HIERARCHICAL CNN-TRANSFORMER BASED MACHINE LEARNING
Clinical prediction models often use structured variables and provide outcomes that are not readily interpretable by clinicians. Further, text medical notes may contain information not immediately available in structured variables. Applicants propose a hierarchical CNN-Transformer model with an explicit attention mechanism as an interpretable, multi-task clinical language model.
1 . A system for conducting machine learning on text inputs having varying sequence lengths, the system comprising:
a computer processor configured to:
receive an input data set representing a text input relating to a healthcare event,
split the input data set into n portions of tokens;
provide the n portions into a transformer-based data model architecture to generate a transformer-based feature extractor data object X;
provide the transformer-based feature extractor data object X into a convolutional neural network to obtain an n×d feature matrix, S, where S=MaxPool(ReLU(Conv(X))), and d feature is a number of output channels of the convolutional neural network;
apply a transformer-based encoder to obtain a matrix, S T =Transformer(S) that shares the same dimension as S;
determine, using at least S T , a plurality of latent attention scores corresponding to each portion of the n portions using a neural network; and
encapsulate the plurality of latent attention scores as output data objects.
2 . The system of claim 1 , wherein the latent attention scores are determined from the n portions of tokens using a position-wise feed-forward network, where given S T , an n-dimensional vector α input is computed as:
a input =FeedForward( S T );
and an attention weight is determined using the relation:
a =Softmax( a input +a mask );
wherein α mask is a n-dimensional vector having values for unmasked positions and values for padding positions, and the attention weight is utilized to establish a n feature -dimensional patient vector p that is computed as a weighted sum of the sentence features.
3 . The system of claim 2 , wherein projection-weighted canonical correlation analysis is utilized to determine a correlation between learned textual features of transformer-based feature extractor data object X and one or more structured variable data.
4 . The system of claim 3 , wherein the projection-weighted canonical correlation analysis is utilized to normalize one or more importance weights that modify the attention weight corresponding to each of latent attention score of the plurality of latent attention scores.
5 . The system of claim 1 , wherein a display interface is controlled based on the plurality of latent attention scores and a visual characteristic associated with each portion of the n portions of the input data set is modified based on a corresponding latent attention score of the plurality of latent attention scores.
6 . The system of claim 5 , wherein the visual characteristic includes at least one of a size factor, a color factor, or an opacity factor.
7 . The system of claim 1 , wherein the transformer-based data model architecture is a neural network maintained on computer memory, and wherein the computer processor is further configured to conduct supervised training of the transformer-based data model architecture based on received data sets representative of evaluator classifications.
8 . The system of claim 1 , wherein the output data objects are utilized to augment one or more electronic medical records with the one or more predicted classifications.
9 . The system of claim 8 , wherein the output data objects include additional metadata representative of the latent attention scores such that the augmented one or more electronic medical records are adapted for improved explainability for the corresponding one or more predicted classifications.
10 . The system of claim 9 , wherein the additional metadata is used to cause rendered visual effects when a computer system renders the augmented one or more electronic medical records for display, the rendered visual effects including a modified visual characteristic controlled based at least on a corresponding latent attention score of the plurality of latent attention scores for each of the n portions of the input data set.
11 . A method for conducting machine learning on text inputs having varying sequence lengths, the method comprising:
receiving an input data set representing a text input relating to a healthcare event, splitting the input data set into n portions of tokens;
providing the n portions into a transformer-based data model architecture to generate a transformer-based feature extractor data object X;
providing the transformer-based feature extractor data object X into a convolutional neural network to obtain an n×d feature matrix, S, where S=MaxPool(ReLU(Conv(X))), and d feature is a number of output channels of the convolutional neural network;
applying a transformer-based encoder to obtain a matrix, S T =Transformer(S) that shares the same dimension as S;
determining, using at least S T , a plurality of latent attention scores corresponding to each portion of the n portions using a neural network; and
encapsulating the plurality of latent attention scores as output data objects.
12 . The method of claim 11 , wherein the latent attention scores are determined from the n portions of tokens using a position-wise feed-forward network, where given S T , an n-dimensional vector α input is computed as:
a input =Feedforward( S T );
and an attention weight is determined using the relation:
a =Softmax( a input +a mask );
wherein α mask is a n-dimensional vector having values for unmasked positions and values for padding positions, and the attention weight is utilized to establish a n feature -dimensional patient vector p that is computed as a weighted sum of the sentence features.
13 . The method of claim 12 , wherein projection-weighted canonical correlation analysis is utilized to determine a correlation between learned textual features of transformer-based feature extractor data object X and one or more structured variable data.
14 . The method of claim 13 , wherein the projection-weighted canonical correlation analysis is utilized to normalize one or more importance weights that modify the attention weight corresponding to each of latent attention score of the plurality of latent attention scores.
15 . The method of claim 11 , wherein a display interface is controlled based on the plurality of latent attention scores and a visual characteristic associated with each portion of the n portions of the input data set is modified based on a corresponding latent attention score of the plurality of latent attention scores.
16 . The method of claim 15 , wherein the visual characteristic includes at least one of a size factor, a color factor, or an opacity factor.
17 . The method of claim 11 , wherein the transformer-based data model architecture is a neural network maintained on computer memory, and wherein the computer processor is further configured to conduct supervised training of the transformer-based data model architecture based on received data sets representative of evaluator classifications.
18 . The method of claim 11 , wherein the output data objects are utilized to augment one or more electronic medical records with the one or more predicted classifications.
19 . The method of claim 18 , wherein the output data objects include additional metadata representative of the latent attention scores such that the augmented one or more electronic medical records are adapted for improved explainability for the corresponding one or more predicted classifications.
20 . A non-transitory computer readable medium storing machine interpretable instructions, which when executed by a computer processor, cause the computer processor to perform steps of a method for conducting machine learning on text inputs having varying sequence lengths, the method comprising:
receiving an input data set representing a text input relating to a healthcare event,
splitting the input data set into n portions of tokens;
providing the n portions into a transformer-based data model architecture to generate a transformer-based feature extractor data object X;
providing the transformer-based feature extractor data object X into a convolutional neural network to obtain an n×d feature matrix, S, where S=MaxPool(ReLU(Conv(X))), and d feature is a number of output channels of the convolutional neural network;
applying a transformer-based encoder to obtain a matrix, S T =Transformer(S) that shares the same dimension as S;
determining, using at least S T , a plurality of latent attention scores corresponding to each portion of the n portions using a neural network; and
encapsulating the plurality of latent attention scores as output data objects.