Efective diabetes management is critical for reducing the risk of complications such as cardiovascular disease, nephropathy, and neuropathy while enhancing patient quality of life. Contemporary technologies like continuous glucose monitoring (CGM) and flash glucose monitoring (FGM) have improved clinical outcomes through realtime data and personalized care. However, these devices' high cost and invasive nature limit their accessibility and acceptance, particularly among uninsured or underinsured populations. This study proposes a non-invasive, cost-efective alternative by examining the relationship between heart rate and blood glucose levels in individuals with type 1 diabetes (T1DM). Machine learning techniques were employed to analyze patient data, including regression analysis, k-nearest neighbors (KNN), neural networks, ensemble bagged trees, and statistical methods such as ANOVA and Tukey's test. Results indicated that 96.3% of the cohort exhibited a statistically significant correlation between heart rate and blood glucose levels, with pronounced variations observed at extreme glycemic values. However, the heart rate was less responsive to moderate glucose fluctuations. These findings suggest that heart rate monitoring may serve as a viable non-invasive proxy for detecting significant glycemic events, ofering a promising alternative to traditional blood glucose monitoring systems and potentially mitigating the economic and physical burdens associated with current technologies.
https://caislab.di.unisa.it/about/ (M. Ruggiero); https://docenti.unisa.it/029717/home (L. D. Biasi); CEUR Workshop
ISSN1613-0073
although less frequent, can provoke dizziness, seizures, cardiac arrhythmias, and even death [
Several glycemic control methods are available. Hemoglobin A1C Testing (H1T) measures average
glucose over 90 days but lacks real-time feedback [
This work investigates noninvasive alternatives, relying on the hypothesis (WH1): “Changes in ECG (Electrocardiogram) patterns are a proxy for glucose level or danger.” Preliminary results [25, 26] support WH1. Higher glucose levels correlate with reduced heart rate variability (HRV) [26], and distinct HR distributions appear before hypoglycemia episodes [25]. Together, these findings suggest that heartbeat patterns reflect blood glucose trends.
Given the noninvasive nature of ECG wearables, heartbeat-based blood glucose monitoring (BCM) could improve patient comfort over traditional BGC. This study aims to answer two questions: “Is it possible to detect and discriminate between hypoglycemia or hyperglycemia events using ECG time series? – RQ1”; “Is it possible to detect modification in a heartbeat and estimate HR trends using glucose monitoring data time-series? – RQ2”.
This section discusses studies on the relationship between glycemic values and heart rate patterns. In [29], the authors examined the relationship between glycemic states (hypoglycemia, hyperglycemia, and normoglycemia) and ECG-derived features. They found a significant increase in heart rate when blood glucose drops below 60 mg/dL (hypoglycemia), with a p-value < 0.0001, but no significant correlation with hyperglycemia. In [30], a study of 148 type 1 diabetes patients revealed that higher HbA1c levels (reflecting higher blood glucose) were associated with increased heart rate, supported by a p-value < 0.004. The study in [31] used data from 31 patients to predict hypoglycemia with machine learning models using smartwatch and CGM data. The model correlated increased heart rate, reduced heart rate variability, and heightened electrodermal activity with hypoglycemia, achieving an ROC AUC of 0.76 ± 0.07. In [32], data from 128 type 2 diabetes patients showed that higher age and blood glucose were positively correlated. Increased parasympathetic activity also correlated with higher blood glucose levels, suggesting that changes in blood glucose may proportionally afect heart rate.
For experimentation, we used various pre-processing and statistical techniques.
A sliding window approach processed sequential data, with a moving mean calculated in each window to smooth fluctuations, reveal trends, and improve stability. ANOVA assessed diferences between group means by partitioning variance. A high F-statistic and p-value<0.05 indicate significant group diferences. In order to identify specific groups, Tukey’s HSD test was applied post-ANOVA, comparing all possible group pairs and controlling error rates. A p-value<0.05 denotes statistical significance. For predictive modeling, using distance metrics like Euclidean or Manhattan, K-Nearest Neighbors (KNN) estimated new instances based on the majority label or mean of K closest neighbors. Weighted K-Nearest Neighbors (WKNN) improved accuracy by giving closer neighbors greater weight via inverse distance. Finally, Ensemble-Bagging Trees (EBT) reduced variance by aggregating decision tree outputs trained on bootstrapped data subsets. While increasing computational cost, EBT yielded more generalized models. Hyperparameter tuning for EBT followed the same process as KNN.
This section outlines the methodologies employed to address the research questions of this study. We utilized a combination of statistical tests and experimental procedures (machine learning) to investigate the relationship between heart rate and glycemia in patients with T1DM. We are aware that we empirically estimated the parameters for sliding window size, time intervals for moving averages, and thresholds for min/max. While these selections yielded strong initial results, future work will focus on systematically optimizing these parameters to enhance performance further. In addition, a comprehensive sensitivity analysis will be undertaken to evaluate the robustness of the proposed approach against variations in parameter values and to ensure stability and generalizability across diferent operating conditions.
This contribution uses a pre-processed version of the HUPA-UCM Diabetes Dataset [27] extended by computing new features from glucose levels and heart rates. The original dataset comprises blood glucose levels, steps, calories, heart rate, and sleep data of 24 T1DM patients, sampled every five minutes, collected by the authors through CGM and Fitbit Ionic smartwatch, and contains up to 144085 rows. HUPA-UCM Diabetes Dataset contains only patients with type 1 diabetes mellitus, and no additional diseases are directly specified. It is important to consider patients 23 and 24, who have more glycemic readings than the remaining patients.
We extended HUPA-UCM by computing moving metrics, statistical features, and labels, as described below. We categorized blood glucose levels (BGL) to compare heart rate (HR) across glycemic groups. Each patient’s data was labeled based on BGL: very low (<60 mg/dL), low (60–89 mg/dL), good (90–179 mg/dL), high (180–249 mg/dL), and very high (>250 mg/dL). Moving averages of HR and BGL were computed using sliding windows to smooth fluctuations and capture trends. Maximum and minimum BGL values were calculated within each 10-observation window. Additional features included moving averages (11-observation centered and 5-observation shifted) and frequency counts for values above or below 1.15% and 1.25% of the max and min values in each window.
We calculated the heart mean (HM) as the average HR over six consecutive observations. The Heart Trend (HT) was then computed as the deviation of the HR sum from the previous HM, with negative HT indicating an increasing HR trend (labeled ”U”) and non-negative HT indicating a stable or decreasing trend (labeled ”D”).
We split the entire EDATA by Patient to predict Patient-specific heart rate trends, generating 24 subdataset (PSD).
Also, for each PSD, we generated two subsets of PSD, holding out 10% of the entire dataset for the test set, to evaluate the model’s performance on unseen data. During model training, a 20-fold cross-validation was also requested. We completed the training process individually for each patient to identify the most accurate models.
For each PSD, we used the features reported in Table 1 as prediction features and heart rate trend as the prediction class.
The ANOVA was used to test whether there was a statistically significant diference in heart rate as a function of diferent glycaemic levels in patients with DM: it was used to test for significant diferences Glucose Rate Moving Average (column_glucose_rate_moving)
A moving average centered on 11 observations for blood glucose levels (BGL).
Heart Rate Moving Average (heart_rate_moving_avg)
A moving average centered on 11 observations for heart rate (HR). (glucose_rate_left_avg)
A left-shifted moving average calculated over five observations for blood glucose levels (BGL). (heart_rate_left_avg)
A left-shifted moving average calculated over five observations for heart rate (HR).
Description Blood glucose levels of the patient.
The caloric intake associated with the patient.
The maximum blood glucose level within a sliding window.
The frequency of glucose values above 1.15% of the maximum value within a sliding window.
The frequency of glucose values above 1.25% of the maximum value within a sliding window.
The minimum blood glucose level within a sliding window.
The frequency of glucose values below 1.15% of the minimum value within a sliding window.
The frequency of glucose values below 1.25% of the minimum value within a sliding window.
Glucose Rate Left Shifted Average Heart Rate Left Shifted Average Max in Window (max_in_window) Frequency Above Threshold 1.15 (count_above_max_1.15) Frequency Above Threshold 1.25 (count_above_max_1.25) Min in Window (min_in_window) Frequency Below Threshold 1.15 (count_below_min_1.15)
Frequency Below Threshold 1.25 (count_below_min_1.25) Calories
Performance metrics of the model during testing between the heart rate averages in the various blood glucose groups defined above. In our case, p-value < 0.05 is considered the threshold for determining significance, suggesting that the variability in heart rate between glycemic groups is not due to chance.
In this T1DM scenario, if the p-value is less than 0.05, there is a significant diference in heart rate between at least two of the glycemic groups. Therefore, the blood glucose level could significantly impact heart rate, warranting further analysis. Also, a high F value indicates that the group variation is much greater than the internal variation, suggesting significant diferences.
The ANOVA Test does not allow for a precise understanding of the efect of blood glucose categories (glucose_label) on heart rate, so Tukey’s posthoc test was applied. Tukey’s test was applied to all patients who showed a p-value < 0.05 after the ANOVA Test to identify which blood glucose groups show significant diferences in heart rate.
Tables 2 and 3 summarize each patient’s validation and testing accuracies. Ensemble Bagged Trees emerged as the top performer in 20 out of 24 patients, while Weighted K-Nearest Neighbors led in the remaining four. Across patients, validation accuracy ranged from 60.4 % to 81.3 %, with a similar spread in test performance. This variability underscores that, although ECG-based predictors can capture glycemic trends, their reliability is patient-dependent.
ANOVA outcomes (Table 4) reveal that 23 of 24 patients (95.8 %) exhibited significant diferences between heart rate distributions across glycemic categories ( < 0.05 ). Patient 22 was the lone exception ( ≥ 0.05 ; = 1.74 ), indicating no clear heart rate stratification by glycemic state and suggesting that, in some individuals, ECG features may lack sensitivity to glucose shifts. The high -values observed in the other patients confirm pronounced efect sizes, particularly when contrasting extreme glycemic bands.
Post hoc Tukey comparisons (Table 4) further clarify these efects. The largest mean diferences occurred between “Very Low” and “Very High” glycemic states (MaD), with confidence intervals excluding zero—evidence of robust heart rate modulation at glycemic extremes. In contrast, adjacent labels such as “Good” versus “High” rarely reached statistical significance, implying that mid-range glucose changes induce subtler cardiovascular responses that may fall below the detection threshold of consumer-grade wearables.
Our findings align with prior work showing that ECG-derived metrics can flag hypo- and hyperglycemia, but add nuance by quantifying patient-level variability. Narasimhan et al. (2023) reported population-average heart rate increases of 5–7 bpm during hypoglycemic episodes, yet noted intersubject standard deviations exceeding 3 bpm—which mirrors our observed 20 % spread in model accuracy. Likewise, González et al. (2021) highlighted the need for personalized calibration to reach clinical-grade sensitivity, a recommendation reinforced by our patient-specific performance gaps.
Clinically, these results suggest two pathways forward. First, the consistent accuracy of Ensemble Bagged Trees in most patients indicates that tree-based ensembles can robustly capture non-linear ECG–glycemia relationships, making them strong candidates for embedded algorithms in wearable platforms. Second, the pronounced failure in certain individuals (e.g., Patient 22) and the modest performance on mid-range glucose shifts underscore the necessity of integrating complementary signals—such as photoplethysmography, activity level, or stress markers—to boost sensitivity and reduce false negatives in critical glycemic ranges.
However, consumer wearables—used here for heart rate capture—can sufer from motion artifacts and variable signal fidelity, particularly during exercise or when there is poor sensor contact. Our moving-average smoothing mitigated some noise, but real-world deployment will demand adaptive ifltering and on-device quality checks. Furthermore, the five-minute sampling interval may miss rapid glycemic excursions; integrating continuous or higher-frequency monitoring could unveil transient ECG patterns predictive of imminent hypo- or hyperglycemia.
Lastly, beyond model optimization, patient engagement and data privacy are pivotal. Personalized model training requires substantial amounts of labeled data, which can be burdensome for users. Federated learning approaches could reconcile the need for individualized calibration with privacy preservation, as has been trialed successfully in diabetes glucose prediction (Li et al., 2024). Future studies should therefore not only refine algorithmic approaches but also develop scalable pipelines for secure data collection, model updating, and clinical validation in diverse T1DM cohorts.
With the advent of modern FGM and CGM sensors, glycemic control in T1DM patients has become faster and more accurate. However, these technologies can be perceived as invasive and sufer from limited user acceptance. This study investigated whether non-invasive ECG tracking via consumer smart devices could serve as a viable alternative by exploring the relationship between blood glucose levels (BGL) and heart rate (HR). Two research questions were formulated (Section 1: Introduction) to examine both the existence of a statistical link and the feasibility of predictive modeling.
To address RQ1, we applied ANOVA and Tukey’s post-hoc tests to data from 24 T1DM patients. ANOVA revealed that 23 of 24 patients exhibited significant HR diferences across glycemic categories (p<0.05), with the most pronounced efects observed when comparing extreme glycemic states. Tukey’s tests localized these diferences, confirming that HR variability tracks BGL fluctuations in a statistically robust manner. Collectively, these findings provide strong evidence of a positive correlation between BGL and HR in T1DM.
RQ2 was explored through supervised machine learning models (KNN, WKNN, and Ensemble Bagged Trees), which exploit features such as glycemic rate of change, threshold counts, and caloric intake. The Heart Trend (HT) parameter successfully classified HR shifts as rising (’U’) or stable/decreasing (’D’), demonstrating that HR dynamics inferred from BGL trends can be predicted with moderate to high accuracy. Ensemble Bagged Trees achieved the best performance in most patients, highlighting the utility of non-linear ensemble methods for capturing complex ECG–glycemia relationships.
Our results underscore the potential of integrating ECG-based monitoring into routine diabetes management. Together, statistical analysis, machine learning and wearable devices could provide continuous, non-invasive alerts for impending hypo- or hyperglycemic events, reducing finger-prick frequency and improving patient comfort.
Although the examination of the relationship between heart rate patterns and glycemic states yields promising results for non-invasive glucose monitoring in patients with T1DM, several limitations should be considered when interpreting the findings and applying the methodology in clinical practice.
First, our cohort consisted of only 24 T1DM patients, without non-diabetic controls, which limited external validity. Also, this study focuses exclusively on patients with T1DM and cannot be generalized T2DM, which accounts for approximately 90–95% of all diabetes cases and involves diferent pathophysiological mechanisms that may afect the heart rate–glucose relationship. Moreover, the lack of control subjects without diabetes limits the ability to isolate diabetes-specific efects on cardiac dynamics.
Second, the five‐minute sampling interval and reliance on consumer‐grade wearables introduce temporal gaps and measurement noise that may obscure rapid glycemic excursions or subtle heart rate variability. Third, model parameters—such as sliding window size, moving‐average intervals, and classification thresholds—were tuned empirically; the absence of a formal sensitivity analysis raises concerns about robustness across diferent populations and operating conditions. Fourth, potential confounders (sleep quality, BMI, physical activity, emotional stress, medication) were recorded but not integrated into the predictive framework, limiting insight into multifactorial influences on HR–BGL dynamics.
To address these issues, future studies should: • Use bigger cohorts (including healthy controls) and balance glycemic reading distributions. • Increase sampling frequency or fuse multi‐sensor data (e.g., continuous ECG, PPG, accelerometry) to capture transient events. • Perform systematic parameter optimization (e.g., grid search, genetic algorithms[33], wavelet transforms) and comprehensive sensitivity analyses. • Develop personalized and privacy‐preserving modeling strategies—such as patient‐specific models, transfer learning (Digital Twins), or federated learning—to accommodate inter‐individual variability. • Incorporate additional physiological and behavioral variables (sleep metrics, BMI[34], stress markers, medication timing) and apply explainability methods (LIME, SHAP) to ensure clinical interpretability.
• Develop a multimodal system that can analyze extra data, like electrodermal activity (EDA). These eforts will be crucial in translating ECG-based approaches into reliable, non-invasive tools for diabetes management.
Despite these limitations, this research aims to establish a preliminary foundation for non-invasive ECG-based glucose monitoring approaches.
D3 4 Health – Digital Driven Diagnostics, Prognostics and Therapeutics for Sustainable Health Care (Project PNC0000001 – CUP: B53C22006090001), funded by the European Union – NextGenerationEU under the National Plan for Complementary Investments to the NRRP.
During the preparation of this work, the author(s) used Grammarly to check grammar and spelling. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [21] J. Wu, Y. Liu, H. Yin, M. Guo, A new generation of sensors for non-invasive blood glucose monitoring, American journal of translational research 15 (2023) 3825. [22] L. Messer, R. Johnson, K. Driscoll, J. Jones, Best friend or spy: a qualitative meta-synthesis on the impact of continuous glucose monitoring on life with type 1 diabetes, Diabetic Medicine 35 (2018) 409–418. [23] J.-P. Le Floch, B. Bauduceau, M. Lévy, H. Mosnier-Pudar, C. Sachon, B. Kakou, Self-Monitoring of Blood Glucose, Cutaneous Finger Injury, and Sensory Loss in Diabetic Patients, Diabetes Care 31 (2008) e73–e73. URL: https://doi.org/10.2337/dc08-1174. doi:10.2337/dc08-1174. arXiv:https://diabetesjournals.org/care/article-pdf/31/10/e73/598535/zdc01008000e73. [24] A. Berg, C. Zachariae, K. Nørgaard, J. Svensson, Skin problems due to treatment with diabetes technology: A narrative review, Medical Research Archives 11 (2023). URL: https://esmed.org/ MRA/mra/article/view/4747. doi:10.18103/mra.v11i11.4747. [25] V. Calcaterra, P. Bosoni, L. Sacchi, G. V. Zuccotti, S. Mannarino, R. Bellazzi, C. Larizza, Continuous glucose and heart rate monitoring in young people with type 1 diabetes: an exploratory study about perspectives in nocturnal hypoglycemia detection, Metabolites 11 (2020) 5. [26] J. P. Singh, M. G. Larson, C. J. O’Donnell, P. F. Wilson, H. Tsuji, D. M. Lloyd-Jones, D. Levy, Association of hyperglycemia with reduced heart rate variability (the framingham heart study), The American Journal of Cardiology 86 (2000) 309–312. URL: https://www.sciencedirect.com/science/ article/pii/S0002914900009206. doi:https://doi.org/10.1016/S0002-9149(00)00920-6. [27] Hidalgo, J. Ignacio; Alvarado, Jorge; Botella, Marta; Aramendi, Aranzazu; Velasco, J. Manuel; Garnica, Oscar (2024), “HUPA-UCM Diabetes Dataset”, Mendeley Data, V1, doi: 10.17632/3hbcscwz44.1 [28] G. B. Drummond, S. L. Vowler, Analysis of variance: variably complex, Advances in Physiology
Education 36 (2012) 85–88. [29] Nguyen, Linh Lan and Su, Steven and Nguyen, Hung T., Identification of Hypoglycemia and Hyperglycemia in Type 1 Diabetic patients using ECG parameters, 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2012, 2716-2719, doi=10.1109/EMBC.2012.6346525 [30] Michael Y Torchinsky and Ricardo Gomez and Jay Rao and Alfonso Vargas and Donald E Mercante and Stuart A Chalew, Poor glycemic control is associated with increased diastolic blood pressure and heart rate in children with Type 1 diabetes, Journal of Diabetes and its Complications, 2004, 220-223, https://doi.org/10.1016/S1056-8727(03)00031-X [31] Lehmann, Vera and Föll, Simon and Maritsch, Martin and van Weenen, Eva and Kraus, Mathias and Lagger, Sophie and Odermatt, Katja and Albrecht, Caroline and Fleisch, Elgar and Zueger, Thomas and Wortmann, Felix and Stettler, Christoph, Noninvasive Hypoglycemia Detection in People With Diabetes Using Smartwatch Data, Diabetes Care, 2023, 993-997, https://doi.org/10.2337/dc22-2290 [32] Novikov, Roman and Zhukova, Liudmila and Novopashin, Maxim, ossibility to Detect Glycemia with Heart Rate Variability in Patients with Type 2 Diabetes Mellitus in a Non-Invasive Glycemic Monitoring System, 2019 Actual Problems of Systems and Software Engineering (APSSE), 2019, 177-181, 10.1109/APSSE47353.2019.00030 [33] Di Biasi, Luigi and De Marco, Fabiola and Auriemma Citarella, Alessia and Barra, Paola and Piotto Piotto, Stefano and Tortora, Genovefa, Hybrid approach for the design of cnns using genetic algorithms for melanoma classification, International Conference on Pattern Recognition, 2022, 514-528, 10.1007/978-3-031-37660-3_36 [34] Islam, U., Mehmood, G., Al-Atawi, A., Khan, F., Alwageed, H. & Cascone, L. NeuroHealth guardian: A novel hybrid approach for precision brain stroke prediction and healthcare analytics. Journal Of Neuroscience Methods. 409 pp. 110210 (2024), https://www.sciencedirect.com/science/article/pii/S0165027024001559