Machine-learned state space model for joint forecasting

Patent metadata

Field	Value
Publication number	US-12217144-B2
Application number	US-202017008338-A
Country	US
Kind code	B2
Filing date	Aug 31, 2020
Priority date	Aug 30, 2019
Publication date	Feb 4, 2025
Grant date	Feb 4, 2025

Abstract

A deep state space generative model is augmented with intervention prediction. The state space model provides a principled way to capture the interactions among observations, interventions, critical event occurrences, true states, and associated uncertainty. The state space model can include a discrete-time hazard rate model that provides flexible fitting of general survival time distributions. The state space model can output a joint prediction of event risk, observation and intervention trajectories based on patterns in temporal progressions, and correlations between past measurements and interventions.

First claim

What is claimed is:

1. A computing system configured for joint prediction of future time-series and time-to-event, the computing system comprising:
one or more processors; and
one or more non-transitory computer-readable media that collectively store:
a machine-learned state space model comprising at least a first machine-learned function that models a relationship between a past state of the machine-learned state space model and a likelihood of a future intervention associated with the machine-learned state space model and a second machine-learned function that models an effect of a past intervention on a future state of the machine-learned state space model, the machine-learned state space model configured to:
receive one or more input time series;
input, responsive to the receiving and based at least in part on the one or more input time series, to the first machine-learned function, data indicative of one or more states of the system at a first time associated with the one or more input time series;
subsequently receive, from the first machine-learned function based on the data indicative of the one or more states, data indicative of one or more probabilities that one or more future interventions will be performed on the system at a second time later than the first time;
input, responsive to the receiving, to the second machine-learned function of the machine-learned state space model, the data indicative of the one or more probabilities that the one or more future interventions will be performed on the system at the second time;
subsequently receive, from the second machine-learned function based on the data indicative of the one or more probabilities that the one or more future interventions will be performed on the system at the second time, data indicative of one or more predicted states of the system at one or more third times later than the second time; and
output, based at least in part on the data indicative of the one or more predicted states, one or more trajectory predictions and a time-to-event prediction, wherein the one or more trajectory predictions comprise a future time series for one or both of future observations of the system or future interventions performed on the system; and

instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:
inputting the one or more input time series into the machine-learned state space model; and
receiving the one or more trajectory predictions and the time-to-event prediction as an output of the machine-learned state space model.

2. The computing system of claim 1 , wherein:
the one or more input time series are extracted from one or more electronic medical records associated with a patient;
the one or more input time series comprise a plurality of clinical measurements of the patient; and
the one or more input time series comprise a plurality of clinical interventions performed on the patient.

3. The computing system of claim 1 , wherein an event associated with the time-to-event prediction comprises at least one of a mortality event or onset of acute condition.

4. The computing system of claim 1 , wherein the time-to-event prediction output by the machine-learned state space model comprises a hazard series defined at one or more points in time associated with the one or more trajectory predictions.

5. The computing system of claim 4 , wherein the operations further comprise determining a survival function based at least in part on the hazard series.

6. The computing system of claim 1 , wherein the machine-learned state space model is a non-linear state space model that comprises one or more neural networks.

7. The computing system of claim 6 , wherein the one or more neural networks comprise one or more multilayer perceptron networks that model state transition, observation, and emission rate functions.

8. The computing system of claim 1 , wherein the operations further comprise a pretraining step, wherein the pretraining step comprises:
identifying one or more generative parameters associated with the machine-learned state space model, the one or more generative parameters identified by variational inference of a variational distribution that approximates a probability distribution associated with the one or more generative parameters;
generating, based at least in part on the one or more generative parameters, a training data set; and
adjusting the one or more generative parameters based at least in part on a loss associated with the training data set.

9. The computing system of claim 8 , wherein the variational distribution is generated using an encoder.

10. The computing system of claim 1 , wherein an intervention associated with a first point in the one or more input time series is related to a state variable associated with a second point in the input time series via a transform equation.

11. The computing system of claim 1 , wherein a state transition associated with the machine-learned state space model is modeled by a probability distribution and wherein a variance of the probability distribution is modeled by a process and measurement noise covariance matrix.

12. The computing system of claim 1 , wherein the machine-learned state space model further comprises:
a third machine-learned function that models a relationship between a state of a machine learned model and a future state of the machine-learned model without external influence;
a fourth machine-learned function that models a relationship between the state of the machine-learned state space model and an observation;
and
a fifth machine-learned function that models a relationship between the state of the machine-learned state space model and a hazard rate.

13. A computer-implemented method of training a machine-learned state space model, the computer-implemented method comprising:
receiving, by one or more computing devices, an input time series comprising a plurality of observations;
inferring, by the one or more computing devices and based at least in part on the input time series, one or more first latent state variables associated with one or more first times of the input time series;
inferring, by the one or more computing devices and based at least in part on an encoder and based at least in part on the one or more first latent state variables, a plurality of probabilities of a plurality of interventions at one or more second times later than the one or more first times;
inferring, by the one or more computing devices and based at least in part on the plurality of probabilities, one or more second latent state variables associated with one or more third times later than the one or more second times;
generating, by the one or more computing devices and based at least in part on one or more generative parameters of the machine-learned state space model and based at least in part on the one or more second latent state variables, a forecast associated with the input time series;
estimating, by the one or more computing devices, a loss associated with the forecast; and
updating, by the one or more computing devices and based at least in part on the loss, the one or more generative parameters of the machine-learned state space model.

14. The method of claim 13 , wherein inferring the one or more latent state variables and inferring the plurality of probabilities of the plurality of interventions comprises:
generating, based at least in part on the encoder, a probability distribution approximating a posterior associated with the one or more latent state variables, the plurality of observations, and th

Assignees

• Google Llc

Inventors

• Xue Yuan
• Zhou Dengyong
• Du Nan
• Dai Andrew Mingbo
• Xu Zhen
• Zhang Kun
• Cui Yingwei

Classifications

• G06N3/0475

  Generative networks · 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

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title