Generating computer models from implicitly relevant feature sets

1. A method, in a data processing system, for training a hybrid machine learning (ML) computer model to simulate a biophysical system of a patient and predict patient classifications based on results of simulating the biophysical system, the method comprising:

executing a mechanistic computer model on an input dataset to generate an output dataset, wherein the input dataset and output dataset are combined to provide a training dataset for training a surrogate ML model;

performing machine learning training of the surrogate ML model, based on the training dataset, to train the surrogate ML model to replicate logic of the mechanistic computer model as part of the surrogate ML model, and generate patient feature outputs based on surrogate ML model input parameters;

performing machine learning training of a transformation ML model to transform patient feature outputs of the surrogate ML model into a distribution of patient features;

performing machine learning training of a generative ML model to train the generative ML model to encode samples from a uniform distribution of input patient data into mechanistic model parameter inputs that are input to the surrogate ML model; and

processing input patient data for a patient through the trained generative ML model, the trained surrogate ML model, and trained transformation ML model to generate a predicted patient classification for the patient, wherein training the generative ML model comprises training the generative ML model as part of an identity transformation comprising a chain of transformations comprising a first transformation, by the generative ML model, from random samples of the uniform distribution of input patient data to mechanistic model parameters, a second transformation from mechanistic model parameters input to the surrogate ML model to outputs of a mechanistic model represented as the patient feature outputs that are output by the trained surrogate ML model, and a third transformation, by the transformation ML model, from the mechanistic model outputs to observed patient features.

2. The method of claim 1 , wherein training the generative ML model further comprises performing a transfer learning operation based on a loss calculation for one or more downstream ML models that are downstream in a data flow from the generative ML model to the one or more downstream ML models.

3. The method of claim 2 , wherein the one or more downstream ML models comprises one or more of the surrogate ML model, the transformation ML model, or a prior ML model, wherein the prior ML model transforms a prior distribution to a uniform distribution.

4. The method of claim 2 , wherein the loss calculation for the one or more downstream ML models comprises a uniformity test based loss calculation on output of the one or more downstream ML models.

5. The method of claim 2 , wherein the loss calculation for the one or more downstream ML models comprises a first loss calculation for the transformation ML model and a second loss calculation for a prior ML model, wherein the generative ML model is trained to encode samples based on both the first loss calculation and the second loss calculation.

6. The method of claim 1 , wherein the surrogate ML model comprises one or more hidden layers of neurons that learn hidden patient pathophysiological states that are not able to be measured non-invasively.

7. The method of claim 6 , wherein the hidden patient pathophysiological states are represented in the surrogate ML model as one or more of indices or biomarkers calculated as functions of mechanistic model parameters.

8. The method of claim 1 , wherein the surrogate ML model models a mechanistic biological system of the patient.

9. The method of claim 1 , wherein the surrogate ML model models a cardiac system of the patient and wherein the surrogate ML model input parameters that are input to the surrogate ML model comprise heart geometry parameters of the patient.

10. A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:

execute a mechanistic computer model on an input dataset to generate an output dataset, wherein the input dataset and output dataset are combined to provide a training dataset for training a surrogate machine learning (ML) model;

perform machine learning training of the surrogate ML model, based on the training dataset, to train the surrogate ML model to replicate logic of the mechanistic computer model as part of the surrogate ML model, and generate patient feature outputs based on surrogate ML model input parameters;

perform machine learning training of a transformation ML model to transform patient feature outputs of the surrogate ML model into a distribution of patient features;

perform machine learning training of a generative ML model to train the generative ML model to encode samples from a uniform distribution of input patient data into mechanistic model parameter inputs that are input to the surrogate ML model; and

process input patient data for a patient through the trained generative ML model, the trained surrogate ML model, and trained transformation ML model to generate a predicted patient classification for the patient, wherein training the generative ML model comprises training the generative ML model as part of an identity transformation comprising a chain of transformations comprising a first transformation, by the generative ML model, from random samples of the uniform distribution of input patient data to mechanistic model parameters, a second transformation from mechanistic model parameters input to the surrogate ML model to outputs of a mechanistic model represented as the patient feature outputs that are output by the trained surrogate ML model, and a third transformation, by the transformation ML model, from the mechanistic model outputs to observed patient features.

11. The computer program product of claim 10 , wherein training the generative ML model further comprises performing a transfer learning operation based on a loss calculation for one or more downstream ML models that are downstream in a data flow from the generative ML model to the one or more downstream ML models.

12. The computer program product of claim 11 , wherein the one or more downstream ML models comprises one or more of the surrogate ML model, the transformation ML model, or a prior ML model, wherein the prior ML model transforms a prior distribution to a uniform distribution.

13. The computer program product of claim 11 , wherein the loss calculation for the one or more downstream ML models comprises a uniformity test based loss calculation on output of the one or more downstream ML models.

14. The computer program product of claim 11 , wherein the loss calculation for the one or more downstream ML models comprises a first loss calculation for the transformation ML model and a second loss calculation for a ML model, wherein the generative ML model is trained to encode samples based on both the first loss calculation and the second loss calculation.

15. The computer program product of claim 10 , wherein the surrogate ML model comprises one or more hidden layers of neurons that learn hidden patient pathophysiological states that are not able to be measured non-invasively.

16. The computer program product of claim 15 , wherein the hidden patient pathophysiological states are represented in the surrogate ML model as one or more of indices or biomarkers calculated as functions of mechanistic model parameters.

17. The computer program product of claim 10 , wherein the surrogate ML model models a mechanistic biological system of the patient.

18. An apparatus comprising:

a processor; and

a memory coupled to the processor, wherein the memory comprises instructions which, when executed by the processor, cause the processor to:

execute a mechanistic computer model on an input dataset to generate an output dataset, wherein the input dataset and output dataset are combined to provide a training dataset for training a surrogate machine learning (ML) model;

perform machine learning training of the surrogate ML model, based on the training dataset, to train the surrogate ML model to replicate logic of the mechanistic computer model as part of the surrogate ML model, and generate patient feature outputs based on surrogate ML model input parameters;

perform machine learning training of a transformation ML model to transform patient feature outputs of the surrogate ML model into a distribution of patient features;

perform machine learning training of a generative ML model to train the generative ML model to encode samples from a uniform distribution of input patient data into mechanistic model parameter inputs that are input to the surrogate ML model; and

process input patient data for a patient through the trained generative ML model, the trained surrogate ML model, and trained transformation ML model to generate a predicted patient classification for the patient, wherein training the generative ML model comprises training the generative ML model as part of an identity transformation comprising a chain of transformations comprising a first transformation, by the generative ML model, from random samples of the uniform distribution of input patient data to mechanistic model parameters, a second transformation from mechanistic model parameters input to the surrogate ML model to outputs of a mechanistic model represented as the patient feature outputs that are output by the trained surrogate ML model, and a third transformation, by the transformation ML model, from the mechanistic model outputs to observed patient features.