Personalizing deep learning models for automatic sleep staging

TL;DR

This paper introduces a personalized deep learning approach for automatic sleep staging that adapts a general model to individual sleep patterns, significantly improving accuracy especially for challenging cases.

Contribution

The authors propose a method to personalize a convolutional neural network for sleep staging by learning individual characteristics from the first night to enhance second-night predictions.

Findings

Personalization improves sleep staging accuracy by about 2 percentage points.

The method benefits subjects with initially low accuracy (<80%).

Combining broad sleep knowledge with subject-specific data yields optimal results.

Abstract

Despite continued advancement in machine learning algorithms and increasing availability of large data sets, there is still no universally acceptable solution for automatic sleep staging of human sleep recordings. One reason is that a skilled neurophysiologist scoring brain recordings of a sleeping person implicitly adapts his/her staging to the individual characteristics present in the brain recordings. Trying to incorporate this adaptation step in an automatic scoring algorithm, we introduce in this paper a method for personalizing a general sleep scoring model. Starting from a general convolutional neural network architecture, we allow the model to learn individual characteristics of the first night of sleep in order to quantify sleep stages of the second night. While the original neural network allows to sleep stage on a public database with a state of the art accuracy, personalizing the model further increases performance (on the order of two percentage points on average, but more for difficult subjects). This improvement is particularly present in subjects where the original algorithm did not perform well (typically subjects with accuracy less than $80\%$). Looking deeper, we find that optimal classification can be achieved when broad knowledge of sleep staging in general (at least 20 separate nights) is combined with subject-specific knowledge. We hypothesize that this method will be very valuable for improving scoring of lower quality sleep recordings, such as those from wearable devices.