Am starting to collect a small set of diabetes relevant (T1) time series data (movement, blood glucose, sleep, etc.) Have talked to many people who are doing such research, but few people seem to know how to process the data. Some have suggested ARIMA modelling, perhaps with ML for a variable window, but looking for any ideas or ideally someone who wants to co-author a research paper. Please get in touch if you have ideas, know relevant papers, or are interested. Thanks, Dmitri
I suggest you to use Principal Component Analysis as applied to the matrix having as statistical units the time points (if you use a costant sampling interval) and as variables the different parmeters you observed.
In this way you will have immediately an idea of what parameters move together from their component loading pattern. On the other side, being the statistical units different observations in time you can plot the relative component scores against time (i.e. having time as X and component score as Y) so you can easily identify the kind of dynamics they are involved (cyclic, trend and so forth). Attached an example..
Hell Alessandro,
thanks for responding.
I have to admit I am not knowledgeable in this area, so any help much appreciated.
My problem is that exercise (one of the lifestyle variables I am tracking) might raise blood glucose during activity, but might stabilize blood glucose(BG) levels for the next 6, 12, or 24 hours later. This might or might not be related to duration of exercise or intensity of exercise.
In addition an individual might typically choose to go running when their BG is high, so there would be a behavioral correlation between high BG and activity, but improved BG values in relation to the starting point.
Do you know if PCA would be suitable for detecting such variable and unknown periods? I am not trying to predict exact BG levels from variables, but rather which factors lead to increased control and BG stability.
Any tips very much appreciated.
Dmitri
ARIMA model is for univariate time series modelling. One of the ideas I know for multivariate time series analysis is ARIMAX model. The model can establish the relationship between different time series. The second I know is generalised additive modelling.
Well, you can also use machine learning techniques to model the multivariate time series. I used to publish papers on infectious time series modelling.
- Badges
- Science topic
- Similar topics
- Medicine
- Disease
- Metabolic Diseases
- Glucose Metabolism Disorders
- Diabetes