Using Machine Learning to Investigate Predictors of Fasting Blood Glucose: Insights into Circadian Timing and Age Interactions

Viktoriya Bu-Dager; Silvia Cirstea

cs.LGSep 26, 2026

Using Machine Learning to Investigate Predictors of Fasting Blood Glucose: Insights into Circadian Timing and Age Interactions

Authors: Viktoriya Bu-Dager, Silvia Cirstea

Organizations: School of Computing and Information Science, Anglia Ruskin University East Road, Cambridge CB1 1PT, U.K.

Abstract

Impaired glucose regulation is a major contributor to metabolic dysfunction and type 2 diabetes. This study developed an interpretable machine-learning framework to predict log-transformed fasting blood glucose using metabolic, hormonal, lifestyle, demographic, nutritional, and circadian variables from the National Health and Nutrition Examination Survey 2017--2020 pre-pandemic dataset. After merging multiple NHANES sub-datasets, data processing used a leakage-resistant pipeline in which imputation, scaling, and one-hot encoding were performed only after dataset splitting and within training folds. Elastic Net, LASSO, and XGBoost models were evaluated using 94 candidate predictors and engineered circadian interaction terms. Performance was assessed using mean absolute error, root mean squared error, coefficient of determination, calibration, and Shapley Additive Explanations. The final interaction-augmented XGBoost model achieved strong performance on the independent test set, with a mean absolute error of 0.0804, a root mean squared error of 0.1148, and a coefficient of determination of 0.7761, using 10 predictors. Glycohemoglobin was the dominant predictor, followed by insulin, diabetes diagnosis, gamma-glutamyl transferase, age, race, and gender. Among the engineered interaction terms, sleep midpoint multiplied by age was consistently retained in repeated random-split analyses, although its contribution remained modest relative to dominant glycaemic predictors. These findings support further investigation of circadian-age interactions in metabolic health.

Figures & tables

Figure 1: Overview of the study workflow. NHANES 2017–2020 pre-pandemic sub-datasets were merged to form the analytic sample ( $n=4{,}016$ ). A total of 94 candidate predictors were assembled across demographic, metabolic, circadian, lifestyle, hormonal, and nutritional domains. Preprocessing included biologically guided winsorization, circular encoding of sleep-related variables, survey coding-based recoding, and correlation-based filtering, after which 91 predictors remained for full-model development. The dataset was then split into a model-development set and an independent test set, with nested cross-validation applied within the development subset. Elastic Net. LASSO, and XGBoost models were trained, followed by interaction-based feature engineering, SHAP-guided feature selection, and construction of the final parsimonious interaction-augmented XGBoost model.

Category	Variables
Metabolic biomarkers	Body mass index (BMI), waist circumference, high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), glycohemoglobin (HbA1c; denoted as GHB in model variables), high-sensitivity C-reactive protein (CRP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transferase (GGT), uric acid, triglycerides, estimated glomerular filtration rate (eGFR), insulin, total bilirubin, systolic blood pressure, diastolic blood pressure, fasting glucose, blood urea nitrogen (BUN), lactate dehydrogenase (LDH), globulin, total protein, total cholesterol, and creatinine
Circadian-related variables	Bedtime, sleep health index, social jetlag hours, sleep apnoea risk, sleep midpoint (weekdays and weekends), sleep schedule category, circadian disruption indicators, and sleep duration variability
Lifestyle and medical history	Physical activity levels (work-related and recreational), family history of diabetes or cancer, smoking status (smoked at least 100 cigarettes in lifetime), sedentary time, lifestyle scores and categories derived from vigorous and moderate work/recreational activity and walking or bicycling, and doctor-diagnosed diabetes
Hormonal profiles	Sex hormone-binding globulin (SHBG), estradiol, androstenedione, luteinizing hormone (LH), and follicle-stimulating hormone (FSH)
Nutritional intake	Macronutrients (energy, carbohydrates, fat, fibre, and protein), micronutrients (vitamins and minerals), caffeine, alcohol, and dietary pattern indicators (e.g., sugar-free, low-carbohydrate, and high-protein diets)
Control variables	Gender, age, race, and family income-to-poverty ratio

Table 1: Variables included from the NHANES 2017–2020 pre-pandemic dataset.

Figure 2: Clinical and biochemical glucose patterns by diabetes status. (a) Violin plots showing fasting glucose distributions in diabetic and non-diabetic individuals. (b) Scatter plot of fasting glucose versus glycohemoglobin (HbA1c), coloured by diabetes status. Linear fits are shown for each group, with a stronger correlation observed in the diabetic cohort (

R^{2}=0.63

) than in the non-diabetic cohort (

R^{2}=0.50

Figure 3: Distribution of sleep-related variables across weekdays and weekends. (a,b) Polar histograms showing weekday and weekend bedtime distributions. (c) Smoothed density plots of sleep duration on weekdays and weekends. (d) Smoothed density plots of sleep midpoint on weekdays and weekends. Weekend distributions show later bedtimes, slightly longer sleep duration, and later sleep midpoint values, consistent with social jetlag patterns.

Model	MAE	RMSE	$R^{2}$ (Dev)	$R^{2}$ (Test)	CV $R^{2}$	Features
Model 1. Elastic Net (Full)	0.0881	0.1228	0.6793	0.7441	$0.672\pm 0.037$	91
Model 1b. LASSO (Full)	0.0914	0.1273	0.6554	0.7251	$0.650\pm 0.041$	91
Model 2. XGBoost (Full)	0.0795	0.1134	0.7863	0.7817	$0.699\pm 0.040$	91
Model 3. XGBoost (Selected)	0.0813	0.1173	0.7667	0.7666	$0.707\pm 0.038$	12
Model 4. XGBoost (+ Interactions)	0.0805	0.1152	0.7867	0.7749	$0.699\pm 0.040$	95
Model 5. XGBoost (+ Interactions, Selected)	0.0805	0.1164	0.7604	0.7701	$0.705\pm 0.037$	12

Table 2: Performance metrics of predictive models for fasting glucose.

Figure 4: Comparison of predictive performance across machine learning models. The blue line (left y-axis) shows test-set

R^{2}

, and the orange line (right y-axis) shows mean absolute error (MAE) for each model variant. Linear baselines (Elastic Net and LASSO) were outperformed by all XGBoost-based models. The full XGBoost model achieved the strongest overall test performance, whereas the final pruned interaction-augmented XGBoost model retained near-comparable accuracy using only 10 predictors, highlighting the trade-off between predictive performance and model parsimony.

Figure 5: Calibration of the final model, Model 5b (XGBoost (+ Interactions, Selected, Pruned)), on the independent test set. The scatterplot compares observed and predicted log-transformed fasting glucose values for individual participants. The dashed black line indicates perfect calibration (predicted = observed). The calibration slope (1.05) and intercept (-0.25) indicate reasonable agreement between predicted and observed values, with limited systematic bias across the prediction range.

Figure 6: SHAP-based feature importance for fasting glucose prediction in the final model, Model 5b (XGBoost (+ Interactions, Selected, Pruned)). (a) SHAP beeswarm plot showing the distribution and direction of feature effects across individual predictions, with points coloured by feature value. (b) Mean absolute SHAP values summarizing each feature’s average contribution to model output. GHB was the dominant predictor, followed by insulin, diabetes diagnosis, GGT, age, race, gender, and the Sleep Midpoint $\times$ Age interaction.

Figure 7: Partial dependence plots for key predictors in the final model, Model 5b (XGBoost (+ Interactions, Selected, Pruned)). Partial dependence plots show the marginal association of (a) GHB, (b) insulin, (c) diabetes diagnosis, and (d) age with predicted fasting glucose. The inset (e) in panel (d) shows SHAP values for the Sleep Midpoint $\times$ Age interaction, highlighting its nonlinear contribution to model predictions.

Figure 8: SHAP dependence plot for the Sleep Midpoint $\times$ Age interaction in the final model, Model 5b (XGBoost (+ Interactions, Selected, Pruned)). Each point represents an individual prediction and is colour-coded by age. Positive SHAP values indicate higher predicted fasting glucose, whereas negative SHAP values indicate lower predicted fasting glucose. The pattern suggests a nonlinear interaction between sleep timing and age in relation to model-predicted fasting glucose.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Sample construction step	N
Participants in the biomarker merge base with fasting glucose and insulin data available for merging	5,090
Excluded: rows with more than 50% missingness across the merged biomarker dataset	145
Remaining biomarker dataset after row-level missingness filtering	4,945
Participants in the sleep questionnaire dataset before filtering	10,195
Excluded: missing critical sleep variables required to derive circadian measures	222
Remaining sleep dataset after critical-variable filtering	9,973

Appendix

Table 3: Construction of the final analytic sample.

Figure 9: Python code used to derive engineered sleep and circadian-related variables, including circular encoding, social jetlag, sleep midpoint, sleep regularity metrics, the composite sleep health index, and an example interaction term.

Figure 10: Clustered correlation matrix of clinical, metabolic, hormonal, circadian, and demographic features. Colours range from blue (negative correlation) to red (positive correlation), centred at zero. Hierarchical clustering of the Pearson correlation matrix reveals groups of related variables across physiological and behavioural domains. Row and column sidebars are colour-coded by feature category, illustrating the multidimensional structure of the dataset and the relationships among metabolic and circadian-related variables.

Model	Hyperparameter search space
Model 1: Elastic Net (Full)	$\alpha\sim\mathrm{Uniform}(0.001,1)$ ; $l1\_ratio\sim\mathrm{Uniform}(0,1)$
Model 1b: LASSO (Full)	$\alpha\sim\mathrm{Uniform}(0.0005,1)$
Models 2–5b: XGBoost	$n\_estimators\in\{1000,1200,1500\}$ ; $learning\_rate\in\{0.01,0.015,0.02\}$ ; $max\_depth\in\{2,3\}$ ; $min\_child\_weight\in\{10,12,15\}$ ; $subsample\in\{0.7,0.8\}$ ; $colsample\_bytree\in\{0.7,0.8\}$ ; $reg\_alpha\in\{0.1,0.2\}$ ; $reg\_lambda\in\{10,15,20\}$ ; $gamma\in\{0.05,0.1\}$

Appendix

Table 4: Hyperparameter search spaces used for model tuning.

Model	Optimal settings
Model 1: Elastic Net (Full)	$\alpha=0.0720$ ; $l1\_ratio=0.0871$
Model 1b: LASSO (Full)	$\alpha=0.0207$
Model 2: XGBoost (Full)	$subsample=0.7$ ; $reg\_lambda=10$ ; $reg\_alpha=0.1$ ; $n\_estimators=1000$ ; $min\_child\_weight=12$ ; $max\_depth=2$ ; $learning\_rate=0.02$ ; $gamma=0.05$ ; $colsample\_bytree=0.8$
Model 3: XGBoost (Selected)	$subsample=0.7$ ; $reg\_lambda=15$ ; $reg\_alpha=0.1$ ; $n\_estimators=1200$ ; $min\_child\_weight=10$ ; $max\_depth=2$ ; $learning\_rate=0.02$ ; $gamma=0.05$ ; $colsample\_bytree=0.7$
Model 4: XGBoost (+ Interactions)	$subsample=0.8$ ; $reg\_lambda=10$ ; $reg\_alpha=0.2$ ; $n\_estimators=1200$ ; $min\_child\_weight=15$ ; $max\_depth=3$ ; $learning\_rate=0.01$ ; $gamma=0.1$ ; $colsample\_bytree=0.8$
Model 5: XGBoost (+ Interactions, Selected)	$subsample=0.7$ ; $reg\_lambda=15$ ; $reg\_alpha=0.2$ ; $n\_estimators=1000$ ; $min\_child\_weight=12$ ; $max\_depth=2$ ; $learning\_rate=0.02$ ; $gamma=0.05$ ; $colsample\_bytree=0.8$

Appendix

Table 5: Optimal hyperparameter settings identified during model-development tuning.

Model	Test $R^{2}$	MAE	RMSE	Tuned CV $R^{2}$
Model 4: XGBoost (+ Interactions)	$0.742\pm 0.036$	$0.0826\pm 0.0022$	$0.1215\pm 0.0071$	$0.7079\pm 0.0105$
Model 5: XGBoost (+ Interactions, Selected)	$0.744\pm 0.033$	$0.0824\pm 0.0023$	$0.1212\pm 0.0065$	$0.7136\pm 0.0100$
Model 5b: XGBoost (+ Interactions, Selected, Pruned)	$0.746\pm 0.033$	$0.0823\pm 0.0023$	$0.1207\pm 0.0068$	$0.7138\pm 0.0096$

Appendix

Table 6: Repeated random-split robustness summary across 10 random seeds for interaction-augmented XGBoost models.

Feature	Importance (mean)	Importance (SD)
GHB	1.0973	0.0252
Insulin (log)	0.0186	0.0018
Diabetes	0.0155	0.0021
GGT (log)	0.0121	0.0035
Age	0.0117	0.0025
Race	0.0097	0.0022

Appendix

Table 7: Permutation feature importance scores for the final model, Model 5b (XGBoost (+ Interactions, Selected, Pruned)).

Feature	Count	Frequency	Importance mean $\pm$ SD
GHB	10	1.0	$1.0585\pm 0.0635$
Insulin (log)	10	1.0	$0.0278\pm 0.0048$
Diabetes	10	1.0	$0.0156\pm 0.0038$
Age	10	1.0	$0.0139\pm 0.0039$
GGT (log)	10	1.0	$0.0103\pm 0.0041$
Race	10	1.0	$0.0093\pm 0.0015$

Appendix

Table 8: Feature-selection stability and across-seed permutation-importance summary for the final compact interaction-augmented model pipeline (Model 5b).

Model / comparison	Test $R^{2}$	MAE	RMSE	Tuned CV $R^{2}$
Model 5b	$0.7459\pm 0.0327$	$0.08230\pm 0.00232$	$0.1207\pm 0.0068$	$0.7138\pm 0.0096$
Model 5b (ablated)	$0.7456\pm 0.0317$	$0.08235\pm 0.00250$	$0.1208\pm 0.0068$	$0.7139\pm 0.0093$
Delta (Model 5b minus ablated)	$0.00030$	$-0.00005$	$-0.00009$	$-0.00010$

Appendix

Table 9: Repeated random-split targeted ablation analysis for Sleep Midpoint

\times

Age .

Model	Test $R^{2}$	MAE	RMSE	Tuned CV $R^{2}$
Restricted Model 4	$0.354\pm 0.016$	$0.1267\pm 0.0047$	$0.1931\pm 0.0112$	$0.3164\pm 0.0071$
Restricted Model 5	$0.338\pm 0.015$	$0.1280\pm 0.0046$	$0.1955\pm 0.0106$	$0.3179\pm 0.0096$
Restricted Model 5b	$0.340\pm 0.016$	$0.1278\pm 0.0049$	$0.1952\pm 0.0111$	$0.3176\pm 0.0093$

Appendix

Table 10: Restricted repeated random-split robustness summary after excluding major glycaemic predictors and related interaction terms.

Feature	Count	Frequency
Age	10	1.0
ACR (log)	10	1.0
GGT (log)	10	1.0
Triglycerides (log)	10	1.0
HDL	10	1.0
SHBG (log)	10	1.0

Appendix

Table 11: Feature selection frequency for the final restricted compact model (restricted Model 5b).

Explore similar work

Beyond the Blood Draw: Explainable Machine Learning for Non-Invasive Dysglycemia Risk Screening
Jun 14, 2026Black Sun, Chenyi Zhang, Kaiyi Ji +1Clinical Risk Prediction
Department of Computer Science, Aarhus University, Aarhus, Denmark · University at Buffalo, SUNY, Buffalo, New York, USA
From Prediction to Practice: A Task-Aware Evaluation Framework for Blood Glucose Forecasting
May 1, 2026Alireza Namazi, Heman ShakeriHealthcare Clinical Decision Support
Department of Computer Science University of Virginia Charlottesville, VA, USA · School of Data Science University of Virginia Charlottesville, VA, USA
GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting
Jun 5, 2026Baiying Lu, Zhaohui Liang, Ryan Pontius +2Zero-Shot Time Series Forecasting Time Series Forecasting
Department of Computer Science Dartmouth College Hanover, NH 03784 · Department of Computer Science Emory University Atlanta, GA 30322 · Quantitative Biomedical Sciences Dartmouth College Hanover, NH 03784

Jun 14, 2026J/K move · Enter open · S save

Beyond the Blood Draw: Explainable Machine Learning for Non-Invasive Dysglycemia Risk Screening

Black Sun, Chenyi Zhang, Kaiyi Ji, Xi Lu

Department of Computer Science, Aarhus University, Aarhus, Denmark · University at Buffalo, SUNY, Buffalo, New York, USA

Clinical Risk Prediction

Open paper AskPDF arXiv

Dysglycemia, encompassing both prediabetes and diabetes, affects huge numbers of adults worldwide, yet many of them remain undiagnosed. We developed and validated machine-learning (ML) models for non-invasive screening of dysglycemia risk that require no laboratory tests. Pooling data from the National Health and Nutrition Examination Survey (NHANES) 2017--2023 (n=14,352), we trained six ML models with stratified 5-fold cross-validation and compared them with two established clinical risk scores. LightGBM achieved the highest area under the receiver operating characteristic curve (AUC=0.820, 95% CI: 0.806--0.835), outperforming the Finnish Diabetes Risk Score (0.745) and American Diabetes Association Risk Test (0.783). SHAP analysis identified age, race/ethnicity, and waist-to-height ratio as the most influential predictors. Subgroup analyses confirmed consistent performance across demographic strata (AUC: 0.735--0.832). These results demonstrate the feasibility of explainable, laboratory-free dysglycemia screening for deployment in community settings and self-tracking health applications.

Abstract

Explore similar work

Beyond the Blood Draw: Explainable Machine Learning for Non-Invasive Dysglycemia Risk Screening

From Prediction to Practice: A Task-Aware Evaluation Framework for Blood Glucose Forecasting

GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting