Emotional state and glucose monitoring typically require subjective questionnaires, invasive blood tests, or specialized sensors that are impractical for frequent or continuous use.​ ECG is widely available and non‑invasive, but existing models struggle to infer higher‑order states (emotion, metabolic status) from noisy, non‑stationary ECG signals, especially when labeled datasets are small.​ Current self‑supervised ECG methods rely on hand‑crafted signal augmentations and often fail to capture richer temporal and contextual information, limiting performance on downstream tasks.

The technology

• The system uses short ECG segments to learn patterns linked to emotions and glucose levels in a data‑efficient way
• It processes 10‑second ECG segments, extracts key signal patterns, and uses an attention‑based model to summarize them into a compact digital “fingerprint” of heart activity.
• The model is first trained on large unlabeled ECG collections by hiding parts of this fingerprint and learning to reconstruct the missing pieces and to recognize when two views come from the same signal.​
• It is then fine‑tuned on smaller labeled datasets so these ECG fingerprints can reliably distinguish emotional states and glucose‑level ranges.​

Fig. 1: Pre-training procedure

Potential applications

• Non‑invasive emotion recognition for mental‑health monitoring, affect‑aware human–computer interaction, and adaptive user interfaces using wearable or consumer ECG devices​
• Continuous or intermittent glucose‑level screening and risk stratification for people with diabetes or pre‑diabetes using ECG from smartwatches, patches, or ECG‑enabled medical wearables
• Foundation model for ECG analytics, serving as a pre‑trained encoder that can be adapted to additional downstream tasks (stress detection, arrhythmia risk, sleep staging) with limited labeled data​
• Enhancement of remote‑patient‑monitoring platforms and telehealth services, enabling multi‑parameter physiological assessment from a single ECG stream

Reference:
Lalzary and L. Wolf, “Dual Contrastive Learning for Self‑Supervised ECG Mapping to Emotions and Glucose Levels,” IEEE Sensors, 2023, DOI: 10.1109/SENSORS56945.2023.10325116.