Neural network system and uses thereof

• US 9,076,107 B2
• Filed: 05/22/2014
• Issued: 07/07/2015
• Est. Priority Date: 08/14/2008
• Status: Active Grant

• Current Assignee: University of Toledo
• Original Assignee: University of Toledo
• Inventors: Cameron, Brent D., Pappada, Scott M.
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): Chubb, Mikayla
• Application Number:
  US14/284,975
• Publication Number:
  US 20140304204A1
• Time in Patent Office: 411 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G16H 20/10 relating to drugs or medications, e.g. for ensuring correct administration to patients
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: G16H 50/50 for simulation or modelling of medical disorders
• CPC Class Codes: G16H 50/70 for mining of medical data, e.g. analysing previous cases of other patients
• CPC Class Codes: G16Z 99/00 Subject matter not provided for in other main groups of this subclass

Abstract

A multifunctional neural network system for prediction which includes memory components to store previous values of data within a network. The memory components provide the system with the ability to learn relationships/patterns existent in the data over time.

Claims

1. A neural network system, comprising:

i) an input layer configured to accept N input signals;

ii) one or more “

i”

hidden layers,

iii) at least one output layer;

iv) at least one neuron “

Y^layer”

within each layer, where “

layer”

is the layer defined as “

input”

, “

hidden_i”

or “

output”

; and

,

v) one or more memory structures configured to;

a) store a recursive memory of input signals past, and

b) allow for at least one time series prediction of a response,

wherein the neural network is configured to include one or more input variables oriented trend analysis post processing algorithms which are configured to analyze one or more predicted outputs of the neural network system for one or more expected trends in predicted analyte values based on previous and current input data presented to the neural network system;

wherein one or more support/post processing algorithms are included in order to modify the neural network system predictive output such that an increased predictive accuracy is achieved; and

wherein the neural network includes a support/post processing algorithm comprising an input variable, or event, oriented trend analysis algorithm.

2. The neural network system of claim 1, wherein the neural network system is configured for both real-time prediction and retrospective prediction.

3. The neural network system of claim 2, wherein one or more support/post processing algorithms are included in order to modify the neural network system predictive output such that an increased predictive accuracy is achieved.

4. The neural network system of claim 3, wherein the neural network includes one or more support/post processing algorithms selected from:

an adaptive analyte threshold based rate of change (ROC) algorithm and input variable, or event, oriented trend analysis algorithm.

5. The neural network system of claim 4, wherein the adaptive analyte threshold based ROC post processing algorithm is configured to track the ROC of analyte data presented to the neural network system.

6. The neural network system of claim 5, wherein, based on a current analyte value and ROC of current and previous analyte values, if the n predicted analyte values do not correlate with the ROC, the predicted output of the neural network system are modified via the post processing algorithm to increase predictive accuracy.

7. The network system of claim 1, wherein at least one trend is quantifiable and is programmed into the neural network system to gauge if predictions generated by the neural network system are accurate.

8. The neural network system of claim 7, wherein if predictions digress from an expected trend, appropriate post-processing is completed to modify predictions and enhance predictive accuracy.

9. The neural network system of claim 8, wherein if the expected trend does not occur in n predicted values, then the post processing algorithm will modify the neural network system output to increase predictive accuracy.

10. The neural network system of claim 1, configured to be a time-lagged feed forward neural network system for predicting analyte levels in a sample or a subject in need thereof.

11. The neural network of claim 1, wherein the network is configured for predicting analyte levels in a sample, the analyte comprising one or more of:

artificial metabolites and/or reaction products.

12. The neural network system of claim 1, wherein the analyte comprises one or more of:

blood and/or urine components proteins;

amino acids;

hormones;

cholesterol;

viruses;

toxins;

antibodies;

antigens;

vitamins;

imaging contrast agents;

illegal drugs;

pharmaceutical compositions; and

steroids.

13. The neural network system of claim 1, wherein one or more of the input signals past comprise:

emotional, lifestyle, and nutritional factors, factors documented during the course of the subject'"'"'s normal everyday life, or during treatment, stay in intensive care or other supervised setting, and medical records.

14. The neural network system of claim 1, wherein the neural network system is configured to be used prospectively and retrospectively to gauge at least one change and/or at least one occurrence of one or more input factors and a corresponding effect of the factors on at least one prediction.

15. The neural network system of claim 1, wherein the neural network system is configured for use as a retrospective or prospective iterative therapeutic, or an educational tool for determination of effect of at least one input variable.

16. The neural network system of claim 1, wherein at least one input variable comprises one or more of:

analyte concentration, analyte levels, analyte states and outcomes.

17. The neural network system of claim 1, for monitoring one or more of:

type I and type II diabetes, hospital/critical care/trauma patients/cardio-thoracic surgical patients, and military personnel.

18. The neural network system of claim 1, configured for predicting one or more outputs selected from:

glucose levels, including discrete and/or CGM values;

glucose states, including classified ranges of glucose values;

glycated hemoglobin A1C values, mortality, morbidities, and complications;

at a point in time n minutes in the future, when applicable, in real-time and retrospectively.

19. The neural network system of claim 1, configured for providing administration of insulin/medication dosage administration, and intelligent therapy guidance, in an integrated or stand-alone manner.

20. A method for predicting analyte levels and medication dosages, comprising:

i) providing a multifunctional neural network system of claim 1 for prediction of one or more of;

analyte concentration values, analyte levels, analyte states, medication dosages, physiologic outcomes and clinical outcomes;

ii) presenting pre-processing of input data to the neural network system;

iii) providing a neural network system output post processing algorithm to enhance predictive accuracy during a time of decreased accuracy;

iv) providing a capability for neural network system re-training for model weight optimization when new data combinations and results are experienced.

21. A neural network system, comprising:

i) an input layer configured to accept N input signals;

ii) one or more “

i”

hidden layers,

iii) at least one output layer;

iv) at least one neuron “

Y^layer”

within each layer, where “

layer”

is the layer defined as “

input”

, “

hidden_i”

or “

output”

; and

,

v) one or more memory structures configured to;

a) store a recursive memory of input signals past, and

b) allow for at least one time series prediction of a response,

wherein one or more memory structures are included in both the input and hidden layers,

wherein the neural network is configured to include one or more input variables oriented trend analysis post processing algorithms which are configured to analyze one or more predicted outputs of the neural network system for one or more expected trends in predicted analyte values based on previous and current input data presented to the neural network system;

wherein one or more support/post processing algorithms are included in order to modify the neural network system predictive output such that an increased predictive accuracy is achieved; and

wherein the neural network includes a support/post processing algorithm comprising an input variable, or event, oriented trend analysis algorithm.

22. The neural network system of claim 21, wherein one or more of the input signals past comprise:

emotional, lifestyle, and nutritional factors, factors documented during the course of the subject'"'"'s normal everyday life, or during treatment, stay in intensive care or other supervised setting, and medical records.

23. The neural network system of claim 21, wherein one or more of the input signals past comprise:

physiologic or medication factors.

24. A neural network system, comprising:

i) an input layer configured to accept N input signals;

ii) one or more “

i”

hidden layers,

iii) at least one output layer;

iv) at least one neuron “

Y^layer”

within each layer, where “

layer”

is the layer defined as “

input”

, “

hidden_i”

or “

output”

; and

,

v) one or more memory structures configured to;

a) store a recursive memory of input signals past, and

b) allow for at least one time series prediction of a response;

wherein the neural network system is configured for both real-time prediction and retrospective prediction;

wherein one or more support/post processing algorithms are included in order to modify the neural network system predictive output such that an increased predictive accuracy is achieved;

wherein the neural network includes one or more support/post processing algorithms selected from;

an adaptive analyte threshold based rate of change (ROC) algorithm and input variable, or event, oriented trend analysis algorithm;

wherein the adaptive analyte threshold based ROC post processing algorithm is configured to track the ROC of analyte data presented to the neural network system;

wherein, based on a current analyte value and ROC of current and previous analyte values, if the n predicted analyte values do not correlate with the ROC, the predicted output of the neural network system are modified via the post processing algorithm to increase predictive accuracy; and

wherein,

n predicted values generated are adjusted to coordinate with real-time ROC to enhance predictive accuracy via Equation [2],

PREDICT_mod=PREDICT_CGM+W_ROC·

ROC_prediect·

Δ

t 

 

Equation[2],

wherein,

PREDICT_CGM is a vector of predicted CGM values with length n, W_ROC is a vector of length n of weights for weighting ROC values based-on the current real time value (threshold),

ROC_predict is a vector of ROC values of length n estimated based on best linear, or nonlinear model of real-time ROC,

Δ

t is a time duration between the two samples or a sampling rate, and

PREDICT_mod is a vector of modified (post-processed) predictions to increase accuracy based on trends in real-time ROC.

25. A neural network system, comprising:

i) an input layer configured to accept N input signals;

ii) one or more “

i”

hidden layers,

iii) at least one output layer;

iv) at least one neuron “

Y^layer”

within each layer, where “

layer”

is the layer defined as “

input”

, “

hidden_i”

or “

output”

; and

,

v) one or more memory structures configured to;

a) store a recursive memory of input signals past, and

b) allow for at least one time series prediction of a response;

wherein the neural network system is configured for both real-time prediction and retrospective prediction;

wherein one or more support/post processing algorithms are included in order to modify the neural network system predictive output such that an increased predictive accuracy is achieved;

wherein the neural network includes one or more support/post processing algorithms selected from;

an adaptive analyte threshold based rate of change (ROC) algorithm and input variable, or event, oriented trend analysis algorithm;

wherein the adaptive analyte threshold based ROC post processing algorithm is configured to track the ROC of analyte data presented to the neural network system;

wherein, based on a current analyte value and ROC of current and previous analyte values, if the n predicted analyte values do not correlate with the ROC, the predicted output of the neural network system are modified via the post processing algorithm to increase predictive accuracy; and

wherein the post processing includes an adaptive analyte threshold based ROC approach;