This paper presents a data-driven approach for analyzing diabetes-related data, with a particular focus on pediatric patients with Type 1 Diabetes Mellitus (T1DM). We propose a methodology that integrates data from multiple wearable devices to examine the impact of physical activity and the use of medical technologies on patient health. Our approach utilizes the JupyterHub framework to combine data analytics and artificial intelligence (AI) within a secure, privacy-preserving environment. We conducted a preliminary case study based on data collected during a pediatric diabetes camp held in September 2024. The analysis employs advanced visualization tools (Plotly and hvPlot) and cross-correlation techniques to uncover patterns between physiological and glycemic parameters. The proposed system enables healthcare professionals to receive graphical and textual insights through an AI assistant, ultimately supporting more informed clinical decision-making and enhancing the quality of care for diabetic children.
To address these challenges, we introduce a methodological framework that integrates physical activity data and glycemic monitoring through a combination of traditional data analysis and artificial intelligence, with a focus on locally deployable learning models.
Methodology
2. Data analysis - evaluate the relationship between physical activity levels and blood glucose lfuctuations to determine the impact of exercise in diabetic patients, 3. Support system development - develop an AI-driven support system that will be able to assist medical specialists by providing real-time insights, graphical representations, and textual explanations based on data analysis, 4. Evaluation of Machine Learning methods - compare the efectiveness of traditional data analysis techniques with AI-driven approaches, accessing the accuracy, usability and practical value of the provided results.
The novelty of our work lies in its focus on pediatric diabetes, where the unique challenges of managing lfuctuating glucose levels, growth-related metabolic changes, and physical activity patterns demand tailored solutions. Furthermore, the integration of local AI-enabled platforms with wearable technologies not only enhances glycemic control in diabetic patients, but also assists healthcare professionals by ofering evidence-based recommendations through an intuitive and user-friendly chatbot interface.
In Section 2 we discuss related work. In Section 3 we present our case-study involving multi-modal sensor data. In Section 4 we describe the data processing pipeline designed for our case-study that involves data fusion, analysis, and visualization and the results obtained with classifical statistical methods. In Section 6 we present preliminary ideas and experiments related toward th integration of an AI agents based on SLM and RAG customized on our domain. In Section 7 we address limitations, conclusions and future work.
During the last decades, we were observing a huge progress both in informatics and medicine spheres.
They became inseparable parts of each other and one of the most vital benefits of such an alliance
is the significant growth of life expectancy. Artificial Intelligence (AI) is nowadays actively used to
support the research on Type 1 Diabetes. According to [
This study focuses specifically on pediatric diabetes, an area where managing blood glucose levels can be particularly challenging. Children and adolescents with diabetes experience unique physiological and behavioral factors that complicate glycemic control, including varying levels of physical activity, growth spurts, and psychosocial influences. The integration of advanced AI algorithms, insulin pumps, and wearable devices ofers a promising solution to these challenges, allowing for continuous monitoring and personalized treatment plans tailored to the specific needs of pediatric patients.
In our project, we leverage JupyterHub as a central platform for acquiring, processing, and analyzing
wearable data in a secure environment. This approach enables collaboration among researchers and
clinicians while ensuring that sensitive health data remains within institutional boundaries. The Jupyter
ecosystem (Hub, Lab and Notebook) [
Our JupyterHub instance is deployed on a virtual machine hosted by the joint laboratory of the Gaslini Pediatric Hospital and our university department. The system integrates data collected from two sources: insulin pumps and glucose monitors, and a textile wearable developed by ComfTech. These datasets are initially stored on proprietary cloud platforms managed by the device manufacturers. We redirect these data to the JupyterHub instance, ensuring local control and privacy. Figure 1 illustrates the system architecture: JupyterHub provides pre-configured environments for data ingestion, fusion, and analysis using the Python ecosystem. The platform allows the generation of analytical reports, which serve as inputs for AI-based assistants. These assistants are designed to support diabetologists by ofering structured insights, thereby reducing the cognitive load of interpreting raw data. JupyterLab also allows professionals to visualize both raw and aggregated data, giving them greater flexibility in exploring patterns related to glycemic control and physical activity.
The data used in this study were collected through a collaborative efort between the Gaslini
Pediatric Hospital and Comftech, a non-participatory spin-of of the Polytechnic University of Milan that
specializes in wearable monitoring devices [
Post-camp, the insulin-related data were downloaded from the Glooko platform, while physical activity data were obtained from Comftech’s proprietary system. Glooko is a diabetes management platform designed to facilitate the visualization, interpretation, and management of diabetes-related data from medical devices. A screenshot of the Glooko interface is shown in Figure 3.
The interface displays the data over a five-day period, including continuous glucose monitoring (CGM), carbohydrate intake, and insulin levels.
All data were anonymized and securely stored in a local repository for further analysis. For this case study, we focus on the dataset from a single anonymized patient to demonstrate our methodology.
Our dataset contains multi-modal time-series data captured over five days. Data formats included CSV, Excel, and JSON, with each device contributing distinct physiological and metabolic metrics. These datasets were organized for preprocessing and analysis within our JupyterHub environment. The dataset includes diferent time-series data on diferent physiological measurements. These parameters, recorded with specific timestamps, are structured as follows: Heart Rate, Breath Rate, Movement. The data encompasses the following fields: date, value, measure type, and timestamp. The dataset is structured in a tabular format with 165,138 entries and 4 columns, with each entry corresponding to a specific timestamp and the corresponding measurement values.
Continuous heart rate variability (HRV). The continuous HRV data includes the HRV score, stress index, and high-frequency (HF) power. The data consists of date, value, measure type, and timestamp ifelds, with a total of 447 records. Certain physiological signals were not directly available for download from the platform and were instead provided as raw data by the Comftech administrator in .csv and .json formats. These additional signals include: ECG trace (tacogram). The ECG data consists of single-lead ECG traces sampled at a rate of 128 Hz. The values are expressed in millivolts (mV). This dataset provides continuous physiological data over a substantial period, with 1,200,256 data points captured.
Respiratory signal. The respiratory data was acquired at a sampling rate of 13 Hz, with the values expressed in analog-to-digital converter (ADC) levels. This data captures respiratory activity over time, represented by 129,766 entries and three columns (date, value, and measure type). Acceleration signals The dataset includes triaxial accelerometer signals, capturing lateral (X-axis), vertical (Y-axis), and antero-posterior (Z-axis) accelerations. These signals were sampled at 25 Hz, and the values are expressed in gravitational units (g). There are 229,300 entries across three axes.
The data from the Glooko platform was downloaded in .csv format. The dataset includes various parameters related to insulin delivery, blood glucose monitoring, and alarms/events. These data are structured as follows: Insulin data The insulin-related data includes two main categories: basal and bolus insulin delivery, as well as overall insulin usage. The dataset provides information about the type of insulin used, the amount administered, and related temporal variables. Basal Insulin Data captures information on basal insulin delivery, which controls overall blood glucose levels, including the date and time of administration, type of insulin used, duration (in minutes), percentage of dosage, frequency, total insulin administered, and serial number. The dataset consists of 466 records. Bolus Insulin Data includes records on insulin bolus injections, which manages spikes caused by eating, with fields for the date and time, type of insulin, pre-meal blood glucose level (in mg/dL), carbohydrate consumption (in grams), carbohydrate-to-insulin ratio, total insulin administered, initial bolus delivery, extended bolus delivery, and serial number. This dataset contains 37 entries. Total Insulin Data includes information on the total bolus and basal insulin administered, as well as the overall insulin usage over time. It consists of six records with date, time, and the corresponding insulin values.
The blood glucose data This data contains entries for manually recorded blood glucose levels. The dataset includes the date and time of each measurement, the glucose value (in mg/dL), and whether the reading was taken manually. This dataset contains 32 records.
Continuous Glucose Monitoring (CGM) Data The CGM data records continuous glucose measurements obtained through a CGM system. The dataset includes the date and time of each glucose reading, the corresponding glucose value (in mg/dL), and the serial number of the device. This dataset comprises 1,067 entries. In summary, the dataset provides a wide and diverse range of both physiological and diabetes related measures, all of which were collected using various sensors mounted on the subject. The comprehensive nature of the dataset enables robust analysis and modeling of the subject’s physiological states over time.
Raw datasets were not immediately compatible with analytical workflows in JupyterLab. To address this, we employed a preprocessing pipeline using the ”Amphi” tool, which supports interaction with CSV files and allows for the construction of custom data workflows.
First, the datasets were separated into nine measurement categories: breath frequency, step count, heart rate, movement index, stress index, heart rate variability, basal insulin, bolus insulin, and CGM data. Each category was processed independently to ensure consistency in formatting and timestamps. Figure 4 illustrates the pipeline used to preprocess the basal and bolus data. The process begins by importing the original .csv file. The first step in the pipeline involves converting the date field from a string format to a Date/Time format. The “Select Columns” block is then employed to retain only the relevant columns—those containing bolus/basal data and corresponding timestamps. In the “Rename Columns” block, the selected columns are renamed for clarity and convenience. Finally, the modified .csv file, containing the preprocessed measurement data, is saved. This pipeline is subsequently applied to the CGM data. In contrast, the preprocessing algorithm for physical data derived from the Comftech platform requires adjustments (Figure 5). For instance, when processing the stress index and heart rate variability measures, the “Filter Rows” block isolates the required parameters from the full dataset.
The “Python Transforms” block is then used to convert the date column from string format into Date/Time format to ensure consistency in date formatting across the dataset. This procedure is repeated for the remaining six parameters. Upon completion of these steps within the Amphi environment, the nine preprocessed .csv files are transferred to an .ipynb notebook for additional processing. More specifically, the columns in each file were reordered to ensure that the date field appears first. Duplicate entries in the “Breath Frequency” file were removed due to the detection of repeated values. Furthermore, in the “Bolus” and “Basal” files, periods in the data were replaced with commas, and the data type was converted to “float64” for consistency. This pipeline resulted in a clean, structured dataset composed of nine CSV files, ready for detailed time-series analysis and machine learning tasks.
This section presents the outcomes of a detailed analysis of the processed dataset, conducted in the Jupyter Notebook environment using Python. We employed a range of data visualization and statistical tools to identify correlations and temporal trends within the physiological and glycemic data. Interactive visualizations were created using Plotly and hvPlot, while cross-correlation and correlation matrix techniques were used to explore interdependencies among variables.
To begin, all nine preprocessed CSV files were loaded into the notebook. A multi-axis scatter plot was generated using Plotly to visualize selected variables—such as glycemia, insulin delivery, heart rate variability (HRV), and stress index—across the five-day camp period. Each data point on the graph is interactive, displaying the associated value, date, and time. The plot includes four drop-down selectors for the axes and a checkbox to toggle between viewing the entire dataset or specific time intervals.
For example, in one configuration (Figure 6), glycemia, insulin levels, and movement index are visualized together. Glycemia levels (blue points) display a cyclical pattern, likely reflecting daily lfuctuations associated with meals, activity, and insulin injections. The corresponding insulin doses (orange points) indicate reactive administration in response to glycemic peaks. The movement index (purple points) suggests a possible relationship between physical activity and blood glucose regulation. The plot is equipped with five interactive widgets: four drop-down menus for selecting the axes and one checkbox for toggling between viewing data for all days or specific days. Users also have the option to exclude axes if fewer parameters need to be visualized.
As an example, the plot can display glycemia alongside administered insulin and movement index (Figure 7). The glycemia values, represented by blue points, exhibit a cyclic pattern with peaks and troughs, suggesting diurnal variations in blood glucose levels. It ranges between 0 and 261 units. Periods of increased glycemic values are interspersed with lower readings, corresponding to physiological or behavioral rhythms, such as meals or insulin administration. The erogated insulin in orange points suggest discrete delivery of insulin in response to elevated glycemia values. The movement index in purple points displays cyclical fluctuations, with periods of higher activity potentially coinciding with glycemic trends. This could indicate a relationship between physical activity and glycemic control.
To enable more flexible visual exploration, a merged dataset was created using timestamp alignment (nearest match method with 1-second resolution). New categorical columns were added to this dataset to represent diferent glycemia and movement levels (e.g., low, normal, high). This merged data was visualized using hvPlot, which provides dynamic plot types tailored for time-series analysis. The hvPlot scatter plot, visualizing all merged measures, is shown in Figure 8. This visualization method is particularly advantageous due to its versatility, allowing data to be explored through various plot types.
We then applied cross-correlation analysis to examine time-lagged relationships between variables. The merged dataset was split into parameters which were then visualized separately from each other.
The plot in Fig. 9 represents the cross-correlation between HeartRate and MovementIndex over a range of time lags, with a maximum lag of 150. Cross-correlation measures the similarity between two signals as one is shifted relative to the other, providing information on potential time-dependent relationships between variables.
Fig. 10 shows the cross-correlation between heart rate and glycemia over a range of time lags reveals the highest positive correlation at a lag of 123, indicating that an increase in heart rate is followed by a corresponding increase in glycemia approximately 123 time units later. In contrast, the strongest negative correlation occurs with a delay of 8, suggesting that an increase in heart rate precedes a decrease in glycemia within this short time frame. The correlation pattern fluctuates over diferent lags, with a significant negative correlation around lag 0, which may indicate an immediate inverse relationship between changes in heart rate and glycemia.
In contrast, the plot, which represents the cross-correlation between movement index and glycemia in Fig. 11, exhibits a more erratic pattern. Unlike the structured periodicity observed in the heart rateglycemia relationship, the correlation between movement and glycemia appears less consistent, with pronounced spikes particularly around lag 0. The presence of positive and negative correlations across diferent time lags suggests that the efect of movement on glycemia is influenced by additional factors, such as insulin administration, food intake, and individual physiological responses. The irregular nature of these fluctuations implies that movement alone may not be a reliable predictor of glycemia changes compared to heart rate.
The cross-correlation values are predominantly positive across all lags, indicating that HeartRate and MovementIndex are generally positively correlated. This aligns with the physiological expectation that higher levels of physical activity are associated with an increase in heart rate. The cross-correlation peaks around positive lags (e.g., 50 to 100), suggesting that changes in the MovementIndex tend to precede corresponding changes in HeartRate. This implies that heart rate increases with some delay after physical activity intensifies. The correlation matrix in Fig. 12 represents the relationships between all the variables. The values in the matrix range from -1 to 1. The strongest correlation is observed between Glycemia and Basal Insulin (0.79), highlighting a strong direct relationship. The stress index demonstrates a strong negative correlation with glycemia (-0.72), suggesting that increased stress levels are associated with lower glycemia, potentially due to metabolic changes or increased energy expenditure under stress. Interestingly, movement index has a moderate positive correlation with HRV score (0.67), suggesting that increased movement is associated with improved heart rate variability, which is often considered a marker of better autonomic function. Breathing frequency correlates positively with insulin delivery (0.59), possibly reflecting an association between metabolic demands and respiration. However, its correlation with stress is weak and negative (-0.17), suggesting that stress levels do not directly influence respiratory rate in a significant way.
There is a moderate positive relationship between HeartRate and MovementIndex (0.42), reflecting the physiological response of increased heart rate during physical activity. These insights highlight the potential of integrated multi-modal data to support more personalized, data-driven diabetes care.
In our preliminary experiments towards the creation of an AI-based assistants, we followed an approach
based on the combination of Small Language Models (SLMs), specialized embeddings and
RetrievalAugmented Generation (RAG) [
• Heart Rate: Number of heartbeats per minute over the dataset period (five days); • Breath Frequency: Number of breaths per minute: • Insulin Administration: Total insulin administered during the monitoring period; • Blood Glucose Levels (Glicemia): Concentration of glucose in the blood; • Physical Activity (Step Count): Number of steps taken during the monitoring period; • Movement Index: Movement intensity during the monitoring period; • Stress Index: Stress levels; • HRV Score: Heart Rate Variability (HRV) score reflecting the variation in time intervals between heartbeats; • Basal Insulin Administration: Basal insulin administered.
For this experiment, the knowledge base was exclusively based on the provided biometric data. During system prompting, additional contextual information was given more precisely, the glycemic values were categorized as follows.
• For a healthy person:
The retrieval mechanism was designed to return relevant data for answering queries. Initially, data was formatted as CSV, but this approach resulted in poor retrieval performance. Switching to a more descriptive format, where each measurement was stated in natural language, e.g. “Heart Rate Max Heart Rate 2024-09-09: 232.0, Min Heart Rate 2024-09-09: 42.0, Mean Heart Rate 024-09-09: 90.37208218293951”, improved the retrieval system’s accuracy. We include some example of PDF generated by the resulting notebooks containing max, min and mean values for all parameters for all days of a given (anonymous) user in Fig. We then formulate the list of 8 questions shown in Fig. 14 based on the considered data with the answers taken as ground truth.
With the help of the LangChain framework and Python libraries for extracting the knowledge based from our documents, we created a ChromaDB vector database to generate a retriever (using the ”as_retriever” method of the ChromaDB vd package) to generate contexts associated to the questions in our test list.
The context associated to a given question can then be used to specialize the answer submitted to a given LLM/SLM. Finally, we combined both the resulting retriever with a compressor to re-rank documents based on their relevance to a given query. The purpose of this combined structure is to streamline the retrieval and ranking process. This two-step process improves the quality of the final results, ensuring that the documents returned are not only relevant but also ranked according to their true relevance to the user’s query.
We then perform a series of tests with diferent SLM models (and diferent sizes/number of parameters) considering an embedding fine-tuned on medical data (more precisely MedEmbed-small-v0.1). For the considered list of questions the context retrieved by our retriever pipeline from the vector database turned out to contain the relevant part with respect to the question. Concerning hallucinations, the best results have been obtained with the Phi-3.5-mini-instruct model (0.25% wrong answers). These results are very preliminary since our dataset is still under construction and we are currently collecting additional data to perform more extensive training sessions and to create pipelines for redirecting queries to custom AI agents depending on the form of considered questions (e.g. descriptive question, query on tabular data, etc). As stated in the previous section, the LLM component should be seen as a conceptual demonstration rather than a final solution. A RAG system tailored specifically for biometric data analysis could be expected to provide more accurate and reliable responses.
This study demonstrates the feasibility of integrating wearable sensor data with AI-assisted analysis to support pediatric diabetes management. Our framework, which combines traditional data processing with machine learning techniques within a Jupyter-based environment, ofers valuable insights into the interplay between physiological signals and glycemic trends. While the current implementation focuses on a single-patient case study, the architecture is designed to scale across larger cohorts and additional device types.
A key challenge remains in bridging structured data analysis with natural language interpretation. Although our JupyterHub environment eficiently handles pre-processing and visualization, translating these findings into actionable recommendations through AI-powered conversational agents is still in early development. Future work will focus on the integration of small language models (SLMs) trained on domain-specific data to facilitate interactive querying and personalized decision support. Furthermore, the integration of the Jupyter ecosystem with data centric architectures such as Spark, Streaming Spark or Dask and data acquisition engine such as Kafka and RabbitMQ are currently under consideration to support large scale processing in real time.
We also plan to evaluate the scalability, usability, and clinical value of the platform in a broader study, incorporating feedback from diabetologists and healthcare providers. Additionally, integrating more advanced time-series modeling techniques and privacy-preserving machine learning approaches will be explored.
In conclusion, this work presents a novel methodology for leveraging wearable technology and AI to enhance pediatric diabetes care. The findings underscore the importance of multi-modal data integration and local computation for privacy and clinical ubtility. With further refinement, this system has the potential to evolve into a robust clinical decision support tool that empowers both healthcare professionals and patients.
Ultimately, integrating such platforms into hospital IT infrastructures and electronic health record systems could bring AI-driven, personalized treatment planning into routine clinical workflows, supporting more adaptive and responsive care for young patients with diabetes.
The paper has been written by the authors. Generative AI has been used for language corrections (DeepL) and to support code generation using Copilot.