RAE-PPG: Duration-Grounded Retain-and-Extend Pretraining for PPG Foundation Models

Suyeong Lee, Hochang Lee, Seokyong Sheem, Daekyum Kim (Korea University)

Summary

Organizes PPG self-supervision around signal duration: a single ~10.76M Transformer is trained successively on 10/30/240 s windows, introducing classical PPG morphology/PRV/waveform targets only when the window can support them, while protecting earlier parameter groups and retaining earlier targets so short-window features remain decodable from longer inputs. Frozen encoder leads on 12 of 18 tasks vs five prior PPG FMs.