A Novel 130.1 pJ/Decision Binary Tree Ensemble Classifier for an Energy Efficient Atrial Fibrillation Detecting ECG Processing System in 22 nm FDSOI.

This paper presents a new classifier architecture for decision tree ensembles (DTEs) based on lookup-tables. The classifier is integrated in an electrocardiogram processing system that detects atrial fibrillation (AF). It targets energy efficiency and achieves best in class performance for AF-detecting systems by a margin of one to two orders of magnitude at 275.9 pJ/heartbeat. The classifier itself only requires 130.1 pJ/decision or 0.93 pJ/node per decision at up to 5 Mdecisions/s inference rate, similar to dedicated analog in-memory classifiers. In addition, a strategy for mapping/synthesizing models to this DTE architecture based on satisfiability modulo theories is proposed.