Skip to main content
2024 Canadian Cardiovascular Congress Main banner
Back

IMAG

POSTER SESSION #5

(P201) PREDICTION OF LEFT VENTRICULAR DIASTOLIC FUNCTION WITH INTERPRETABLE MACHINE LEARNING AND CONTINUUM MAPPING

Saturday, October 26, 2024
12:00 PM – 12:30 PM PT
Location: BOARD 7

    Presenting Author(s)

  • Ricky Hu, MD, MASC

    Internal Medicine Resident / AI Engineer
    University of British Columbia, British Columbia, Canada

    • Co-Author(s)

  • HV

    Hooman Vaseli, BASc

    PhD Student
    Electrical and Computer Engineering
    Vancouver, British Columbia, Canada

  • Jeff Yim, MD

    Cardiology fellow
    CSE
    Vancouver, British Columbia, Canada

  • Darwin F. Yeung, MD, FRCPC (he/him/his)

    Clinical Assistant Professor
    University of British Columbia
    University of British Columbia
    Vancouver, British Columbia, Canada

  • Christina Luong, MD, MHSc, FRCPC, FASE

    Clinical Associate Professor
    University of British Columbia
    University of British Columbia
    Vancouver, British Columbia, Canada

  • MT

    Michael Y. C. Tsang, MD, FRCPC, FASE

    Clinical Associate Professor
    CSE
    The University of British Columbia
    Vancouver, British Columbia, Canada

    • Senior Author(s)

  • PA

    Purang Abolmaesumi, PhD, PhD

    Professor
    Department of Electrical and Computer Engineering
    Vancouver, British Columbia, Canada

    • Supervisor(s)

  • PA

    Purang Abolmaesumi, PhD, PhD

    Professor
    Department of Electrical and Computer Engineering
    Vancouver, British Columbia, Canada

    • Senior Author(s)

  • Teresa S.M Tsang, MD, FRCPC, FACC, FASE

    Professor & Associate Dean, Faculty of Medicine, UBC; Director of VGH-UBC Echo Lab and AI Echo Core Lab, VCHRI AI Hub; Executive Director, VCH Research Institute
    University of British Columbia
    Vancouver, British Columbia, Canada

    • Supervisor(s)

  • Teresa S.M Tsang, MD, FRCPC, FACC, FASE

    Professor & Associate Dean, Faculty of Medicine, UBC; Director of VGH-UBC Echo Lab and AI Echo Core Lab, VCHRI AI Hub; Executive Director, VCH Research Institute
    University of British Columbia
    Vancouver, British Columbia, Canada

Disclosure(s):
Ricky Hu, MD, MASC: No financial relationships to disclose
Teresa S.M Tsang, MD, FRCPC, FACC, FASE: No relevant disclosure to display


Background: Diastolic function classification currently employs manually selected variables with discrete output. The limitation is that this may lack statistical quantification of important variables and discrete output limits classification resolution. Machine learning (ML) has potential to provide a continuous predicted classification score and to quantify variable weights to identify predictiveness.

METHODS AND RESULTS: Echocardiography studies were obtained from February 1, 2010 to March 31, 2016 along with the 8 echocardiographic variables used in the original American Society of Echocardiography (ASE) classification algorithm. 5 ML models were built to predict ASE diastolic function class: Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), Neural Network (NN), and CatBoost. Hyperparameters were optimized with gridsearch and outputs were mapped as a regression to provide a continuous score. Feature importances were computed with perturbation importance algorithms to identify predictive variables. 5-fold cross-validation was employed with the ASE algorithm as ground truth to generate confidence intervals from multiple trials of testing the models.

17,538 patients were included in the study. The models obtained high prediction accuracies (DT: 0.856 ± 0.009, SVM: 0.780 ± 0.018, RF: 0.994 ± 0.030, NN: 0.869 ± 0.033, CatBoost: 0.995 ± 0.003), with CatBoost and RF above 0.990. Feature importances for CatBoost identified top predictive features (left ventricular ejection fraction: 12.108 ± 0.304, left atrial volume index: 27.897 ± 0.907, tricuspid regurgitation velocity: 22.555 ± 0.346). Although this likely represents correlation to the ASE algorithm, this can be programmed for other outcomes in the future. The algorithm is real-time and automated with runtime under 0.01s per new patient input.

Conclusion: The ML models predicted ASE-defined diastolic function with high accuracy with strengths of interpretable feature importances and continuous output for high resolution classification. This study supports usage of such ML architectures, which can be modified to improve causality of diastolic modeling by using outcomes such as heart failure in the future as ground truth to improve diastolic function prediction in a quantified way.