Prediction of glucose dynamics by Bayesian modeling

Nowadays, we can witness the widespread use of Continuous Glucose Monitoring (CGM) portable devices. This project aims to build an almost real-time automatic alerting system to be deployed on glucose time-series.
We would like to leverage the Bayesian statistical framework of Gaussian Processes (GPs) for analyzing those glucose time-series. In particular, Bayesian models do have the advantage of updating dynamically themselves to the incoming new pieces of information. A second advantage relies on their intrinsic ability to be interpretable by their credible intervals. These credible intervals appeal clinician’s needs because they offer an intuitive quantification of the uncertainty of the statistical model (“Should I trust it?”). The third advantage relies on exploiting credible intervals as automatic alarms: each data-point which lies outside a credible interval, will trigger an alarm. Lastly, GPs are perfect for physiological time-series since they are non-stationary and non-linear.
However, the main disadvantage of GPs includes their computationally intensive nature. This issue hinders their adoption within in a tele-medicine setting since an alarm may be triggered after the pathological event had taken place, because the training of the algorithm is excessively slow. For that reason, we will try to use parallelized GPs which rely on multiple Graphics Processing Units (GPUs), thus mitigating the computational burden.
In a nutshell, we will try to make GPs parallel by leveraging mixture-of-expert and sequential Monte Carlo. We will hope to use JAX Python library which enables GPUs computations. If everything works, we should provide a clinical meaningfulness to GPs (still to be established their use "on-line"), since at the moment GPs are just academic exercises (demonstrated use "off-line"). In other words, we want to transfer GPs as a statistical paradigm from academia towards the clinic. Indeed, GPs have already 20 years of literature showing its usefulness for biomedical time-series. Nevertheless, there is no use of GPs in the daily clinical practice, as yet. The main hindrance for such transfer of knowledge is the intrinsic computational complexity of GPs.
In conclusion, the project aims to build an automatic alarm system that may help in better stabilize blood sugar level in people wearing CGM devices. The main challenge will be to translate GPs’ statistical framework into a GPU-compatible Python library in order to speed up computation.