Systems, methods and devices for predicting personalized biological state, predicting personalized behavior, and recommending personalized behavior with models produced with meta-learning
An exemplary method can include using meta-learning on various biological and/or behavior related data sets to generate model parameters for predicting biological and/or behavior predictions. Meta-learned model parameters can configure learning algorithms to rapidly train model/functions for predicting user biological and/or behavioral responses. In some embodiments, recommendations can be generated for a user based on predicted biological and/or one or more behavioral predictions. Corresponding systems are also disclosed.
1 . A method for personalized blood glucose monitoring for a test subject, comprising:
(a) acquiring a plurality of task data sets from a plurality of subjects, each of the plurality of task data sets comprising input values and output values from a different source associated with a subject in the plurality of subjects,
wherein the acquiring comprises (i) using a continuous glucose monitor (CGM) sensor to obtain time-series blood glucose measurements of the subject, and (ii) using a heart rate monitor (HRM) sensor to obtain time-series heart rate monitor values of the subject;
wherein the task data sets comprise time series data, with a value for one time being an input value, and a value for a subsequent time being an output value;
(b) performing a meta-learning operation comprising:
(1) processing at least a portion of the input values with a biology model that generates predicted output values comprising at least one blood glucose level corresponding to the input values, wherein the biology model comprises a neural network configured to process the time series data, such that a value for one time of the time series data is processed as an input value to the neural network and another value for a subsequent time of the time series data is produced as an output value of the neural network,
(2) generating meta-learning error values at least in part by comparing each predicted output value to the output value corresponding to the respective input value,
(3) training the biology model based at least in part on the meta-learning error values, wherein the training comprises adjusting parameters for the neural network of the biology model, such that the biology model reaches convergence with a meta-error target of the biology model, wherein
the parameters are stored as meta-learned parameters after all task data sets have been processed by the meta-learning operation;
(c) training a population-based meta-learned model, at least in part by configuring the neural network of the biology model with the meta-learned parameters;
(d) acquiring test subject data from the test subject by (i) using the CGM sensor to obtain time-series blood glucose measurements of the test subject, and (ii) using the HRM sensor to obtain time-series heart rate monitor values of the test subject, wherein the test subject data was not used for performing the meta-learning operation in (c), and wherein the test subject is not among the plurality of subjects in (a); and
(e) subsequent to the training of (c), training a personalized subject prediction model to predict at least one blood glucose level for the test subject, at least in part by re-training the population-based meta-learned model with the test subject data, the test subject data set comprising input values and output values and being different from each of the task data sets,
wherein the re-training comprises adjusting the meta-learned parameters of the population-based meta-learned model, such that the personalized subject prediction model reaches convergence with an error target of the personalized subject prediction model,
thereby producing a personalized subject prediction model that is optimized for the test subject; and
(f) predicting a blood glucose level for the test subject with the personalized subject prediction model.
2 . The method of claim 1 , wherein the task data sets comprise food consumption events.
3 . The method of claim 1 , wherein the task data sets comprise data for different populations.
4 . The method of claim 1 , wherein the test subject data set comprises time series data, with a value for one time being an input value, and a value for a subsequent time being an output value corresponding to the input value.
5 . The method of claim 1 , wherein the neural network comprises a long short-term memory network (LSTM).
6 . A method for personalized behavior monitoring of a test subject, comprising:
(a) acquiring a plurality of task data sets from a plurality of subjects, each of the plurality of task data sets comprising input values and output values of human behaviors from a different source associated with a subject in the plurality of subjects,
wherein the acquiring comprises (i) using a continuous glucose monitor (CGM) sensor to obtain time-series blood glucose measurements of the subject, and (ii) using a heart rate monitor (HRM) sensor to obtain time-series heart rate monitor values of the subject;
wherein the task data sets comprise time series data, with a value for one time being an input value, and a value for a subsequent time being an output value;
(b) performing a meta-learning operation comprising:
(1) processing at least a portion of the input values with a behavior model that generates predicted output values comprising at least one behavior corresponding to the input values, wherein the behavior model comprises a neural network configured to process the time series data, such that a value for one time of the time series data is processed as an input value to the neural network and another value for a subsequent time of the time series data is produced as an output value of the neural network,
(2) generating meta-learning error values at least in part by comparing a predicted output value to the output value corresponding to the respective input value,
(3) training the behavior model based at least in part on the meta-learning error values, wherein the training comprises adjusting parameters for the neural network of the behavior model, such that the behavior model reaches convergence with a meta-error target of the behavior model, wherein
the parameters are stored as meta-learned parameters after all task data sets have been processed by the meta-learning operation;
(c) training a population-based meta-learned model, at least in part by configuring the neural network of the behavior model with the meta-learned parameters;
(d) acquiring test subject data from a test subject by (i) using the CGM sensor to obtain time-series blood glucose measurements of the test subject, and (ii) using the HRM sensor to obtain time-series heart rate monitor values of the test subject, wherein the test subject data was not used for performing the meta-learning operation in (c), and wherein the test subject is not among the plurality of subjects in (a);
(e) subsequent to the training of (c), training a personalized subject prediction model for the test subject to predict at least one behavior for the test subject, at least in part by re-training the population-based meta-learned model with test subject data set, the test subject data set comprising input values and output values and being different from each of the task data sets,
wherein the re-training comprises adjusting the meta-learned parameters of the population-based meta-learned model, such that the personalized subject prediction model reaches convergence with an error target of the personalized subject prediction model,
thereby producing a personalized subject prediction model that is optimized for the test subject; and
(f) predicting a behavior of the test subject with the personalized subject prediction model.
7 . The method of claim 6 , wherein the task data sets comprise food consumption events.
8 . The method of claim 6 , wherein the task data sets comprise physical activities.
9 . The method of claim 6 , wherein the task data sets comprise population data.
10 . The method of claim 6 , wherein predicting the predicted behavior comprises a plurality of behaviors.
11 . The method of claim 6 , wherein the predicted behavior comprises at least one food consumption event.
12 . The method of claim 6 , wherein the predicted behavior comprises at least one physical activity.
13 . The method of claim 6 , wherein the neural network comprises a long short-term memory network (LSTM).