key: cord-0261624-7zox46x6
authors: Srivastava, S.
title: Classifying REM Sleep Behavior Disorder through CNNs with Image-Based Representations of EEGs
date: 2022-04-05
journal: nan
DOI: 10.1101/2022.04.03.22273365
sha: db5b45d2d5f20bd0060e5e550cef5a36e7256b4a
doc_id: 261624
cord_uid: 7zox46x6

Objective: Rapid Eye Movement Sleep Behavior Disorder (RBD) is a parasomnia with a high conversion rate to -synucleinopathies such as Parkinson's Disease (PD), dementia with Lewy bodies (DLB), and multiple system atrophy (MSA). The objective of this paper is to classify RBD in patients through a convolutional neural network utilizing image-based representations of electroencephalogram (EEG) channels. Methods: Analysis was conducted on polysomnography data from 22 patients with RBD and 12 healthy controls acquired from the Cyclic Alternating Pattern (CAP) Sleep Database. EEG channels were split into four frequency bands (theta, alpha, beta, and gamma). Power spectrum density was calculated through a Fast Fourier Transformation (FFT) and converted into 2D images through bicubic interpolation. RBD classification was accomplished through a pre-trained VGG16 CNN with layer weights being fine-tuned. Results: The model was successful in classifying RBD patients over non-RBD patients and achieved 97.92% accuracy over 40 epochs. Accuracy increased dramatically with increased data generated from FFT and interpolation (62.63% to 97.92%). Conclusions: This project proposes a novel approach toward an automatic classification of RBD and highlights the importance of deep learning models in the field. The proposed transfer learning model outperforms state-of-the-art models and preliminarily accounts for the lack of computational resources in clinical spaces, thereby increasing the accessibility of automatic classification. Significance: By leveraging transfer learning and raw data, the results demonstrate that a similar model for the classification of RBD patients could easily be translated to the clinical atmosphere, drastically accelerating the classification pipeline. The proposed methods are also applicable to -synucleinopathies, including PD, DLB, and MSA.

Rapid Eye Movement (REM) Sleep Behavior Disorder (RBD) is a prodrome of neurodegeneration that is characterized by a loss of normal muscle atonia and dream-enactment motor activity during REM sleep (Roguski et al. (2020) ; Högl and Stefani (2016) ). RBD is present in about 0.5 to 1.25% of the general population (Haba-Rubio et al. (2017) ; Kang et al. (2013) ; Pujol et al. (2017) , but given that it is difficult to gauge accurately (Roguski et al., 2020) and because screening is not always conducted, most cases go unrecognized (Atassi et al., 2021) . Recent evidence has indicated that RBD is an early predictor for Parkinson's Disease (PD) (Postuma et al., 2019) -as well as other α-synucleinopathies such as dementia with Lewy bodies (DLB), and multiple system atrophy (MSA) (Stefani and Högl, 2019) . At a 14-year follow-up, the "the estimated risk of overt neurodegeneration is around 97% at a 14-year follow-up, while the conversion rate to PD is about 90%" (Rechichi et al., 2021) . However, rapid development in the drug industry has yielded advancements in new neuroprotective drugs that could only be effective if administered in pre-clinical stages of a pathology (Cruz-Vicente et al. (2021) ; Devos et al. (2020) ). Recognizing RBD early and accurately then becomes invaluable in treatment and prevention of onset and prognosis of future neurodegenerative diseases that may have severe consequences. Polysomnography (PSG) evidence is the gold standard for diagnosis of RBD, among other sleep disorders and neurological diseases. These PSGs often contain electroencephalography (EEG), electromyography (EMG) and electrooculography (EOG) data, where EEG recordings are frequently used to identify RBD. However, even with ample data, different physicians have been recorded making different assessments about patients (Teplan (2002) ; Klok et al. (2018) ; Siuly and Li (2015) ; Islam et al. (2018) ). Thus, there is need to create an automatic classification pipeline that accounts for variation and can learn from past diagnoses. Machine learning is the optimal tool towards approaching this automatic classification problem. Literature shows numerous studies that utilize conventional machine learning algorithms -such as Support Vector Machine (SVM), k -Nearest Neighbors (k -NN), and Random Forests (RF) -to classify RBD (Islam et al. (2018) ; Rechichi et al. (2021) ; Cooray et al. (2019) ; Cooray et al. (2021) ; Buettner et al. (2020) ). However, almost all of them utilize massive amounts of features, which can be incredibly computationally expensive. Further, many of these previous approaches do not consider the spatial, temporal, and spectral properties of EEG signals to the fullest extent. Analyzing these channels may be incredibly fruitful to produce a model that achieves high accuracy and maintains robustness. Thus, image-based representations of EEGs, as portrayed in (Arasteh et al., 2021) , and their value is important to adequately consider all properties of EEG signals and to maximize robustness (Phan et al., 2019) . The advent of deep learning has deepened interests in applying deep learning to the field of RBD detection. Artificial neural networks, modeled after human neural networks, are powerful in that they can automatically recognize features within data, and deviate from hand-crafted features -which is crucially necessary. In this paper, I propose an automated pipeline for RBD detection through image-based representations, which to the best of my knowledge, has not been explored in this field yet. This novel approach to classify RBD attempts to further improve RBD detection through the power of deep learning, while steering away from conventional machine learning algorithms. (Terzano et al., 2001) . Managed by the MIT Laboratory for Computational Physiology on PhysioNet (Goldberger et al., 2000) , the CAP database contains a collection of 108 polysomnography recordings of both healthy patients and patients with varying neurological disorders -ranging from bruxism to narcolepsy to RBD. Data from 16 healthy subjects and 22 RBD patients was acquired, however, three healthy subjects were removed due to the presence of ECG artifacts and another two healthy subjects were excluded from analysis due to the absence of EMG recording. Three RBD patients were removed from analysis due to the presence of ECG artifacts. Each recording was subsequently scanned to ensure that at least 5 minutes of REM episodes are present, as handled in (Rechichi et al., 2021) .

Age # of Subjects # of F # of M Healthy 31.5 ± 4.3 12 6 6 RBD 70.4 ± 6.4

19 2 17 Table 1 . Summary of the data utilized in the paper. Comparatively, the healthy subjects are far younger than the patients in the RBD cohort. There are also predominately more males than females in the RBD cohort.

The recordings were pre-processed to reduce noise and minimize artifacts. The pre-processing standards used in this paper follow trends established in (Cooray et al., 2019) and (Cooray et al., 2021) . All data -EMG, EEG, and ECG -were first resampled at 200Hz. The EEG signal was pre-processed with a 500th order band pass finite impulse response (FIR) filter from 0.3-40Hz. The EMG signal was filtered with a 500th order notch filter from 50-60Hz, as well as a 500th order band pass FIR filter from 10-100Hz. The ECG signal was filtered with a 10th order Butterworth band pass filter between 5-45Hz.

As established in (Lee et al., 2019) , data was then bandpass filtered into four conventional EEG frequency bands (Groppe et al., 2013 ) bands corresponding to wavelength: theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-45 Hz). Epoch size was calculated so that there are numerous cycles at the bandwidth's frequency, resulting in 10-second "mini-epochs."

Fast Fourier Transformation (FFT) was applied on the time series with windows equal to 1 second to receive 10 different measures per band (10 second mini-epoch sliced into 1 second intervals per band). This resulted in 40 different measures total per subject. These measurements are known as the power spectral density (PSD) estimation of the EEG signal. The Welch's method was utilized to calculate this PSD, which is given by the following equation:

The measurements for each electrode were then converted into 2D images through a bi-cubic interpolation technique, similar to that of (Fadel et al., 2020) . The problem of classification can now be tackled through image recognition. Thus, these images were then fed into a convolutional neural network (CNN).

The power of transfer learning was harnessed to build the model. Transfer learning utilizes a pre-trained CNN on a large dataset, rather than training a CNN from scratch. Although often utilized for when there is not enough data to train an accurate CNN, transfer learning is increasingly being used in other scenarios due to its powerful, comparable results to individually trained CNNs (2014)). Thus, this paper leverages this tremendous power by deploying a pretrained VGG-16 model [cite] on a small dataset to solve the binary classification of RBD patients versus healthy patients.

In accordance with the model developed by [cite] , this paper modified the base VGG-16 model architecture by utilizing a fully connected layer (1 × 1 × 512) after the last max-pooling layer (7 × 7 × 512), a fully connected layer (1 × 1 × 64) with activation function of "Relu," and a fully connected layer (1 × 1 × 2) with the activation function of "Softmax" for the classification task consists of two categories, RBD and healthy controls. Figure 2 shows the model architecture for the trained network and Table 2 reports the trained model's summary. . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted April 5, 2022. 

CNNs are known to be "data hungry" (Roy et al., 2018) , often requiring large amounts of data to achieve high, meaningful accuracy. To explore the effect of a larger dataset, the model was tested four times, each time with a different number of images per patient: 4, 8, 20, and 40. The accuracy for each test are reported in Table 3 . As hypothesized, RBD classification increases exponentially with an increase in data. Figure 

With the implementation of a novel method in this field, it is important to acknowledge that this paper does not come without limitations. The cohort of data utilized in this study was heavily male dominant and RBD patients were significantly older than healthy subjects. This could have led to overfitting of the model (not just to male or older patients, but the entire dataset itself too). To maintain consistent dimensionality for images, the number of cycles in each epoch may have differed, which could translate to biases in prediction. Further, due to a lack of sensor data and information, a complete Azimuthal Equidistant Projection (AEP) (Anderson, 1974) and Clough-Tocher interpolation (Alfeld, 1984) could not be conducted as performed in (Fadel et al., 2020) . Thus, this preliminary study may not be able to validate the complete power of convolutional neural networks in this field. Additionally, the ICSD-3 criteria for diagnosis of RBD relies on both REM Sleep Without Atonia (RSWA) -lack of regular muscle atonia during REM sleep -as well as dream enacting and violent behavior during REM sleep, which is reported through a questionnaire often filled out by a sleeping partner. This paper is unable to automatically assess sleep for this dream enacting. However, this paper still provides ample evidence and reason to explore an automatic classification of RBD through neural networks to better accuracy levels and simplify the process for RBD classification for neurologists. This is the first work, to the best of my knowledge, that goes beyond standard machine learning algorithmss to consider multimodal neural networks for RBD classification. Future developments in this field should thus explore the usage of convolutional neural networks for detection on a larger dataset. As acknowledged above, the skewed dataset could have resulted in overfitting and there is a need to utilize larger and more diverse datasets with RBD patients and healthy subjects. A concatenation of numerous datasets, such as the MASS (Montreal Archive of Sleep Studies) cohort, the CAP Sleep Database, and other private datasets, could create this potential, large dataset. However, this dataset could still face a problem that is relevant to this paper: sensor positions should also be taken into consideration to utilize the AEP and Clough-Tocher interpolation techniques. The lack of adequate annotation of different sensors, as well as an unequal number of electrodes present per subject, made it difficult to employ these algorithms. Future developments should try to incorporate these techniques to ensure a closer relationship with the image and the Raw data. Further, a new CNN model can be designed -one that traverse through each 10 second mini-epoch during REM sleep and calculates only one measurement through FFT. Although this could be both time and computationally intensive, this would ensure that the full Raw data is being utilized, while creating numerous images per patient. There is potential to create more than 40 images, which could improve accuracy. This would also eliminate the potential for an unequal number of cycles at the bandwidth's frequency, as all the data is being utilized.

the disorder (St Louis and Boeve, 2017). There is a necessity to not just simplify this process but improve it to ensure that RBD is detected in patients at an early stage. This paper provides a framework for the automatic classification of RBD through polysomnogram (PSG) recordings alone, which would drastically speed up diagnosis. Its low computational cost would also be beneficial for applications to clinics that may not have the computational power to channel through numerous models with billions of endpoints. Further, given that there is high accuracy, patients, if experiencing violent sleep problems, have the potential to get screened for RBD instantly which would be extremely valuable.

Although PSG recordings are heralded as the gold-standard sleep assessment technique, they are impractical for usage outside a laboratory or clinic, due to "its relatively high cost, the specialized training, and the time burden required to conduct and interpret studies" (Chinoy et al., 2020) . Sleeptracking devices have already been developed for consumer usage and attempt to cut costs, while providing ample data for interpretation. Given the data that these devices output, the framework proposed in this paper can also be utilized and determine whether a patient is exhibiting RBD-like phenotypes during their sleep. With the advent of 'digital' meetings for doctors -especially during the COVID-19 pandemic -there is strong potential for these devices to be delivered to patients, and upon adequate testing at home, diagnosis can be sped up without the need for the patient to endure burdensome PSG testing at a clinic.

Alfeld, P. A trivariate clough-tocher scheme for tetrahedral data. 

Rem behavior disorder and sleep related hallucinations, a case study

High-performance diagnosis of sleep disorders: A novel, accurate and fast machine learning approach using electroencephalographic data

Performance of seven consumer sleep-tracking devices compared with polysomnography

Detection of rem sleep behaviour disorder by automated polysomnography analysis

Proof of concept: Screening for rem sleep behaviour disorder with a minimal set of sensors

Recent developments in new therapeutic agents against alzheimer and parkinson diseases: In-silico approaches

Seven solutions for neuroprotection in parkinson's disease

Multi-class classification of motor imagery eeg signals using image-based deep recurrent convolutional neural network

Shortcut learning in deep neural networks. CoRR, abs

Dominant frequencies of resting human brain activity as measured by the electrocorticogram

Prevalence and determinants of rapid eye movement sleep behavior disorder in the general population

Rem sleep behavior disorder (rbd)

Adversarial examples are not bugs, they are features

Optimal imf selection of emd for sleep disorder diagnosis using eeg signals

Sleep Behavior Disorder in the Korean Elderly Population: Prevalence and Clinical Characteristics

A new fully automated random-forest algorithm for sleep staging

Abnormal phase coupling in parkinson's disease and normalization effects of subthreshold vestibular stimulation

Joint classification and prediction cnn framework for automatic sleep stage classification

Risk and predictors of dementia and parkinsonism in idiopathic REM sleep behaviour disorder: a multicentre study

Idiopathic rem sleep behavior disorder in the elderly spanish community: a primary care center study with a two-stage design using videopolysomnography

Assessing rem sleep behaviour disorder: From machine learning classification to the definition of a continuous dissociation index

A neurologist's guide to rem sleep behavior disorder

A deep recurrent neural network for abnormal eeg identification

Very deep convolutional networks for large-scale image recognition

Designing a robust feature extraction method based on optimum allocation and principal component analysis for epileptic eeg signal classification

Rem sleep behavior disorder: Diagnosis, clinical implications, and future directions

Sleep in parkinson's disease

Fundamentals of eeg measurement

Atlas, rules, and recording techniques for the scoring of cyclic alternating pattern (cap) in human sleep