Analyzing health events using recurrent neural networks

• US 9,652,712 B2
• Filed: 07/27/2015
• Issued: 05/16/2017
• Est. Priority Date: 07/27/2015
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Corrado, Gregory Sean, Dean, Jeffrey Adgate
• Primary Examiner(s): Hill, Stanley K
• Assistant Examiner(s): BUSS, BENJAMIN J
• Application Number:
  US14/810,384
• Publication Number:
  US 20170032243A1
• Time in Patent Office: 659 Days
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 3/10 Interfaces, programming languages or software development kits, e.g. for simulating neural networks
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: G16H 50/30 for calculating health indices; for individual health risk assessment

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for using recurrent neural networks to analyze health events. One of the methods includes: processing each of a plurality of initial temporal sequences of health events to generate, for each of the initial temporal sequences, a respective network internal state of a recurrent neural network for each time step in the initial temporal sequence; storing, for each of the initial temporal sequences, one or more of the network internal states for the time steps in the temporal sequence in a repository; obtaining a first temporal sequence; processing the first temporal sequence using the recurrent neural network to generate a sequence internal state for the first temporal sequence; and selecting one or more initial temporal sequences that are likely to include health events that are predictive of future health events in the first temporal sequence.

Claims

1. A method comprising:

obtaining a plurality of initial temporal sequences of health events,

wherein each of the plurality of initial temporal sequences of health events is associated with a different patient,

wherein each of the initial temporal sequences comprises respective health-related data associated with the respective patient that is associated with the initial temporal sequence at each of a plurality of time steps, and

wherein, for one or more of the time steps in each of the initial temporal sequences, the health-related data at the time step is a respective token from a predetermined vocabulary of tokens, each token in the vocabulary representing a different health event;

processing each of the plurality of initial temporal sequences of health events using a recurrent neural network to generate, for each of the initial temporal sequences, a respective network internal state of the recurrent neural network for each time step in the initial temporal sequence,

wherein the recurrent neural network has been trained to receive input temporal sequences and, for each time step in each input temporal sequence, generate a network internal state for the time step and predict future events occurring after the health event identified at the time step from the network internal state for the time step;

storing, for each of the plurality of initial temporal sequences, one or more of the network internal states for the time steps in the temporal sequence in an internal state repository;

obtaining a first temporal sequence of health events that is associated with a current patient;

processing the first temporal sequence of health events using the recurrent neural network to generate a sequence internal state for the first temporal sequence; and

comparing the sequence internal state for the first temporal sequence to the network internal states for the initial temporal sequences that are stored in the internal state repository to determine a plurality of network internal states from the network internal states for the initial temporal sequences that are similar to the sequence internal state;

selecting, as temporal sequences that are likely to include health events that are predictive of future health events that may become associated with the current patient, the initial temporal sequences corresponding to the similar network internal states.

2. The method of claim 1, wherein determining network internal states in the internal state repository that are similar to the sequence internal state comprises:

computing, for each of the network internal states in the internal state repository, a respective similarity measure between the network internal state and the sequence internal state; and

determining the similar network internal states from the similarity measures.

3. The method of claim 1, further comprising:

associating each network internal state in the internal state repository with a respective time step and a respective initial temporal sequence for which the network internal state was generated.

4. The method of claim 3, further comprising:

providing, for presentation to a user, data identifying, for each of the selected initial temporal sequences, the health events in the selected initial temporal sequences that are at time steps subsequent to the time step for which the corresponding network internal state was generated.

5. The method of claim 3, further comprising:

computing, from the health events in the selected initial temporal sequences that are at time steps subsequent to the time step for which the corresponding network internal state was generated, a statistic for a particular health event identifying a frequency of occurrence of the particular health event; and

providing the computed statistic for presentation to a user.

6. The method of claim 1, wherein the recurrent neural network is trained to generate, for each of the plurality of time steps in each input training sequence, a respective score for each of a plurality of possible health events from the network internal state for the time step, wherein the respective score for each of the possible health events represents a likelihood that the possible health event is a health event at a time step subsequent to the time step in the input training sequence.

7. The method of claim 1, wherein processing the first temporal sequence of health events using the recurrent neural network to generate a sequence internal state for the first temporal sequence comprises:

for each time step in the first temporal sequence, processing the data identifying the health event for the time step using the recurrent neural network to generate a network internal state for the time step; and

selecting the network internal state for a last time step in the first temporal sequence as the sequence internal state for the first temporal sequence.

8. The method of claim 1, wherein, for one or more of the time steps in each of the initial temporal sequences, the health-related data at the time step is other health-related data classified as impacting the health of the respective patient.

9. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

obtaining a plurality of initial temporal sequences of health events,

wherein each of the plurality of initial temporal sequences of health events is associated with a different patient,

wherein each of the initial temporal sequences comprises respective health-related data associated with the respective patient that is associated with the initial temporal sequence at each of a plurality of time steps, and

wherein, for one or more of the time steps in each of the initial temporal sequences, the health-related data at the time step is a respective token from a predetermined vocabulary of tokens, each token in the vocabulary representing a different health event;

processing each of the plurality of initial temporal sequences of health events using a recurrent neural network to generate, for each of the initial temporal sequences, a respective network internal state of the recurrent neural network for each time step in the initial temporal sequence,

wherein the recurrent neural network has been trained to receive input temporal sequences and, for each time step in each input temporal sequence, generate a network internal state for the time step and predict future events occurring after the health event identified at the time step from the network internal state for the time step;

storing, for each of the plurality of initial temporal sequences, one or more of the network internal states for the time steps in the temporal sequence in an internal state repository;

obtaining a first temporal sequence of health events that is associated with a current patient;

processing the first temporal sequence of health events using the recurrent neural network to generate a sequence internal state for the first temporal sequence; and

comparing the sequence internal state for the first temporal sequence to the network internal states for the initial temporal sequences that are stored in the internal state repository to determine a plurality of network internal states from the network internal states for the initial temporal sequences that are similar to the sequence internal state;

selecting, as temporal sequences that are likely to include health events that are predictive of future health events that may become associated with the current patient, the initial temporal sequences corresponding to the similar network internal states.

10. The system of claim 9, wherein determining network internal states in the internal state repository that are similar to the sequence internal state comprises:

computing, for each of the network internal states in the internal state repository, a respective similarity measure between the network internal state and the sequence internal state; and

determining the similar network internal states from the similarity measures.

11. The system of claim 9, the operations further comprising:

associating each network internal state in the internal state repository with a respective time step and a respective initial temporal sequence for which the network internal state was generated.

12. The system of claim 11, the operations further comprising:

providing, for presentation to a user, data identifying, for each of the selected initial temporal sequences, the health events in the selected initial temporal sequences that are at time steps subsequent to the time step for which the corresponding network internal state was generated.

13. The system of claim 11, the operations further comprising:

computing, from the health events in the selected initial temporal sequences that are at time steps subsequent to the time step for which the corresponding network internal state was generated, a statistic for a particular health event identifying a frequency of occurrence of the particular health event; and

providing the computed statistic for presentation to a user.

14. The system of claim 9, wherein the recurrent neural network is trained to generate, for each of the plurality of time steps in each input training sequence, a respective score for each of a plurality of possible health events from the network internal state for the time step, wherein the respective score for each of the possible health events represents a likelihood that the possible health event is a health event at a time step subsequent to the time step in the input training sequence.

15. The system of claim 9, wherein processing the first temporal sequence of health events using the recurrent neural network to generate a sequence internal state for the first temporal sequence comprises:

for each time step in the first temporal sequence, processing the data identifying the health event for the time step using the recurrent neural network to generate a network internal state for the time step; and

selecting the network internal state for a last time step in the first temporal sequence as the sequence internal state for the first temporal sequence.

16. The system of claim 9, wherein, for one or more of the time steps in each of the initial temporal sequences, the health-related data at the time step is other health-related data classified as impacting the health of the respective patient.

17. A computer program product encoded on one or more non-transitory computer readable media, the computer program product comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

obtaining a plurality of initial temporal sequences of health events,

wherein each of the plurality of initial temporal sequences of health events is associated with a different patient,

wherein each of the initial temporal sequences comprises respective health-related data associated with the respective patient that is associated with the initial temporal sequence at each of a plurality of time steps, and

wherein, for one or more of the time steps in each of the initial temporal sequences, the health-related data at the time step is a respective token from a predetermined vocabulary of tokens, each token in the vocabulary representing a different health event;

processing each of the plurality of initial temporal sequences of health events using a recurrent neural network to generate, for each of the initial temporal sequences, a respective network internal state of the recurrent neural network for each time step in the initial temporal sequence,

wherein the recurrent neural network has been trained to receive input temporal sequences and, for each time step in each input temporal sequence, generate a network internal state for the time step and predict future events occurring after the health event identified at the time step from the network internal state for the time step;

storing, for each of the plurality of initial temporal sequences, one or more of the network internal states for the time steps in the temporal sequence in an internal state repository;

obtaining a first temporal sequence of health events that is associated with a current patient;

processing the first temporal sequence of health events using the recurrent neural network to generate a sequence internal state for the first temporal sequence; and

comparing the sequence internal state for the first temporal sequence to the network internal states for the initial temporal sequences that are stored in the internal state repository from the network internal states for the initial temporal sequences to determine a plurality of network internal states that are similar to the sequence internal state;

selecting, as temporal sequences that are likely to include health events that are predictive of future health events that may become associated with the current patient, the initial temporal sequences corresponding to the similar network internal states.

18. The computer program product of claim 17, wherein selecting one or more initial temporal sequences comprises:

determining network internal states in the internal state repository that are similar to the sequence internal state; and

selecting the initial temporal sequences for which the similar network internal states were generated as the initial temporal sequences from the plurality of initial temporal sequences that are likely to include health events that are predictive of future health events in the first temporal sequence.