Power of two fine-tuning for patient-specific cardiac arrhythmia recognition
Résumé
Deep learning is one of the main solution in cardiac arrhythmia classifier for its excellent performances. Research aims to embed these classifiers as near as possible from the sensors to process the data in real time and get ride of the wireless communication drawbacks. This work investigates the performance of a hardware-friendly quantified learning framework that addresses the issue of data confidentiality. It achieves the same performance than the state of the arts without data augmentation. It is also compatible with fine-tuning increasing the state of the art performances by up to 11 points with a model ×25000 smaller.