TARNA: A 1.18-μW/Channel Topology-Aware Residual Recurrent CIM Processor for Zero-Shot Seizure Detection

Zero-shot seizure detection avoids patient-specific training but remains challenging under strict power constraints and patient variability. This work presents TARNA, a topologyaware recurrent network accelerator implemented as an ultra-low-power computing-in-memory (CIM) zero-shot seizure detector. TARNA combines graph convolutions with masked residual recurrent updates to eliminate nonlinear gating, while a sparsity-aware CIM macro further reduces energy. Fabricated in 65−nm CMOS, TARNA achieves 100 % zeroshot event-based sensitivity and a 0.048/hr false-alarm rate on the CHB-MIT dataset, and an AUROC of 0.88 on the TUSZ dataset, operating at the lowest reported power of 22.5μ W.