Selecting speech features for building models for detecting medical conditions

• US 10,152,988 B2
• Filed: 05/07/2018
• Issued: 12/11/2018
• Est. Priority Date: 05/05/2017
• Status: Active Grant

• Current Assignee: Canary Speech, LLC
• Original Assignee: Canary Speech, LLC
• Inventors: Kim, Jangwon, Kwon, Namhee, O'Connell, Henry, Walstad, Phillip, Yang, Kevin Shengbin
• Primary Examiner(s): Shah, Paras D
• Application Number:
  US15/973,498
• Publication Number:
  US 20180322894A1
• Time in Patent Office: 218 Days
• CPC Class Codes: A61B 5/1123 Discriminating type of movement, e.g. walking or running A61B5/1116, A61B5/112 take precedence
• CPC Class Codes: A61B 5/4088 Diagnosing of monitoring cognitive diseases, e.g. Alzheimer, prion diseases or dementia
• CPC Class Codes: A61B 5/4803 Speech analysis specially adapted for diagnostic purposes
• CPC Class Codes: A61B 5/7267 involving training the classification device
• CPC Class Codes: G06F 2111/10 Numerical modelling
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 25/66 for extracting parameters related to health condition detecting or measuring for diagnostic purposes A61B5/00
• CPC Class Codes: G16H 10/20 for electronic clinical trials or questionnaires
• CPC Class Codes: G16H 40/67 for remote operation
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: G16H 50/50 for simulation or modelling of medical disorders
• CPC Class Codes: G16H 80/00 ICT specially adapted for facilitating communication between medical practitioners or patients, e.g. for collaborative diagnosis, therapy or health monitoring

Abstract

A mathematical model may be trained to diagnose a medical condition of a person by processing acoustic features and language features of speech of the person. The performance of the mathematical model may be improved by appropriately selecting the features to be used with the mathematical model. Features may be selected by computing a feature selection score for each acoustic feature and each language feature, and then selecting features using the scores, such as by selecting features with the highest scores. In some implementations, stability determinations may be computed for each feature and features may be selected using both the feature selection scores and the stability determinations. A mathematical model may then be trained using the selected features and deployed. In some implementations, prompts may be selected using computed prompt selection scores, and the deployed mathematical model may be used with the selected prompts.

Claims

1. A system for training a mathematical model for detecting a medical condition, the system comprising at least one computer configured to:

obtain a training corpus comprising speech data items, wherein each speech data item is labelled with a diagnosis value;

obtain speech recognition results for each speech data item using automatic speech recognition, wherein the speech recognition results for a speech data item comprise a transcription of the speech data item;

compute a plurality of acoustic features for each speech data item in the training corpus, wherein the plurality of acoustic features is computed from the speech data item and wherein computation of the plurality of acoustic features does not use the speech recognition results of the speech data item;

compute a plurality of language features for each speech data item in the training corpus by processing the speech recognition results;

compute a feature selection score for each feature of the plurality of acoustic features and each feature of the plurality of language features, wherein;

the feature selection score for a feature indicates a usefulness of the feature for detecting the medical condition, and

the feature selection score is computed using, for each speech data item, a value of the feature and the diagnosis value corresponding to the speech data item;

select a plurality of features from the plurality of acoustic features and the plurality of language features using the feature selection scores;

train the mathematical model for detecting the medical condition using the selected plurality of features for each speech data item of the training corpus;

deploy a computer program product or computer service for detecting the medical condition using the mathematical model;

present, by the computer program product or computer service, a prompt to a person;

receive, by the computer program product or computer service, a speech data item corresponding to speech of a person in response to the prompt;

compute a medical diagnosis score by processing the received speech data item using the mathematical model; and

display, by the computer program product or computer service, one or more of the medical diagnosis score or a medical diagnosis based on the medical diagnosis score.

2. The system of claim 1, wherein each speech data item of the training corpus corresponds to a prompt of a plurality of prompts, the plurality of prompts comprising the presented prompt, and wherein the at least one computer is configured to:

compute a medical diagnosis score for each speech data item of the training corpus by processing the speech data items with the mathematical model;

compute a prompt selection score for each prompt of the plurality of prompts using the medical diagnosis scores;

select a subset of prompts from the plurality of prompts using the prompt selection scores, the subset of prompts comprising the presented prompt;

deploy the computer program product or computer service for detecting the medical condition using the mathematical model and the subset of prompts;

receive a speech data item corresponding to speech of the person for each prompt of the subset of prompts; and

compute the medical diagnosis score for the person by processing the received speech data items using the mathematical model.

3. The system of claim 1, wherein the at least one computer is configured to:

divide the training corpus into a plurality of folds; and

compute a statistic for each feature and each fold of the plurality of folds.

4. The system of claim 3, wherein the at least one computer is configured to:

compute a stability determination for each feature of the plurality of acoustic features and the plurality of language features using the statistics for each feature and each fold of the plurality of folds; and

select the plurality of features using the stability determinations.

5. The system of claim 1, wherein the mathematical model comprises a neural network or a support vector machine.

6. The system of claim 1, wherein the plurality of acoustic features comprises at least one of spectral features, prosodic features, or voice quality features.

7. A computer-implemented method for training a mathematical model for detecting a medical condition, the method comprising:

obtaining a training corpus comprising speech data items, wherein each speech data item is labelled with a diagnosis value;

obtaining speech recognition results for each speech data item using automatic speech recognition, wherein the speech recognition results for a speech data item comprise a transcription of the speech data item;

computing a plurality of acoustic features for each speech data item in the training corpus, wherein the plurality of acoustic features is computed from the speech data item and wherein computation of the plurality of acoustic features does not use the speech recognition results of the speech data item;

computing a plurality of language features for each speech data item in the training corpus by processing the speech recognition results;

computing a feature selection score for each feature of the plurality of acoustic features and each feature of the plurality of language features, wherein;

the feature selection score for a feature indicates a usefulness of the feature for detecting the medical condition, and

the feature selection score is computed using, for each speech data item, a value of the feature and the diagnosis value corresponding to the speech data item;

selecting a plurality of features from the plurality of acoustic features and the plurality of language features using the feature selection scores;

training the mathematical model for detecting the medical condition using the selected plurality of features for each speech data item of the training corpus;

deploying a computer program product or computer service for detecting the medical condition using the mathematical model;

presenting, by the computer program product or computer service, a prompt to a person;

receiving, by the computer program product or computer service, a speech data item corresponding to speech of a person in response to the prompt;

computing a medical diagnosis score by processing the received speech data item using the mathematical model; and

displaying, by the computer program product or computer service, one or more of the medical diagnosis score or a medical diagnosis based on the medical diagnosis score.

8. The computer-implemented method of claim 7, wherein the medical condition is a concussion or Alzheimer'"'"'s disease.

9. The computer-implemented method of claim 7, wherein the plurality of language features comprises one or more of a number of pause fillers over a period of time, a number of pause fillers over a number of words, word difficulty, or speaking rate.

10. The computer-implemented method of claim 7, wherein computing a feature selection score for a feature comprises generating a pair of numbers for each speech data item of the training corpus, and wherein a first number of the pair corresponds to a feature value and a second number of the pair corresponds to a diagnosis value.

11. The computer-implemented method of claim 7, comprising:

dividing the training corpus into a plurality of folds; and

computing a statistic for each feature and each fold of the plurality of folds.

12. The computer-implemented method of claim 11, comprising:

computing a stability determination for each feature of the plurality of acoustic features and the plurality of language features using the statistics for each feature and each fold of the plurality of folds; and

selecting the plurality of features using the stability determinations.

13. The computer-implemented method of claim 7, comprising:

selecting a plurality of prompts using the mathematical model; and

training a second mathematical model using the selected plurality of prompts and the speech data items of the training corpus.

14. One or more non-transitory computer-readable media comprising computer executable instructions that, when executed, cause at least one processor to perform actions comprising:

obtaining a training corpus comprising speech data items, wherein each speech data item is labelled with a diagnosis value;

obtaining speech recognition results for each speech data item using automatic speech recognition, wherein the speech recognition results for a speech data item comprise a transcription of the speech data item;

computing a plurality of acoustic features for each speech data item in the training corpus, wherein the plurality of acoustic features is computed from the speech data item and wherein computation of the plurality of acoustic features does not use the speech recognition results of the speech data item;

computing a plurality of language features for each speech data item in the training corpus by processing the speech recognition results;

computing a feature selection score for each feature of the plurality of acoustic features and each feature of the plurality of language features, wherein;

the feature selection score for a feature indicates a usefulness of the feature for detecting a medical condition, and

the feature selection score is computed using, for each speech data item, a value of the feature and the diagnosis value corresponding to the speech data item;

selecting a plurality of features from the plurality of acoustic features and the plurality of language features using the feature selection scores;

training a mathematical model for detecting the medical condition using the selected plurality of features for each speech data item of the training corpus;

deploying a computer program product or computer service for detecting the medical condition using the mathematical model;

presenting, by the computer program product or computer service, a prompt to a person;

receiving, by the computer program product or computer service, a speech data item corresponding to speech of a person in response to the prompt;

computing a medical diagnosis score by processing the received speech data item using the mathematical model; and

displaying, by the computer program product or computer service, one or more of the medical diagnosis score or a medical diagnosis based on the medical diagnosis score.

15. The one or more non-transitory computer-readable media of claim 14, wherein computing a first acoustic feature of the plurality of acoustic features comprises:

computing a value for each short-time segment of an audio signal to obtain a plurality of values; and

computing the first acoustic feature using the plurality of values.

16. The one or more non-transitory computer-readable media of claim 14, wherein the feature selection score comprises an adjusted Rand index, adjusted mutual information, an absolute Pearson correlation, or an absolute Spearman correlation.

17. The one or more non-transitory computer-readable media of claim 14, wherein the actions comprise:

computing a stability determination for each feature of the plurality of acoustic features and the plurality of language features; and

selecting the plurality of features using the stability determinations.

18. The one or more non-transitory computer-readable media of claim 14, wherein each speech data item of the training corpus corresponds to a prompt of a plurality of prompts, the plurality of prompts comprising the presented prompt, and wherein the actions comprise:

computing a medical diagnosis score for each speech data item of the training corpus by processing the speech data items with the mathematical model;

computing a prompt selection score for each prompt of the plurality of prompts using the medical diagnosis scores;

selecting a subset of prompts from the plurality of prompts using the prompt selection scores, the subset of prompts comprising the presented prompt; and

deploying the computer program product or computer service for detecting the medical condition using the mathematical model and the subset of prompts.

19. The one or more non-transitory computer-readable media of claim 14, wherein the actions comprise:

obtaining a plurality of non-speech features for each speech data item in the training corpus;

computing a feature selection score for each feature of the plurality of non-speech features;

selecting the plurality of features using the feature selection scores for the plurality of non-speech features.