Skip to main content
Back

UVA's DIAX Turns a Week of Diabetes Data Parsing Into an Afternoon

A JSON schema and open-source toolkit stitches six major diabetes trial datasets, over 10 million patient-hours, into one format that ML models can actually consume.

Pro

  • DIAX is an open-source JSON schema unifying diabetes time-series data across trials and devices.
  • Ships conversion scripts for six major datasets: DCLP3, DCLP5, IOBP2, PEDAP, T1DEXI, Loop.
  • Covers over 10 million patient-hours of CGM, insulin, meal, and wearable signals.
  • Stores raw timestamps without forced alignment, preventing interpolation artifacts across irregular sample rates.
  • Bundles Python and MATLAB utilities for AGP plots, glycemic metrics, and time alignment.
  • Positioned as a research harmonization layer, not a clinical exchange standard or data host.

Anyone who has tried to train a blood glucose forecasting model knows the pain: every diabetes dataset ships in its own idiosyncratic CSV layout, with different column names, timestamp conventions, insulin units, and meal annotation schemes. Before you write a single line of model code, you burn days writing bespoke parsers. A new open-source project from the University of Virginia's Center for Diabetes Technology wants to end that ritual.

DIAX (DIAbetes eXchange) is a standardized JSON format and conversion toolkit that acts as a translational resource rather than a data host, and currently supports major trials including DCLP3, DCLP5, IOBP2, PEDAP, T1Dexi, and Loop, totaling over 10 million patient-hours of data. Think of it as a research harmonization layer sitting on top of existing releases, not a new data silo.

The fragmentation tax

Modern diabetes care generates dense multivariate time series. Continuous Glucose Monitoring (CGM), Smart Insulin Pens, and Automated Insulin Delivery systems each spit out their own streams of glucose readings, basal rates, boluses, and meal announcements. Public clinical trial releases from the Jaeb Center for Health Research are gold mines for machine learning, but every trial packages its data differently. That inconsistency is a silent tax on reproducibility: two papers claiming state-of-the-art glucose prediction on "the DCLP3 dataset" may actually be training on subtly different preprocessed versions.