Strain Curve Classification Using Supervised Machine Learning Algorithm with Physiologic Constraints.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Speckle tracking echocardiography (STE) quantifies myocardial deformation using time-strain curves (TSCs).
• Current clinical practice primarily uses peak global longitudinal strain due to accuracy concerns with STE.
• Uncertainty in STE accuracy limits the full utilization of detailed myocardial deformation data.

Purpose of the Study:

• To develop and validate a supervised machine learning algorithm for classifying time-strain curves (TSCs) from STE.
• To differentiate between physiologic and artifactual TSCs automatically.
• To establish a foundation for analyzing subtle, layer-specific myocardial deformation patterns.

Main Methods:

• A physiologically constrained, fully automatic machine learning algorithm was developed.
• The algorithm was trained on a dataset of 415 healthy patients, with three 2-D longitudinal views per patient.
• Data was processed using the in-house K-SAD STE software.

Main Results:

• The algorithm achieved an overall accuracy of 86.4% in classifying TSCs as physiologic, artifactual, or undetermined.
• The positive predictive value for identifying a physiologic strain curve was 89%.
• This demonstrates a significant advancement in the reliable interpretation of STE-derived myocardial deformation.

Conclusions:

• The developed machine learning algorithm reliably classifies TSCs, addressing current limitations in STE accuracy.
• This automated classification is a crucial step towards utilizing the full temporal information in segmental TSCs.
• Enabling accurate separation of physiologic and artifactual curves paves the way for identifying pathological conditions using STE.