Tensor amplification-based data processing
A method of generating an assessment of medical condition for a patient includes obtaining a patient data tensor indicative of a plurality of tests conducted on the patient, obtaining a set of tensor factors, each tensor factor of the set of tensor factors being indicative of a decomposition of training tensor data for the plurality of tests, the decomposition amplifying low rank structure of the training tensor data, determining a patient tensor factor for the patient based on the obtained patient data tensor and the obtained set of tensor factors, applying the determined patient tensor factor to a classifier such that the determined further tensor factor establishes a feature vector for the patient, the classifier being configured to process the feature vector to generate the assessment, and providing output data indicative of the assessment.
1. A method of generating an assessment of medical condition for a patient, the method comprising:
obtaining, by a processor, a patient data tensor indicative of a plurality of tests conducted on the patient;
obtaining, by the processor, a set of tensor factors, each tensor factor of the set of tensor factors being indicative of a decomposition of training tensor data for the plurality of tests, the decomposition amplifying low rank structure of the training tensor data;
determining, by the processor, a patient tensor factor for the patient based on the obtained patient data tensor and the obtained set of tensor factors;
applying, by the processor, the determined patient tensor factor to a classifier such that the determined tensor factor establishes a feature vector for the patient, the classifier being configured to process the feature vector to generate the assessment; and
providing, by the processor, output data indicative of the assessment.
2. The method of claim 1 , wherein the set of tensor factors comprises a tensor factor indicative of a correlation of the decomposition to multiple predetermined values of a noise level parameter.
3. The method of claim 1 , wherein the set of tensor factors comprises a tensor factor indicative of a correlation of the decomposition to timing windows for the plurality of tests.
4. The method of claim 1 , wherein determining the patient tensor factor comprises solving a least squares problem to minimize a difference between the patient data tensor and a tensor product of the obtained tensor factors and the patient tensor factor.
5. The method of claim 1 , further comprising forming the feature vector for the patient by combining the determined patient tensor factor with electronic health record data for the patient.
6. The method of claim 1 , wherein the obtained set of tensor factors is indicative of training data processed via a procedure in which the training data is clustered in accordance with presence of the medical condition.
7. The method of claim 1 , wherein the plurality of tests comprises an electrocardiogram (ECG) test, a blood pressure test, and a photoplethysmography (PPG) test conducted during a time period window.
8. The method of claim 1 , wherein the low rank structure comprises a plurality of rank 1 tensors.
9. A method of generating an assessment of medical condition for a patient, the method comprising:
obtaining, by a processor, tensor data comprising one or more patient-based tensors for patient data indicative of a plurality of tests conducted on the patient and first and second class tensors for first and second patient classes with and without the medical condition, respectively;
adjusting, by the processor, dimensionality of either the one or more patient-based tensors or the first and second class tensors via a tensor decomposition so that the one or more patient-based tensors or the first and second class tensors have a same dimensionality, the tensor decomposition generating a set of tensor factors that amplify low rank structure;
computing, by the processor, first and second similarity scores for first and second tensor pairings of the one or more patient-based tensors relative to the first and second patient classes, respectively;
selecting, by the processor, the first patient class or the second patient class for the assessment based on the computed first and second similarity scores; and
providing, by the processor, output data indicative of the assessment.
10. The method of claim 9 , wherein adjusting the dimensionality comprises replacing a tensor factor generated via the tensor decomposition with a matrix that minimizes a distance between tensors for which the first and second similarity scores are computed.
11. The method of claim 9 , wherein obtaining the tensor data comprises stacking the patient data onto the first and second class tensors to create first and second patient-based tensors.
12. The method of claim 9 , wherein computing the first and second similarity scores further comprises determining a normalized inner product of tensors for which the first and second similarity scores are computed.
13. The method of claim 9 , wherein computing the first and second similarity scores further comprises implementing a canonical polyadic (CP) decomposition procedure on tensors for which the first and second similarity scores are computed.
14. The method of claim 9 , wherein selecting the first patient class or the second patient class comprises determining which of the first patient class and the second patient class has a higher similarity score.
15. The method of claim 9 , wherein the low rank structure comprises a plurality of rank 1 tensors.
16. A method of de-noising tensor data in preparation for a medical condition assessment, the method comprising:
obtaining, by a processor, the tensor data;
estimating, by the processor, a noise level of the tensor data using tensor amplification-based dimension reduction;
implementing, by the processor, a decomposition by iteratively amplifying low rank structure of the tensor data while an error term of the decomposition is greater than the estimated noise level; and
providing, by the processor, tensor factors of the decomposition when the error term is equal to or less than the estimated noise level, or a de-noised version of the tensor data based on the tensor factors.
17. The method of claim 16 , wherein implementing the decomposition comprises implementing a tensor amplification-based tensor decomposition.
18. The method of claim 16 , wherein the low rank structure comprises a plurality of rank 1 tensors.
19. The method of claim 16 , wherein estimating the noise level using tensor amplification-based dimension reduction comprises generating a set of tensor factors, each tensor factor of the set of tensor factors being indicative of a decomposition of the tensor data.
20. The method of claim 16 , wherein the tensor data is indicative of physiological data.