Quantum Machine Learning for Classification of Left Ventricular Ejection Fraction Phenotypes from Echocardiograms

Pierre Decoodt, Muhammad Waqas Arshad, Marielle Morissens, David Q. Liu

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: quantum computing, machine learning, ensemble learning, medical imaging, video recording, cardiology

TL;DR: Quantum Machine Learning for Left Ventricular Ejection Fraction

Abstract: End-to-end video classification by transfer learning allows one to categorize left ventricular ejection fraction (LVEF) into reduced EF (rEF), midrange EF (mEF), and preserved EF (pEF) from echocardiographic recordings, avoiding delineation of the LV cavity by a human expert or an AI algorithm. We developed a PyTorch implementation using MoViNet. Classical and PennyLane-based classical-quantum models were created. Fine-tuning involved the top four of the five model blocks. We tested our models on Stanford's EchoNet dataset with the original train-val-test split. We tuned the models on the validation set. We developed ternary classifiers for distinguishing between rEF, mEF and pEF, and binary classifiers for rEF vs. rest and not pEF vs. rest. We used the output probabilities of all the classifiers as features subjected to a soft-voting ensemble algorithm consisting of random forest, Gaussian naive Bayes, and logistic regression. For the test set, the extension of the receiver operating characteristic (ROC) to one-vs.-rest multiclass showed a microaveraged ROC AUC score of 0.96. The ROC AUC score and the balanced accuracy were 0.96 and 0.89 for rEF vs. rest, 0.94 and 0.86 for not pEF vs. rest. After optimization of the decision threshold, the sensitivity and specificity of rEF vs. rest and not pEF vs. rest were always above 0.85.

Track: 3. Imaging Informatics

Registration Id: BMN6QR48D6S

Submission Number: 35