Surfacing insights into left and right ventricular dysfunction through deep learning
Inventors
VAID, Akhil • NADKARNI, Girish N. • Glicksberg, Benjamin S.
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12642512-B2
Publication Date
2026-06-02
Expiration Date
Abstract
Introduced here approaches to developing, training, and implementing algorithms to cardiac dysfunction through automated analysis of physiological data. As an example, a model may be developed and then trained to quantify left and right ventricular dysfunction using electrocardiogram waveform data that is associated with a population of individuals who are diverse in terms of age, gender, ethnicity, socioeconomic status, and the like. This approach to training allows the model to predict the presence of left and right ventricular dysfunction in a diverse population. Also introduced here is a regression framework for predicting numeric values of left ventricular ejection fraction.
Core Innovation
The invention provides a method for predicting right ventricular status or function from electrocardiogram (ECG) waveform data by using transthoracic echocardiogram (TTE) echo reports as a source of labeled clinical observations. TTE data include echo reports containing text representing medical observations about right ventricular status or function, and a natural language processing (NLP) algorithm is applied to identify indicators of right ventricular status or function and assign labels to the echo reports based on the identified indicators.
Electrocardiogram (ECG) data comprising ECG waveforms are obtained for patients that include a third plurality also included in the first plurality. The TTE data and the ECG data are matched to identify data pairs of an echo report and an ECG waveform that are generated within a predetermined interval of time, and training data are generated by labeling the ECG waveforms of the identified data pairs with the labels assigned to the respective echo reports.
A neural network model is trained using the training data to predict right ventricular status or function of an individual from an ECG waveform of the individual. In dependent configurations, the neural network model can be implemented as a deep convolutional neural network and can predict right ventricular status or function using binary labels for right ventricular systolic dysfunction (RVSD) and/or right ventricular dilation (RVD) and/or multiclass labels indicating severity.
Claims Coverage
The document provides two independent claims, a method and a system, that share the same core pipeline. The main inventive features are the NLP-driven labeling of TTE echo report text, the temporal matching between TTE echo report pairs and ECG waveform data, and the training of a neural network to predict right ventricular status or function from ECG waveform data.
NLP-derived labels from TTE echo report indicators
Accessing transthoracic echocardiogram (TTE) data for a first plurality of patients comprising a plurality of echo reports containing text representing medical observations about right ventricular status or function; applying a natural language processing (NLP) algorithm to the echo reports to identify indicators of the right ventricular status or function; and assigning labels to the echo reports based on the identified indicators.
Temporally matching TTE echo reports to ECG waveform data
Obtaining electrocardiogram (ECG) data comprising electrocardiogram waveforms for a second plurality of patients including a third plurality that is also included in the first plurality; and matching the TTE data with the ECG data to identify data pairs of an echo report and an electrocardiogram waveform generated within a predetermined interval of time.
Training a neural network on labeled ECG waveforms to predict right ventricular status
Generating training data by labeling the electrocardiogram waveforms of the identified data pairs with labels assigned to the respective echo reports; and training a neural network model, using the training data, to predict right ventricular status or function of an individual from an electrocardiogram waveform of the individual.
Across both independent claims, the inventive coverage centers on deriving right ventricular status or function labels from unstructured TTE echo report text via NLP, pairing those labels to ECG waveform data through temporal matching within a predetermined interval, and training a neural network to predict right ventricular status or function from ECG waveforms. Dependent claim refinements include specific neural network architecture and label types.
Stated Advantages
Documented Applications
No documented applications found
Interested in licensing this patent?