Code for BGLP2020 paper on glucose prediction using OpenAPS and OhioT1DM data
Usage:
In the terminal run the following commands twice. First with dataset='oaps' and then with dataset='ohio' in preprocess.sh and run.sh files:
chmod +x ./preprocess.sh
./preprocess.sh
chmod +x ./run.sh
./run.sh
chmod +x generate_bglp_results.sh
./generate_bglp_results.sh
In shell files, set the following parameters according to the experiment you
are running.
dataset='ohio' (ohio or oaps)
root_directory = "../../../../PHI/PHI_OHIO/" (the root folder)
data_directory= $root_directory"data/" (containing raw data)
output_directory="OHIO_models/" (folder to save the results in)
model_directory='OAPS_models/' (folder containing model pretrained on OAPS data
filter_data = 'True' (True or False - use median filter to smooth training data)
normalize_data = False (True or False - normalize all feature values to be in the range of the BG levels)
threshold = 15 (integer value - removes data for which BG levels are below this value)
history_window = 12 (integer value - number of past glucose values to use. 12 samples means an hour of previous data (frequency = 5 minutes)
prediction_window = 60 (30 or 60 minutes - prediction horizon for BG forecasting (in minutes))
dimension = multivariate (univariate or multivariate)
prediction_type = single (single or multi. This refers to single step or multioutput forecasting)
model_name = RNN (['LSTM', 'RNN'])
save_results = False (True or False. It will replace old output files in the output directory)
mimic_pipeline = "student" (student, teacher, retrain, teacher_student)
test_extrapolation = True (True or False - to extrapolate test data or not. Read report.docx for more info)