Interpretable deep learning framework for mining and predictive modeling of health care data

Patent metadata

Field	Value
Publication number	US-11144825-B2
Application number	US-201715829768-A
Country	US
Kind code	B2
Filing date	Dec 1, 2017
Priority date	Dec 1, 2016
Publication date	Oct 12, 2021
Grant date	Oct 12, 2021

Abstract

A method for creating an interpretable model for healthcare predictions includes training, by a deep learning processor, a neural network to predict health information by providing training data, including multiple combinations of measured or observed health metrics and corresponding medical results, to the neural network. The method also includes determining, by the deep learning processor and using the neural network, prediction data including predicted results for the measured or observed health metrics for each of the multiple combinations of the measured or observed health metrics based on the training data. The method also includes training, by the deep learning processor or a learning processor, an interpretable machine learning model to make similar predictions as the neural network by providing mimic data, including combinations of the measured or observed health metrics and corresponding predicted results of the prediction data, to the interpretable machine learning model.

First claim

What is claimed is:

1. A method for computational phenotyping, comprising:
training a neural network having a prediction layer using an input including at least one of user provided input corresponding to health metrics or sensor input detected by a sensor and corresponding to the health metrics, and using a target corresponding to a desired output for the input;
determining at least one of soft prediction scores of the prediction layer of the neural network or activations of a highest hidden layer of the neural network corresponding to layer inputs of the highest hidden layer; and
training a gradient boosting tree model to mimic the neural network based on the input to the neural network and the at least one of the soft prediction scores of the prediction layer of the neural network or the activations of the highest hidden layer of the neural network.

2. The method of claim 1 further comprising providing a new input to the gradient boosting tree model to predict a new output.

3. The method of claim 1 wherein the neural network further includes at least two hidden layers and training the gradient boosting tree model to mimic the neural network further includes training the gradient boosting tree model using the activations of the highest hidden layer of the at least two hidden layers.

4. The method of claim 3 wherein training the gradient boosting tree model using the activations of the highest hidden layer of the at least two hidden layers further includes:
training a classifier using the activations of the highest hidden layer and the target;
determining classifier soft prediction scores based on the trained classifier; and
training the gradient boosting tree model to mimic the neural network based on the input to the neural network and the classifier soft prediction scores.

5. The method of claim 4 wherein the classifier includes logistic regression.

6. The method of claim 1 wherein training the gradient boosting tree model to mimic the neural network further includes training the gradient boosting tree model using the soft prediction scores of the prediction layer of the neural network.

7. The method of claim 1 wherein the neural network includes at least one of:
a feedforward network having multiple nonlinear layers such that a layer output of each nonlinear layer is used as a layer input for each subsequent nonlinear layer;
a stacked autoencoder network that includes encoder networks and decoder networks with tied weights to minimize a squared reconstruction loss to the input; or
a long short-term memory network that includes multiple blocks that each determine a cell state and a block output based on a previous cell state, a previous block output, and a current time series input.

8. The method of claim 7 wherein the long short-term memory network further includes a memory prediction layer configured to predict a memory output based on the block output of each of the multiple blocks.

9. The method of claim 1 wherein the gradient boosting tree model includes a plurality of weak learners each including at least one of a classifier or a regression tree, and includes a prediction function that predicts a tree output based on tree inputs and the plurality of weak learners.

Assignees

• Univ Southern California

Inventors

• Liu Yan
• Che Zhengping
• Purushotham Sanjay

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G16H50/70 Primary

  for mining of medical data, e.g. analysing previous cases of other patients · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title