Feature matching as improved transfer learning technique for wearable EEG

With the rapid rise of wearable sleep monitoring devices with non-conventional electrode configurations, there is a need for automated algorithms that can perform staging on configurations small amounts labeled data. Transfer learning has ability to adapt neural network weights from source modality (e.g. standard configuration) new target configuration). We propose feature matching, transfer strategy as an alternative commonly used finetuning approach. This method consists training model larger data and few paired samples modality. For those samples, extracts features modality, matching these corresponding compare three different domains, two architectures, varying Particularly cohorts (i.e. 2 - 5 recordings in recording setting), systematically outperforms mean percentage point improvements accuracy ranging 0.3% 3.0% scenarios datasets. Our findings suggest approach, especially very low regimes. As such, we conclude promising novel devices.