ECGluFormer: glucose prediction from ECG via multi-loss, transformer-based aggregation

Mu-Ruei Tseng, Ricardo Gutierrez-Osuna

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Diabetes, deep learning, ECG analytics, non- invasive glucose prediction, hypoglycemia.

TL;DR: We present a transformer-based model that can estimate glucose levels from intermittent ECG signals.

Abstract:

Continuous glucose monitors (CGMs) have become ubiquitous in diabetes care but are unreliable in the hypoglycemic
range, where they are most critical. We present ECGluFormer, a deep-learning (DL) model that estimates glucose levels non-invasively from single-lead ECG. Our model addresses two critical problems in ECG-based glucose prediction: (1) conventional DL models tend to under-report hypoglycemia due the rarity of those events, and (2) ECG signals can be intermittent in free-living conditions (e.g., motion artifacts, packets drops). To address these issues, ECGluFormer uses a multi-objective loss function that ensures the distribution of glucose predictions is consistent with ground-truth, and a Transformer-based model to aggregate beat-level glucose predictions when significant data losses (over 30% of all beats) are missing. We validate ECGluFormer on ambulatory data containing up to 17 days of synchronized ECG and CMG data from patients with type-1 diabetes. Our multi-objective loss function outperforms alternative loss functions across regression and classification metrics. ECGluFormer also consistently outperforms five baseline models that also aggregate beat-level predictions.

Track: 1. Biomedical Sensor Informatics

Registration Id: 9VNYZPGGJ2L

Submission Number: 226