Estimating glucose
from the heartbeat.

The problem

Blood glucose monitoring typically relies on finger-prick sampling or a sensor inserted under the skin. Cost, discomfort, and inconvenience limit how often people test, and many people with diabetes remain undiagnosed.

Our approach

Daibeats is investigating whether physiological information in the ECG — heart rate variability and waveform morphology — can be used to estimate glucose. Because ECG is already widely recorded, this would require software rather than a new sensor.

Data & research

Models are developed and evaluated on public clinical and wearable datasets — MIMIC-IV, eICU, AI-READI, and D1NAMO — with subject-independent splits, data-leakage controls, and external validation.

Current work

Model development → validation → product development. Models have been developed on retrospective data; external validation is in progress; a product prototype, including Claude-generated reports for clinicians, comes next.

Cardiac-Glycemic Coupling

Blood glucose concentration directly modulates cardiac electrophysiology. Hyperglycemia prolongs QTc intervals through potassium channel disruption; hypoglycemia activates sympathetic responses measurable in HRV. These effects are weak, overlapping, and confounded — but they are consistent enough to learn from at scale.

Temporal Alignment Problem

ECG-CGM pairing is not trivial. CGM readings reflect interstitial glucose with a 10–20 minute physiological lag behind blood glucose. ECG morphology reflects instantaneous cardiac state. Correctly aligning these two timeseries requires careful window selection, lag correction, and validation against simultaneous blood draws.

Validation Across Datasets

Small wearable datasets (D1NAMO: 9 subjects) reveal feasibility. ICU databases (MIMIC-IV: 131K+ patients) reveal robustness under pharmacological confounding, metabolic extremes, and device variability. We evaluate on both — and report Clarke Error Grid Zone distributions, not just R² or MAE.

Safety-Aware Loss Functions

Standard regression losses (MSE, MAE) are clinically naive — they penalize a 50 mg/dL error equally at euglycemia and at a critical hypoglycemic threshold. SAGE-Net's loss function engineering weights errors by their clinical zone, trading overall RMSE for dramatically improved sensitivity in dangerous ranges.