Superclass-conditional gaussian mixture model for personalized prediction on dialysis events
A computer-implemented method for model building is provided. The method includes receiving a training set of medical records and model hyperparameters. The method further includes initializing an encoder as a Dual-Channel Combiner Network (DCNN) and initialize distribution related parameters. The method also includes performing, by a hardware processor, a forward computation to (1) the DCNN to obtain the embeddings of the medical records, and (2) the distribution related parameters to obtain class probabilities. The method additionally includes checking by a convergence evaluator if the iterative optimization has converged. The method further includes performing model personalization responsive to model convergence by encoding the support data of a new patient and using the embeddings and event subtype labels to train a personalized classifier.
1 . A computer-implemented method for model building, comprising:
receiving a training set of medical records and model hyperparameters;
initializing an encoder, as a Dual-Channel Combiner Network (DCCN), and distribution related parameters;
building a model using a forward computation to (1) the encoder to obtain embeddings of the medical records by using a multilayer perceptron (MLP) of the DCCN to encode static patient profiles of the medical records and one or more Long Short-Term Memories (LSTMs) of the DCCN to encode temporal patient status features of the medical records, and (2) the distribution related parameters to obtain membership probabilities of a plurality of classes;
manipulating the embeddings of the medical records with the distribution related parameters to enable adaptation of the model to a fine-grained multi-class task;
performing iterative optimization of the model between (1) a step for obtaining posterior probabilities indicative of membership probabilities of the embeddings in one or more subclasses of each class of the plurality of classes, and (2) a step for obtaining updated encoder and distribution related model parameters while fixing one or more of the obtained posterior probabilities;
personalizing the model responsive to convergence of the iterative optimization with a final set of updated encoder and distribution related model parameters by encoding support data of a new patient to obtain new embeddings and using the new embeddings to train a personalized classifier for the new patient that is configured to perform the fine-grained multi-class task; and
performing model testing by encoding test data of the new patient to obtain test embedding and using the test embeddings and the personalized classifier to predict event subtypes.
2 . The computer-implemented method of claim 1 , wherein the forward computation comprises computing a loss function with the embeddings and the membership probabilities of the plurality of classes for optimization.
3 . The computer-implemented method of claim 1 , wherein the predicted event subtypes are used to support decision making for dialysis treatment planning.
4 . The computer-implemented method of claim 1 , further comprising performing a dialysis event on a patient responsive to the predicted event subtypes.
5 . The computer-implemented method of claim 1 , wherein outputs of the static channel and the temporal channel are concatenated and projected to a compact embedding used for prediction by a combination layer of the encoder.
6 . A non-transitory computer program product for model building, the non-transitory computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
receiving a training set of medical records and model hyperparameters;
initializing an encoder, as a Dual-Channel Combiner Network (DCCN), and distribution related parameters;
building, using a hardware processor, a model using a forward computation to (1) the encoder to obtain embeddings of the medical records by using a multilayer perceptron (MLP) of the DCCN to encode static patient profiles of the medical records and one or more Long Short-Term Memories (LSTMs) of the DCCN to encode temporal patient status features of the medical records, and (2) the distribution related parameters to obtain membership probabilities of a plurality of classes;
manipulating the embeddings of the medical records with the distribution related parameters to enable adaptation of the model to a fine-grained multi-class task;
performing iterative optimization of the model between (1) a step for obtaining posterior probabilities indicative of membership probabilities of the embeddings in one or more subclasses of each class of the plurality of classes, and (2) a step for obtaining updated encoder and distribution related model parameters while fixing one or more of the obtained posterior probabilities;
personalizing the model responsive to convergence of the iterative optimization with a final set of updated encoder and distribution related model parameters by encoding support data of a new patient to obtain new embeddings and using the new embeddings to train a personalized classifier for the new patient that is configured to perform the fine-grained multi-class task; and
performing model testing by encoding test data of the new patient to obtain test embedding and using the test embeddings and the personalized classifier to predict event subtypes.
7 . The non-transitory computer program product of claim 6 , wherein the forward computation comprises computing a loss function with the embeddings and the membership probabilities of the plurality of classes for optimization.
8 . The non-transitory computer program product of claim 6 , further comprising performing a dialysis event on a patient responsive to the predicted event subtypes.
9 . The non-transitory computer program product of claim 6 , wherein outputs of the static channel and the temporal channel are concatenated and projected to a compact embedding used for prediction by a combination layer of the encoder.
10 . A computer processing system for model building, comprising:
a memory device for storing program code; and
a hardware processor operatively coupled to the memory device for storing program code to
receive a training set of medical records and model hyperparameters;
initialize an encoder, as a Dual-Channel Combiner Network (DCCN), and distribution related parameters;
build, by the hardware processor, a model using a forward computation to (1) the encoder to obtain embeddings of the medical records by using a multilayer perceptron (MLP) of the DCCN to encode static patient profiles of the medical records and one or more Long Short-Term Memories (LSTMs) of the DCCN to encode temporal patient status features of the medical records, and (2) the distribution related parameters to obtain membership probabilities of a plurality of classes;
manipulate the embeddings of the medical records with the distribution related parameters to enable adaptation of the model to a fine-grained multi-class task;
perform iterative optimization of the model between (1) a step for obtaining posterior probabilities indicative of membership probabilities of the embeddings in one or more subclasses of each class of the plurality of classes, and (2) a step for obtaining updated encoder and distribution related model parameters while fixing one or more of the obtained posterior probabilities;
personalize the model responsive to convergence of the iterative optimization with a final set of updated encoder and distribution related model parameters by encoding support data of a new patient to obtain new embeddings and using the new embeddings to train a personalized classifier for the new patient that is configured to perform the fine-grained multi-class task; and
perform model testing by encoding test data of the new patient to obtain test embedding and using the test embeddings and the personalized classifier to predict event subtypes.
11 . The computer processing system of claim 10 , wherein the forward computation comprises computing a loss function with the embeddings and the membership probabilities of the plurality of classes for optimization.
12 . The computer processing system of claim 10 , wherein the processor further runs the program code to control performing a dialysis event on a patient responsive to the predicted event subtypes.
13 . The computer processing system of claim 10 , wherein outputs of the static channel and the temporal channel are concatenated and projected to a compact embedding used for prediction by a combination layer of the encoder.