We propose a parameter-efficient continual learning method for wearable human activity recognition. By freezing pretrained features and learning lightweight channel-wise gates, the model adapts to new users while reducing catastrophic forgetting. On PAMAP2, it improved final accuracy from 56.7% to 77.7% and reduced forgetting from 39.7% to 16.2%, while training under 2% of the model’s parameters and requiring no replay buffer.
We present GluBox, a multimodal forecasting system that leverages continuous glucose monitoring (CGM) as a core wearable sensing modality, together with behavioral and clinical data that influence blood glucose patterns, to predict long-term blood glucose patterns in individuals with type 1 diabetes while prioritizing clinically consequential errors.
This paper introduces Patch-TACA, a multimodal transformer that forecasts long-term blood glucose in healthy individuals by fusing CGM data with physiological and behavioral signals via time-aware cross-attention and self-supervised pretraining. On twelve participants, it achieved 14.26 ± 3.48 mg/dL RMSE at a 90-minute horizon and 93.9% hyperglycemia prediction accuracy, outperforming GlySim and Gluformer baselines. The results show that multimodal sensor fusion enables accurate long-horizon glucose forecasting for proactive metabolic health monitoring before disease onset.