Systems and methods for predicting post-operative right ventricular failure using echocardiograms
Inventors
Arora, Rohan Shad • Hiesinger, William • Quach, Nicolas Tung-Da
Assignees
Leland Stanford Junior University
Publication Number
US-12307666-B2
Publication Date
2025-05-20
Expiration Date
2041-04-26
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Systems and methods for incorporating machine learning to predict post operative right ventricular failure using echocardiograms are described. In an embodiment, the system obtains echocardiography video data describing a patient's heart, generates several dense trajectory descriptors based on the echocardiography video data, reduces the dense trajectory descriptors to a bag-of-words representation, generates a first prediction metric of RV failure based on the bag-of-words representation, generates a second prediction metric based on the echocardiography video data using a neural network, and generates an output prediction metric by applying a weighted classifier to the first prediction metric and the second prediction metric.
Core Innovation
The invention provides systems and methods that incorporate machine learning to predict post-operative right ventricular (RV) failure using echocardiograms. The system obtains echocardiography video data describing a patient's heart, generates dense trajectory descriptors based on this video data, and reduces these descriptors to a bag-of-words representation. A first prediction metric of RV failure is generated using an unsupervised neural network based on the bag-of-words representation, and a second prediction metric is generated using a supervised neural network directly on the video data. A final output prediction metric is generated by applying a weighted classifier to these two metrics.
The problem addressed is the need for accurate prediction of post-operative RV failure, which remains a significant cause of short-term mortality in patients receiving left ventricular assist devices (LVADs). Traditional methods rely on clinical judgment, lab parameters, and qualitative echocardiogram assessments, which may be insufficient for reliable outcomes. The present invention aims to overcome these limitations by leveraging advanced machine learning techniques applied to echocardiography data.
The machine learning system described processes echocardiographic video data through parallel streams, such as greyscale video channels and optical flow, within a neural network architecture to predict the risk of RV failure. The approach enables extraction and analysis of subtle motion, shape, and appearance parameters from the echocardiogram, providing data-driven and potentially more accurate predictions than traditional assessment methods.
Claims Coverage
The independent claim covers a system with several inventive features centered on automated echocardiography analysis for predicting post-operative RV failure.
Automated echocardiography application for prediction of RV failure
A processor and memory system containing an application that: - Obtains echocardiography video data describing a patient's heart. - Generates a plurality of dense trajectory descriptors based on the echocardiography video data. - Reduces these descriptors to a bag-of-words representation. - Generates a first prediction metric using an unsupervised neural network based on the bag-of-words representation. - Generates a second prediction metric using a supervised neural network on the video data. - Applies a weighted classifier to the first and second prediction metrics to generate an output prediction metric for RV failure.
The claim establishes an automated system capable of processing echocardiogram video data with a multi-step machine learning pipeline to predict post-operative right ventricular failure, combining features from unsupervised and supervised neural networks with a weighted classifier.
Stated Advantages
The machine learning system enables prediction of post-operative RV failure in LVAD patients using pre-operative echocardiograms, potentially outperforming existing human-conducted risk scoring methods and experts.
The system can extract and analyze subtle regional aberrations in myocardial motion that traditional manually extracted echocardiographic measures may fail to capture.
The method can track features of importance without requiring human supervision such as outlines or labels, allowing rapid deployment to diverse echocardiography-based problems in an unbiased, structure-agnostic manner.
The system is able to directly analyze spatiotemporal information from cardiac musculature and valves, facilitating characterization of features relevant to cardiac diseases.
The AI echocardiography system can serve as a clinical decision support system for instituting effective RV rescue treatments.
For real-time inference, the ML system can make predictions on a single study within 500 ms, supporting fast clinical workflows.
Documented Applications
Prediction of post-operative right ventricular failure in patients considered for left ventricular assist device (LVAD) implant using pre-operative echocardiograms.
Use as a clinical decision support system for instituting right ventricular rescue treatments based on prediction of high risk of post-operative RV failure.
Early detection of heart failure and disease phenotyping where treatment is guided by echocardiography assessments.
Application to a multitude of cardiac clinical decision support scenarios in which predictions are made based on qualitative echocardiography assessments.
Integration within pre-operative clinical workflows to support patient selection and randomization to right ventricular rescue trials.
Interested in licensing this patent?