US-12616379-B2 - System and method for filtering time-varying data for physiological signal prediction

US12616379B2US 12616379 B2US12616379 B2US 12616379B2US-12616379-B2

Abstract

Systems and methods for filtering time-varying data for filtering and extracting a predicted physiological signal. A method including: segmenting the time-varying data into temporal windows; using a trained filter machine learning model, predicting an error for each prediction of the physiological signal for each window of time-varying data, the filter machine learning model trained using physiological signal predictions based on training time-varying data and known values of the physiological signal for the training time-varying data; discarding each window of time-varying data when the predicted error for such window is greater than a threshold; and where the window of time-varying data is not discarded, outputting at least one of the window of time-varying data and the predicted error for each prediction of the physiological signal.

Inventors

Daniyal Liaqat
Mohamed ABDALLA
Eyal de Lara
Frank RUDZICZ
Andrea GERSHON
Robert Wu

Assignees

THE GOVERNING COUNCIL OF THE UNIVERSITY OF TORONTO
UNIVERSITY HEALTH NETWORK

Dates

Publication Date: 20260505
Application Date: 20200527
Priority Date: 20190912

Claims (18)

1 . A computer-implemented method for filtering continuous time-varying data for physiological signal prediction, the method comprising: receiving the continuous time-varying data from a physical physiological sensor; segmenting the continuous time-varying data into temporal windows; using a trained filter machine learning model, predicting an error of predictions of the physiological signal for each temporal window of the time-varying data, the filter machine learning model is trained using physiological signal predictions based on training continuous time-varying data and known values of the physiological signal for the training continuous time-varying data; discarding, each temporal window of time-varying data when the predicted error for such window is greater than a threshold; and where the temporal window of time-varying data is not discarded, using an extraction machine learning model, determining a physiological signal prediction for each temporal window of the time-varying data, the extraction machine learning model trained using a training set of continuous time-varying data and known values of the physiological signal for the training set of continuous time-varying data, and outputting the physiological signal prediction.
2 . The method of claim 1 , wherein the threshold is tunable by a user.
3 . The method of claim 1 , wherein input to the filter machine learning model comprises one or more aggregate measures of the time-varying data.
4 . The method of claim 1 , wherein the portion of training time-varying data used to train the extraction machine learning model is different than the portion of training time-varying data used to make predictions by the extraction machine learning model for training of the filter machine learning model.
5 . The method of claim 4 , wherein the portion of training time-varying data used to train the extraction machine learning model comprises uncontrolled data collected while a participant conducted an unspecified activity.
6 . The method of claim 1 , wherein the physiological signal comprises respiratory rate and the time-varying data comprises continuous motion sensor data from one or more movement sensors.
7 . The method of claim 6 , wherein the continuous motion sensor data is received from at least one of an accelerometer, a magnetometer and a gyroscope.
8 . The method of claim 7 , wherein input to the filter machine learning model comprises a plurality of aggregate measures for each axis of the continuous motion sensor data.
9 . The method of claim 8 , wherein training of the filter machine learning model comprises performing principal component analysis (PCA) on the aggregate measures and passing the PCA output to the filter machine learning model for training.
10 . The method of claim 7 , wherein an Inertial Measurement Unit (IMU) in a smartwatch comprises at least one of the accelerometer, the magnetometer, and the gyroscope.
11 . A system for filtering time-varying data for physiological signal prediction, the continuous time-varying data received from a physical physiological sensor, the system comprising one or more processors and a data storage device, the one or more processors in communication with the data storage device and configured to execute: an interface module to receive the time-varying data and segment the time-varying data into temporal windows; a filter module to: using a trained filter machine learning model, predict an error of predictions of the physiological signal by a trained extraction machine learning model for each temporal window of the time-varying data, the filter machine learning model trained using physiological signal predictions based on training time-varying data and known values of the physiological signal for the training time-varying data; discard the window of time-varying data when the predicted error for such window is greater than a threshold; and an extractor module to: where the window of the time varying data is not discarded, use the extraction machine learning model to determine a physiological signal prediction for each temporal window of the time-varying data, the extraction machine learning model trained using a training set of continuous time-varying data and known values of the physiological signal for the training set of continuous time-varying data; output the physiological signal prediction.
12 . The system of claim 11 , wherein the threshold is tunable by a user via input to the interface module.
13 . The system of claim 11 , wherein input to the filter machine learning model comprises one or more aggregate measures of the time-varying data.
14 . The system of claim 11 , wherein the portion of training time-varying data used to train the extraction machine learning model is different than the portion of training time-varying data used to make predictions by the extraction machine learning model for training of the filter machine learning model.
15 . The system of claim 11 , wherein the physiological signal comprises respiratory rate and the time-varying data comprises continuous motion sensor data from one or more movement sensors.
16 . A computer-implemented method for training two interoperable machine learning models using continuous time-varying training data from a plurality of sources, each source comprising a physical physiological sensor, an extraction machine learning model predicting an outcome on the continuous time-varying data and a filter machine learning model predicting an error of the outcome prediction of the extraction machine learning model, the continuous time-varying training data comprising continuous time-varying data with known outcomes, the method comprising: training the extraction machine learning model comprising passing a majority portion of the training continuous time-varying data through the extraction machine learning model with the known outcomes as labels for training; and training the filter machine learning model comprising: passing a minority portion of the training continuous time-varying data through the extraction machine learning model to output predicted outcomes; determining an error for each predicted outcome compared to the known outcome; and passing the minority portion of the training continuous time-varying data through the filter machine learning model with the determined errors as labels for training.
17 . The method of claim 16 , wherein the time-varying training data comprises time-varying data from a plurality of sources, and wherein the division of majority portion and minority portion applies to the time-varying training data from each of the plurality of sources.
18 . The method of claim 16 , wherein a portion of the training time-varying data that is separate from the majority portion and the minority portion is used for testing both the filter machine learning model and the extraction machine learning model.

Description

TECHNICAL FIELD The following relates generally to real-time data processing; and more specifically, to a system and method for filtering time-varying data for physiological signal prediction. BACKGROUND Receiving and processing physiological signals in real-time can provide an important window into the health and well-being of an individual. For example, respiratory rate is a vital physiological signal that may be useful for a multitude of clinical applications, especially if measured outside of controlled clinical settings (referred to as “in-the-wild”). Some approaches to in-the-wild physiological signal monitoring can require specialized, expensive, invasive, and/or cumbersome devices or procedures. In addition, some approaches may be overly susceptible to noise in the readings, such that meaningful signals received in-the-wild may not be practical or useful due to insufficient accuracy or computational burden. SUMMARY In an aspect, there is provided a computer-implemented method for filtering time-varying data for physiological signal prediction, the physiological signal predicted based on the time-varying data, the method comprising: receiving the time-varying data; segmenting the time-varying data into temporal windows; using a trained filter machine learning model, predicting an error for each prediction of the physiological signal for each window of time-varying data, the filter machine learning model trained using physiological signal predictions based on training time-varying data and known values of the physiological signal for the training time-varying data; discarding each window of time-varying data when the predicted error for such window is greater than a threshold; and where the window of time-varying data is not discarded, outputting at least one of the window of time-varying data and the predicted error for each prediction of the physiological signal. In a particular case of the method, the threshold is tunable by a user. In another case, input to the filter machine learning model comprises one or more aggregate measures of the time-varying data. In yet another case, the physiological signal prediction comprises using an extraction machine learning model to predict the physiological signal using the time-varying data as input, the extraction machine learning model trained using the training time-varying data and known values of the physiological signal for the training time-varying data. In yet another case, training of the filter machine learning model comprises performing principal component analysis (PCA) on the predictions of the extraction machine learning model and passing the PCA output to the filter machine learning model for training. In yet another case, the portion of training time-varying data used to train the extraction machine learning model is different than the portion of training time-varying data used to make predictions by the extraction machine learning model for training of the filter machine learning model. In yet another case, the physiological signal comprises respiratory rate and the time-varying data comprises continuous motion sensor data from one or more movement sensors. In yet another case, the continuous motion sensor data is received from at least one of an accelerometer, a magnetometer, and a gyroscope. In yet another case, input to the filter machine learning model comprises a plurality of aggregate measures for each axis of the continuous motion sensor data. In yet another case, an Inertial Measurement Unit (IMU) in a smartwatch comprises at least one of the accelerometer, the magnetometer, and the gyroscope. In another aspect, there is provided a system for filtering time-varying data for physiological signal prediction, the physiological signal predicted based on the time-varying data, the system comprising one or more processors and a data storage, the one or more processors in communication with the data storage device and configured to execute: an interface module to receive the time-varying data and segment the time-varying data into temporal windows; and a filter module to: using a trained filter machine learning model, predict an error for each prediction of the physiological signal by the trained extraction machine learning model for each window of time-varying data, the filter machine learning model trained using physiological signal predictions based on training time-varying data and known values of the physiological signal for the training time-varying data; discard the window of time-varying data when the predicted error for such window is greater than a threshold; and where the window of time-varying data is not discarded, output at least one of the window of time-varying data and the predicted error for each prediction of the physiological signal. In a particular case of the system, the threshold is tunable by a user via input to the interface module. In another case, input to the filter machine learning model comprises one or more aggre