Blood Glucose Level Prediction usingPhysiological Models and Support Vector Regression

Presented By RAVI KANT KUMAR

PREVIOUS WORK SVR model was trained to predict the BG levels of a T1D patient. Pivot date Training sets Testing sets Trained with a linear kernel Two separate SVR model is used For 30 minutes RMSE is 18.0 when base line 25.0 For 60 minutes RMSE is 30.9 when base line 43.2

Contd..

Subsequently, SVR was refined to use a more comprehensive set of features.

The feature templates used followed 8 steps.

For the next model design Pivot date, Training sets, Testing sets SVR model trained with a Gaussian kernel. Compared against two baselines : the simple t0 baseline that uses the

present BG level as the prediction and ARIMA model trained on 4 days of CGM data

Result

PHYSICIAN PERFORMANCE DATA

Result

The first doctor, used both the CGM data and daily events to perform prediction, outperform the ARIMA model that is using only CGM data.

The third doctor overestimated the increase or decrease in BG levels. To address this problem- a threshold of τ = 5mg/dl for 30 minute

prediction and τ = 10mg/dl for 60 minute prediction was chosen accordingly.

PHYSIOLOGICAL MODEL

Three compartments: meal absorption dynamics, insulin dynamics, and glucose dynamics

Equations based on Duke’s PhD thesis. State variables X, input variables U, and a state transition functionthat computes the next state given the current state and input variables

i.e.

Contd...

Meal Absorption Dynamics:• Cg1 (t) = carbohydrate consomption (g).• Cg2 (t) = carbohydrate digestion (g).

2) Insulin Dynamics:• IS(t) = subcutaneous insulin (μU).• Im(t) = insulin mass (μU).• I(t) = level of active plasma insulin (μU/ml).

3) Glucose Dynamics:• Gm(t) = blood glucose mass (mg).• G(t) = Blood glucose concentration (mg/dl).

  The state transition function captures dependencies among variables in the model,

Contd..

meal absorption compartment, the equations are:

The equations for the insulin compartment are:

Contd..

The general equation for the glucose compartment is

where:

Finally, the glucose and insulin concentrations depend deterministically on their mass equivalents

Contd..

Extended Kalman filter (EKF) model is used to reduce noise, Since the CGM data comes every 5 minutes, the Kalman filter ran state

prediction step every 1 minute and a correction step every 5 minutes. Results from initial experiments on the reference dataset from Physician

performance data were not better than the simple t0 baseline

So use of grid search which employs α The parameters α used in the physiological model were tuned to match

published behaviour and further refined based on feedback from the doctors

It is not feasible to use of grid search on every parameters. So alternative model is used.

SVR MODEL WITH PHYSIOLOGICAL FEATURE The extended Kalman filter was first run up to the training/test point t0,

with a correction step every 5 minutes, including a correction at t0. This resulted in a state vector X(t0).

The EKF model was then run in prediction mode for 60 more minutes, and the state vectors at 30 minutes X(t0 + 30) and 60 minutes X(t0 + 60) were selected for feature generation

The actual physiological features were as follows: all the state variables in the vectors X(t0 + 30), X(t0 + 60), and the difference vectors

X(t0 + 30) − X(t0), X(t0 + 60) − X(t0).

EXPERIMENTAL RESULTS

The results demonstrate the significant utility of physiological modeling for the engineering of features in machine learning models for BG level prediction.

Alternative Method The Bergman’s minimal model can adequately describe the dynamics of glucose and insulin

interaction in the blood system. Based on sum of squares technique, a linear state feedback controller has been obtained to

regulate the blood’s insulin and glucose concentration levels around the basal levels The model equations are nonlinear and are given by

G(t), the glucose concentration in the blood plasma. X(t), the insulin’s effect on the net glucose disappearance. I(t), the insulin concentration in plasma at time. Ib, the basal preinjection level of insulin. p1, the insulin-dependant rate constant of glucose consumption. p2, the rate for decrease in tissue glucose. p3, the insulin-dependant increase in glucose uptake ability in tissue per unit of insulin

concentration. n, the first order decay rate for insulin in blood. h, the threshold value of glucose.

In Fig 1 the glucose concentration levels of the patients are finally stabilized at the basal level

Figures 2 and 3 show the insulin profile and the control function for the three patients, respectively

In Fig. 3, the control signal doesn’t increase rapidly, but it goes up steadily without giving shock to the patient

The controller structure is a linear state feedback , and smoothness and dampness of control response, makes it easy to regulate blood glucose

Thank you