Back to results
US 9,792,823 B2Granted
Multi-view learning in detection of psychological states
AssigneeRaytheon BBN Technologies Corp.
InventorsXiaodan Zhuang, Viktor Rozgic, Michael Roger Crystal
Life of this patent
Estimating the term…
Abstract
Systems and methods for using multi-view learning to leverage highly informative, high-cost, psychophysiological data collected in a laboratory setting, to improve post-traumatic stress disorder (PTSD) screening in the field, where only less-informative, low-cost, speech data are available. Partial least squares (PLS) methods can be used to learn a bilinear factor model, which projects speech and EEG responses onto latent spaces, where the covariance between the projections is maximized. The systems and methods use a speech representation based on a combination of audio and language descriptors extracted from spoken responses to open-ended questions.
Claims17
A system for detecting psychological states, the system comprising:
- a device to obtain sensor data corresponding to a subject's response to stimuli;
- a feature extraction processor to extract one or more features from the sensor data to generate feature representation data, the feature representation data being associated with a first view;
- a memory device configured to store a projection matrix modeling a relationship between the first view and a second view, and to store a detection model trained using training data associated with the first view transformed to a latent space using the projection matrix, wherein the first view is associated with low-cost data and the second view is associated with high-cost data;
- and
- a detection module to transform the feature representation data to the latent space using the projection matrix, and to apply the transformed feature representation data to the detection model to detect one or more psychological states of the subject,
- wherein the device to obtain sensor data comprises a first device to obtain first sensor data corresponding to a training subject's response to the stimuli and a second device to obtain second sensor data corresponding to the training subject's response to the stimuli, wherein the feature extraction processor is a first feature extraction processor, wherein the feature representation data is first feature representation data, the system further comprising:
- a second feature extraction processor to extract one or more features from the second sensor data to generate second feature representation data, the second feature representation data being associated with the second view;
- a relationship learning module to receive the first and second feature representation data and to learn the projection matrix for projecting the first feature representation to the latent space related to the first and second views;
- and
- a training module to transform the first feature representation data to the latent space using the projection matrix, and to train the detection model using the transformed feature representation data.
7 more elements
A method for detecting psychological states, the method comprising:
- obtaining sensor data associated with a first view, the sensor data corresponding to a subject's response to stimuli;
- performing feature extraction to generate feature representation data, a projection matrix modeling a relationship between the first view and a second view, wherein the first view is associated with low-cost data and the second view is associated with high-cost data;
- transforming the feature representation data to a latent space using the projection matrix;
- and
- applying a detection model to the transformed feature representation data to detect one or more psychological states of the subject,
- wherein obtaining sensor data associated with a first view comprises obtaining first sensor data using a first device, the first sensor data corresponding to a training subject's response to the stimuli, wherein the feature representation data is first feature representation data, and obtaining second sensor data using a second device, the second sensor data corresponding to the training subject's response to the stimuli, the method further comprising:
- performing feature extraction on the first sensor data to generate first feature representation data, the first feature representation data being associated with the first view:
- performing feature extraction on the second sensor data to generate second feature representation data, the second feature representation data being associated with a second view;
- learning the projection matrix using the first and second feature representation data;
- transforming the first feature representation data to the latent space using the projection matrix;
- and
- training the detection model using the transformed first feature representation data.
9 more elements
Dependencies are inferred from the claim text.