DETECTING LOW EJECTION FRACTION FROM ELECTROCARDIOGRAM DATA USING ARTIFICIAL INTELLIGENCE

1 . A method, comprising:

obtaining electrocardiogram (ECG) data for a subject;

providing the ECG data as input to a plurality of trained artificial intelligence (AI) models, wherein each trained AI model is trained to determine a score of low ejection fraction (EF) based on the ECG data;

determining, using each trained AI model, a score of low EF based on the ECG data;

determining an overall score of low EF based on a combination of the scores of low EF determined by each of the trained AI models;

selecting, from a plurality of operating point thresholds that correspond to combinations of subject characteristics, an operating point threshold for the subject based on a plurality of subject characteristics of the subject;

comparing the overall score of low EF to the operating point threshold; and

determining an indication of low EF for the subject based on the comparison.

2 . The method of claim 1 , wherein selecting the operating point threshold comprises:

selecting the operating point threshold that minimizes a difference between a sensitivity and a specificity of indications of low EF determined using the plurality of trained AI models for subjects having the plurality of subject characteristics.

3 . The method of claim 1 , wherein selecting the operating point threshold for the subject comprises:

selecting the operating point threshold for the subject based on the plurality of subject characteristics, wherein the plurality of subject characteristics includes a representation of sex of the subject.

4 . The method of claim 1 , wherein selecting the operating point threshold for the subject comprises:

selecting the operating point threshold for the subject based on the plurality of subject characteristics, wherein the plurality of subject characteristics includes a representation of age of the subject.

5 . The method of claim 1 , wherein selecting the operating point threshold comprises:

selecting the operating point threshold that minimizes a difference between a sensitivity and a specificity of indications of low EF determined using the plurality of trained AI models, wherein the operating point threshold is based on the plurality of subject characteristics that includes a representation of sex of the subject and a representation of age of the subject.

6 . The method of claim 1 , wherein each trained AI model has a same architecture, and wherein each trained AI model was trained using a different initial seed for weight initialization.

7 . The method of claim 1 , wherein determining the overall score of low EF based on the combination of the scores of low EF comprises:

determining the overall score of low EF based on an average of the scores of low EF.

8 . The method of claim 1 , wherein each trained AI model in the plurality of trained AI models includes a convolutional neural network (CNN), wherein the CNN includes one or more convolutional blocks and one or more fully connected blocks.

9 . The method of claim 1 , wherein a trained AI model of the plurality of trained AI models is trained by:

obtaining transthoracic echocardiogram (TTE) data for a subject;

obtaining ECG data for the subject;

generating training data by labeling the ECG data based on whether the TTE data indicates that the subject has low EF; and

training the trained AI model using the training data.

10 . The method of claim 9 , wherein obtaining the ECG data for the subject comprises:

determining a TTE result date corresponding to a date of an earliest TTE result in the TTE data that indicates low EF; and

obtaining the ECG data for the subject that is within a time window of the TTE result date.

11 . The method of claim 1 , the method further comprising:

automatically scheduling the subject for a transthoracic echocardiogram based on the indication of low EF.

12 . The method of claim 1 , wherein the indication of low EF prediction is determined while the ECG data is being obtained from the subject.

13 . The method of claim 9 , wherein obtaining the ECG data for the subject comprises:

determining that an earliest TTE result in the TTE data indicates low EF;

obtaining a TTE result date of the earliest TTE result; and

obtaining the ECG data for the subject that includes an ECG result having an ECG result date closest to the TTE result date among ECG results of the ECG data, wherein the ECG result date is within a time window of the TTE result date.

14 . The method of claim 9 , wherein generating the training data by labeling the ECG data based on whether the corresponding TTE data indicates low EF comprises:

generating the training data by labeling the ECG data based on whether an EF value of the corresponding TTE data is below a threshold EF value.

15 . The method of claim 1 , wherein obtaining the ECG data for the subject comprises:

obtaining non-paced ECG data for the subject.

16 . The method of claim 9 , wherein between five percent and twenty percent of samples in the training data are labeled as indicating low EF.

17 . A system, comprising:

one or more processors; and

one or more non-transitory computer-readable memories storing instructions executable by the one or more processors to cause the system to:

obtain electrocardiogram (ECG) data for a subject;

provide the ECG data to an ensemble of trained artificial intelligence (AI) models, wherein each trained AI model of the ensemble is trained to determine a score of low ejection fraction (EF) based on the ECG data;

determine a combined score of low EF for the subject based on a weighted combination of the scores of low EF determined by each trained AI model;

select, from a plurality of operating point thresholds that correspond to combinations of subject characteristics, an operating point threshold for the subject based on a representation of age of the subject and a representation of sex of the subject;

compare the combined score of low EF to the operating point threshold; and

determine an indication of low EF for the subject based on the comparison.

18 . The system of claim 17 , wherein the one or more processors select the operating point threshold for the subject by being further configured to:

select the operating point threshold that minimizes a difference between a sensitivity and a specificity of indications of low EF determined using the ensemble of trained AI models for other subjects having the age of the subject and the sex of the subject.

19 . A non-transitory computer-readable medium storing instructions executable by one or more processors to cause actions to be performed, the actions comprising:

obtaining electrocardiogram (ECG) data for a subject;

providing the ECG data to an ensemble of trained artificial intelligence (AI) models, wherein each trained AI model of the ensemble is trained to determine a risk of low ejection fraction (EF) based on the ECG data;

determining, using one or more fully connected layers, a combined risk of low EF for the subject based on the risks of low EF determined by each trained AI model;

selecting, from a plurality of operating point thresholds that correspond to combinations of subject characteristics, an operating point threshold for the subject based on a representation of age of the subject and a representation of sex of the subject;

comparing the combined risk of low EF to the operating point threshold; and

determining an indication of low EF for the subject based on the comparison.

20 . The non-transitory computer-readable medium of claim 19 , wherein selecting the operating point threshold for the subject comprises:

selecting the operating point threshold that minimizes a difference between a sensitivity and a specificity of indications of low EF determined using the ensemble of trained AI models for other subjects having the sex of the subject and the age of the subject.