Paper deep dive
GlucoTune: A Unified Framework for Blood Glucose Preprocessing, Forecasting, and Benchmarking in Diabetes
Davide Marelli, Giorgia Rigamonti, Mirko Paolo Barbato, Paolo Napoletano
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/24/2026, 2:39:25 AM
Summary
The paper introduces GlucoTune, a unified and extensible framework for blood glucose time-series data preprocessing, forecasting, and benchmarking in diabetes management. It addresses reproducibility issues by providing standardized, YAML-configurable preprocessing pipelines that handle missing data, feature alignment, and normalization without redistributing sensitive preprocessed data. The framework includes a curated library of state-of-the-art statistical, machine learning, and deep learning models for blood glucose prediction, along with a graphical user interface (GUI) and a benchmarking leaderboard to facilitate fair comparison and reproducible experimentation across public datasets like OhioT1DM and DiaTrend.
Entities (12)
Relation Signals (9)
GlucoTune → addresses → Type 1 Diabetes
confidence 95% · Preprocessing blood glucose time-series data is a critical yet often overlooked step in developing data-driven methods for diabetes management, particularly for type 1 diabetes.
GlucoTune → supports → DiaTrend
confidence 95% · (2) DiaTrend [18] features longitudinal data... This step also uniforms the naming conventions... Currently, the framework supports four major state-of-the-art Type 1 Diabetes (T1D) datasets.
GlucoTune → supports → OhioT1DM
confidence 95% · Currently, the framework supports four major state-of-the-art Type 1 Diabetes (T1D) datasets. (1) OhioT1DM
GlucoTune → uses → YAML
confidence 92% · Reproducible preprocessing is achieved through configurable pipelines defined in portable YAML configuration files
GlucoTune → includes → SMOTE
confidence 90% · It currently supports Synthetic Minority Over-sampling Technique (SMOTE) [27]
GlucoTune → provides → Graphical User Interface
confidence 90% · A graphical user interface (GUI) further improves accessibility by allowing users to configure, execute, and compare experiments
GlucoTune → includes → Transformer
confidence 85% · Transformer-based architectures (e.g., Informer, Autoformer, FEDformer)
GlucoTune → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Preprocessing blood glucose time-series data is a critical yet often overlooked step in developing data-driven methods for diabetes management, particularly for type 1 diabetes. The lack of standardized preprocessing workflows and evaluation protocols hinders reproducibility and complicates fair comparison across studies. These challenges are further exacerbated by data-sharing restrictions, as privacy and licensing constraints often prevent the redistribution of preprocessed medical datasets. To address these limitations, we present GlucoTune, a comprehensive and extensible framework for reproducible experimentation with blood glucose time-series data. The framework standardizes the entire experimental workflow, from preprocessing to model evaluation, enabling reproducible experiments directly from the original datasets. Reproducible preprocessing is achieved through configurable pipelines defined in portable YAML configuration files, ensuring consistent data handling without distributing sensitive preprocessed data. Beyond preprocessing, GlucoTune provides a unified interface for implementing, training, and evaluating blood glucose prediction models. The framework integrates public datasets through standardized wrappers and provides a curated collection of state-of-the-art blood glucose prediction and general time-series forecasting methods, while remaining readily extensible to additional datasets, preprocessing strategies, and forecasting models. To promote transparent and consistent evaluation, GlucoTune includes a benchmarking leaderboard that reports results across datasets, preprocessing configurations, and forecasting methods, enabling systematic comparison of experimental settings. We demonstrate the effectiveness of GlucoTune through comprehensive experiments and assess its usability in a user study.
Tags
Links
- Source: https://arxiv.org/abs/2607.21117v1
- Canonical: https://arxiv.org/abs/2607.21117v1
Trouble viewing inline? Open PDF directly →
Full Text
43,719 characters extracted from source content.
Expand or collapse full text
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS1 GlucoTune: A Unified Framework for Blood Glucose Preprocessing, Forecasting, and Benchmarking in Diabetes Davide Marelli, Giorgia Rigamonti, Mirko Paolo Barbato, Paolo Napoletano, Senior Member, IEEE Abstract — Preprocessing blood glucose time-series data is a critical yet often overlooked step in developing data-driven methods for diabetes management, particularly for type 1 diabetes. The lack of standardized preprocessing workflows and evaluation protocols hinders reproducibility and complicates fair comparison across studies. These challenges are further exacerbated by data-sharing restric- tions, as privacy and licensing constraints often prevent the redistribution of preprocessed medical datasets. To address these limitations, we present GlucoTune, a com- prehensive and extensible framework for reproducible ex- perimentation with blood glucose time-series data. The framework standardizes the entire experimental workflow, from preprocessing to model evaluation, enabling repro- ducible experiments directly from the original datasets. Reproducible preprocessing is achieved through config- urable pipelines defined in portable YAML configuration files, ensuring consistent data handling without distribut- ing sensitive preprocessed data. Beyond preprocessing, GlucoTune provides a unified interface for implementing, training, and evaluating blood glucose prediction models. The framework integrates public datasets through stan- dardized wrappers and provides a curated collection of state-of-the-art blood glucose prediction and general time- series forecasting methods, while remaining readily exten- sible to additional datasets, preprocessing strategies, and forecasting models. To promote transparent and consistent evaluation, GlucoTune includes a benchmarking leader- board that reports results across datasets, preprocessing configurations, and forecasting methods, enabling system- atic comparison of experimental settings. A graphical user interface (GUI) further improves accessibility by allowing users to configure, execute, and compare experiments with minimal programming effort. We demonstrate the effective- ness of GlucoTune through comprehensive experiments and assess its usability in a user study. Index Terms— Diabetes, Blood glucose prediction, Data Preprocessing, Baseline construction, Reproducibility. I. INTRODUCTION D IABETES mellitus is a chronic condition character- ized by insufficient or absent insulin production, lead- ing to impaired regulation of Blood Glucose Concentration The authors are with the Department of Informatics, Systems and Communication, University of Milano-Bicocca, Milano 20126, Italy (e-mail: davide.marelli@unimib.it; giorgia.rigamonti@unimib.it; mirko.barbato@unimib.it; paolo.napoletano@unimib.it). Correspondin author: Davide Marelli. (Davide Marelli and Giorgia Rigamonti con- tributed equally to this work.) (BGC) [1]. Persistent hyperglycemia (BGC > 180 mg/dL) is associated with an increased risk of long-term vascular complications [2], whereas hypoglycemia (BGC < 70 mg/dL) may result in severe acute adverse events [3]. The widespread adoption of Continuous Glucose Moni- toring (CGM) devices has enabled the development of a broad range of Machine Learning (ML) and Deep Learning (DL) approaches for BGC prediction. Existing methods span classical statistical models, such as ARIMA [4], traditional machine learning approaches, such as Random Forests [5], and modern deep learning architectures including CNNs [6], RNNs [7], [8], and Transformers [9], [10]. As the number and diversity of predictive models continue to grow, ensuring fair comparisons and reproducible experi- mental evaluation has become increasingly important. A major source of experimental variability lies in the preprocessing stage, where methodological choices differ substantially across studies. These differences include the selection of external features [11], [12], strategies for handling missing CGM data [13], [14], and the application of filtering, smoothing, and data augmentation techniques [15], [16]. Moreover, public datasets often use inconsistent naming conventions for similar variables and include heterogeneous data modalities beyond glucose measurements, requiring dataset-specific preprocess- ing procedures [17], [18]. These challenges are further exac- erbated by the sensitive nature of BGC data, as privacy regula- tions often prevent the redistribution of preprocessed datasets. Consequently, reproducing published experiments and estab- lishing standardized evaluation protocols become considerably more difficult, underscoring the need for standardized prepro- cessing workflows and transparent experimental pipelines. Several software tools have been proposed to support BGC prediction and benchmarking. Among these, GluPredKit [19] provides an open-source framework integrating preprocessing, model training, and evaluation. Although it represents an im- portant step toward structured benchmarking, its flexibility is constrained by a predefined preprocessing workflow and tight coupling with specific ML models, reducing its adaptability across different datasets and research objectives. Other tools primarily focus on CGM-derived metric compu- tation and exploratory analysis rather than configurable prepro- cessing for predictive modeling. For instance, iglu [20] offers a comprehensive set of CGM-derived metrics and visualiza- tions for assessing glucose control and variability, whereas arXiv:2607.21117v1 [cs.LG] 23 Jul 2026 2IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS Glucostats [21] enables efficient computation and visualization of a wide range of standardized glucose metrics. Similarly, Continuous Glucose Monitoring Time Series Data Analysis (CGMTSA) [22] approaches CGM as time-series data by pro- viding methods for missing-data imputation, outlier detection, metric computation, and interactive temporal visualizations. Although these tools provide valuable support for statistical analysis and clinical interpretation, they are not designed as modular and extensible frameworks for reproducible end-to- end machine learning experimentation. To address these limitations, we present GlucoTune 1 , a mod- ular and extensible framework that standardizes the complete workflow for reproducible experimentation with BGC time- series data, from preprocessing to model evaluation. Con- figurable preprocessing pipelines are fully specified through portable YAML configuration files, enabling experiments to be reproduced directly from the original datasets without requiring the redistribution of sensitive preprocessed data. The preprocessing layer supports configurable preprocessing operations, including missing-data handling, feature selection, glycemic thresholding, data partitioning, augmentation, and normalization. GlucoTune includes publicly available datasets through standardized wrappers, providing a consistent inter- face across heterogeneous data sources. Building upon this preprocessing layer, it provides a unified environment for im- plementing, training, and evaluating blood glucose prediction models. The framework integrates a curated collection of sta- tistical, ML, and DL models, including both diabetes-specific predictors and general-purpose time-series forecasting meth- ods, while remaining readily extensible to additional datasets, preprocessing strategies, and forecasting models. This archi- tecture enables researchers to establish strong experimental baselines and systematically compare forecasting approaches under consistent preprocessing and evaluation settings. Its modular architecture facilitates the seamless integration of additional datasets, preprocessing strategies, and forecasting models while preserving a unified and reproducible experi- mental workflow. To further improve accessibility, an intuitive Graphical User Interface (GUI) supports the complete workflow, from prepro- cessing configuration to model selection, training, evaluation, and result comparison, while integrated visualizations facilitate data exploration and interpretation. To evaluate the proposed framework, we present a bench- marking leaderboard reporting results across different pre- processing configurations and baseline forecasting models, enabling systematic and reproducible comparison under a unified experimental setting. We further assess the usability of GlucoTune through a dedicated user study. Together, these evaluations demonstrate the effectiveness of the framework and its potential to improve the reproducibility and compa- rability of blood glucose prediction research. I. METHODS This section presents the architecture of GlucoTune, includ- ing its configurable preprocessing pipeline, integrated frame- 1 § https://unimib-islab.github.io/glucotune Dataset Selection Missing Data Handling Features Alignment and Cleanup Events to Continuity Time Series Cropping Prediction Horizon Segmentation Data Splitting (train, val, test) Smoothing Data Augmentation Normalization Statistics and Charts generation Preprocessed Dataset Data InputPreprocessingOptional PreprocessingData Output Fig. 1: Workflow of GlucoTune data preprocessing; dashed lines indicate optional components that execute only in specific user configurations. work for baseline modeling and evaluation, and Graphical User Interface (GUI). GlucoTune is distributed as a collection of cross-platform Python packages covering data preprocessing, baseline fore- casting methods, and the GUI. Its modular architecture allows users to install and use only the components relevant to their needs while providing a standardized interface for integrating additional datasets, preprocessing strategies, and forecasting models. This design facilitates straightforward extension of the framework while preserving a unified and reproducible experimental workflow. To promote reproducible experimentation, preprocessing workflows are fully specified through portable YAML con- figuration files. This approach allows researchers to share complete preprocessing configurations without redistributing preprocessed datasets, whose redistribution is often restricted by privacy regulations and dataset usage policies. Conse- quently, other researchers can independently gain access to the original datasets and reproduce the entire experimental setup under identical conditions. A. Preprocessing Pipeline As illustrated in Fig. 1, this section presents the data preparation workflow implemented in GlucoTune. The prepro- cessing pipeline standardizes blood glucose time-series data through a configurable sequence of preprocessing operations. It currently supports four public datasets and integrates a broad range of preprocessing strategies derived from state-of-the-art methods, providing a flexible and extensible framework for preparing data for blood glucose prediction. The following subsections provide a detailed overview of its workflow. 1) Dataset Selection: Currently, the framework supports four major state-of-the-art Type 1 Diabetes (T1D) datasets. (1) OhioT1DM [17] contains eight weeks of data from 12 in- dividuals with T1D released for the 2018 and 2020 challenges. It includes CGM measurements sampled every 5 minutes, insulin delivery records (basal and bolus), self-reported car- bohydrate intake, event annotations (e.g., exercise and stress), and optional wearable sensor data such as heart rate and activity signals. (2) DiaTrend [18] features longitudinal data from 54 individuals with T1D. It comprises 27,561 days of CGM data at 5-minute intervals, 8,220 days of insulin pump MARELLI et al.: GLUCOTUNE: A UNIFIED FRAMEWORK FOR BLOOD GLUCOSE PREPROCESSING, FORECASTING, AND BENCHMARKING IN DIABETES3 data (including basal insulin for 17 subjects), bolus doses, carbohydrate intake, pump settings, and detailed demographic and clinical profiles. (3) T1DiabetesGranada [23] is one of the largest and longest open-access longitudinal datasets for type 1 diabetes research. It includes 257,780 days of 15- minute CGM recordings collected over four years from 736 patients (4) T1DEXI [24] contains data of 497 adults with type 1 diabetes participating in a four-week structured at- home exercise study. The dataset includes continuous glucose monitoring data sampled every 5 minutes, together with heart rate, insulin dosing, and self-reported exercise and dietary intake collected via a custom smartphone application. This step also uniforms the naming conventions of available data across datasets to standardize the subsequent processing. 2) Missing Data Handling: During this phase, users can decide whether to handle missing data by ensuring a uniform interval between observations and filling in gaps caused by missing BGC readings via interpolation. Additionally, users can specify the maximum allowable gap in CGM data before splitting the time series. If a missing interval exceeds this threshold, the time series will be segmented. 3) Feature Alignment and Cleanup: Depending on the cho- sen dataset, users can pick different features. First, GlucoTune synchronizes the selected features with CGM readings, en- suring that each BGC timestamp corresponds to observations within a maximum delay defined by the observation frequency of the dataset. This process eliminates potential information gaps before the BGC measurements, improving data consis- tency. Next, it applies filtering techniques to both features and CGM readings to correct discrepancies and remove outliers caused by recording errors. 4) Events to Continuity: If bolus insulin or carbohydrate intake data is selected, users can enable an optional continuity process to ensure a more consistent representation of these features, following the approach in [25]. 5) Time Series Cropping: Given that some features may span a more limited period than CGM readings, if the users selected features that are present in a limited time span the system crops all-time series to ensure that all features remain temporally consistent. 6) Prediction Horizon Segmentation: The time series is di- vided into samples according to the user-defined PH, the selected number of past observations (input), and the specified number of future observations (target). For example, if the PH is 30 minutes and CGM readings occur every 5 mins, each time series segment consists of 30 samples (24 input + 6 target). 7) Data Splitting: Users can choose how to split the time series and define the percentage of data allocated to the train, validation, and test sets. The framework supports three strate- gies: (1) Temporal split, where data is divided chronologically to train on past data, validate on intermediate data, and test on future data for a realistic performance assessment; (2) Stratified split, which ensures a balanced distribution of hypo- /normo-/hyper-glycemic events across train, validation, and test sets; and (3) Hybrid split, combining both approaches. For this stage, GlucoTune allows users to customize hypoglycemia and hyperglycemia thresholds which impacts the data distribu- tion in the splits. This flexibility supports varying definitions in research. For example, while some studies classify hypo- glycemia as below 70 mg/dL, others use 80 mg/dL. Similarly, hyperglycemia can be defined as above 180 mg/dL or 250 mg/dL, depending on study criteria (e.g., [13], [26]). 8) Smoothing: Users can optionally apply smoothing filters to CGM data to reduce noise. The framework uses a Gaussian filter with configurable sigma value, allowing fine control over the smoothing process. 9) Data Augmentation: GlucoTune incorporates data aug- mentation techniques which can be optionally used to address class imbalance. It currently supports Synthetic Minority Over- sampling Technique (SMOTE) [27], which users can cus- tomize by applying a multiplier factor or specifying the target number of hypo- and hyperglycemia events. It can be applied globally across the dataset or individually for each subject. 10) Statistics and Charts Generation: At this stage the pre- processing is mostly complete and statistics on the data are extracted as well as plots about the data split, glucose data, and sample time series are generated. 11) Normalization: Lastly, users can choose whether to apply data normalization, with the option to select Min-Max Normalization or Max Normalization. B. Model Library and Baseline Construction Once the preprocessing stage is completed, GlucoTune enables users to construct baseline models, facilitating straight- forward comparison with state-of-the-art approaches. In this section, we describe the model library and the baseline con- struction module of the framework. This functionality stream- lines experimental evaluation, allowing custom methods to be benchmarked within a unified and reproducible environment. The current model library spans multiple methodological families, including classical statistical approaches, traditional machine learning algorithms, and a wide range of neural archi- tectures. Specifically, it comprises classical statistical models (e.g., Naive, ARIMA), tree-based machine learning methods (e.g., XGBoost), and several neural model families such as recurrent networks (e.g., RNN, LSTM, GRU, BiGRU), con- volutional and temporal convolutional networks (e.g., TCN), Transformer-based architectures (e.g., Informer, Autoformer, FEDformer), linear and decomposition-based models (e.g., N- BEATS), as well as more recent approaches including MLP- Mixer variants, probabilistic models, graph-based models, and LLM-based approaches. Most of the implemented models have been selected for their relevance to blood glucose predic- tion tasks. In addition, the library includes well-established general-purpose time series forecasting models adapted to the glucose prediction setting, thereby broadening the range of comparative baselines. Fig. 2 illustrates the hierarchical organization of the model library, highlighting the different methodological families and their associated implementations. As with the preprocessing component, is designed to be highly modular, extensible, and flexible. Its architecture en- ables the seamless integration of new models with minimal effort, allowing the framework to evolve alongside advances in time series forecasting and glucose prediction methodologies. 4IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS Fig. 2: General overview of the models currently available in the GlucoTune model library. C. Graphical User Interface The GUI is available as a separate cross-platform Python package, making it an optional companion to GlucoTune (see Fig. 3). While GlucoTune is fully operable via the command-line interface, the GUI enhances accessibility for non-developers by guiding users through the whole workflow. Users can start with the preprocessing pipeline, selecting and customizing the configuration that best fits their experimental needs. They may then export the selected configuration for future reuse or sharing, or directly execute the newly defined strategy. The GUI also supports importing, modifying, and running existing configuration files. After completing the preprocessing step, users can continue within the GlucoTune GUI to build a custom baseline. The framework supports the selection and configuration of different modeling ap- proaches within the same environment, facilitating straight- forward benchmarking against state-of-the-art methods under consistent preprocessing and evaluation settings. 1) Data visualization functionality in the GUI: To support data understanding before and after preprocessing, the GUI includes a visualization page that displays glucose trends, distribution across predefined glucose ranges, and data avail- ability over time (Fig. 3). These plots offer an intuitive overview that aids in interpreting and validating glucose data. I. RESULTS We assess GlucoTune through a predictive modeling bench- mark and a usability study, highlighting its technical utility and user experience. A. Predictive Modeling Benchmark To validate the effectiveness of GlucoTune, including both preprocessing pipeline and integrated model library, we design (a) (b) Fig. 3: Graphical User Interface of GlucoTune. a unified benchmarking framework covering all currently supported models. The evaluation begins by preprocessing the selected datasets using the configurable pipeline of GlucoTune, ensuring consistency across experiments. We conduct the analysis on two representative datasets among those currently supported, namely OhioT1DM and DiaTrend, considering PH of 30 and 60 minutes, and com- paring configurations with and without SMOTE-based data augmentation. All models are trained and evaluated under the same protocol to ensure fair comparison across heterogeneous families: we adopt a temporal split, apply max normalization, and optimize using Mean Squared Error (MSE) loss. This setup reduces variability due to preprocessing and training choices, helping isolate differences attributable to the models. The full leaderboard 2 reports a comprehensive set of evalu- ation metrics commonly adopted in blood glucose prediction, including Root Mean Squared Error (RMSE, overall prediction error), Mean Absolute Error (MAE, average absolute devi- ation), Mean Absolute Relative Difference (MARD, relative prediction error), Time Gain (TG, temporal prediction advan- tage), sensitivity and specificity to hypo- and hyperglycemic events (Sen hypo , Sen hyper , Spec hypo , Spec hyper ), Pearson Correlation Coefficient (PCC, linear agreement), Prediction in clinically critical zones D and E (PDE, probability of clinically wrong treatment) [28], and time lag (τ lag , systematic prediction delay) [29]. These metrics are jointly considered to provide a standardized and comprehensive evaluation, address- ing the fragmented use of individual metrics in prior work, and capturing complementary aspects of predictive performance, 2 https://unimib-islab.github.io/glucotune/ leaderboard MARELLI et al.: GLUCOTUNE: A UNIFIED FRAMEWORK FOR BLOOD GLUCOSE PREPROCESSING, FORECASTING, AND BENCHMARKING IN DIABETES5 including accuracy, temporal alignment, and clinical relevance. For clarity, Table I reports the top 10 models using a subset of widely adopted metrics (RMSE, MAE, MARD, TG, sensi- tivity and specificity to hypo- and hyperglycemia), providing a concise yet informative summary of predictive performance across datasets and prediction horizons [13], [30]. To facilitate comparison across different classes of forecasting methods, the table also reports the number of trainable parameters and Floating Point Operations (FLOPs), allowing readers to jointly assess predictive accuracy and computational com- plexity. Model selection is performed at the family level: all statistical and traditional machine learning models are included due to their limited number, while neural architectures are ranked within each family using a composite score derived from normalized key metrics (RMSE, TG, and sensitivity), retaining only the top-performing models. In this case, model selection is performed at the family level: all statistical and traditional machine learning models are included due to their limited number, while neural architectures are ranked within each family using a composite score derived from normalized key metrics (RMSE, TG, and sensitivity), retaining only the top-performing models. In addition to quantitative performance metrics, GlucoTune automatically generates a set of visual evaluation plots (Fig. 4) to facilitate qualitative comparison of forecasting methods. Among these, the Clarke Error Grid [31] and the Parkes Error Grid [32] assess the clinical relevance of glucose predictions by classifying each forecast into risk zones (A–E). Ideally, predictions should fall within Zone A, indicating clinically accurate estimates, whereas Zones B and C correspond to progressively higher clinical risk and Zones D and E rep- resent potentially dangerous predictions that may lead to inappropriate treatment decisions. In addition, the confusion matrix summarizes the percentages of correctly and incorrectly classified glycemic events, providing further insight into the classification performance of each forecasting method. To further analyze the benchmarking results, we employ bubble plots that relate clinically relevant performance metrics to model complexity, quantified by the number of trainable parameters and FLOPs, across all evaluated forecasting mod- els. These visualizations facilitate the analysis of the trade- off between predictive performance and computational cost (Fig. 5). B. System usability study To assess usability of GlucoTune’s GUI, we conducted a study with participants from the University of Milano- Bicocca. Each user performed a standardized preprocessing task using GlucoTune, involving the OhioT1DM dataset, se- lection of CGM, bolus, and carbohydrate features, smoothing (σ=1), max normalization, and a 30-minute PH. All users performed the task in the same controlled environment without external disturbances, no time limit was imposed. After the task, participants completed a questionnaire combining the standard System Usability Scale (SUS, questions 1–10) [33] with five custom items targeting GlucoTune-specific features: (1) I found it easy to configure GlucoTune according to the requirements; (2) I found it useful to preprocess the dataset without writing any code; (3) The preprocessing steps were clearly explained and easy to follow; (4) The graphical inter- face supported me well throughout the task; (5) I would feel confident repeating this preprocessing workflow in the future. A total of 11 participants completed the study. The results indicate that GlucoTune is highly usable, with an average SUS score of 80.91 and an additional 22.18 out of 25 from the custom questions. The average usability score set from a study on 500 systems is 68 [34]. Participants described the tool as intuitive, appreciated the no-code approach, and felt confident completing the task autonomously. A few suggestions were made to improve usability: users recommended adding tooltips to explain key features, making the interface more self- explanatory and accessible. These insights will inform future refinements, aimed at enhancing user guidance through clearer, more interactive support in GlucoTune. IV. DISCUSSION The experimental evaluation demonstrates the practical value of GlucoTune as a unified framework for reproducible blood glucose prediction research. Rather than emphasizing differences between forecasting models, the benchmarking results show that predictive performance is strongly influenced by the experimental setting, including dataset characteristics, preprocessing choices, and prediction horizon. These findings highlight the importance of standardized preprocessing and evaluation protocols to enable fair and meaningful comparison across studies. By standardizing the complete experimental workflow, from preprocessing to model evaluation, GlucoTune reduces the variability introduced by heterogeneous data preparation and implementation-specific preprocessing and evaluation choices. Consequently, differences in predictive performance can be interpreted with greater confidence as reflecting the fore- casting models themselves rather than inconsistencies in the experimental pipeline. This directly addresses one of the main challenges in blood glucose prediction research, where heterogeneous preprocessing procedures and restrictions on sharing preprocessed datasets often hinder reproducibility and objective benchmarking. Beyond improving reproducibility, GlucoTune simplifies the development and evaluation of forecasting methods through configurable preprocessing pipelines, portable YAML config- uration files, standardized dataset wrappers, and an integrated collection of baseline forecasting methods. Together, these components reduce the engineering effort required to repro- duce experiments, establish strong baselines, and systemati- cally compare forecasting approaches under identical prepro- cessing and evaluation settings, while allowing the framework to be readily extended with additional datasets and models. Another important contribution of GlucoTune is the pro- motion of comprehensive model evaluation. In addition to conventional predictive metrics, the framework incorporates clinically relevant evaluation measures and model complexity indicators, including the number of trainable parameters and FLOPs. This multidimensional benchmarking strategy enables 6IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS TABLE I: Benchmarking of forecasting performance and computational complexity across datasets, forecasting models, and prediction horizons. PH = 30 minPH = 60 min DatasetModelTypeAug.Params.FLOPsRMSE↓MAE↓MARD↓TG↑Sen hypo ↑Sen hyper ↑Spec hypo ↑Spec hyper ↑RMSE↓MAE↓MARD↓TG↑Sen hypo ↑Sen hyper ↑Spec hypo ↑Spec hyper ↑ PH=30 minPH=30 min(mg/dL)(mg/dL)(%)(min)(%)(%)(%)(%)(mg/dL)(mg/dL)(%)(min)(%)(%)(%)(%) OhioT1DM Naive Statistical –0016.1810.657.107.6174.1385.9199.2598.5625.7417.1611.448.3360.1875.9698.8897.70 ARIMA–37K16.289.316.1514.1386.1790.6798.7598.2425.8315.6710.3521.0574.7082.4198.1397.21 XGBoostML–115K37K13.618.645.8315.5857.6284.9699.6799.0928.7618.7511.1227.8527.3480.9099.8996.65 Transformer Neural ✓293K14M13.178.295.5917.1572.9687.9599.3998.9222.0714.579.9228.8234.0773.2299.7198.63 NBEATS✓3M5M13.288.285.5617.1571.1386.6499.4899.0622.4014.8910.2727.6226.1074.2799.8098.49 BiGRU✓702K33M12.968.155.4617.1070.4185.8799.5299.2221.4213.869.2830.2635.4274.7499.7198.75 SOFTS✓7M14M13.568.485.6515.1681.4588.9599.1798.8822.7314.749.8423.8362.3178.7298.8898.17 TCN✓150K9M13.458.505.7116.3370.0886.4099.4899.0521.9914.709.9128.2733.9772.2599.7098.57 KAN✓369K645K13.618.485.7415.5966.4286.2999.5199.0722.2014.509.7827.9045.0472.7899.4998.67 TimeLLM–177M99G14.099.296.2112.4574.4786.4599.3998.9322.8815.3310.2722.6951.4373.6599.2598.50 DiaTrend Naive Statistical –0021.3013.878.057.8960.7688.1899.4096.8733.9122.3012.948.9946.1380.8399.1894.93 ARIMA–37K21.7812.427.1313.9779.4091.8198.8096.6234.1420.7811.9120.2464.7585.3898.2794.74 XGBoostML–148K48K17.4311.106.5315.6447.9289.7699.8097.7328.7618.7511.1227.8527.3480.9099.8996.65 Transformer Neural –293K14M16.9410.696.2716.9858.8890.7299.6697.5928.2118.2610.7231.2433.9381.5299.7996.63 NBEATS✓2M5M17.0710.766.3016.2955.3890.5499.7397.6228.5418.4510.8629.4730.5281.9499.8696.48 BiGRU✓702K33M16.9210.746.3316.7050.0890.1199.8197.7728.2218.3510.8130.7825.8380.7999.9296.88 SOFTS✓7M14M17.7511.176.4914.1673.7591.3599.2997.2329.8419.2911.2924.2152.2083.7799.1995.60 TCN✓150K9M16.9710.806.3516.6554.6390.9099.7397.5028.2618.3610.8530.2328.6082.2099.8796.43 KAN✓369K645K17.1310.866.3816.5854.3790.5299.7497.6028.5018.5410.9031.1529.4181.2399.8596.60 TimeLLM–177M99G18.5112.317.1612.1659.7088.7099.5397.3930.5020.7412.2123.6134.3178.7899.6796.27 050100150200250300350400 Reference glucose concentration (mg/dl) 0 50 100 150 200 250 300 350 400 Predicted glucose concentration (mg/dl) A B BC C D D E E Clarke Error Grid 0.02 0.04 0.06 0.08 0.10 Probability (%) (a) Clarke Error Grid 050100150200250300350400450500550 Reference glucose concentration (mg/dl) 0 50 100 150 200 250 300 350 400 450 500 550 Predicted glucose concentration (mg/dl) A A B B C C D DE Parkes Error Grid 0.02 0.04 0.06 0.08 0.10 Probability (%) (b) Parkes Error Grid HypoNormoHyper Reference glucose concentration Hypo Normo Hyper Predicted glucose concentration 70.4%0.5%0.0% 29.6% 98.7%14.1% 0.0%0.8% 85.9% Confusion matrix (c) Confusion Matrix Fig. 4: Plots generated by the GlucoTune Baselines module for the BiGRU model on the OhioT1DM dataset with a 30-minute prediction horizon and data augmentation enabled. researchers to analyze the trade-offs between predictive per- formance, clinical relevance, and computational complexity. The benchmarking results further illustrate that improvements in predictive performance are not necessarily accompanied by proportional increases in computational complexity, em- phasizing the importance of jointly considering accuracy and efficiency when selecting forecasting models. The usability study complements the technical evaluation by demonstrating that the proposed framework is accessible in practice. Participants reported a mean SUS score well above the commonly accepted benchmark, indicating that the graphi- cal interface effectively supports reproducible experimentation for users with different levels of programming expertise. These results suggest that standardized experimental workflows can be made accessible without sacrificing flexibility. V. CONCLUSION In this work, we presented GlucoTune, a unified and exten- sible framework for reproducible experimentation with blood glucose time-series data. This standardization enables more reliable comparison of forecasting methods and provides a solid foundation for future methodological developments in blood glucose prediction research. Our experimental evaluation highlights the importance of standardized preprocessing and evaluation protocols for blood glucose prediction research. By combining configurable pre- processing pipelines, standardized dataset wrappers, a com- prehensive collection of baseline forecasting methods, inte- grated benchmarking, and an intuitive graphical user interface, GlucoTune enables transparent, reproducible, and comparable evaluation of forecasting approaches. The usability study fur- ther demonstrates that these capabilities are accessible to users with different levels of programming expertise, supporting the adoption of reproducible experimental workflows across a broad research community. The modular architecture of GlucoTune is designed to accommodate future methodological advances. Additional datasets, preprocessing strategies, forecasting models, and evaluation metrics can be integrated through standardized interfaces while preserving compatibility with existing work- flows. Future work will focus on expanding the collection of supported datasets and forecasting methods, incorporating additional preprocessing and evaluation techniques, further enhancing the framework’s visualization and analysis capabil- ities, and broadening the applicability of GlucoTune beyond Type 1 Diabetes to other diabetes management scenarios, including Type 2 Diabetes. GlucoTune addresses a key need in blood glucose pre- diction research by providing a standardized infrastructure for rigorous experimentation. Rather than introducing another forecasting architecture, it establishes a unified framework for the rigorous and systematic evaluation of both existing and future prediction methods. This standardization enables more reliable comparison of forecasting methods and provides a solid foundation for future methodological developments in blood glucose prediction research. MARELLI et al.: GLUCOTUNE: A UNIFIED FRAMEWORK FOR BLOOD GLUCOSE PREPROCESSING, FORECASTING, AND BENCHMARKING IN DIABETES7 (a)(b) (c)(d) Fig. 5: Bubble plots illustrating the relationship between sensitivity to adverse glycemic events and model complexity, quantified by the number of trainable parameters (a, c) and FLOPs (b, d), on the OhioT1DM and DiaTrend datasets. ACKNOWLEDGMENT This work was funded by the National Plan for NRRP Com- plementary Investments (PNC, established with the decree-law 6 May 2021, n. 59, converted by law n. 101 of 2021) in the call for the funding of research initiatives for technologies and innovative trajectories in the health and care sectors (Directo- rial Decree n. 931 of 06-06-2022) - project n. PNC0000003 - AdvaNced Technologies for Human-centrEd Medicine (project acronym: ANTHEM) 3 . This work reflects only the authors’ views and opinions, neither the Ministry for University and Research nor the European Commission can be considered responsible for them. Computationalresourcesprovidedbyhpc- ReGAInS@DISCo (https://w.disco.unimib.it). REFERENCES [1] A. D. Association, “Diagnosis and classification of diabetes mellitus,” Diabetes care, vol. 33, no. Supplement 1, p. S62–S69, 2010. 3 https://fondazioneanthem.it/ [2] M. Kotagal, R. G. Symons, I. B. Hirsch, G. E. Umpierrez, E. P. Dellinger, E. T. Farrokhi, D. R. Flum et al., “Perioperative hyperglycemia and risk of adverse events among patients with and without diabetes,” Annals of surgery, vol. 261, no. 1, p. 97–103, 2015. [3] P. E. Cryer, S. N. Davis, and H. Shamoon, “Hypoglycemia in diabetes,” Diabetes care, vol. 26, no. 6, p. 1902–1912, 2003. [4] R. H. Shumway, D. S. Stoffer, R. H. Shumway, and D. S. Stoffer, “Arima models,” Time series analysis and its applications: with R examples, p. 75–163, 2017. [5] A. Bhimireddy, P. Sinha, B. Oluwalade, J. W. Gichoya, and S. Purkayastha, “Blood glucose level prediction as time-series modeling using sequence-to-sequence neural networks,” in KDH@ECAI. CEUR Workshop Proceedings, 2020. [6] J. Freiburghaus, A. Rizzotti, and F. Albertetti, “A deep learning approach for blood glucose prediction of type 1 diabetes,” in Proceedings of the Proceedings of the 5th International Workshop on Knowledge Discovery in Healthcare Data co-located with 24th European Conference on Artificial Intelligence (ECAI 2020), 29-30 August 2020, Santiago de Compostela, Spain, vol. 2675, 2020. [7] J. Martinsson, A. Schliep, B. Eliasson, and O. Mogren, “Blood glucose prediction with variance estimation using recurrent neural networks,” Journal of Healthcare Informatics Research, vol. 4, p. 1–18, 2020. [8] K. Li, J. Daniels, C. Liu, P. Herrero, and P. Georgiou, “Convolutional recurrent neural networks for glucose prediction,” IEEE journal of biomedical and health informatics, vol. 24, no. 2, p. 603–613, 2019. [9] Y. Xue, S. Guan, and W. Jia, “Bgformer: An improved informer model to enhance blood glucose prediction,” Journal of Biomedical Informatics, vol. 157, p. 104715, 2024. [10] Q. Bian, A. As’ arry, X. Cong, K. A. b. M. Rezali, and R. M. 8IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS K. b. Raja Ahmad, “A hybrid transformer-lstm model apply to glucose prediction,” PLoS One, vol. 19, no. 9, p. e0310084, 2024. [11] M. M. H. Shuvo and S. K. Islam, “Deep multitask learning by stacked long short-term memory for predicting personalized blood glucose concentration,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 3, p. 1612–1623, 2023. [12] G. Rigamonti, M. P. Barbato, D. Marelli, and P. Napoletano, “Improving detection of type-1 diabetes adverse events using gru networks,” in 2024 IEEE 8th Forum on Research and Technologies for Society and Industry Innovation (RTSI). IEEE, 2024, p. 79–84. [13] X. Zheng, S. Ji, and C. Wu, “Predicting adverse events for patients with type-1 diabetes via self-supervised learning,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, p. 1526–1530. [14] C. Rodriguez Leon, O. Banos, O. Fernandez Mora, A. Martinez Bedmar, F. Rufo Jimenez, and C. Villalonga, “Prediction of blood glucose levels in patients with type 1 diabetes via lstm neural networks,” in International Work-Conference on Artificial Neural Networks. Springer, 2023, p. 563–573. [15] H. N. Mhaskar, S. V. Pereverzyev, and M. D. Van der Walt, “A deep learning approach to diabetic blood glucose prediction,” Frontiers in applied mathematics and statistics, vol. 3, p. 14, 2017. [16] M. De Bois, M. A. El Yacoubi, and M. Ammi, “Adversarial multi- source transfer learning in healthcare: Application to glucose prediction for diabetic people,” Computer Methods and Programs in Biomedicine, vol. 199, p. 105874, 2021. [17] B. R. Marling C., “The OhioT1DM Dataset for Blood Glucose Level Prediction: Update 2020,” CEUR workshop proceedings, p. 71–74, 2020. [18] T. Prioleau, A. Bartolome, R. Comi, and C. Stanger, “Diatrend: A dataset from advanced diabetes technology to enable development of novel analytic solutions,” Scientific Data, vol. 10, no. 1, p. 556, 2023. [19] M. K. Wolff, S. Royston, A. L. Fougner, H. G. Schaathun, M. Steinert, and R. Volden, “Glupredkit: Development and user evaluation of a standardization software for blood glucose prediction,” arXiv preprint arXiv:2406.08915, 2024. [20] S. Broll, J. Urbanek, D. Buchanan, E. Chun, J. Muschelli, N. M. Punjabi, and I. Gaynanova, “Interpreting blood glucose data with r package iglu,” PloS one, vol. 16, no. 4, p. e0248560, 2021. [21] P. Peiro-Corbacho, F. J. Lara-Abelenda, D. Chushig-Muzo, A. M. W ̈ agner, C. Granja, and C. Soguero-Ruiz, “Glucostats: an efficient python library for glucose time series feature extraction and visual analysis,” BMC bioinformatics, vol. 26, no. 1, p. 230, 2025. [22] J. Shao, Z. Liu, S. Li, B. Wu, Z. Nie, Y. Li, and K. Zhou, “Continuous glucose monitoring time series data analysis: a time series analysis pack- age for continuous glucose monitoring data,” Journal of Computational Biology, vol. 30, no. 1, p. 112–116, 2023. [23] C. Rodriguez-Leon, M. D. Aviles-Perez, O. Banos, M. Quesada- Charneco, P. J. Lopez-Ibarra Lozano, C. Villalonga, and M. Munoz- Torres, “T1diabetesgranada: a longitudinal multi-modal dataset of type 1 diabetes mellitus,” Scientific Data, vol. 10, no. 1, p. 916, 2023. [24] M. C. Riddell, Z. Li, R. L. Gal, P. Calhoun, P. G. Jacobs, M. A. Clements, C. K. Martin, F. J. Doyle I, S. R. Patton, J. R. Castle et al., “Examining the acute glycemic effects of different types of structured exercise sessions in type 1 diabetes in a real-world setting: the type 1 diabetes and exercise initiative (t1dexi),” Diabetes care, vol. 46, no. 4, p. 704–713, 2023. [25] H. Butt, I. Khosa, and M. A. Iftikhar, “Feature transformation for efficient blood glucose prediction in type 1 diabetes mellitus patients,” Diagnostics, vol. 13, no. 3, p. 340, 2023. [26] Y. Deng, L. Lu, L. Aponte, A. M. Angelidi, V. Novak, G. E. Karniadakis, and C. S. Mantzoros, “Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients,” NPJ Digital Medicine, vol. 4, no. 1, p. 109, 2021. [27] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: synthetic minority over-sampling technique,” Journal of artificial intel- ligence research, vol. 16, p. 321–357, 2002. [28] A. Neumann, Y. Zghal, M. A. Cremona, A. Hajji, M. Morin, and M. Rekik, “A data-driven personalized approach to predict blood glucose levels in type-1 diabetes patients exercising in free-living conditions,” Computers in biology and medicine, vol. 190, p. 110015, 2025. [29] T. Zhu, K. Li, J. Chen, P. Herrero, and P. Georgiou, “Dilated recurrent neural networks for glucose forecasting in type 1 diabetes,” Journal of Healthcare Informatics Research, vol. 4, no. 3, p. 308–324, 2020. [30] M. P. Barbato, G. Rigamonti, D. Marelli, and P. Napoletano, “Lightweight sequential transformers for blood glucose level prediction in type-1 diabetes,” IEEE Journal of Biomedical and Health Informatics, 2025. [31] W. Clarke, C. DC, L. Gonder-Frederick, W. Carter, and S. Pohl, “Evaluating clinical accuracy of systems for self-monitoring of blood glucose,” Diabetes care, vol. 10, p. 622–8, 09 1987. [32] A. Pf ̈ utzner, D. C. Klonoff, S. Pardo, and J. L. Parkes, “Technical aspects of the parkes error grid,” Journal of Diabetes Science and Technology, vol. 7, no. 5, p. 1275–1281, 2013. [33] J. Brooke et al., “Sus-a quick and dirty usability scale,” Usability evaluation in industry, vol. 189, no. 194, p. 4–7, 1996. [34] J. Sauro, “Measuring usability with the system usability scale (sus),” 2011. [Online]. Available: http://w.measuringusability.com/sus.php