Artificial intelligence based cardiac event predictor systems and methods
A method and system for determining cardiac disease risk from electrocardiogram trace data is provided. The method includes receiving electrocardiogram trace data associated with a patient, the electrocardiogram trace data having an electrocardiogram configuration including a plurality of leads. One or more leads of the plurality of leads that are derivable from a combination of other leads of the plurality of leads are identified, and a portion of the electrocardiogram trace data does not include electrocardiogram trace data of the one or more leads. The portion of the electrocardiogram data is provided to a trained machine learning model, to evaluate the portion of the electrocardiogram trace data with respect to one or more cardiac disease states. A risk score reflecting a likelihood of the patient being diagnosed with a cardiac disease state within a predetermined period of time is generated by the trained machine learning model based on the evaluation.
1 . A method for determining cardiac disease risk from electrocardiogram trace data, comprising:
receiving electrocardiogram trace data associated with a subject, the electrocardiogram trace data having an electrocardiogram configuration including a plurality of leads and a time interval and comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval;
identifying, from the plurality of leads of the electrocardiogram trace data, one or more first leads of the plurality of leads including electrocardiogram trace data that is derivable from the voltage data associated with a combination of other leads of the plurality of leads, wherein second leads of the plurality of leads are other than the one or more first leads;
determining a first set of electrocardiogram trace data based on a first subset of the second leads;
determining a second set of electrocardiogram trace data based on a second subset of the second leads, wherein the first subset is different from the second subset;
inputting the electrocardiogram trace data associated with the second leads into a trained machine learning model comprising a plurality of branches, the machine learning model being configured to receive the first set of electrocardiogram trace data at a first branch of the plurality of branches and receive the second set of electrocardiogram trace data at a second branch of the plurality of branches, wherein the first branch of the plurality of branches has a different size than the second branch of the plurality of branches, the machine learning model being trained to generate predictions associated with cardiac disease risk based on input data; and
generating, by the trained machine learning model and based on evaluating portions of the electrocardiogram trace data associated with the second leads at respective branches of the trained machine learning model, a risk score reflecting a likelihood of the subject being diagnosed with a cardiac disease state within a predetermined period of time from when the electrocardiogram trace data associated with the subject was generated, the predetermined period of time being less than or equal to one year; and
outputting the risk score to at least one of a memory or a display.
2 . The method of claim 1 , further comprising:
generating, based on the risk score and an additional risk score associated with a second cardiac disease state, a composite prediction, the composite prediction reflecting a likelihood that the subject will be diagnosed with either of the cardiac disease state or the second cardiac disease state within the predetermined period of time from when the electrocardiogram trace data associated with the subject was generated.
3 . The method of claim 1 , wherein the electrocardiogram trace data associated with the subject does not include features indicating that the subject has the cardiac disease state when the electrocardiogram trace data associated with the subject was generated.
4 . The method of claim 1 , further comprising generating, based on the risk score, a treatment recommendation; and
outputting the treatment recommendation to at least one of the memory or the display,
wherein the treatment recommendation includes a recommendation to perform additional cardiac monitoring for the subject.
5 . The method of claim 1 , further comprising:
training a convolutional neural network on a first plurality of subjects and a second plurality of subjects, wherein a time between a date of a recorded ECG and a diagnosis of a cardiac disease for each of the first plurality of subjects is less than a diagnosis threshold, and a time between a date of a recorded ECG and a diagnosis of the cardiac disease for each of the second plurality of subjects is greater than the diagnosis threshold, and wherein a time between a date of a particular recorded ECG and a diagnosis of the cardiac disease that is less than the diagnosis threshold is indicative of the cardiac disease being active at the date of the particular recorded ECG; and
providing the trained convolutional neural network as the trained machine learning model.
6 . The method of claim 5 , further comprising:
refining the trained convolutional neural network using only the second plurality of subjects, wherein the diagnosis threshold is selected from a number of days.
7 . The method of claim 5 , wherein the cardiac disease is aortic stenosis.
8 . The method of claim 1 , wherein the trained machine learning model is selected based at least in part on a severity of the cardiac disease state the generated risk score represents.
9 . The method of claim 1 , wherein the trained machine learning model is based on training data associated with a plurality of clinical sites, and wherein the method further comprises training the machine learning model using subject data associated with one site of the plurality of clinical sites and testing the trained machine learning model on the remaining sites of the plurality of clinical sites.
10 . The method of claim 1 , wherein the electrocardiogram trace data associated with the subject comprises ECG data selected from one or more of acute myocardial infarction, atrial fibrillation, atrial flutter, complete block, early repolarization, fascicular block, first-degree atrioventricular block, intraventricular conduction block, left bundle branch block, right bundle branch block, ischemia, left anterior descending artery ischemia, right bundle branch block, low QRS, left ventricular hypertrophy, non-specific ST-T wave, Non-specific T wave, other bradycardia, premature atrial contractions, pacemaker, poor tracing, prior infarction, prior myocardial infarction anterior, prolonged QT, premature ventricular contractions, right axis deviation, second degree atrioventricular block, sinus bradycardia, supraventricular tachycardia, tachycardia, tachyarrhythmia, T inversion, or ventricular tachycardia.
11 . The method of claim 1 , wherein the trained machine learning model is a composite model, and wherein training the composite model comprises:
generating a subject timeline for each subject;
anchoring each respective subject timeline to a date of occurrence of an echocardiogram; and
labeling each respective subject as having a positive or negative ECG based at least in part on a date of an ECG with respect to the date of occurrence of the echocardiogram.
12 . The method of claim 11 , further comprising excluding subjects from training after a censoring event is detected in the subject timeline.
13 . The method of claim 1 , wherein the cardiac disease state is one of aortic stenosis, aortic regurgitation, mitral stenosis, mitral regurgitation, tricuspid regurgitation, reduced ejection fraction, increased interventricular septal thickness, cardiac amyloidosis, or stroke.
14 . The method of claim 1 , wherein the trained machine learning model is trained to evaluate subject data of an electronic health record with respect to one or more cardiac disease states.
15 . The method of claim 1 , wherein the electrocardiogram trace data evaluated at the first branch is acquired over a first time period, the first time period being twice as long as time periods associated with the electrocardiogram trace data evaluated at each of the other branches of the plurality of branches.
16 . The method of claim 1 , wherein a portion of the electrocardiogram trace data evaluated at the first branch is reconstructed, at least in part, from the one or more first leads.
17 . A system comprising:
a computer including a processing device, the processing device configured to:
receive electrocardiogram trace data associated with a subject, the electrocardiogram trace data having an electrocardiogram configuration including a plurality of leads and a time interval and comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval;
identify, from the plurality of leads of the electrocardiogram trace data, one or more first leads of the plurality of leads including electrocardiogram trace data that is derivable from the voltage data associated with a combination of other leads of the plurality of leads, wherein second leads of the plurality of leads are other than the one or more first leads;
determining a first set of electrocardiogram trace data based on a first subset of the second leads;
determining a second set of electrocardiogram trace data based on a second subset of the second leads, wherein the first subset is different from the second subset;
input the electrocardiogram trace data associated with the second leads into a trained machine learning model comprising a plurality of branches, the machine learning model being configured to receive the first set of electrocardiogram trace data at a first branch of the plurality of branches and receive the second set of electrocardiogram trace data at a second branch of the plurality of branches, wherein the first branch of the plurality of branches has a different size than the second branch of the plurality of branches, the machine learning model being trained to generate predictions associated with cardiac disease risk based on input data;
generate, by the trained machine learning model and based on evaluating portions of the electrocardiogram trace data associated with the second leads at respective branches of the trained machine learning model, a risk score reflecting a likelihood of the subject being diagnosed with a cardiac disease state within a predetermined period of time from when the electrocardiogram trace data associated with the subject was generated, the predetermined period of time being up to one year; and
output the risk score to at least one of a memory or a display.
18 . The system of claim 17 , wherein the cardiac disease state is one of aortic stenosis, aortic regurgitation, mitral stenosis, mitral regurgitation, tricuspid regurgitation, reduced ejection fraction, increased interventricular septal thickness, cardiac amyloidosis, or stroke.
19 . A non-transitory computer readable medium, comprising instructions that, when executed by a processing device, cause a computer to:
receive electrocardiogram trace data associated with a subject, the electrocardiogram trace data having an electrocardiogram configuration including a plurality of leads and a time interval and comprising, for each lead included in the plurality of leads, voltage data associated with at least a portion of the time interval;
identify, from the plurality of leads of the electrocardiogram trace data, one or more first leads of the plurality of leads including electrocardiogram trace data that is derivable from the voltage data associated with a combination of other leads of the plurality of leads, wherein second leads of the plurality of leads are other than the one or more first leads;
determining a first set of electrocardiogram trace data based on a first subset of the second leads;
determining a second set of electrocardiogram trace data based on a second subset of the second leads, wherein the first subset is different from the second subset;
input the electrocardiogram trace data associated with the second leads into a trained machine learning model comprising a plurality of branches, the machine learning model being configured to receive the first set of electrocardiogram trace data at a first branch of the plurality of branches and receive the second set of electrocardiogram trace data at a second branch of the plurality of branches, wherein the first branch of the plurality of branches has a different size than the second branch of the plurality of branches, the machine learning model being trained to generate predictions associated with cardiac disease risk based on input data;
generate, by the trained machine learning model and based on evaluating portions of the electrocardiogram trace data associated with the second leads at respective branches of the trained machine learning model, a risk score reflecting a likelihood of the subject being diagnosed with a cardiac disease state within a predetermined period of time from when the electrocardiogram trace data associated with the subject was generated, the predetermined period of time being less than or equal to one year; and
output the risk score to at least one of a memory or a display.
20 . The non-transitory computer readable medium of claim 19 , wherein the cardiac disease state is one of aortic stenosis, aortic regurgitation, mitral stenosis, mitral regurgitation, tricuspid regurgitation, reduced ejection fraction, increased interventricular septal thickness, cardiac amyloidosis, or stroke.