Comparison of deep transfer learning algorithms and transferability measures for wearable sleep staging

Abstract

Background Obtaining medical data using wearable sensors is a potential replacement for in-hospital monitoring, but the lack of data for such sensors poses a challenge for development. One solution is using in-hospital recordings to boost performance via transfer learning. While there are many possible transfer learning algorithms, few have been tested in the domain of EEG-based sleep staging. Furthermore, there are few ways for determining which transfer learning method will work best besides exhaustive testing. Measures of transferability do exist, but are typically used for selection of pre-trained models rather than algorithms and few have been tested on medical signals. We tested several supervised transfer learning algorithms on a sleep staging task using a single channel of EEG (AF7-Fpz) captured from an in-home commercial system. Results Two neural networks—one bespoke and another state-of-art open-source architecture—were pre-trained on one of six source datasets comprising 11,561 subjects undergoing clinical polysomnograms (PSGs), then re-trained on a target dataset of 75 full-night recordings from 24 subjects. Several transferability measures were then tested to determine which is most effective for assessing performance on unseen target data. Performance on the target dataset was improved using transfer learning, with re-training the head layers being the most effective in the majority of cases (up to 63.9% of cases). Transferability measures generally provided significant correlations with accuracy (up to $$r = -0.53$$ r = - 0.53 ). Conclusion Re-training the head layers provided the largest performance boost. Transferability measures are useful indicators of transfer learning effectiveness.

Domain Adaptation Using Large Scale Databases for Sleep Staging from a Novel In-Ear Sensor

“…-61] and close to or better than approaches using wearable scalp EEG sensors [62][63][64]. Furthermore, results were on par with the 79.1% accuracy and κ of .704 achieved in our previous work using the same neural network and transfer learning method on a wearable scalp EEG sensor [44]. Results were also consistent with other approaches using ear-EEG, which have achieved Cohen's κ ranging from .61-.767 [35,36,[38][39][40][41]; and using cEEGrid, which have achieved Cohen's κ ranging from .20-.762 [31][32][33][34].…”

“…We used a convolutional neural network architecture which we previously found to be effective for transfer learning on sleep staging tasks [44]. The architecture consists of…”

“…For our analysis we used the across ear-canals channel, which contains a stronger EEG signal than the cymba channels or same-ear channels. As with the architecture, we used the pre-processing steps which we fond effective in our previous sleep staging paper [44]. EEG for both the source and target datasets was downsampled to 100Hz using polyphase resampling with zero-phase lowpass FIR anti-aliasing filter, after which the short-time Fourier transforms of 30-second epochs were taken and used as input for the neural network.…”

Survey of Transfer Learning Approaches in the Machine Learning of Digital Health Sensing Data

“…On the other hand, model-based interpretation focuses on the development of accurate prediction models in the target task by transferring knowledge based on the model control strategy, parameter control strategy, and model ensemble strategy [34]. There are various TL approaches that have been developed and examined to improve healthcare services and patients' health, such as fine-tuning [35][36][37][38][39][40], feature extraction [41][42][43][44], multitask learning [45,46], domain adaptation [40,47,48], federate learning [49][50][51], as well as meta learning methods (such as zero-shot [52], one-shot [53], and few-shot learning [53,54]). In this paper, we discuss twenty-seven studies in detail, distributed as presented in Figure 5, to highlight the applications of TL to enhance healthcare services and outcomes based on DH sensing data, as shown in Figure 6.…”

Sleep-Deep-Net learns sleep wake scoring from the end-user and completes each record in their style

“…Similar approaches have been used in human data and wearable technology. 19, 20 However, this approach has not been applied to data from mice, the most commonly used animal model in pre-clinical sleep research, as far as we are aware. It is important to note that no automated sleep scoring method works perfectly on every file.…”