Re-evaluating Heart Rate Variability Biomarkers for Glucose Sensing: The Impact of Age Normalisation and Subject-Independent Validation
1st Md Basit Azam
Department of Computer Science & Engineering
Tezpur University
Napaam - 784 028, Tezpur, Assam, INDIA
📧 mdbasit@tezu.ernet.in
2nd Sarangthem Ibotombi Singh
Department of Computer Science & Engineering
Tezpur University
Napaam - 784 028, Tezpur, Assam, INDIA
📧 sis@tezu.ernet.in
| Metric | HbA1c cohort (n = 29) | FBG cohort (n = 38) |
|---|---|---|
| Best model | Extra Trees | Extra Trees |
| R² | 0.222 | 0.086 |
| MAE (original scale) | 1.18 percentage points | 2.27 mmol/L (41 mg/dL) |
| Pearson r (p) | 0.476 (0.009) | 0.344 (0.034) |
| Permutation test p | 0.002 | 0.002 |
| Bootstrap 95% CI for R² | [0.13, 0.82] | [0.10, 0.72] |
| Age normalisation benefit | None (19/20 combinations ≤ baseline; one trivial exception ΔR²=+0.0001 for HbA1c) | None (all 20 combinations worse) |
Clinical context: Bootstrap CIs exclude zero for both cohorts, providing statistical evidence for genuine HRV–glycemic associations. However, the FBG lower bound (0.10) is close to zero, and both CIs are wide due to small sample sizes (n = 29–38), reflecting substantial uncertainty. A clinically viable non-invasive glycemic estimator would require R² > 0.7 with errors confined to Clarke Error Grid zones A and B a threshold not reached in this study. These findings should be interpreted as hypothesis-generating preliminary evidence only.
pip install -r requirements.txtPlace your dataset in the working directory with this structure:
./Dataset_on_electrocardiograph/dataset_ecg/
├── clinical_indicators.xlsx
├── objective_sleep_quality.xlsx
├── subjective_sleep_quality.xlsx
├── ECG/
│ ├── 20200101.mat
│ └── ...
└── RR_interval/
├── 20200101.mat
└── ...
python complete_preprocessing.pyThis creates processed_data_v2/ with extracted features, separated cohort targets, and LOSO fold definitions.
python run_complete_analysis.pyThis executes three steps sequentially:
| Step | Script | Description |
|---|---|---|
| 1 | comprehensive_baseline.py |
20-model baseline comparison with LOSO + 6 age-adjustment methods + 20-parameter sensitivity grid |
| 2 | ablation_study.py |
13-configuration feature domain ablation |
| 3 | validation_framework.py |
Permutation testing (n=500), bootstrap CIs (n=500), residual diagnostics, learning curves |
Note: The runner checks that
processed_data_v2/exists before proceeding. All scripts must be run in order as each depends on the previous step's outputs.
Or run individual components:
# Step 1: Preprocessing only
python run_complete_analysis.py --only-preprocessing
# Step 2: Baseline comparison only (requires preprocessing)
python run_complete_analysis.py --skip-preprocessing --only-baselines
# Step 3: Ablation study only
python run_complete_analysis.py --skip-preprocessing --only-ablation./
├── processed_data_v2/ # From complete_preprocessing.py
│ ├── features.csv # 105 extracted features
│ ├── signal_specifications.json # ECG/HRV signal documentation
│ ├── ecg_scaling_logs.json # ECG amplitude scaling audit
│ ├── targets/
│ │ ├── hba1c_cohort.csv # HbA1c targets (n=29)
│ │ └── fbg_cohort.csv # FBG targets (n=38)
│ └── loso_splits/
│ ├── hba1c_cohort/ # LOSO fold definitions
│ └── fbg_cohort/
│
├── analysis_results_v3/ # From comprehensive_baseline_revised.py
│ ├── hba1c_cohort/
│ │ ├── baseline_comparison.csv # 20-model R², MAE, correlation
│ │ ├── baseline_comparison.png # Model comparison bar charts
│ │ ├── age_adjustment_comparison.csv # 6 methods × R² results
│ │ ├── prediction_scatter.png # Predicted vs actual plot
│ │ └── summary.json # Cohort-level summary
│ ├── fbg_cohort/
│ │ └── ...
│ ├── dual_cohort_model_comparison.png # Side-by-side cohort comparison
│ ├── feature_importance_by_domain.png # Domain contribution analysis
│ ├── feature_selection_stability_heatmap.png # Fold-by-feature binary heatmap
│ └── age_sensitivity_heatmap.png # 5×4 parameter grid heatmap
│
├── ablation_results_v3/ # From ablation_study_revised.py
│ ├── hba1c_cohort/
│ │ ├── ablation_results.csv # 13 configurations × metrics
│ │ └── ablation_figure.png # Domain ablation bar chart
│ └── fbg_cohort/
│ └── ...
│
└── validation_results_v3/ # From validation_framework_revised.py
├── hba1c_cohort/
│ ├── validation_report.json # Full statistical report
├── fbg_cohort/
└── ...
| File | Purpose |
|---|---|
complete_preprocessing.py |
Loads raw clinical, ECG, and sleep data from the Mendeley dataset; extracts 105 features across 6 domains; validates ECG signal amplitude and documents scaling; creates separated HbA1c/FBG cohort targets and LOSO fold splits |
comprehensive_baseline_revised.py |
Runs 20 models under LOSO with within-fold SelectKBest (k=15) and StandardScaler; compares 6 age-adjustment methods; performs 20-combination sensitivity analysis; generates publication figures |
ablation_study_revised.py |
Evaluates 13 feature-domain configurations using Bayesian Ridge under LOSO with within-fold preprocessing; quantifies contributions of clinical, ECG, HRV, sleep, and demographic feature groups |
validation_framework_revised.py |
Permutation testing (n=500), bootstrap 95% CIs (n=500, subject-level resampling), residual diagnostics (Shapiro-Wilk, bias, heteroscedasticity), learning curve analysis |
run_complete_analysis.py |
Sequential runner for the three analysis steps; checks processed_data_v2/ exists before proceeding |
This study analyses cross-sectional associations between ECG-derived features and glycemic status across subjects — not real-time glucose sensing or within-subject temporal prediction. The ECG–glucose relationship examined is correlational and cross-sectional, using spot measurements (HbA1c and FBG) obtained during hospitalisation. This constraint reflects dataset availability and represents a conservative analytical approach less susceptible to overfitting on within-subject temporal autocorrelation.
HbA1c (reflecting 3-month average glycemic control) and fasting blood glucose (FBG; reflecting acute metabolic status) are analysed as strictly separate cohorts, preventing the common methodological error of combining fundamentally different glucose metrics. Both targets were log-transformed to address distributional skewness and improve regression stability.
This separation prevents the common error of combining fundamentally different glucose metrics, which confounds the physiological interpretation.
All preprocessing occurs strictly within each LOSO fold:
For each held-out subject:
1. SelectKBest(f_regression, k=15) fitted on training subjects only
2. StandardScaler fitted on training subjects only
3. Held-out subject transformed using training-derived parameters
4. Model fitted and prediction recorded
This prevents information leakage from held-out test subjects into feature selection or scaling — the single most impactful methodological correction in this study.
| Category | Models |
|---|---|
| Naïve (2) | Mean predictor, Median predictor |
| Linear (7) | OLS, Ridge (α=0.1, 1.0), Lasso (α=0.1), ElasticNet, Bayesian Ridge, Huber Regressor |
| Tree ensembles (4) | Random Forest, Extra Trees, Gradient Boosting, AdaBoost |
| SVM (3) | SVR with RBF, linear, and polynomial (degree 2) kernels |
| Neural networks (4) | MLP (32), MLP (64,32), MLP (128,64,32), MLP (64,32) tanh |
All models use minimally configured hyperparameters (scikit-learn defaults with no nested tuning), deliberately providing conservative baselines on small samples. Neural network results should be interpreted as performance under these specific constraints — small tabular data with no architecture search — rather than as a general assessment of neural architectures for physiological prediction tasks. The extreme negative R² values observed for MLPs reflect numerical instability under LOSO on very small samples, not a general unsuitability of these models.
| Method | Description |
|---|---|
| No adjustment | Baseline (raw features) |
| Proposed formula | HRV / (age/65 + 0.1), threshold from Umetani et al. (1998) |
| Residualisation | Regress age out of HRV features via linear regression |
| Age-bin z-score | Z-score within age quartiles (young/middle/senior/elderly) |
| Polynomial interaction | Age² + HRV × age interaction terms |
| Simple division | HRV / age |
Additionally, a sensitivity analysis tests 20 parameter combinations (5 age thresholds: 55, 60, 65, 70, 75 × 4 stability constants: 0.05, 0.1, 0.15, 0.2). For HbA1c, 19/20 combinations performed at or below baseline; one trivial exception (threshold 65, ε=0.15, ΔR²=+0.0001) was observed. For FBG, all 20 combinations worsened performance. No combination provided clinically meaningful improvement.
Bayesian Ridge was used as the ablation model because its linear coefficient structure makes the effect of domain removal directly visible in R² changes. Note that Bayesian Ridge shows poor performance in the baseline comparison (R² = −0.035 for HbA1c) due to feature selection instability on small samples — this is expected and does not contradict the ablation results, where its linear structure is specifically exploited for interpretability. Tree-based ensembles were deliberately excluded from ablation because their robustness to irrelevant features attenuates the measurable impact of domain removal.
| Configuration | Features included |
|---|---|
| Full Model | All 105 features (baseline) |
| No Age Normalisation | All except age-normalised HRV |
| Only Age-Normalised + Demographics | Age-normalised HRV + demographics only |
| No Sleep-Stage HRV | All except per-stage HRV features |
| HRV Only | Stage-specific HRV + age-normalised HRV + demographics |
| ECG Only | ECG morphology + demographics |
| Clinical Only | Clinical measurements + demographics |
| No ECG | All except ECG morphology features |
| No Clinical | All except clinical measurement features |
| Demographics Only | Age, height, weight only |
| Only Deep Sleep HRV | Deep sleep HRV + age-normalised + demographics |
| Only REM HRV | REM sleep HRV + age-normalised + demographics |
| Only Rapid Sleep HRV | Rapid sleep HRV + age-normalised + demographics |
Key ablation findings: For HbA1c, Clinical Only (R² = 0.163) was the best-performing configuration, outperforming the Full Model (R² = −0.035 under Bayesian Ridge) — indicating that adding ECG/HRV features to clinical data introduces noise under strict within-fold feature selection on small samples. For FBG, Demographics Only and ECG Only each achieved R² = 0.110, the best ablation result for that cohort.
- Permutation testing: n = 500 permutations; both cohorts p = 0.002
- Bootstrap CIs: n = 500 subject-level resamples; 95% confidence intervals
- Residual diagnostics: Shapiro-Wilk normality, mean-bias test, heteroscedasticity analysis
- Learning curves: Sample-size adequacy assessment
| Domain | Count | Examples |
|---|---|---|
| Demographics | 3 | Age, height, weight |
| Clinical measurements | 20 | Blood pressure, lipid panel, renal/liver function, haematology |
| ECG morphology | 24 | Signal statistics (mean, SD, range, SNR) for 24h / sleep / daytime |
| HRV Features | 33 | Mean RR, SDNN, RMSSD, pNN50, CV per sleep stage (DS, REM, RS) |
| Age-normalised HRV | 3 | Mean RR normalised by age factor per sleep stage |
| Sleep quality | 22 | PSQI components (11), CPC-derived metrics (11) |
If you use this work, please cite our paper:
@article{azam_re-evaluating_2026,
title = {Re-evaluating heart rate variability biomarkers for glucose sensing: the impact of age normalisation and subject-independent validation},
issn = {1472-6947},
url = {https://doi.org/10.1186/s12911-026-03455-8},
doi = {10.1186/s12911-026-03455-8},
journal = {BMC Medical Informatics and Decision Making},
author = {Azam, Md Basit and Singh, Sarangthem Ibotombi},
month = apr,
year = {2026},
}The dataset used and analyzed during this study is publicly available in the Mendeley Data repository:
- Repository Name: Dataset on electrocardiograph, sleep and metabolic function of male type 2 diabetes mellitus
- Access Link: Mendeley Data [1]
We welcome contributions!
This project uses an MIT License. See the LICENSE file for details.
This project is for academic research purposes. Please contact the author for commercial use.
The authors acknowledge support from the Google Cloud Research Credits program under Award GCP19980904 and partial computing resources from Google’s TPU Research Cloud (TRC), both of which provided critical infrastructure for this research.
This study did not receive any specific grants from public, commercial, or not-for-profit funding agencies.
[1] Cheng, Wenquan; Chen, Hongsen; Tian, Leirong; Ma, Zhimin; Cui, Xingran (2023), “Dataset on electrocardiograph, sleep and metabolic function of male type 2 diabetes mellitus ”, Mendeley Data, V4, doi: 10.17632/9c47vwvtss.4