Dementia has become serious social and economic issue throughout the world. Over 50 million persons live with dementia globally, the number of patients is estimated to be threefold by 2050 based on the World Alzheimer Report 20191. This report also estimates the current annual cost of dementia is 1 trillion US dollars, and a figure will increase doubled by 2030.
Accurate differential diagnosis of dementia subtypes is essential to optimize clinical care, for instance, by carefully avoiding medication with anticholinergic properties and promoting aggressive identification of therapeutic targets such as orthostatic hypotension and constipation2. Moreover, differential diagnosis would improve the quality of life of both patients and caregivers and enhance their active participation in the treatment program3.
Alzheimer's disease (AD) and dementia with Lewy bodies (DLB) are the most common types of neurodegenerative dementia. AD is characterized by progressive cognitive and functional deficits, while the core features of DLB are, besides progressive cognitive decline, parkinsonism, fluctuating cognition, and recurrent visual hallucinations4. There is growing clinical interest in distinguishing these diseases, however, DLB and other types of dementia, including AD, share similar clinical and neuropsychological features, which complicates the differential diagnosis. In addition, not all of clinical symptoms of DLB were present throughout the disease course, making the diagnosis more difficult5.
On the other hand, the disease attracting recent medical attention as the leading cause of reversible dementia is idiopathic normal pressure hydrocephalus (iNPH)6. iNPH is characterized by symptoms of gait disturbance, cognitive deterioration, urinary incontinence without any preceding diseases. Brain imaging shows typical morphologic abnormalities such as ventriculomegaly, dilation of the Sylvian fissures and narrowing of the sulci and subarachnoid spaces over the high-convexity area of the brain, indicating excessive accumulation of cerebrospinal fluid (CSF) in the ventricles7. Patients with iNPH were often examined in specialized medical institutions for dementia where neurologists, neurosurgeons, and psychiatrists worked together and they hold the uniform biological background confirmed by brain magnetic resonance imaging (MRI). The prevalence of possible iNPH was reported to be 1–2% in the elderly population by recent community or population-based studies7,8. However, many older patients with iNPH remain undiagnosed and untreated, yielding the urgent need for developing noninvasive biomarkers for diagnosis6.
Clinical features and neuropsychological profiles can overlap among dementia diseases and shift in accordance with the disease stages of dementia. Differential diagnosis of dementia largely depended on clinical features, neuropsychological profiles, and various biomarkers, for instance, structural MRI, Single-photon emission computed tomography (SPECT), amyloid PET, CSF markers, meta-iodobenzylguanidine (MIBG) myocardial scintigraphy, Dopamine transporter (DAT) uptake in basal ganglia, and Polysomnography (PSG) confirmation of rapid eye movement (REM) sleep without atonia4.
The investigation of above-mentioned biomarkers can be invasive with certain degree, and be conducted heavily costly and in limited institutions with leading-edge equipment9. One of the biomarkers in clinical practice to conquer these limitations is electroencephalograph (EEG).
EEG signals reflect the superposition of electromagnetic fields generated from cortical neurons' interaction at a macroscopic level3, thus EEG is considered to be the prime candidate of functional biomarker of synapse dysfunction and loss in dementia diseases10.
EEG offers a noninvasive technique that is inexpensive, highly available, and sensitive to functional state of human brain11. Recently, EEG have been utilized as a promising examination to screen and assist diagnosing dementia12 and neurophysiological findings associated with neurodegenerative diseases have been accumulated11. For instance, numerous studies have reported that patients with AD exhibited the disruptions in EEG activities, such as generalized slowing activities, reduced global synchronization, and anteriorization of fast-frequency oscillation13. On the other hand, the EEG features as a potential biomarker of DLB have been well acknowledged14 and international clinical diagnosis criteria of DLB referred to the prominent posterior slow-wave EEG activity with periodic fluctuations in the pre-alpha/theta range as a supportive biomarker of DLB 4. Regarding iNPH, patients with iNPH exhibited decreased occipital alpha rhythm and altered resting EEG network compared with healthy controls15.
These findings can be promising, however, exhaustive research sessions and experienced researchers must be needed to obtain these research results. These limitations can lead to the necessity of accurate and automatic assessment equipment available to general hospitals for screening and assist diagnosing of dementia diseases under world-wide rapid increasing of patients with dementia. The classification for multiple dementia diseases based on automatic EEG machine-learning analysis were reported in the past studies5,17, archiving the high discriminations for multiple dementia diseases. However, these studies included small patients’ numbers of each dementia diseases (about 20 persons in each dementia diseases at least). In addition, more than 20 EEG arbitrarily set parameters were analyzed for discrimination of dementia diseases, then it was shown that the EEG parameters with high differentiation ability differed depending on each dementia diseases, yielding limitations on practical application with too complicated analysis conditions. As mentioned above, accurate differential diagnosis of dementia subtypes is essential to optimize clinical care, moreover, especially for undiagnosed patients with iNPH, developing noninvasive biomarkers for screening iNPH is urgently needed for finding the treatable dementia to lead to medical treatment.
In this study, we evaluated deep-learning based automatic classification of electroencephalography for discrimination of multiple dementia diseases (AD, DLB, iNPH) with one algorithm based on close to clinical EEG data in anticipation of social application on larger sample, which held high accessibility and cost effectiveness derived from EEG measurements. Deep learning methods have successfully solved classification tasks and provided new insights in various fields, however, decision processes of their predictions have been hardly understandable and interpretable, hampering the clinical acceptance in medical applications17. Thus, we also analyzed the power-based EEG analysis that has been widely analyzed in EEG analysis for dementia diseases11 and considered the interpretability of the results from deep learning.
This study included DLB and iNPH to which deep-learning EEG analysis has not ever been applied and had a wider range of clinical applications while compromising the past auto-diagnosis of EEG research for dementia diseases to date.