI) is a common cardiac scan for diagnosing coronary artery disease and assessing cardiovascular risk.
• Most MPI patients exhibit normal results, necessitating further risk stratification methods.
• Unsupervised machine learning can potentially identify distinct patient subgroups within normal MPI findings.

Purpose of the Study:

• To determine if unsupervised machine learning can identify unique patient phenotypes among those with normal MPI scans.
• To investigate whether these identified phenotypes are associated with an increased risk of death or myocardial infarction.

Main Methods:

• Utilized a large international multicenter MPI registry with 9,849 patients in the training cohort and 12,528 in the external testing cohort.
• Applied unsupervised cluster analysis to identify distinct patient phenotypes based on MPI data.
• Evaluated clinical and imaging features of identified clusters and their association with adverse cardiovascular outcomes.

Main Results:

• Unsupervised learning identified four distinct patient phenotypes from normal MPI scans.
• Cluster 4, comprising 20.2% of the population, showed a significantly higher risk of death or myocardial infarction (HR 6.17) compared to pharmacologic stress (HR 3.03) or prior myocardial infarction (HR 1.82).
• Patients in Clusters 1 and 2 predominantly underwent exercise stress, while Clusters 3 and 4 primarily used pharmacologic stress.

Conclusions:

• Unsupervised learning successfully identified four distinct phenotypes in patients with normal MPI scans.
• A significant proportion of these patients, particularly in Cluster 4, are at a very high risk of myocardial infarction or death.
• Patient phenotyping using machine learning may enhance risk stratification for individuals with normal MPI results.