Electrocardiogram-based assessment of diastolic function
Diastolic function may be assessed by operating one or more machine-learned computational models on electrocardiograms or electrocardiogram-derived features to compute quantitative diastolic indicators, including estimates of echocardiography parameters conventionally measured by echocardiography. In various embodiments, parameters derived from time-frequency transforms of the electrocardiograms are used as input to the model(s) and/or computed within the model(s).
1. A method for quantifying diastolic function, the method comprising:
receiving one or more electrocardiograms measured for a patient;
converting the one or more electrocardiograms, using time-frequency transform, into time-frequency features;
operating one or more machine-learned computational models on input comprising the time-frequency features to compute an estimate or estimates of one or more echocardiogram parameters indicative of diastolic function, the one or more machine-learned computational models having been trained in a supervised manner using values of the one or more echocardiogram parameters obtained by echocardiography as ground-truth outputs.
2. The method of claim 1 , further comprising computing one or more additional indicators of diastolic function based at least in part on the estimate or estimates of the one or more echocardiogram parameters.
3. The method of claim 2 , wherein the one or more additional indicators of diastolic function are computed by operating one or more second machine-learned computational models on input comprising the estimate or estimates of the one or more echocardiogram parameters.
4. The method of claim 3 , wherein the one or more second machine-learned computational models comprise one or more ensemble models.
5. The method of claim 2 , wherein the one or more additional indicators of diastolic function comprise at least one of a left ventricular relaxation risk score, a lateral left ventricular relaxation index, a septal left ventricular relaxation index, or a composite left ventricular relaxation index.
6. The method of claim 2 , wherein the one or more additional indicators of diastolic function comprise a categorical diastolic function indicator.
7. The method of claim 6 , wherein the categorical diastolic function indicator has a value range comprising normal, abnormal, and borderline diastolic function.
8. The method of claim 6 , wherein the categorical diastolic function indicator has a value range comprising low possibility, possible, borderline, probable, and highly probable left ventricular relaxation abnormality.
9. The method of claim 6 , wherein the categorical diastolic function indicator is determined by comparison of the estimate or estimates of the one or more echocardiogram parameters against one or more thresholds.
10. The method of claim 1 , wherein the one or more machine-learned computational models result from training on pairs of input feature sets and a ground-truth outputs for a plurality of patients, the input feature sets comprising the time-frequency features.
11. The method of claim 1 , wherein a first neural network is used to convert the electrocardiograms into the time-frequency features, wherein the one or more machine-learned computational models comprise one or more second neural networks, and wherein the time-frequency features output by the first neural network are provided as inputs to the one or more second neural networks.
12. The method of claim 11 , wherein weights of the first neural network are initialized to implement a time-frequency transform and are subsequently adjusted during end-to-end training of the combined first and second neural networks.
13. The method of claim 12 , wherein the one or more second neural networks are trained with fixed values of the weights of the first neural network prior to the end-to-end training of the combined first and second neural networks.
14. The method of claim 1 , wherein the one or more machine-learned computational models comprise one or more regression models.
15. The method of claim 14 , wherein the one or more regression models comprise at least one of a random forest model or a least squares model.
16. The method of claim 1 , wherein the time-frequency features derived from the time-frequency maps comprise extrema across frequency at one or more points in time associated with the P, Q, R, S, or T waves.
17. The method of claim 1 , wherein the input to the one or more machine-learned computational models further comprises at least one of one or more patient demographic parameters or one or more time-domain features derived directly from the one or more electrocardiograms.
18. The method of claim 17 , wherein the one or more time-domain features derived directly from the electrocardiograms comprise Glasgow-derived parameters.
19. A system comprising:
one or more hardware processors; and
memory storing instructions which, when executed by the one or more hardware processors, perform operations comprising:
receiving one or more electrocardiograms measured for a patient;
converting the one or more electrocardiograms, using time-frequency transform, into time-frequency features;
operating one or more machine-learned computational models on input comprising the time-frequency features to compute an estimate or estimates of one or more echocardiogram parameters indicative of diastolic function, the one or more machine-learned computational models having been trained in a supervised manner using values of the one or more echocardiogram parameters obtained by echocardiography as ground-truth outputs.
20. A non-transitory computer-readable medium storing processor-executable instructions which, when executed by one or more computer processors, cause the one or more computer processors to perform operations comprising:
receiving one or more electrocardiograms measured for a patient;
converting the one or more electrocardiograms, using time-frequency transform, into time-frequency features;
operating one or more machine-learned computational models on input comprising the time-frequency features to compute an estimate or estimates of one or more echocardiogram parameters indicative of diastolic function, the one or more machine-learned computational models having been trained in a supervised manner using values of the one or more echocardiogram parameters obtained by echocardiography as ground-truth outputs.