Result: Strain analysis in mitral regurgitation using interpretable machine learning

Title:
Strain analysis in mitral regurgitation using interpretable machine learning
Contributors:
Lupi, Amalia, Nobile, Marco Salvatore, Bianco, Roberto, Gerosa, Gino, Quaia, Emilio, Pepe, Alessia
Publication Year:
2025
Collection:
Università Ca’ Foscari Venezia: ARCA (Archivio Istituzionale della Ricerca)
Document Type:
Academic journal article in journal/newspaper
Language:
unknown
Relation:
volume:27; journal:JOURNAL OF CARDIOVASCULAR MAGNETIC RESONANCE; https://hdl.handle.net/10278/5095390
DOI:
10.1016/j.jocmr.2024.101411
Rights:
info:eu-repo/semantics/openAccess
Accession Number:
edsbas.4F19479C
Database:
BASE

Further Information

Background: Cardiovascular Magnetic Resonance (CMR) is considered the gold standard technique for ventricular volumes, function, and kinesis evaluation, using Steady State Free Precession (SSFP) cine sequences. The analysis of SSFP for myocardial deformation quantification, by feature tracking is gaining more space in the clinical arena and has shown high sensitivity to detect early contractile dysfunction in several cardiovascular diseases. In patients with mitral regurgitation (MR), global systolic function (expressed as left ventricle ejection fraction, LVEF) seems often to be preserved, due to valvular disease, despite strain alterations. The aim of this study was to predict LVEF starting from strain analysis through a machine learning approach based on fuzzy reasoning, without the need for arbitrary thresholds. Methods: Patients with MR consecutively referred to our center to perform CMR were enrolled. Our protocol included long and short axis SSFP sequences, used to perform strain analysis through circle cvi software. With extracted strain data (global longitudinal strain, GLS, global circumferential strain, GCS, long and short axis global radial strain, GRS LAX and SAX), we developed an intrinsically interpretable model using the pyFUME python library, which generates a fuzzy inference system (FIS) using a data-driven approach. A FIS relies on fuzzy set theory, in which elements can belong to multiple sets with varying degrees of membership. pyFUME determines the different types of patients that are present in the data, partitions them into fuzzy clusters, and calculates fuzzy rules that predict the output for each cluster. The final LVEF value is a weighted interpolation of all rules: the more a patient matches the rules’ logic description, the higher its weight. Results: 58 patients were used to develop the model. 75% of the samples was used for training and 25% was left for testing. We assumed the existence of 2 clusters. According to results, the FIS can accurately predict the LVEF, with a MAE of ...