Towards a Flexible Deep Learning Method for Automatic Detection of Clinically Relevant Multi-Modal Events in the Polysomnogram

TL;DR

This paper introduces a deep learning model for automatic detection of sleep-related events like arousals and leg movements, demonstrating promising results and potential for broader sleep analysis applications.

Contribution

A novel deep learning approach incorporating RNNs for detecting sleep events, with optimized configurations for arousal and leg movement detection.

Findings

Arousal detection achieved F1 score of 0.75 with RNN and dynamic window.

Leg movement detection achieved F1 score of 0.65 with static window.

Model shows potential for general sleep analysis, with room for improvement.

Abstract

Much attention has been given to automatic sleep staging algorithms in past years, but the detection of discrete events in sleep studies is also crucial for precise characterization of sleep patterns and possible diagnosis of sleep disorders. We propose here a deep learning model for automatic detection and annotation of arousals and leg movements. Both of these are commonly seen during normal sleep, while an excessive amount of either is linked to disrupted sleep patterns, excessive daytime sleepiness impacting quality of life, and various sleep disorders. Our model was trained on 1,485 subjects and tested on 1,000 separate recordings of sleep. We tested two different experimental setups and found optimal arousal detection was attained by including a recurrent neural network module in our default model with a dynamic default event window (F1 = 0.75), while optimal leg movement detection was attained using a static event window (F1 = 0.65). Our work show promise while still allowing for improvements. Specifically, future research will explore the proposed model as a general-purpose sleep analysis model.

Tables

Table 1. TABLE I: MrOS subset demographics. Significant p -values at α = 0.05 𝛼 0.05 \alpha=0.05 are shown in bold.

	train	eval	test	p-value
N	1,485	165	1000	-
Age (years)	$76.4\pm 5.5$	$76.6\pm 4.9$	$76.4\pm 5.6$	0.631
BMI ($\mathrm{kg}{s}^{-2}$)	$27.2\pm 3.8$	$27.2\pm 3.4$	$27.1\pm 3.7$	0.879
AHI ($h^{-1}$)	$12.8\pm 12.9$	$10.6\pm 11.8$	$11.9\pm 12.8$	0.029
AI ($h^{-1}$)	$23.6\pm 11.5$	$24.1\pm 12.2$	$23.4\pm 11.8$	0.607
PLMI ($h^{-1}$)	$34.8\pm 37.0$	$37.8\pm 38.9$	$37.3\pm 38.0$	0.204

Table 2. TABLE II: Proposed network architecture. ϕ C subscript italic-ϕ 𝐶 \phi_{C} , linear mixing module; ϕ T subscript italic-ϕ 𝑇 \phi_{T} , temporal feature extraction module; ϕ R subscript italic-ϕ 𝑅 \phi_{R} , recurrent neural network module; ψ clf subscript 𝜓 clf \psi_{\mathrm{clf}} , event classification module; ψ loc subscript 𝜓 loc \psi_{\mathrm{loc}} , event localization module; C , number of input channels; T , number of samples in segments; C ~ = 2 2 + n max ~ 𝐶 superscript 2 2 subscript 𝑛 \tilde{C}=2^{2+n_{\max}} , number of output channels; K , number of event classes; N d subscript 𝑁 𝑑 N_{d} , number of default events in segment; T ~ = T / 2 n max ~ 𝑇 𝑇 superscript 2 subscript 𝑛 \tilde{T}=\nicefrac{{T}}{{2^{n_{\max}}}} , reduced temporal dimension; bGRU, bidirection gated recurrent unit; ReLU, rectified linear unit.

Module	Input dim.	Output dim.	Type	Kernel size	No. kernels	Stride	Activation
${\varphi }_C$	$(C,T)$	$(C,T)$	1D convolution	$C$	$C$	1	linear
${\varphi }_{T,\mathrm{init}}$	$(C,T)$	$(8,T)$	1D convolution	$3$	8	$1$	–
	$(8,T)$	$(8,T)$	Batch norm.	–	8	–	ReLU
	$(8,T)$	$(8,T/2)$	1D max. pool.	$2$	–	$2$	–
${\varphi }_{T,k}$ $n=2,\dots ,n_{\max }$	$(2^{n+1},T/2^{n-1})$	$(2^{n+2},T/2^{n-1})$	1D convolution	$3$	$2^{n+2}$	$1$	–
	$(2^{n+2},T/2^{n-1})$	$(2^{n+2},T/2^{n-1})$	Batch norm.	–	$2^{n+2}$	–	ReLU
	$(2^{n+2},T/2^{n-1})$	$(2^{n+2},T/2^n)$	1D max. pool.	$2$	–	$2$	–
${\varphi }_R$	$(\tilde{C},\tilde{T})$	$(2\times \tilde{C},\tilde{T})$	bGRU	$\tilde{C}$	–	–	–
${\psi }_{\mathrm{clf}}$	$(\tilde{C},\tilde{T})$	$(\left(K+1\right)N_d,1)$	1D convolution	$\tilde{T}$	$\left(K+1\right)N_d$	$\tilde{T}$	softmax for each $K+1$ kernel
${\psi }_{\mathrm{loc}}$	$(\tilde{C},\tilde{T})$	$(2N_d,1)$	1D convolution	$\tilde{T}$	$2N$	$\tilde{T}$	linear

Table 3. TABLE III: Application of optimized models on test data. Data are shown as subject-averaged F1, precision (Pr) and recall (Re) with associated standard deviations. Top four rows correspond to Experiment A, while bottom two rows correspond to Experiment B. AR: arousal; LM: leg movement; RNN: recurrent neural network.

Model	F1	Pr	Re
LM, static	$0.648\pm 0.148$	$0.631\pm 0.181$	$0.720\pm 0.141$
AR, static	$0.727\pm 0.102$	$0.706\pm 0.113$	$0.771\pm 0.132$
LM, dynamic	$0.647\pm 0.148$	$0.627\pm 0.181$	$0.722\pm 0.14$
AR, dynamic.	$0.729\pm 0.102$	$0.699\pm 0.115$	$0.785\pm 0.131$
LM, RNN	$0.639\pm 0.147$	$0.606\pm 0.180$	$0.727\pm 0.126$
AR, RNN	$0.749\pm 0.105$	$0.772\pm 0.107$	$0.748\pm 0.138$

Equations

$$\mathrm{Pr}$$

$$F1$$

Full text

I INTRODUCTION

Analysis of sleep patterns is performed manually by experts in sleep clinics using rules and guidelines defined by the American Academy of Sleep Medicine recently updated in 2018 [1]. These guidelines outline technical and clinical best practices when performing routine polysomnography (PSG), which is an overnight recording of electroencephalography (EEG), electrooculography (EOG), electromyography (EMG) electrocardiography (ECG), respiratory effort and peripheral limb activity. Expert technicians and somnologists use these physiological variables to analyse sleep patterns and diagnose sleep disorders based on key metrics and indices, such as total sleep time, amount of sleep spent in various sleep stages, and the observed number of discrete events per hour of sleep. Specifically, the number of arousals (short awakenings during sleep, $<$15 s), non-periodic and periodic leg movements (PLM), and the number of apnea events per hour of sleep are summarized in the arousal index (AI), periodic leg movements index (PLMI) and apnea/hypopnea index (AHI), the latter of which is a combination of apneic (no/obstructed respiratory effort) and hypopneic (reduced respiratory effort) events. Excessive amounts of these events are disruptive to normal sleep, which can lead to patient complaints of excessive daytime sleepiness [2], which in turn is linked to an increase in e.g. automotive accidents and reduced quality of life [3]. Increased number of PLMs is also linked to other sleep disorders such as restless legs syndrome, and periodic leg movement disorder [4, 5].

Correct diagnosis of sleep disorders is predicated on precise scoring of sleep stages as well as accurate scoring of these discrete sleep events. However, the current gold standard of manual analysis by experienced technicians is inherently biased and inconsistent.Several studies have shown low inter-rater reliability on both the scoring of sleep stages [6, 7, 8], arousals [9], and respiratory events [10]. Furthermore, manual analysis of PSGs is time-consuming and prone to scorer fatigue. Thus, there is a need for efficient systems that provide deterministic and reliable scorings of sleep studies.

Several recent studies have already explored automatic classification of sleep stages in large cohorts with good results [11, 12, 13, 14, 15], however, the reliable and consistent detection and classification of discrete PSG events in large cohorts remain largely unexplored.

Recent studies on certain microevents in sleep have indicated that sleep spindles and K-complexes can be reliably detected and annotated with start time and duration using deep learning methods [16, 17]. Specifically, these studies proposed a single-shot event detection algorithm, that parallels the YOLO and SSD algorithms used for object detection in 2D images [18, 19], however, they were limited in scope by detecting events only at the EEG level, and did not explicitly take advantage of the temporal connection of the detected events. Additionally, experiments were carried out on a small-scale database [16].

In this study, we focused on the detection of arousals (AR) and leg movements (LM). These events arise from highly distinct physiological sources, EEG and leg EMG, while ARs are also visible in the EOG and chin EMG. These events are important for the precise characterization of sleep patterns and possible diagnosis of sleep disorders, and an accurate detection is therefore of high interest. We extend previous work in [16, 17] by 1) preprocessing and analysing multiple input signals at the same time, and 2) taking into account important temporal context using recurrent neural networks. Furthermore, we apply our model on a larger database than previous studies.

II DATA

II-A MrOS Sleep Study

The MrOS Sleep Study is a part of the larger Osteoporotic Fractures in Men Study with the objective of researching the links between sleep disorders, fractures, cardiovascular disease and mortality in older males ($>65$ years) [20, 21, 22]. Between 2003 and 2005, 3,135 of the original 5,994 participants were recruited to undergo full-night PSG recording at six centers in the US at two separate visits (visit 1 and visit 2) with following 3 to 5-day actigraphy studies at home. The resulting PSG studies were subsequently scored by experienced sleep technicians for standard sleep variables including sleep stages, leg movements, arousals, and respiratory events.

II-B Included events and signals

In this study, we only considered the detection of two PSG events, arousals and leg movements. These events are characterized by a start time and a duration, which we extracted from 2,907 PSG studies from visit 1 available from the National Sleep Research Resource repository [23, 24]. From each PSG study, we extracted left and right central EEG, left and right EOG, chin EMG, and EMG from the left and right anterior tibialis. EEG and EOG channels were referenced to the contralateral mastoid process, while a leg EMG channel was synthesized by referencing left to right. Any PSG without the full set of channels or without any event scoring was eliminated from further analysis.

II-C Subset demographics and partitioning

In total, 2,650 out of the 2,907 PSGs available from visit 1 were included in this study. These were partitioned into train, eval, and test sets of sizes 1,485, 165, and 1,000 studies, respectively. A subset of key demographic and PSG variables are presented in Table I.

III METHODS