Deep-learning detection of mild cognitive impairment from sleep electroencephalography for patients with Parkinson’s disease

PLOS ONE RESEARCH ARTICLE Deep-learning detection of mild cognitive impairment from sleep electroencephalography for patients with Parkinson’s disease Madan Parajuli1, Amy W. Amara2, Mohamed Shaban ID1* 1 Electrical and Computer Engineering, University of South Alabama, Mobile, Alabama, United States of America, 2 Movement Disorders Center, University of Colorado, Aurora, Colorado, United States of America a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Parajuli M, Amara AW, Shaban M (2023) Deep-learning detection of mild cognitive impairment from sleep electroencephalography for patients with Parkinson’s disease. PLoS ONE 18(8): e0286506. https://doi.org/10.1371/journal. pone.0286506 Editor: Mohammad Amin Fraiwan, Jordan University of Science and Technology, JORDAN Received: March 6, 2023 Accepted: May 16, 2023 Published: August 3, 2023 Copyright: © 2023 Parajuli et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting information files. Funding: Research reported in this manuscript was supported by the National Center for Advancing Translational Research of the National Institutes of Health under award number UL1TR003096-03. Competing interests: The authors have declared that no competing interests exist. * Abstract Parkinson’s disease which is the second most prevalent neurodegenerative disorder in the United States is a serious and complex disease that may progress to mild cognitive impairment and dementia. The early detection of the mild cognitive impairment and the identification of its biomarkers is crucial to support neurologists in monitoring the progression of the disease and allow an early initiation of effective therapeutic treatments that will improve the quality of life for the patients. In this paper, we propose the first deep-learning based approaches to detect mild cognitive impairment in the sleep Electroencephalography for patients with Parkinson’s disease and further identify the discriminative features of the disease. The proposed frameworks start by segmenting the sleep Electroencephalography time series into three sleep stages (i.e., two non-rapid eye movement sleep-stages and one rapid eye movement sleep stage), further transforming the segmented signals in the timefrequency domain using the continuous wavelet transform and the variational mode decomposition and finally applying novel convolutional neural networks on the time-frequency representations. The gradient-weighted class activation mapping was also used to visualize the features based on which the proposed deep-learning approaches reached an accurate prediction of mild cognitive impairment in Parkinson’s disease. The proposed variational mode decomposition-based model offered a superior accuracy, sensitivity, specificity, area under curve, and quadratic weighted Kappa score, all above 99% as compared with the continuous wavelet transform-based model (that achieved a performance that is almost above 92%) in differentiating mild cognitive impairment from normal cognition in sleep Electroencephalography for patients with Parkinson’s disease. In addition, the features attributed to the mild cognitive impairment in Parkinson’s disease were demonstrated by changes in the middle and high frequency variational mode decomposition components across the three sleep-stages. The use of the proposed model on the time-frequency representation of the sleep Electroencephalography signals will provide a promising and precise computer-aided diagnostic tool for detecting mild cognitive impairment and hence, monitoring the progression of Parkinson’s disease. PLOS ONE | https://doi.org/10.1371/journal.pone.0286506 August 3, 2023 1 / 23 PLOS ONE Deep-learning for PD-MCI diagnosis 1. Introduction Parkinson’s Disease (PD) [1, 2] is a complex neurodegenerative disease that is characterized by motor symptoms such as: slowness of movement and tremor as well as non-motor symptoms including cognitive and memory changes, anxiety, depression and sleep problems. According to Parkinson’s foundation, 1 million patients were diagnosed with the disease in the U.S. and 10 million individuals suffer from the disease worldwide [3]. The disease is depicted as challenging for physicians and specialists to diagnose and grade. Observation of motor system abnormalities is the current means of clinical diagnosis despite being subjective and prone to human error. It was also reported in [4] that the accuracy of the clinical diagnosis performed by movement disorders experts is unsatisfactory (79.6% initial assessment accuracy and 83.9% follow-up assessment accuracy). Hence, earlier and precise detection of PD and initiation of neuroprotective treatments are crucial to improve the disease prognosis and possibly slow down its progression. Several state-of-the-art deep-learning techniques have been recently proposed for PD diagnosis, staging and biomarkers detection based on Electroencephalography (EEG), Magnetic Resonance Imaging (MRI), speech tests, handwriting exams and sensory data [5]. It was reported in [5] that the majority of the deep-learning techniques exploit either resting state EEG or handwriting/sensory data. As compared to CT, PET and MRI, EEG is relatively an inexpensive tool used for the diagnosis of several brain diseases including epilepsy, tumors and stroke. In addition, EEG derived from polysomnography is the gold standard measure for evaluation of sleep. Even if imaging modalities such as CT, MRI or PET were attempted during sleep, EEG would still be required to distinguish between sleep and wake and to distinguish between sleep stages. There have been several studies that have shown that subjects with PD exhibit unique EEG biomarkers including decreased β (12–35 Hz) and γ (> 35 Hz) powers [6, 7], slowing of resting-state oscillatory brain activity [8, 9] and significant changes in phase-amplitude coupling when compared to healthy controls (HC) [10, 11]. Patients with PD frequently experience sleep disorders, including insomnia, rapid eye movement (REM) sleep behavior disorder (RBD), and excessive daytime sleepiness [12, 13]. In addition, PD is characterized by alterations in sleep architecture, including reductions in REM sleep which plays a vital role in consolidating procedural memory and motor skills [14]. Recent studies have shown that both REM and non-REM (NREM) sleep exhibit unique features in PD and PD with dementia as compared to healthy controls (HC), including lower stability, higher slowing ratio, an increase in spectral power in the δ (1–4 Hz) and θ (4–8 Hz) bands during REM, as well as lower baseline power in σ waves (12–15 Hz) within the parietal regions during the NREM sleep stages [15 (...truncated)