Before:
The dataset used in this work consisted of simultaneous glucose measurements and multi-modal wearable data. Glucose was sampled every 5 minutes using a Dexcom G6 CGM sensor. Blood Volume Pulse (BVP), Tri-axial Accelerometer (acc_x, acc_y, acc_z), Electrodermal Activity (EDA), Skin Temperature (Temp), and Heart Rate data were sampled at the following sampling resolutions from the Empatica E4:
BVP: 64 Hz acc_x, acc_y, acc_z: 32 Hz EDA: 4 Hz Temp: 4 Hz HR: 1 Hz
The following pipeline was applied for initial pre-processing:
Tri-axial accelerometry data was used to compute the vector magnitude of the acceleration (ACC). Outlier removal and filtering i) HR and Temp signals were filtered in the time-domain by removing infeasible values. ii) EDA, ACC, and BVP signals were filtered in the frequency-domain using the following filters: EDA: low-pass filter, cutoff = [0.5] ACC: band-pass filter, cutoff = [0.29, 10] BVP: band-pass filter, cutoff = [0.5, 5] Epoch-wise segmentation Each signal was segmented into 5-minute epochs based on the available glucose timestamps. Missing value imputation
Now:
I computed all available Cross Entropies between pairs of signals and calculated the epoch-wise cross-correlation/pearson’s correlation between all combinations of signal pairs. In order to interpret the calculation, I compared male and female subjects, as well as subjects with HbA1C levels in the High-Normal (HN) range (5.2–5.6) and subjects classified as Prediabetes (PD) with HbA1C levels of 5.7–6.4. I stated low: 0.00−0.33, moderate: 0.34−0.66, high: 0.67 and above (including values greater than 1.0)
Future:
transforming the signal to normal distribution before calculating pearson correlation calculating pearson correlation separted by glucose level longer epoch size - more than 5 mins - to get a more significant change when graining the signal