System and method for attentional multimodal pain estimation
A computer-based system and method for generating a pain score of a subject using one or more sensory signals extracted from an AV signal of the subject. The AV signal may comprise one or more sensory signals including a face sensory signal, a body sensory signal and an audio sensory and wherein one or more of the sensory signals is missing from the AV signal.
1 . A method for generating a pain score for a subject, the method comprising:
receiving an audio/visual (AV) signal of a subject;
extracting one or more sensory signals from the AV signal, wherein the one or more sensory signals are selected from a face sensory signal, a body sensory signal and an audio sensory signal and wherein one or more of the sensory signals is missing from the AV signal;
reconstructing the one or more missing sensory signals from one or more of the sensory signals present in the AV signal to generate one or more reconstructed sensory signals; and
performing attentional fusion to generate a pain score for the subject from the one or more sensory signals extracted from the AV signal and the one or more reconstructed sensory signals, wherein performing attentional fusion comprises generating attentive features for the one or more sensory signals extracted from the AV signal and the one or more reconstructed sensory signals and concatenating the generated attentive features to generate the pain score.
2 . The method of claim 1 , wherein extracting the one or more sensory signals from the AV signal further comprises:
training a multimodal autoencoder to extract spatio-temporal features from AV signals under different missing sensory signal conditions;
extracting spatio-temporal features from the one or more sensory signals of the AV signal using the trained multimodal autoencoder; and
generating a spatio-temporal latent space from the extracted spatio-temporal features using the trained multimodal autoencoder.
3 . The method of claim 2 , wherein reconstructing the one or more missing sensory signals to generate one or more reconstructed sensory signals further comprises:
generating a joint probability latent space from the spatio-temporal latent space, wherein the joint probability latent space comprises latent features; and
reconstructing the one or more missing sensory signals from the joint probability latent space.
4 . The method of claim 1 , wherein the pain score comprises an intensity estimation.
5 . The method of claim 1 , further comprising, recording, with an audio/video (AV) recorder, facial expressions, body movements, and sounds of the subject, wherein said AV recorder comprises a video camera for recording video of the facial expressions and body movements and a microphone for recording sounds of the subject to generate the AV signal of the subject.
6 . A system for estimating pain that may be experienced by a subject, the system comprising:
an audio/video (AV) recorder comprising a video camera for recording video of the facial expressions and body movements of a subject and a microphone for recording sounds of a subject to generate an AV signal of the subject,
a processor running a machine learning algorithm for,
extracting one or more sensory signals from the AV signal, wherein the one or more sensory signals are selected from a face sensory signal, a body sensory signal and an audio sensory signal and wherein one or more of the sensory signals is missing from the AV signal;
reconstructing the one or more missing sensory signals from one or more of the sensory signals present in the AV signal to generate one or more reconstructed sensory signals; and
performing attentional fusion to generate a pain score for the subject from the one or more sensory signals extracted from the AV signal and the one or more reconstructed sensory signals, wherein performing attentional fusion comprises generating attentive features for the one or more sensory signals extracted from the AV signal and the one or more reconstructed sensory signals and concatenating the generated attentive features to generate the pain score.
7 . The system of claim 6 , wherein extracting the one or more sensory signals from the AV signal further comprises:
training a multimodal autoencoder to extract spatio-temporal features from AV signals under different missing sensory signal conditions;
extracting spatio-temporal features from the one or more sensory signals of the AV signal using the trained multimodal autoencoder; and
generating a spatio-temporal latent space from the extracted spatio-temporal features using the trained multimodal autoencoder.
8 . The system of claim 7 , wherein reconstructing the one or more missing sensory signals to generate one or more reconstructed sensory signals further comprises:
generating a joint probability latent space from the spatio-temporal latent space, wherein the joint probability latent space comprises latent features; and
reconstructing the one or more missing sensory signals from the joint probability latent space.
9 . The system of claim 6 , wherein the pain score comprises an intensity estimation.
10 . The system of claim 6 further comprising an output device for outputting the pain score.
11 . A non-transitory computer-readable medium storing a set of instructions configured for being executed by at least one processor for performing a method for generating a pain score for a subject, the method comprising:
receiving an audio/visual (AV) signal of a subject;
extracting one or more sensory signals from the AV signal, wherein the one or more sensory signals are selected from a face sensory signal, a body sensory signal and an audio sensory signal and wherein one or more of the sensory signals is missing from the AV signal;
reconstructing the one or more missing sensory signals from one or more of the sensory signals present in the AV signal to generate one or more reconstructed sensory signals;
and to generate one or more reconstructed sensory signals; and
performing attentional fusion to generate a pain score for the subject from the one or more sensory signals extracted from the AV signal and the one or more reconstructed sensory signals, wherein performing attentional fusion comprises generating attentive features for the one or more sensory signals extracted from the AV signal and the one or more reconstructed sensory signals and concatenating the generated attentive features to generate the pain score.
12 . The medium of claim 9 , wherein extracting the one or more sensory signals from the AV signal further comprises:
training a multimodal autoencoder to extract spatio-temporal features from AV signals under different missing sensory signal conditions;
extracting spatio-temporal features from the one or more sensory signals of the AV signal using the trained multimodal autoencoder; and
generating a spatio-temporal latent space from the extracted spatio-temporal features using the trained multimodal autoencoder.
13 . The medium of claim 12 , wherein reconstructing the one or more missing sensory signals to generate one or more reconstructed sensory signals further comprises:
generating a joint probability latent space from the spatio-temporal latent space, wherein the joint probability latent space comprises latent features; and
reconstructing the one or more missing sensory signals from the joint probability latent space.
14 . The medium of claim 9 , wherein the pain score comprises an intensity estimation.