Recurrent neural network to decode trial criteria

1. A method for correcting errors in textual data, comprising:

training a first recurrent neural network (RNN) configured as encoder and a second RNN configured as a decoder to improve output accuracy based on an unrefined input criteria containing one or more errors and a corresponding already-curated output criteria, comprising:

generating a partially-curated output criteria by processing the unrefined input criteria using the first RNN and the second RNN; and

refining a set of weights of the first RNN and the second RNN based on a difference between the partially-curated output criteria and the already-curated output criteria;

identifying a first input criterion for the clinical trial, wherein the first input criterion comprises a first textual statement pertaining to a medical condition of a patient, and wherein the first input criterion contains one or more errors;

encoding the first input criterion using the trained first RNN configured as an encoder, wherein the encoder extracts key features of the medical condition of the patient and converts the key features to a numerical representation;

generating a curated output criterion by processing the encoded first input criterion with the trained second RNN configured as a decoder, wherein the curated output criterion comprises a second textual statement pertaining to the medical condition of a patient derived from the key features, and wherein the curated output criterion does not contain the one or more errors; and

improving qualification scoring of a machine learning model by ingesting the curated output criterion into the machine learning model to identify the one or more candidates for the clinical trial, wherein ingesting comprises applying natural language processing to the curated output criterion to extract concepts and logical relationships pertaining to the medical condition of the patient.

2. The method of claim 1 , wherein the curated output criterion adheres to a preferred set of logical rules and format.

3. The method of claim 1 , wherein the key features comprise a set of attributes of the first statement comprising textual alphanumerical words, and punctuation, and a set of logical rules pertaining to the set of attributes.

4. The method of claim 1 , further comprising iteratively training the encoder and the decoder using a plurality of unrefined input criteria associated with a plurality of clinical trials and a corresponding plurality of already-curated output criteria associated with the plurality of clinical trials.

5. The method of claim 4 , further comprising:

identifying the plurality of unrefined input criteria, wherein each of the unrefined input criteria comprises an unrefined textual statement pertaining to one or more medical conditions of one or more patients; and

identifying the plurality of already-curated output criteria, wherein each of the already-curated output criterial comprises a curated textual statement pertaining to the one or more medical conditions of the one or more patients.

6. The method of claim 5 , wherein iteratively training the encoder and the decoder comprises:

encoding and decoding, using the encoder and the decoder, each of the plurality of unrefined input criteria into a corresponding plurality of partially curated output criteria, respectively;

determining one or more textual differences between each of the partially curated output criteria and a corresponding curated output criteria;

applying a training algorithm to the encoder and the decoder to incrementally adjust weights of nodes within the encoder and the decoder for each of the partially curated output criteria; and

determining whether the weights have converged based on the determined one or more textual differences for the plurality of partially curated output criteria falling below a predetermined threshold.

7. The method of claim 6 , further comprising repeating said encoding, said first determining, said applying, and said second determining using the partially curated output criteria and the already-curated output criteria to re-adjust the weights until the weights converge.

8. The method of claim 6 , wherein the training algorithm is a back propagation neural network training algorithm.

9. The method of claim 1 , further comprising periodically retraining the encoder and the decoder using a collection of unrefined input criteria and curated output criteria already processed by the encoder and the decoder associated with a plurality of clinical trials.

10. The method of claim 1 , further comprising providing a minimum size window of text surrounding each word of the input criterion to provide the encoder with a context for the input criterion.

11. A system, comprising:

a processor; and

a memory storing program code, which, when executed on the processor, performs operations for correcting errors in textual data, the operations comprising:

training a first recurrent neural network (RNN) configured as encoder and a second RNN configured as a decoder to improve output accuracy based on an unrefined input criteria containing one or more errors and a corresponding already-curated output criteria, comprising:

generating a partially-curated output criteria by processing the unrefined input criteria using the first RNN and the second RNN; and

refining a set of weights of the first RNN and the second RNN based on a difference between the partially-curated output criteria and the already-curated output criteria;

identifying a first input criterion for the clinical trial, wherein the first input criterion comprises a first textual statement pertaining to a medical condition of a patient, and wherein the first input criterion contains one or more errors;

encoding the first input criterion using the trained first RNN configured as an encoder, wherein the encoder extracts key features of the medical condition of the patient and converts the key features to a numerical representation;

generating a curated output criterion by processing the encoded first input criterion with the trained second RNN configured as a decoder, wherein the curated output criterion comprises a second textual statement pertaining to the medical condition of a patient derived from the key features, and wherein the curated output criterion does not contain the one or more errors; and

improving qualification scoring of a machine learning model by ingesting the curated output criterion into the machine learning model to identify the one or more candidates for the clinical trial, wherein ingesting comprises applying natural language processing to the curated output criterion to extract concepts and logical relationships pertaining to the medical condition of the patient.

12. The system of claim 11 , further comprising iteratively training the encoder and the decoder using a plurality of unrefined input criteria associated with a plurality of clinical trials and a corresponding plurality of curated output criteria associated with the plurality of clinical trials.

13. The system of claim 12 , further comprising:

identifying the plurality of unrefined input criteria, wherein each of the unrefined input criteria comprises an unrefined textual statement pertaining to one or more medical conditions of one or more patients; and

identifying the plurality of curated output criteria, wherein each of the curated output criterial comprises a curated textual statement pertaining to the one or more medical conditions of the one or more patients.

14. The system of claim 13 , wherein iteratively training the encoder and the decoder comprises:

encoding and decoding, using the encoder and the decoder, each of the plurality of unrefined input criteria into a corresponding plurality of partially curated output criteria, respectively;

determining one or more textual differences between each of the partially curated output criteria and a corresponding curated output criteria;

applying a training algorithm to the encoder and the decoder to incrementally adjust weights of nodes within the encoder and the decoder for each of the partially curated output criteria; and

determining whether the weights have converged based on the determined one or more textual differences for the plurality of partially curated output criteria falling below a predetermined threshold.

15. The system of claim 14 , further comprising repeating said encoding, said first determining, said applying, and said second determining using the partially curated output criteria and the curated output criteria to re-adjust the weights until the weights converge.

16. A computer program product for correcting errors in textual data, the computer program product comprising:

a computer-readable storage medium having computer-readable program code embodied therewith, the computer-readable program code executable by one or more computer processors to perform an operation comprising:

training a first recurrent neural network (RNN) configured as encoder and a second RNN configured as a decoder to improve output accuracy based on an unrefined input criteria containing one or more errors and a corresponding already-curated output criteria, comprising:

generating a partially-curated output criteria by processing the unrefined input criteria using the first RNN and the second RNN; and

refining a set of weights of the first RNN and the second RNN based on a difference between the partially-curated output criteria and the already-curated output criteria;

identifying a first input criterion for the clinical trial, wherein the first input criterion comprises a first textual statement pertaining to a medical condition of a patient, and wherein the first input criterion contains one or more errors;

encoding the first input criterion using the trained first RNN configured as an encoder, wherein the encoder extracts key features of the medical condition of the patient and converts the key features to a numerical representation;

generating a curated output criterion by processing the encoded first input criterion with the trained second RNN configured as a decoder, wherein the curated output criterion comprises a second textual statement pertaining to the medical condition of a patient derived from the key features, and wherein the curated output criterion does not contain the one or more errors; and

improving qualification scoring of a machine learning model by ingesting the curated output criterion into the machine learning model to identify the one or more candidates for the clinical trial, wherein ingesting comprises applying natural language processing to the curated output criterion to extract concepts and logical relationships pertaining to the medical condition of the patient.

17. The computer program product of claim 16 , the operation further comprising iteratively training the encoder and the decoder using a plurality of unrefined input criteria associated with a plurality of clinical trials and a corresponding plurality of curated output criteria associated with the plurality of clinical trials.

18. The computer program product of claim 17 , the operation further comprising:

identifying the plurality of unrefined input criteria, wherein each of the unrefined input criteria comprises an unrefined textual statement pertaining to one or more medical conditions of one or more patients; and

identifying the plurality of curated output criteria, wherein each of the curated output criterial comprises a curated textual statement pertaining to the one or more medical conditions of the one or more patients.

19. The computer program product of claim 18 , wherein iteratively training the encoder and the decoder comprises:

encode and decode, using the encoder and the decoder, each of the plurality of unrefined input criteria into a corresponding plurality of partially curated output criteria, respectively;

determine one or more textual differences between each of the partially curated output criteria and a corresponding curated output criteria;

apply a training algorithm to the encoder and the decoder to incrementally adjust weights of nodes within the encoder and the decoder for each of the partially curated output criteria; and

determine whether the weights have converged based on the determined one or more textual differences for the plurality of partially curated output criteria falling below a predetermined threshold.

20. The computer program product of claim 19 , the operation further comprising repeating said encoding, said first determining, said applying, and said second determining using the partially curated output criteria and the curated output criteria to re-adjust the weights until the weights converge.