Can Breath Biomarkers Causally Influence Blood Glucose? Investigating VOC-Mediated Modulation in Diabetes
Varsha Sharma, Prasanta K. Guha, Avik Ghose · 2026
Read Can Breath Biomarkers Causally Influence Blood Glucose? Investigating VOC-Mediated Modulation in Diabetes by Varsha Sharma, Prasanta K. Guha, Avik Ghose (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-22075.txt.
Introduction
Varsha Sharma, Prasanta Guha, and Avik Ghose explore whether breath biomarkers can causally influence blood glucose in diabetes. Diabetes burdens global health, and early detection matters for timely intervention. Using causal inference, they estimate the impact of volatile organic compounds such as acetone, isopropanol, isoprene, and ethanol on blood glucose levels, alongside lifestyle variables. They also build a classifier and a risk-based ranking system for people in the gray zone, finding specific compounds show strong causal influence on glucose and models can reliably classify high-risk individuals.
Catalog record
| CREATOR | Varsha Sharma, Prasanta K. Guha, Avik Ghose |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19743 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.22075
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Can Breath Biomarkers Causally Influence Blood Glucose? Investigating VOC-Mediated Modulation in Diabetes Varsha Sharma, Prasanta K. Guha, Avik Ghose Abstract Diabetes is a global health burden, and early detection is critical for timely intervention. This study explores a non-invasive, data-driven framework to identify individuals at risk of diabetes using Volatile Organic Compounds (VOCs) and lifestyle variables. We use causal inference techniques to estimate the impact of VOCs such as acetone, isopropanol, isoprene, and ethanol on blood glucose levels. Additionally, we designed a classifier to distinguish diabetics from non-diabetics using non-invasive markers. We created a risk-based ranking system for individuals in the “gray zone,” and identified natural clusters in the population using Gaussian Mixture Model. Our results suggest that specific VOCs exhibit a strong causal influence on glucose levels and that machine learning models can reliably classify and stratify individuals at high risk. This integrated causal-explainable analysis can support the development of tool for non-invasive early screening of diabetes. 1 Introduction Diabetes mellitus represents a growing global health challenge, with an estimated 537 million adults affected worldwide as of 2021 16 . Early detection and intervention are critical to prevent complications, yet current diagnostic methods such as fasting blood glucose tests and HbA1c measurements are invasive, intermittent, and often fail to capture early metabolic dysregulation 17 , 1 . There is an urgent need for non-invasive, real-time biomarkers that can identify at-risk individuals before clinical onset and provide insights into the underlying physiological mechanisms driving glucose dysregulation 18 . Recent advances in metabolomics have highlighted breath volatile organic compounds (VOCs) as promising candidates for metabolic monitoring. Breath VOCs like acetone 19 , 2 , isoprene 5 , ethanol 2 , and isopropanol 4 reflect distinct metabolic pathways (ketosis, cholesterol metabolism, and gut microbiome activity) 19 , 7 , 13 and collectively enhance non-invasive glucose monitoring 12 . While prior studies have established correlations between specific VOCs and diabetes, The causal relationship between these biomarkers and blood glucose modulation has not yet been fully elucidated. Correlation alone does not imply causation, and confounding factors such as medication use, lifestyle, and comorbidities may obscure true biological mechanisms. Several prior studies have examined the association between lifestyle factors and diabetes risk using machine learning and statistical methods. For example, 20 modeled diabetes risk from daily habits and diet, while 8 focused on Mediterranean diet adherence. 3 highlighted the role of physical activity in glycemic control, and 9 used wearable sensors to capture behavioral patterns for risk prediction. Others, such as 11 and 10 , explored causal modeling and interpretable ML with lifestyle and exposome data. However, these approaches largely overlook the causal role of VOCs, focusing instead on behavioral and physiological variables. The potential for VOCs as non-invasive causal indicators of diabetes onset remains underexplored. This study addresses that gap by applying causal inference techniques to determine whether breath-based VOCs actively influence blood glucose levels. In addition, we combine this with supervised and unsupervised machine learning approaches to identify high-risk individuals and uncover latent subgroups. Our key contributions are: 1. A causal framework for estimating the effect of VOCs on glucose levels. 2. A classification and risk-ranking system to flag high-risk non-diabetic subjects (latent subgroups). 3. A clustering analysis for segmenting diabetic and non-diabetic subjects. 2 Methods To explore the underlying cause-effect relationships between breath VOCs and blood glucose levels, we employed a structured causal inference framework. The goal was to identify whether individual VOCs or their combination causally influence glucose concentration, supporting their use as non-invasive biomarkers for early diabetes detection. 2.1 Data Description and Preprocessing The study involved 94 participants: 32 healthy, 8 newly diagnosed diabetics, 27 with well-controlled diabetes, and 27 with poorly controlled diabetes. Ethical approval was obtained from the Institutional Ethics Committee of IPGMER Hospital, India. Participant recruitment, clinical evaluation, and sample collection were conducted by medical professionals under expert supervision. Informed consent was obtained from all subjects, who refrained from alcohol and tobacco for at least 7 days and rested 30 minutes prior to sampling. Key variables such as age, gender, BMI, blood pressure, lifestyle habits (sleep, fruits per day, alcohol and tobacco consumption behavior), stress, comorbidities, and medications (metformin, etc.) were documented. Blood samples were analyzed for glucose and ketone levels, while breath samples (0.5L) were collected using the isothermal re-breathing method and analyzed via Gas Chromatography-Mass Spectrometry (GC-MS) for VOC profiling. Details of participant characteristics, sampling procedures and preprocessing steps are available in 15 . 2.2 Causal Association We investigated the following causal hypotheses: 1. Does individual VOCs causally influence blood glucose levels? 2. What is the combined causal effect of VOCs on blood glucose? 3. Are VOCs responsive to changes in blood glucose (reverse causality)? 2.2.1 Causal Inference Framework The framework begins with preprocessing and encoding of treatment variables, followed by constructing causal graphs and using propensity-based techniques to adjust for confounding. For each VOC (acetone, isopropanol, isoprene, ethanol), we modeled it as the treatment variable and blood glucose as the outcome. We employed the DoWhy model to estimate the average treatment effect (ATE) using backdoor criterion-based adjustment and linear regression as the base estimator. To validate these effects, we performed placebo-based refutation tests by permuting treatment assignments and re-estimating the effect, assessing whether the original causal estimate could be attributed to spurious correlations. In addition to individual VOCs, we also modeled the joint effect of all four VOCs as a multivariate treatment vector to assess cumulative influence on glucose levels. Furthermore, we applied reverse causality models to test whether glucose could be influencing VOC levels instead essential to rule out bidirectional dependencies and confirm causal directionality. All causal models were refuted using placebo interventions and observed to maintain effect significance. 2.2.2 Explainability and Confounding Checks To enhance the interpretability of our causal models and validate the relative importance of features influencing blood glucose, we integrated SHapley Additive exPlanations (SHAP) 6 into our analytical pipeline. Using a tree-based ensemble model (XGBoost) trained on VOCs and lifestyle variables, we computed SHAP values to visualize features that strongly influenced the model’s glucose predictions. Further, to assess the robustness of our causal estimates against varying confounding conditions, we systematically experimented with multiple sets of potential confounders and re-estimated the causal effects of individual VOCs (acetone, isopropanol, isoprene, and ethanol) as well as their combined influence on blood glucose. 2.3 Classification and Risk Ranking To evaluate the predictive power of non-invasive biomarkers identified through causal modeling, we trained classification models to distinguish diabetic and non-diabetic subjects using only VOCs and lifestyle data (excluding blood glucose and ketone). Models such as logistic regression, XGBoost, and random forest were employed. Furthermore, we generated a risk continuum by ranking predicted probabilities of diabetes, thereby identifying high-risk individuals in the non-diabetic group, forming a clinically important “gray zone” for early intervention. To validate the causal influence of breath VOCs on blood glucose levels and their potential in identifying high-risk non-diabetic individuals (gray zone), we created a synthesized glucose marker (synthetic glucose) as a linear combination of four VOCs (acetone, isopropanol, isoprene, and ethanol). The coefficients for the synthetic glucose were derived directly from the causal effect estimates obtained through our causal inference analysis, thereby embedding physiological relevance into the marker. We then used