Time-Aware Cross-Attention for Multi-Modal Sensor-Based Blood Glucose Forecasting

Aashritha Machiraju, Ebrahim Farahmand, Shovito Barua Soumma, Asiful Arefeen, Carol Johnston, Hassan Ghasemzadeh

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

Keywords: Prediabetes, Attention Mechanism, Wearable sensors, Digital Health, Deep Learning, Metabolic Health

TL;DR: A time-aware cross-attention + LSTM model fuses raw wearable streams to forecast glucose 90 min ahead, slashing RMSE by ≈18 % versus state-of-the-art.

Abstract: Accurate blood glucose forecasting is essential for proactive management of metabolic health, particularly when leveraging data from wearable sensors. However, existing models struggle with integrating multimodal time-series data with inconsistent sampling rates. This paper proposes a novel forecasting framework that incorporates a time-aware cross-attention mechanism with an LSTM architecture to predict blood glucose levels using continuous glucose monitoring (CGM) data alongside physiological and behavioral signals, such as heart rate(HR), electrodermal activity, accelerometry, and dietary intake. The proposed method dynamically encodes temporal features without the need for preprocessing and employs gated multi-head cross-attention layers to fuse sensor modalities effectively. We evaluated our approach on a newly collected dataset from 12 healthy individuals. Our method outperforms the baseline and the state-of-the-art GlySim model across multiple prediction horizons (5–90 minutes), achieving up to 17.8\% improvement in RMSE and 11.2\% gain in correlation

Track: 1. Digital Health Solutions (i.e. sensors and algorithms) for diagnosis, progress, and self-management

NominateReviewer: Hassan Ghasemzadeh- [e-mail]

Submission Number: 133