Spectro-spatial features in distributed human intracranial activity proactively encode peripheral metabolic activity

CGM Decoder Demo for submission to Nature Communciations.

Authors: Huang, Y.*, Wang, J.B.*, Parker, J.J., Shivacharan, R., Lal, R.A., Halpern. C.H.

Dependencies

• h5py >= 2.10.0
• numpy >= 1.19.2
• scikit_learn >= 1.1.2
• scipy >= 1.5.4

Installation

Runs out of the box assuming dependencies have been installed. Installation time is negligible.

Running the Code

CGM_Decoder.ipynb provides a minimal example of the decoder trained and tested on data for a single patient without any time delay.

The core of the decoder runs out of the GlucoseDecoder class in Decoder.py. It has several key functions:

Constructor

Decoder(model, linear_correction=True,normalize=True)

Inputs:

• model is the choice of model, in our case a LASSO as implemented in SKLearn
• linear_correction applies a second linear model after the LASSO fit
• normalize will normalize all inputs to unit variance prior to model application.

Fitting a model

(predictions,predictions_corrected,rvalue) = fit_model(linkedTF,linkedBG)

Inputs:

• linkedTF is a $ N\times M$ array where $N$ is the number of time points, and $M$ is the number of features (e.g. combinations of electrodes and frequency band)
• linkedBG is a $N\times1$ array listing the blood glucose corresponding to each time point.

Outputs:

• predictions is an $N\times1$ array listing the predicted blood glucose prior to linear correction discussed above
• predictions_corrected is an $N\times1$ array listing the predicted blood glucose after linear correction discussed above. Assuming default settings this should be the prediction used
• rvalue is the Pearson Correlation between corrected predicted and actual blood glucose values

Predicting with the model

predictions = predict(self,linkedTF,linkedBG=None)

Inputs:

• linkedTF is a $ N\times M$ array where $N$ is the number of time points, and $M$ is the number of features (e.g. combinations of electrodes and frequency band)
• linkedBG is an optional $N\times1$ array listing the blood glucose corresponding to each time point

Outputs:

• predictions is an $N\times1$ array listing the predicted blood glucose after linear correction discussed above
• rvalue is the Pearson Correlation between corrected predicted and actual blood glucose values if linkedBG is provided