Transcript insulin pump

Machine Learning Based Algorithm for Automatic Insulin Regulation in Insulin Pumps
Final project submitted by Talia Melamed1 and Einav Yemini1 under the supervision of Dr. Ofer Yodfat2 , Mr. Iddo Michael Gescheit2
(1) Tel Aviv University (2) Medingo Ltd.
4. Results
1. Introduction
2. Objectives
BG
(-150(
BG
(-135)
BG
(-120)
BG
(-90)
BG
(-75)
BG
(-60)
BG
(-45)
BG
(-30)
CARB
Last meal
bolus
+ + + + + + + + +
+
+ + + + + + +
• Influence of previous Blood Glucose Concentrations (BGC):
basal
weights
1
BG (-30)
2
BG (-30-45)
3
BG (-30-60)
4
BG (-30-75)
5
BG (-30-90)
RMSE of trained ANN with additional remote
BGCs
7 BG (-30-120)
8 BG (-30-135)
0.27
Last meal
basal
insulin on
board
Slope (45-75)
-
+
-
-
0.22
Model
Test Set RMSE
0.17
ANN
3.47
Linear Regression
3.59
0
1
9 BG (-30-150)
2
3
4
5
6
7
8
9
10
BGC(-30) with additional remote BGCs
• Number of layers:
•Most errors are for extreme values
of the target:
Valalidation Performance
Validation Performance
Validation RMSE
Unexpected correlation
Final ANN performance:
•3.5% reduction in error, comparing
to Linear Regression:
0.32
6 BG (-30-105)
Rule based methods, such as the "Bolus Wizard", currently used to calculate the required insulin dose. These
methods have proved to be inaccurate, since significant parameters are taken into consideration.
Combining the two aspects of insulin delivery and glucose sensing encourages us to take advantage of these
technologies and improve the clinical outcome of insulin delivery, operating in a semi-open mode, laying down
the foundations for future closed loop systems – the "artificial pancreas“. In the scope of this project, we intend
to investigate the practical use of ANNs, preferably for use in conjunction with the Solo™ MicroPump Insulin
Delivery System.
BG
(-105)
bolus
weights
RMSE
• Type 1 Diabetes Mellitus is a chronic metabolic disease caused by insufficient secretion of insulin by the beta cells in the pancreas,
resulting in an inability to control blood glucose concentrations. Type 1 diabetes patients must have constant and tight supervision of
their blood glucose levels via multiple blood glucose measurements and insulin administration to avoid short and long-term
complications.
• Insulin administration is typically carried out by Multiple Daily Insulin Injections (MDII) or Continuous Subcutaneous
Insulin Infusion (CSII), i.e., insulin pump therapy. An insulin pump is a processor-controlled medical device that
continuously delivers insulin to the subcutaneous tissue through a soft cannula replaced every 2-3 days.
• Further development in Continuous Glucose Monitoring (CGM) technology offers the patients tighter control over BG levels.
• Recent academic work has been researching the implementation of machine learning methods, e.g., Artificial Neural Network (ANN),
in diabetes care, capable of learning physiological patterns based on a set of inputs and required outputs.
Expected correlation
Feature Analysis
• Weight analysis in perceptron - sign of each input parameter:
3.7
3.6
3.5
3.4
0
1
2
3
4
5
Number of layers
•ANN predicts output for unseen data based on learning patterns
•Learning based on 3 examples sets:
1. Training set – used for set weights between neurons
2. Development set – used for choose the network
architecture avoiding overfitting of the ANN to the training example
3. Test set - used for evaluation of the ANN
• ANN is composed of an input layer, one or more hidden layers
(a network with no hidden layers is called a perceptron) and an
output layer.
• Each layer consists of one or more neurons and weights
connecting the layers
9 equally-spaced
measurements of Blood
Glucose Concentrations
(BGC) (30-150 minutes
before a meal) [mg/dl]
• Number of neurons in hidden layer –
ANN Structure
Amount of carbohydrates
to be consumed [grams]
Last meal bolus insulin
dose [Units]
Last meal basal insulin
dose [Units]
“Insulin on board"
BGC "tendency"
•The learning algorithm updates the weights in order to minimize the error over the training set. In this way ANN
learns a function connecting between input and output variables.
• Data generated via AIDA software - simulating virtual diabetes patience.
• Data sets based on one AIDA patient profile, as our objective is "tailored" insulin administration per patient.
• Feature Analysis in a Perceptron:
∆ Goal: input feature analysis and selection
∆ Weight analysis
∆ Influence of previous BGCs
• ANN Architecture
∆ Number of neurons in each hidden layer
∆ Number of layers
∆ Transfer function
simplest model that gives good performance:
MSE for different number of neurons in one hidden layer
with purelin TF
13.5
Bolus
insulin
dose
Basal
insulin
dose
13
MSE
3. Methods
12.5
12
11.5
0
2
4
6
8
10
12
14
16
18
20
Number of neurons
5. Discussion
• Data collection - Simulated data has limited features and questionable fit to real
life data. Moreover, creating a benchmark is necessary for scientific progress.
• Our ANN model is simple with reasonable error. Results show that complexity
does not improve performance. One main reason for this could be the
limitations of the synthetic data.
6. Future Work
• Examining the effect of increasing the size of the training set.
• Employ different combinations of the input parameters.
• Adding a second ANN for predicting the corresponding blood-glucose
concentrations according to the first ANN output.
• Introducing the ANN with real life data.