Continuous glucose monitors have transformed diabetes care, streaming a glucose reading every few minutes and generating, over the course of a clinical study, millions of data points that promise an unprecedented window into metabolic health. Yet behind every elegant ambulatory glucose profile chart lies an unglamorous and surprisingly stubborn problem: the raw exports that manufacturers produce are messy, inconsistent, and largely incompatible with one another. A research team at the Medical University of Graz, working with colleagues in Bern, has now tackled this bottleneck head-on with GluCINDa, short for Glucose: Coherent Import. Neat Data, an open-source Python application described in BMC Bioinformatics that turns heterogeneous continuous glucose monitoring exports into clean, analysis-ready datasets without ever sending patient data off the local machine.
The problem GluCINDa addresses is more subtle than it first appears. Continuous glucose monitoring is now a central data source in both clinical practice and diabetes research, and while dedicated analysis packages exist for computing metrics such as time in range, glucose management indicator, and coefficient of variation, the workflows that convert manufacturer exports into analysable datasets remain heterogeneous and largely manual. Each vendor structures its files differently. Abbott’s LibreView exports open with metadata rows before the header; Dexcom’s Clarity files place the header first and then insert a variable number of metadata rows whose count depends on device settings; Medtronic’s CareLink exports contain multiple data blocks separated by repeated header rows, and those headers can shift column positions mid-file. Existing import tools, the authors note, are limited to specific devices or platforms, or require uploading sensitive data to a remote server, which is a non-starter for many ethics boards and multi-centre collaborations.
GluCINDa’s design philosophy is deliberately conservative and transparent. The application runs entirely locally, preserving data privacy, and is operated through a graphical user interface that lowers the barrier for researchers who are not comfortable with command-line tools. It supports parallel processing so that large cohorts can be processed efficiently. At its heart, the core algorithm processes input files row by row, which allows it to absorb structural quirks such as multiple embedded tables, mid-file header shifts, and the mixture of numeric and non-numeric glucose values that vendors sprinkle through their exports. Rather than relying on statistical inference or machine learning to guess what a file contains, the software implements manufacturer-specific detection as a fixed rule set applied to header structure, keeping the entire detection process deterministic and auditable independent of whatever analysis comes downstream.
Privacy considerations shaped another key design decision. Subject identifiers in the harmonised output are derived exclusively from file names, and the file content itself is never parsed for personal metadata. This means that names, birthdates, and sensor identifiers embedded in export headers are simply discarded rather than carried into the research dataset. The validation study underlying the paper was approved by the Medical University of Graz Ethics Committee, with consent waived because the work used only retrospective clinical data, and the authors report that no identifiable information was extracted or retained in the harmonised output.
The validation effort was unusually thorough for a data-ingestion tool. The team assembled 87 real-world continuous glucose monitoring exports spanning three manufacturers, four download platforms, and two languages, then pushed them through the pipeline under four alternative directory structures, three operating systems, and multiple Python versions. The outputs were identical in every configuration, a result that speaks to the reproducibility that the tool is designed to guarantee. Line-count verification provided a mathematical proof of completeness: the 87 input files contained 3,166,529 lines in total, and the single-file documentation files generated by GluCINDa summed to exactly double that number, 6,333,058, confirming that no input row was silently skipped anywhere in the dataset. A fitted regression line comparing input and documentation line counts had a slope of exactly 2 and a correlation of 1.
The authors also verified that the software was identifying the right streams within each file. In LibreView exports, for example, the tool correctly discarded the leading metadata rows, mapped columns from the header, and distinguished genuine sensor glucose values from scan values and from later rows containing insulin and meal data that are not continuous glucose readings at all. In CareLink files, which interleave pump data, aggregated insulin data, and sensor data in separate blocks, GluCINDa correctly tracked the shifting headers and even flagged a suspicious entry with a value warning indicating a possible outlier or non-value. Heuristic-based stream selection was independently checked against a manual reference, and file-by-file summaries confirmed plausible stream identification across all 87 validation files.
Clinical plausibility was assessed by computing standard ambulatory glucose profile metrics from the harmonised data using the iglu package and comparing them against the manufacturer portal reports that clinicians normally rely on. The observed deviations were small and traceable to known differences between platform-level aggregation and the raw export content, which is precisely the kind of discrepancy researchers need to understand when they choose to compute metrics themselves rather than trust a vendor’s summary. Two researchers manually entered the portal PDF metrics into a spreadsheet and cross-checked each other’s entries, adding a layer of human verification to the comparison.
Benchmarking against alternatives delivered perhaps the most striking finding: almost none of the generic parsers and existing continuous glucose monitoring tools could parse the validation files directly. Standard CSV readers in both R and Python stumbled over the metadata blocks and structural irregularities, and established analysis packages such as cgmanalysis, CGManalyzer, CGMprocessing, CGM Shiny, and iglu were similarly limited in what raw exports they could ingest. This gap between what analysis tools expect and what manufacturers actually deliver is the hidden labour that GluCINDa formalises, and it explains why so many research groups have historically resorted to bespoke, undocumented cleaning scripts that no one else can reproduce.
By separating data ingestion from statistical analysis and retaining explicit provenance information for each extracted value, GluCINDa reduces manual effort, enhances reproducibility, and supports multi-centre and real-world research settings in which heterogeneous device data must be consolidated reliably. The provenance layer is particularly important for scientific integrity: every value in the harmonised output can be traced back to its source column and row in the original export, so reviewers and collaborators can audit exactly how each number arrived in the analysis dataset. The team is already extending the framework to insulin data, with planned outputs for basal, bolus, and total daily insulin that will unify records across manufacturers, including the tricky case of Medtronic exports where basal rates must be converted into delivered doses using recorded durations.
For a field racing toward large-scale real-world evidence, the arrival of a deterministic, privacy-preserving, vendor-agnostic ingestion layer is quietly significant. Multi-centre diabetes trials increasingly pool sensor data from patients using different devices and different download portals, and the credibility of any pooled analysis rests on the assumption that the underlying data were harmonised identically for every participant. GluCINDa makes that assumption verifiable rather than hopeful, and by releasing the tool as open-source software the Graz team has invited the research community to inspect, test, and extend it. The messy plumbing of glucose data may never make headlines on its own, but tools like this are what allow the headline findings above them to be trusted.
Subject of Research: An open-source tool for consolidating and harmonising continuous glucose monitoring data for diabetes research
Article Title: GluCINDa (Glucose: Coherent Import. Neat Data.) an open-source tool for the consolidation of continuous glucose monitoring data for research
Article References: Baumann, P. M., Buhlheller, C., Grebien, A., Taucher, C., Posch, F., Bally, L., & Mader, J. K. (2026). GluCINDa (Glucose: Coherent Import. Neat Data.) an open-source tool for the consolidation of continuous glucose monitoring data for research. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06643-5
Image Credits: AI Generated
DOI: 10.1186/s12859-026-06643-5
Keywords: continuous glucose monitoring, CGM, open-source software, data harmonisation, diabetes research, reproducibility, Python, data integration, LibreView, Dexcom Clarity, Medtronic CareLink, Glooko
News Source: Drew Townsend. (October 13, 2026). New Open-Source Tool Tames the Chaos of Continuous Glucose Monitoring Data. Scienmag.
