A Resource-Efficient Algorithm–Hardware Co-Design Toward Semi-Supervised Neurological Symptoms Prediction

Recent advancements in neuroscience, coupled with the rise of deep neural networks (DNNs), have markedly enhanced the ability to predict physiological disorders. Typically, these models are trained on cloud servers using labeled datasets and subsequently deployed to devices for inference. However, this conventional cloud-based training and edge inference approach is not feasible for real-life scenarios owning to issues such as data privacy, the laborious data labeling process, and variations in conditions resulting from electrode shifts and changes in physiological status. In response to these challenges, we present REAL, a resource-efficient algorithm-hardware co-design training framework. This transformer-based framework incorporates direct feedback alignment (DFA) to optimize hardware and computational efficiency during training process. Moreover, it utilizes semi-supervised learning along with bespoke data augmentation techniques to reduce the dependency on extensive data labeling. Based on the proposed algorithmic model, we design and implement a dedicated hardware architecture. We tested REAL on both EEG-based seizure prediction and ECG-based arrhythmia classification tasks. For the EEG task, our framework enhances prediction sensitivity by an average of 7.24% and decreases the false positive rate by 0.04/h, demonstrating robust performance across varying levels of annotated data. With minimal labeled data, REAL reached an AUC of 0.91 and a sensitivity of 89.15%, outperforming all previous works with similar task settings. For the ECG task, REAL achieved an accuracy of 82.81% and a macro F1 score of 0.827 under a 20% labeling rate, demonstrating strong generalization to a different biomedical modality. Additionally, REAL is the first hardware design targeting semi-supervised neurological symptom prediction, achieving an energy efficiency of 3.97 TOPS/W and an area efficiency of 2.39 GOPS/mm², both of which exceed all previous related works.